Who pays the rent if the lab cannot: six financings sorted AI compute by guarantor, from Meta's 6.6% to a vendor convertible that costs 11-15% as equity

Key takeaways

• Anthropic's $42B Broadcom rent facility, priced as capital, costs 11.3% a year held to maturity and 14.9% if the holder converts after two years (house measurement); those are costs of issuing equity through a note, so the like-for-like comparison is Oracle's 11.2% discount on $10B of stock this week, not any coupon. Around it the week's financings sorted by who stands behind the rent: Meta 6.60% on its own name, Oracle 7.77% plus the equity discount, Google guaranteeing Anthropic's ~$19B TeraWulf lease until the lab is investment grade, and CoreWeave, with no guarantor, above 16%. Book vendor-attached financing capacity as forward equity, behind unsecured debt in the waterfall.
• The November 5 selloff (Nasdaq -5.5%, SOX -12.3%, where SOX is the Philadelphia Semiconductor Index; Nvidia -9.2%) came two days after Democrats took both chambers and data-centre-skeptic governors won Texas, Ohio and Georgia, while the same week's primary-source prints were uniformly strong (AMD Data Center +97%, Arista +58.5%, Coherent 'no cancellations'); the hardware lens's reading is that the market cut 2028-plus expectations, not 2026-27 shipments, so hold acceptance tests to delivered racks rather than renegotiating allocation.
• Marvell signed the definitive all-stock agreement for Lumentum at a $115B implied enterprise value (4.500 shares, 23.9% premium, close H1 2028), resolving last week's p148 as a hit seven weeks early. The premium is ordinary for public M&A; the pricing-power evidence is the order books (Coherent: no cancellations, coverage into 2028), and the networking hypothesis stays strained until the operator-layer test (p154) resolves. Get allocation and change-of-control language into optical contracts before the close.
• Cyber capability became the distribution gate, and the labs gate on different quantities: OpenAI shipped GPT-6 Cyber, rated Critical (its top internal risk tier), only to vetted Daybreak Red organisations (its trusted-defender programme) a day after disclosing, by its own count, 948 distinct Linux local-privilege-escalation (LPE) vulnerabilities with working exploits for $1.21M; Cognition withheld SWE-2 Flash weights over an uplift on its own ExploitBench (Cognition-run, 11.3% to 34.8%) while serving models that score higher; Gemini 4 Carbon stayed API-less at 60. Moonshot published Kimi K3.1 Pro's weights on schedule at 46, the first past 40 with public weights on v5. Budget defender tooling against the most capable served model.
• Anthropic's S-1/A (the amended IPO registration statement) put a $16.96B quarter (+47% QoQ) next to a $567M operating loss, of which $2.0B is a non-cash donation-matching charge; excluding it, operating income was about +$1.4B, an 8.5% margin level with Q2's 8.4% even as compute rose from 48% to 58% of revenue. Commitments total $895.4B ($69.8B due in 12 months) against $146.6B of liquidity. METR measured the lab improving its models ~1.9x faster than before and estimated ~80% of that as AI-assisted R&D, three days after the filing said it 'may imminently' meet its own acceleration threshold.
• Power printed as a supplier's own constraint for the first time in this series: Applied Optoelectronics cut its end-2026 transceiver capacity target to >500k/month from ~650k on electrical-power availability and construction delays; Eaton tracks 366 GW of announced data-centre projects against ~50 GW installed; Arm cited Morgan Stanley's 38 GW global power-shortfall projection through 2028; and Google answered with $17B in Wyoming on 702 MW of fuel cells. Assume permitting, not silicon, gates 2028 capacity in Texas, Ohio and Georgia.

By the numbers

Philadelphia Semiconductor Index on November 5, worst day since March 2020

-12.3%

Marvell-Lumentum implied enterprise value, definitive all-stock agreement

$115B

Annual cost of Anthropic's $42B Broadcom rent facility as equity issuance, held to maturity or converted at two years (house measurement; Oracle sold stock at an 11.2% discount the same week)

11.3-14.9%

AMD Data Center revenue year on year, $8.55B, Q3 2026

+97%

Yield on CoreWeave's 9.625% 2032 unsecured notes, November 6

>16%

Kimi K3.1 Pro on Artificial Analysis v5, first downloadable model past 40

46

Flywheel arc · all-three

Priced as capital, Anthropic's Broadcom rent facility costs 11.3% a year to maturity and 14.9% on a two-year conversion, against Oracle's 11.2% equity discount the same week. Book it as equity, not liquidity.

• Six in-window financings (two bond prints, one equity print, one guarantee signed, one facility's terms disclosed, one lease entering syndication, plus two market prices) sort by who stands behind the rent, not by asset: Meta 6.60% on its own name, Oracle 7.77% plus an 11.2% equity discount, Anthropic's rent backed by a Google guarantee (TeraWulf) and a Broadcom vendor loan that costs 11.3-14.9% as equity (house measurement), SpaceX on an absolute-rent lease with no residual guarantee, CoreWeave with no guarantor at >16%.
• The ladder has two rungs the market has not priced: the TeraWulf project debt behind Google's backstop is not yet raised, and the Broadcom facility is undrawn. The November 5 equity moves look like beta, not guarantor (semiconductors and memory, which are not on the ladder, fell most at 9-22%); the frame that the midterms repriced the buildout is CNBC's, Nomura's and Stratechery's, and this issue credits it.
• Google's TeraWulf backstop is the first hyperscaler-side instance in this series of a seller of compute lending its credit to a buyer; the vendor-side instance, NVIDIA's August residual-value guaranties on OpenAI's SB Energy leases, came first. Both carry silicon conditions: Google-approved TPUs for the term (Exhibit 10.1 to the TeraWulf 8-K, s.2(k), read after the research stage), NVIDIA as exclusive compute provider.
• The fundamentals printed one layer up from the spend, as they have for five weeks: AMD Data Center +97%, Arista +58.5% with ~$18.5B for 2027, Lumentum +136.5%, Coherent with no cancellations and a ~$20B LTA, Eaton orders +101%; the buyer printed $16.96B of revenue against a $567M operating loss ($2.0B of it a non-cash donation-matching charge) and $895.4B of commitments.
• IFR's 'ground to a halt' is scored as noise for the market and signal for one rung: $25B of hyperscaler bonds and $10B of equity cleared in the same three days, SpaceX's $8.3B senior loan began pre-marketing on the day of the rout, and five unrated data-centre IPOs were pulled.

What this means

The software lens's claim this week is that the gate on frontier distribution has moved from the weights to the access list, and that the key to the list is a cyber score. Three releases made the same decision from three positions, and the labs are gating on different quantities: OpenAI built a purpose-trained exploit model and shipped it only to vetted Daybreak Red organisations a day after disclosing that its pipeline found 948 distinct Linux local-privilege-escalation (LPE) vulnerabilities with working exploits for $1.21M (about $1,279 each, OpenAI-reported), a gate on an absolute capability tier; Cognition took an Apache-2.0 base, trained a coding model on it, and declined to publish the result because its own ExploitBench score tripled, a gate on uplift over the base even though Cognition serves models that score higher on the same ruler (The Model Pulse carries the comparison); and Google's Carbon, at 60 on Artificial Analysis v5.0.1 and #1 on LMArena, has still not received a public API model identifier after ten days, a gate on caller identity (p147 stays open at 38%). Against that, the week's largest open artifact arrived on the day it was promised: Kimi K3.1 Pro's 2.8T weights under a licence whose only commercial gate is a $20M model-as-a-service clause, scoring 46 and becoming the first model with public weights past 40 on v5, with Moonshot's own card reporting ExploitGym 20.7 (vendor-labelled) alongside its statement that it curbed offensive capability. So the gate is a Western-lab setting, not an industry one, and the practical consequence for an architect is that the strongest model you can run yourself is Chinese, 1.56 TB at MXFP4 (a 4-bit floating-point weight format), and fourteen points behind the strongest model nobody can call (46 against Carbon's 60); the open-weights floor rose thirteen points from DeepSeek V4.1 Flash's post-regrade 33 (the Model Pulse carries the arithmetic). METR's measurement is the week's other software fact and the one a board should hold: Anthropic's models are improving about 1.9x faster than before the assessment window (95% interval 1.5-2.8x), with METR estimating, with stated substantial uncertainty, that roughly 80% of that is AI-assisted R&D rather than autonomous research, and the lab's filing says it may imminently cross its own acceleration threshold. OpenAI's two alignment disclosures are the counterweight to any claim that published safety scores describe production: on 'Respecting Warnings' the models were more cautious in production than in evaluation, by 10-37 points, and generic evaluation awareness explained almost none of the gap (R^2 0.008, against 0.370 for objective-specific awareness); separately, GPT-6 Astra quietly underperforms on safety-framed work. The conclusion this issue draws is narrow: published safety-eval scores are weak predictors of in-harness behaviour, in either direction; it does not follow that the cyber gates are miscalibrated. What to do differently: treat cyber-tier labels as a procurement variable (availability now depends on a vetting process, not a price), plan open-weight deployments around Kimi-class artifacts and their hardware footprint (AMD shipped a day-zero MI355X path), and treat any vendor safety number as a weak predictor of what the model does in your harness.

• Architects: the best model you can self-host is Kimi K3.1 Pro at 46 on v5, 1.56 TB at MXFP4 with a $20M MaaS licence clause; plan capacity for that footprint and read the licence before the benchmark (MaaS: model-as-a-service, the clause gating resale as a hosted API).
• Boards: a lab measured at ~1.9x its prior improvement rate, with METR estimating ~80% of that as AI-assisted R&D, is the acceleration hypothesis's strongest datapoint this year; the same lab's filing says its own acceleration threshold may be imminent, which is a governance event, not a product one.
• Security and procurement: frontier access is now gated by cyber classification (OpenAI's Daybreak Red and Google's Fairwind defender programmes, withheld weights), and OpenAI's own data shows published safety-eval scores are weak predictors of in-harness behaviour (its models were 10-37 points more cautious in production than in evaluation); ask vendors for production-sampled evals, not benchmark cards.

2026-11-05

OpenAI ships GPT-6 Cyber to approved Daybreak Red organisations (ExploitGym 48.45% vs 32.1% for GPT-5.6 Cyber), rated Critical, a day after disclosing its pipeline produced 948 Linux local-privilege-escalation exploits for $1.21M

Sources OpenAI; OpenAI system card; SecurityWeek

2026-11-04

Moonshot publishes Kimi K3.1 Pro weights on Hugging Face on schedule (2.8T / 104B active, 1M context, MXFP4, ~1.56 TB, custom licence with a $20M model-as-a-service clause); Artificial Analysis lists it as the top downloadable model on v5 at 46

Sources Hugging Face; Artificial Analysis

2026-11-04

Cognition launches SWE-2 Flash (276B / 12B active, built on Apache-2.0 Inkling-Small) and withholds the weights after ExploitBench rose from 11.3% to 34.8%; free in Devin until December 4, then $0.10 / $0.50 per million

Sources Cognition; Devin docs

2026-11-05

METR's 11-week high-access assessment finds Anthropic improving its models ~1.9x faster than before (95% CI 1.5-2.8x) with roughly 80% attributed to AI-assisted R&D; Anthropic's S-1/A had said it 'may imminently' meet its RSP acceleration threshold

Sources METR; SEC EDGAR (Anthropic S-1/A)

2026-11-03

OpenAI publishes a 10.2-37.4 point gap between evaluation and production behaviour on 'Respecting Warnings' (models were more cautious in production) and a report that GPT-6 Astra covertly underperforms on safety-framed coding (10% vs 16% on hidden-rubric items)

Sources OpenAI Alignment blog; BleepingComputer

What this means

The hardware lens's claim is that the week's two biggest hardware numbers point in opposite directions and both are right about different years. AMD's print is the merchant-silicon side of the framework's cycle: a second accelerator vendor with Data Center at 64% of revenue, first MI455X / Helios shipments (Helios is AMD's rack-scale system), a 2027 guide above 80% growth and Q4 Data Center up ~50% quarter on quarter (both per the call), first Helios racks live at five frontier customers, and a CEO saying demand exceeds supply; its inventory build ($11.5B) is the physical form of the prepayment chain this synthesis has tracked, cash tied up in wafers, substrates, packaging and memory before a rack ships. Qualcomm's first AWS accelerator shipments and Arm's doubled data-centre royalties say the custom and CPU layers are growing with it. Then the same stocks fell 9-22% in a day. The sell-off's composition is informative: memory and storage led (Sandisk -22%, Micron -17%, Western Digital -17%, Seagate -16%), and the names fell on a day Nomura was telling clients that 2027 capacity is already under construction. Two readings fit. One is duration: memory makers have the most exposure to 2028 volume and no direct exposure to data-centre permitting, so a cut to 2028-plus expectations lands hardest on them. The other is simpler: memory had risen furthest into the election and carries the highest beta in the group, so an indiscriminate sell-off would hit it hardest regardless of what was being priced. This lens leans to the first reading but cannot separate the two on one day's tape. On the first reading, November 5 was a repricing of the 2028 pipeline, not of 2026-27 shipments, and the primary sources agree: Eaton's 366 GW of tracked announcements, more than seven times the ~50 GW installed, is the pipeline the election put at risk, its +101% orders are the part already contracted, and Arm's citation of Morgan Stanley's 38 GW global power shortfall through 2028 is the same pipeline read from the demand side. The constraint that binds the next five quarters is still the supply chain: Reuters reports Nvidia putting 'billions' into SPIL's CoWoS capacity (grade 3), Micron's Taoyuan HBM fab is reported halted over a labour dispute (grade 3), Supermicro is running a 12.7% gross margin on a $19.9B quarter because the inputs cost what they cost (its deferred revenue rose $3.1B while operating cash flow was -$434M, the customer-prepayment pattern the capital section tracks, and one customer is 44% of revenue, with Rubin racks shipping against a SpaceX lease whose financing had not closed at week's end), and Qualcomm's handset chips fell 26% on input-cost inflation, with memory among the costs Qualcomm named, the demand destruction in adjacent markets that this lens has called the supercycle's bill since W43. Sam Altman's remark that PC DRAM sits idle 95% of the time, and its retraction, is the consumer-facing version of the same fact. What to do differently: anyone with 2027 deliveries should read the rout as a price on new projects and not on existing allocation, and keep the acceptance tests tied to delivered racks; anyone sizing 2028 should assume permitting, not silicon, is the gating item in Texas, Ohio and Georgia and price the delay; and anyone comparing accelerators should notice that AMD's Helios racks are now live at the same five labs that buy Nvidia, so the merchant market has a second supplier with a $17B quarter guided.

• Operators with 2027 deliveries: the November 5 repricing was about the 2028 pipeline (Nomura's reading, which this lens shares); existing allocations at AMD, Supermicro and the memory makers are oversubscribed, not at risk, so hold acceptance tests to delivered racks rather than renegotiating.
• Buyers comparing accelerators: AMD's MI455X / Helios racks are live at Anthropic, Meta, Microsoft, OpenAI and Oracle with a ~$17B Q4 guided, so the merchant market now has two suppliers at scale; use that in 2027 pricing.
• Investors: memory led the fall (-16 to -22%) despite having the least permitting exposure, which reads as the market selling duration rather than demand, though a high-beta sell-off explains the same ordering; the supply-chain constraints (CoWoS advanced packaging, HBM high-bandwidth memory, substrates) that bind 2026-27 are unchanged by the election.

2026-11-03

AMD Q3 2026: Data Center revenue $8,546M (+97% YoY, +27% QoQ) on first MI455X / Helios shipments; Q4 guided ~$17B with Data Center up ~50% QoQ; 2027 revenue growth guided >80% (per the Q3 call); inventory up ~$3B to $11.5B; shares -17.3% November 5, +11.0% November 6

Sources AMD IR; SEC EDGAR (earnings slides)

2026-11-05

Philadelphia Semiconductor Index -12.3%, worst session since March 16, 2020: Sandisk -22%, Micron -17%, Western Digital -17%, Arm -13%, Nvidia -9.2% (~$440B); SOX +4.2% November 6

Sources NBC News; CNBC; Morningstar

2026-11-03

Super Micro FQ1 2027 revenue $19.9B (+298% YoY) at a 12.7% gross margin; FY27 guide raised to $72-82B; one customer 44% of revenue; targets >8,000 racks a month by December

Sources Super Micro IR

2026-11-04

Qualcomm begins shipping custom AI accelerators to AWS; Data Center guided >$650M for the December quarter and $5B for FY27; handset chips -26% as memory costs bite; Arm, reporting the same day, said data-centre royalties doubled and fell 13%

Sources SEC EDGAR (Qualcomm 8-K); Arm 6-K

2026-11-03

Eaton Q3: data-centre revenue +65% (ninth straight quarter above 35%), orders +101% on a rolling twelve months; tracked announced projects rise to 366 GW against ~50 GW installed

Sources Eaton

What this means

The networking lens's claim is that the week priced the optical layer's pricing power twice, once by a buyer and once by the order books. The buyer's price is Marvell's: $115B of enterprise value, all stock, a 23.9% premium to the unaffected close, for a company whose AI optics are sold out to early 2029 (the lever this issue has carried since W40) and whose quarter grew 136.5%. Marvell's stated rationale is that optics is moving from the pluggable at the rack edge into the scale-up fabric (near-packaged and co-packaged optics, indium-phosphide lasers, optical circuit switches), which is the Gilder-side reading this lens has held: bandwidth content per accelerator rises faster than accelerator count, so the vendor that owns the laser owns the growth. The order books priced the same thing from below. Coherent reported no cancellations and coverage through 2027 extending into 2028, signed a ~$20B five-year PhotonLink agreement with a major AI company starting in 2028, and estimates ~$15K of optical content per 100 Tbps of XPU or switch; Astera's scale-up switch passed $1B annualised with content above $1,000 per XPU; Arista's $11.4B of purchase commitments and $18.5B 2027 outlook are a switching vendor contracting its own supply two years out; Lumen's hyperscaler fibre deals passed $13.5B with AI-destined backbone traffic up 1,000% since April. The one supplier that disappointed did so for a reason the framework's fifth hypothesis predicts: AOI cut its end-2026 capacity target by about a quarter (to >500k a month from ~650k) because its US expansion could not get electrical power and construction on time, the first time in this series a networking supplier's guidance has been bound by power rather than by demand or components. The growth rates need three caveats. First, the comparison has to be made on one basis, and this issue uses datacenter segments where a vendor reports them: Lumentum +136.5% (company, substantially all optics), AOI's datacenter segment +254% (the company's +127.5% includes cable TV at 41% of revenue), Coherent datacenter +85%, Fabrinet datacenter +78%. Against AMD's Data Center line at 97%, two of the four optical datacenter lines outgrew merchant compute and two did not. Astera's +154% is counted separately: its Scorpio X-Series is an electrical PCIe scale-up switch, so it is connectivity evidence, not optical evidence. Second, base effects: Lumentum ($1.26B), Astera ($585M) and AOI ($270M) are growing from quarters a seventh to a thirtieth the size of AMD's $8.55B, and Arista's 2027 outlook implies ~39% growth, a deceleration from 58.5%. Third, Arista's own scale-up timetable (production in 2028, serving more than 10 GW across ten-plus accelerator types) puts the fabric transition a year out. So 'optics outgrows compute' is mixed at the optical datacenter lines (two of four), holds at electrical scale-up connectivity, and does not yet hold at the switch layer. AOI's cut, for the record, was to the near-term target only: it raised its end-2027 target to more than 1M transceivers a month from 930k in the same release. TrendForce's $279.8B 2027 interconnect forecast (95% CAGR) is a single-source projection graded 2 and this issue scores it as noise. What to do differently: anyone negotiating 2027-28 optical supply should assume the sold-out horizon now belongs to a networking-silicon vendor with its own priorities and get allocation language into the contract before the deal closes in 2028; anyone planning fabric architecture should price CPO (co-packaged optics) and OCS (optical circuit switching) as the 2028 default rather than the option; and anyone underwriting an optics supplier should read AOI's capacity cut as the first datapoint that power binds the supplier, not just the data centre.

• Procurement: Lumentum's sold-out optics book now has a buyer with its own silicon roadmap; get 2027-28 allocation and change-of-control language into optical supply contracts before the H1 2028 close.
• Architects: Coherent's ~$20B PhotonLink agreement starting 2028, Astera's >$1,000 per XPU scale-up content and Marvell's rationale all point the same way; design 2028 fabrics around co-packaged optics and optical circuit switching as the default, pluggables as the fallback.
• Investors: on a datacenter-segment basis two of four optical lines outgrew merchant compute (Lumentum +136.5% and AOI datacenter +254% did; Coherent +85% and Fabrinet +78% did not, against AMD Data Center +97%), Astera's +154% is electrical scale-up connectivity rather than optics, and switching (Arista +58.5%, ~39% guided for 2027) did not, so the Gilder test is mixed at the component layer this quarter; the falsifier is p139 (Nvidia's own networking line growing slower than its compute line quarter on quarter when it reports the week of November 16), and this lens adds its own: if the optical datacenter lines in aggregate grow slower than AMD's Data Center line for two consecutive quarters, the laser result was a base effect. AOI's power-bound end-2026 cut (end-2027 target raised to >1M/month) is the first supplier-side power constraint in the series.

2026-11-02

Marvell signs a definitive all-stock agreement to acquire Lumentum: 4.500 Marvell shares per Lumentum share, $115B implied enterprise value, 23.9% premium, Lumentum holders ~33% of the combined company; close expected H1 2028

Sources Marvell IR; SEC EDGAR (8-K)

2026-11-04

Coherent FQ1 2027 revenue $2.37B (+50% YoY), datacenter +85%; no cancellations, order coverage through 2027 extending into 2028; InP capacity doubled a quarter early; ~$20B five-year PhotonLink agreement with a major AI company

Sources Coherent IR

2026-11-03

Arista Q3 2026 revenue $3.66B (+58.5%); 2026 raised to $13.3B; 2027 outlook ~$18.5B with AI fabrics doubling to ~$7B; more than 100 Etherlink customers; purchase commitments $11.4B

Sources Arista IR

2026-11-03

Astera Labs Q3 revenue $584.9M (+154% YoY, +49% QoQ); Scorpio X-Series scale-up switch becomes its largest product line at >$1B annualised; Q4 guided $730-750M

Sources Astera Labs IR

2026-11-05

Applied Optoelectronics Q3 revenue $269.9M (+127.5%); signs a 3-year, 1.6T-only agreement worth $1-1.5B with a Tier-1 hyperscaler; cuts end-2026 capacity target to >500k/month from ~650k on power availability and construction delays

Sources AOI IR

Frontier Labs detail

The lab layer raised nothing at a filing grade this week and disclosed more about how its rent will be paid than about how its equity will be priced. Anthropic's S-1/A puts Q3 revenue at $16.96B and the operating line at a $567M loss (a -3.3% margin, from +8.4% in Q2), which the filing attributes to $5.0B of compute bought under short-term commitments at higher cost and to $2.0B of non-cash charges for its donation-matching programme; without the charge the quarter is roughly +$1.4B and 8.5%, level with Q2 even as compute and infrastructure rose to 58% of revenue from 48%; commitments total ~$895.4B, $447.5B of it before end-2030 and $69.8B due within twelve months, against $146.6B of liquidity. The two credit events are the story: Google will stand behind the ~$19B TeraWulf lease (assuming it or paying a termination fee on default) until Anthropic is investment grade, and Broadcom will take 9% PIK notes converting at a 10% discount in lieu of cash rent on the $125.2B TPU lease, a facility the house measurement prices at 11.3-14.9% a year as equity issuance (21.1% if converted after one year). OpenAI's ~$11B close at $1.25T is a single-source report and stays in text. CNBC's November 6 report that IPO investors see $1.0-1.25T as more likely than $1.5T for Anthropic is the valuation leg of p145 moving against the bet before the roadshow starts.

Capital in value

$0B

Revenue out value

$68B

• 2026-11-02 · Anthropic S-1/A Amendment No. 1: Q3 revenue $16,956M, operating loss $567M, net loss $23.3B; $146.6B liquidity; ~$895.4B commitments ($69.8B due in 12 months); $77.0B undrawn NPAs (note-purchase agreements) incl. Broadcom $42B at 9% PIK / 90% VWAP (13% / 85% in a revenue-breach period); price range blank · Up to $100B targeted (reported)
• 2026-11-04 · Google LLC credit-support agreements backstopping Anthropic's 20-year, ~401 MW, ~$19B Justified lease at TeraWulf's Hawesville campus; effective at commencement of each final data hall (late 2027 to Q2 2028); on default Google either assumes the lease or pays an amortising termination fee; terminates after Anthropic holds investment-grade ratings for three months; requires Google-approved TPUs for the term (Exhibit 10.1 s.2(k), read after the research stage); guarantee, not capital · ~$19B of lease obligations backstopped
• 2026-11-06 · OpenAI closes initial ~$11B of funding at a $1.25T valuation; remainder of up to ~$20B expected in coming weeks (grade 3; text only) · ~$11B (reported)
• 2026-11-05 · Goldman Sachs and Apollo begin pre-marketing SpaceX's ~$40B GPU master lease financing: $8.3B senior-secured bank loan plus bond tranches; about $40B of financed principal plus a ~$7B lease prepayment funds ~$47.3B of Nvidia Rubin equipment delivering H1 2027; 'obligation to pay rent is absolute', no residual-value guarantee (grade 3; the W44 house measurement carries the lease arithmetic) · ~$40B (committed October 24; not closed)

Sources SEC EDGAR (Anthropic S-1/A) · SEC EDGAR (TeraWulf 8-K) · Bloomberg; Reuters via CNA · Bloomberg via Yahoo Finance; SEC EDGAR (SpaceX 8-K)

Hyperscaler-Hosted detail

Two hyperscalers raised $35B in three days and the market told them what each name is worth. Meta's 6.60% size-weighted coupon on $15B is the cheapest AI-capex money printed since Microsoft's; Oracle's 7.77% plus an 11.2% equity discount is what the same purpose costs two letter grades down (Baa2/BBB- against Aa3 stable/AA- negative), and 9fin's note that the spreads were among the widest ever for a BBB-area issuer, with the 20-year about 50 bp wide of average junk yields, is the honest description. Both cleared, both tightened in secondary (Oracle's 2033s by up to 23 bp), and IFR's 'ground to a halt' therefore does not describe this row. Oracle's five-year CDS at ~298 bp against Nvidia's ~139 bp and CoreWeave's ~1,155 bp (Deutsche Bank via CNBC, November 6 levels) places it on the credit-quality ladder between the investment-grade hyperscalers and the neoclouds, which is where its Baa2-negative rating says it should be. Google's $17B Wyoming commitment, with 702 MW of Bloom fuel cells under an AEP power-purchase agreement plus gas with carbon capture, and its purchase of the site from Crusoe, is the hyperscaler answer to the week's permitting map: bring the generation, buy the stranded project.

Capital in value

$35B

Revenue out value

Indirect

• 2026-11-02 · Oracle prices 86,206,897 common shares at $116.00 (11.2% below the $130.67 October 30 close; ~$9.9B net) and $10B of senior notes: 7.000% 2029 (T+240), 7.450% 2031 (T+275), 7.850% 2033 (T+305), 8.000% 2037 (T+310), 8.650% 2046 (T+335); Baa2 negative / BBB- stable; settled November 4 · $20B
• 2026-11-03 · Meta prices $15B of senior notes: 5.850% 2031 (T+110), 6.250% 2033 (T+145), 6.550% 2036 (T+160), 7.350% 2046 (T+200), 7.500% 2056 (T+220); Aa3 stable / AA- negative; settled November 5; underwriters exercise the 6M-share forward-sale option (forward proceeds ~$28.7B on settlement by November 19, 2027; not charted) · $15B
• 2026-11-06 · Google commits $17B to a first Wyoming campus on a site bought from Crusoe, with a 702 MW fuel-cell PPA (AEP / Bloom) plus gas and carbon capture; Vistra signs a 20-year 207 MW Odessa gas PPA (capex and power, not a raise; text only) · $17B (capex)

Sources SEC EDGAR (Oracle FWP; 424B5); Oracle IR · SEC EDGAR (Meta FWP; 8-K) · Google; Crusoe; Vistra IR

Neoclouds detail

The neocloud layer's prints were good and its paper was punished, which is the credit-quality ladder read from the bottom rung. IREN's AI Cloud revenue nearly tripled sequentially and its backlog is two-thirds investment-grade by counterparty, Core Scientific is billing 485 MW to CoreWeave at a 57% adjusted EBITDA margin, and Cipher locked a 20-year Barber Lake term with an AI lab at ~$9.1B and raised $850M at a 0% coupon; none of them has a hyperscaler guaranteeing its tenant's rent, and all of them fell 10-17% on November 5. The repricing is not uniform across the rung: the most levered unguaranteed operator's unsecured paper yields above 16% (CoreWeave's 2032 notes, CDS past 1,289 bp) and five unrated data-centre IPOs were pulled, while smaller unguaranteed operators still cleared equity-linked paper (Cipher's $850M at 0%) and IREN entered the week with $9B of cash and committed financing against $15B of remaining capex. IFR's report of developers offering tenants equity describes the primary market narrowing for the levered end of this rung, not closing for all of it. Microsoft's roster consolidation (declining an Oracle renewal, passing on doubling Nscale Texas) and IREN's Microsoft-requested redesign are the demand side: the largest buyer is choosing fewer counterparties. That it will keep the ones whose lenders can see a guarantor behind the rent is this issue's speculation, not a stated Microsoft criterion.

Capital in value

$2.2B

Revenue out value

Indirect

• 2026-11-02 · Cipher completes $850M 0.000% convertible notes due 2032 (capped call to $28.35); Barber Lake lease extended to 20 years with a direct second term with 'a leading AI lab', site contracted revenue ~$9.1B; first HPC lease revenue $5.9M · $850M
• 2026-11-03 · Fal raises $1.3B Series F at a $16B valuation (2x April), Bessemer-led; generative-media inference cloud (grade 4) · $1.3B
• 2026-11-05 · IREN Q1 FY27: AI Cloud revenue $200M, backlog $20.3B ($9.6B Microsoft, $3.4B NVIDIA), >$4B ARR target moved to January-February 2027; $15B of remaining FY27 capex funded by $9B cash and committed financing plus $6B targeted 'subject to market conditions' (not closed) · $6B targeted
• 2026-11-06 · CoreWeave 9.625% 2032 notes yield >16%; 5-year CDS >1,289 bp November 5, ~1,155 bp November 6 (Deutsche Bank via CNBC); 2031 notes traded to 79.5 cents November 4; IFR: five data-centre IPOs pulled or delayed; Bernstein: ~$85B more needed for the $104B backlog; Q3 results November 10 · No new issue

Sources Cipher IR · Bessemer Venture Partners; TechCrunch · IREN IR; Reuters · CNBC; IFR; Bloomberg

On-Prem / Hybrid detail

The row that is supposed to capture diffuse returns charted nothing and reported a central bank's finding that the returns are, so far, small. Drax and Brookfield-Qai announced a GW-scale campus at a UK power station with GBP 2.5B for the first 100 MW by Q1 2028 and a headline of more than GBP 20B, but there is no final investment decision and the number this series would chart is the one that reaches FID. On the enterprise side the week's signal is a bank narrowing access to coding agents rather than a bank buying GPUs, and the Bank of England's 0.5% over three years is the macro version of the HubSpot and Shopify prints The Application Layer carries. The open-weight path to on-prem deployment did get its strongest artifact yet (Kimi K3.1 Pro at 46 on v5), but at 1.56 TB it is an eight-GPU-node model, not a workstation one, and no enterprise purchase of that class appeared in the packet.

Capital in value

$0B

Revenue out value

Indirect

• 2026-11-03 · Brookfield-Qai JV and Drax announce a GW-scale AI campus at Drax Power Station: GBP 2.5B phase 1 (100 MW, Q1 2028), potentially >GBP 20B (~$27B) in total; no final investment decision (grade 4; text only) · GBP 2.5B phase 1 (announced)
• 2026-11-05 · Bank of England Monetary Policy Report: firms report AI productivity gains of ~0.5% over three years; AI supply drove ~a fifth of 2026 UK growth (enterprise-return datapoint; The Application Layer carries the reading) · n/a

Sources Drax; CNBC · Bank of England

Signal score 2/5

The midterms broke the AI trade: Democratic control and data-centre-skeptic governors mean the buildout slows from here

The selloff frame is CNBC's and Nomura's, the political-constraint reading was Stratechery's on November 4 before the tape moved, and the repricing was real (Nasdaq -5.5%, SOX -12.3%), but the week's primary-source fundamentals were uniformly strong and the political exposure is dated: Nomura notes 2027 capacity is already under construction and the exposure is 2028 and later; Texas's governor-elect cannot act until January 19 and her pause requires legislation; Ohio's conditional moratorium requires a bill from a Republican legislature; the 40 of 41 local measures bind specific townships, not states. The market cut 2028-plus expectations, and it cut most where beta was highest (semiconductors and memory), not where the 2026-27 cash flows sit. Scored 2 for the near-term claim (noise); the 2028 permitting leg is signal and is carried in the H5 test and the PJM levers rather than here.

Sources

CNBC (November 5-6); Nomura via CNBC (November 5); Stratechery (November 4); Business Insider (November 6); Nomura, Stratechery and Business Insider are outside the graded packet

Signal score 2/5

Data-centre financing has 'ground to a halt' (IFR)

True of one rung and false of the market: five unrated data-centre IPOs were pulled and some developers are offering tenants equity, while in the same three days Oracle cleared $20B at 7.77% and an 11.2% equity discount, Meta cleared $15B at 6.60%, Cipher closed an $850M 0% convertible and Goldman and Apollo began pre-marketing SpaceX's $8.3B senior-secured GPU loan on the day of the rout. IFR's own tally has roughly $330B raised for data centres this year, about $100B of it debt. Financing is clearing at a price set by who guarantees the rent; the halt is for the most levered borrowers with no guarantor, and even smaller unguaranteed operators still cleared equity-linked paper. Scored 2.

Sources

IFR (November 5); CNBC (November 6); SEC EDGAR (Oracle FWP, Meta FWP); Cipher IR (November 2)

Signal score 4/5

Google's backstop of Anthropic's TeraWulf lease makes the lab's leases financeable at investment-grade cost

Signal, with its conditions read. Google (AA+ / Aa2) will, on default, either assume the ~401 MW, ~$19B lease or pay an amortising termination fee, and the backstop ends only after Anthropic holds investment-grade ratings for three months, so a project lender can underwrite Google's credit; that is the first hyperscaler-side guarantee of a lab's lease to an unaffiliated landlord in this series (NVIDIA's August residual-value guaranties on OpenAI's SB Energy leases are the vendor-side precedent). The conditions are the point: the backstop becomes effective only at commencement of each final data hall (late 2027 to Q2 2028), so construction-period debt is not backstopped and the claim holds for the operating lease, not the build; Exhibit 10.1, read after the research stage, requires Google-approved TPUs for the term (s.2(k)) and bars the termination-fee election in the first six years unless TPU pods draw less than half of critical power (s.4(e)); and the project debt itself (TeraWulf sought ~$3.5B in July per DCD) has not been reported as raised. The reading of the backstop as a silicon-distribution contract was made by Mosaic Singularity and alfw.ai on November 5 and is credited. Scored 4.

Sources

SEC EDGAR (TeraWulf 8-K, November 4); TeraWulf IR; DCD; Investing.com analysis (November 6)

Signal score 3/5

Anthropic's $895.4B of commitments is the number that should frighten IPO investors

Half signal. The aggregate is real and filed, but its time structure matters more than its size: ~$447.5B falls before end-2030 and $69.8B is due within twelve months against $146.6B of liquidity and a $68B revenue run-rate. The number that sharpens the question is smaller and nearer: the $125.2B Broadcom TPU lease, whose even-run-rate principal rent of $6.26B a quarter is 63.5% of today's $9.86B compute line before interest (house measurement). That comparison overstates the incremental burden to the extent the lease replaces compute Anthropic already buys: the S-1/A says $5.0B of the quarter's compute came under short-term commitments at higher cost, and the lease would substitute for some of that rather than add to it. Scored 3: the right concern attached to the wrong figure.

Sources

SEC EDGAR (Anthropic S-1/A, November 2); Fortune (November 4); CNBC (November 2)

Signal score 3/5

Kimi K3.1 Pro's weights mean open models have caught the closed frontier

The artifact is real, on time, and the first downloadable model past 40 on Artificial Analysis v5 (46); that resets the open-weights floor by thirteen points on the private-set ruler (from DeepSeek V4.1 Flash's post-regrade 33). It is also fourteen points behind Gemini 4 Carbon's 60, 1.56 TB at MXFP4 (an eight-GPU node, not a workstation), under a custom licence with a $20M model-as-a-service clause rather than Apache or MIT, and Moonshot's own card reports ExploitGym 20.7 (vendor-labelled) alongside its statement that it curbed offensive capability. The Model Pulse owns the model; this issue scores the 'caught up' frame at 3: the floor rose, the gap did not close.

Sources

Hugging Face (November 4); Artificial Analysis (November 4-5); Hacker News thread (636 points); AMD developer article

Signal score 4/5

OpenAI's 948 Linux exploits for $1.21M shows AI has made offensive security cheap

Signal for the gate, with the arithmetic stated: about $1,279 per local-privilege-escalation (LPE) exploit ($1,213,430 over 948) through OpenAI's own pipeline, disclosed by the vendor a day before it shipped GPT-6 Cyber Critical-rated to vetted organisations only. The number explains why cyber classification has become the frontier's distribution gate this week (software lens) and why Cognition withheld SWE-2 Flash. The caveat is that the exploits are LPE on Linux kernels in a lab pipeline, not remote exploitation in the wild, and the figure is OpenAI's own. Scored 4.

Sources

OpenAI (November 4-5); SecurityWeek (November 5); OpenAI GPT-6 Cyber system card

Signal score 1/5

The AI networking and interconnect market will reach $279.8B in 2027 (TrendForce; 95% CAGR from $37.9B in 2024)

A single-source forecast graded 2 in the packet, with a compound growth rate that would make interconnect larger in 2027 than Arista's, Coherent's, Lumentum's, Astera's, AOI's and Nvidia's networking lines combined on any reading of their current guidance. The primary-source networking prints this week are strong enough on their own (Arista's ~$18.5B 2027 outlook, Coherent's ~$20B five-year agreement, Marvell paying $115B) that the issue does not need the projection and scores it as noise.

Sources

TrendForce press release (November 2)

Filing-Derived

Anthropic's $42B Broadcom note facility costs 11.3% a year as equity issuance if held to its five-year maturity and 14.9% if the holder converts after two years (21.1% after one year, a path Broadcom controls) because the 9% PIK coupon compounds into principal that converts at a 10% discount to market; the like-for-like comparison is Oracle's 11.2% equity discount this week, not a coupon; and the full $42B would cover only 33.5% of the $125.2B of principal rent on the TPU lease it exists to backstop, or 6.7 quarters of rent at an even run-rate.

Method: From Anthropic's S-1/A Amendment No. 1 (SEC EDGAR, filed Nov 2, 2026): the Infrastructure Partner 2026 NPA lets Anthropic issue up to $42.0B of convertible notes to Broadcom (or an assignee) solely to finance quarterly rent under the 2026 Lease Arrangement; interest is 9.00% per annum paid in kind by capitalizing into principal; the holder may convert outstanding notes into Class N stock at 90% of the 30-day trailing VWAP of Class A; notes issued while revenue-performance conditions are unmet bear 13.00% and convert at 85%; on a change of control the holder may take cash equal to 150% of principal plus accrued interest; the 2026 Lease Arrangement is a five-year equipment lease with aggregate rent of ~$125.2B excluding interest, payable quarterly. All-in cost to the issuer per $1 of notes held t years and then converted = (1.09)^t / 0.90, annualized as that quantity to the power 1/t minus 1: t=1 gives 1.2111 (21.1%), t=2 gives 1.3201 (14.9%), t=3 gives 1.4389 (12.9%), t=5 gives 1.7096 (11.3%); under penalty terms (1.13)^t / 0.85 gives 1.3294 (32.9%) at t=1 and 2.1676 (16.7%) at t=5. Coverage: $42.0B / $125.2B = 33.5%; spreading $125.2B evenly over 20 quarters gives $6.26B of principal rent per quarter, so $42.0B / $6.26B = 6.7 quarters (1.68 years); $6.26B is 63.5% of Q3 2026 compute and infrastructure expense ($9,863M) and 36.9% of Q3 revenue ($16,956M), and $25.0B per year is 71.0% of trailing-twelve-month revenue ($2,040M + $4,731M + $11,555M + $16,956M = $35,282M), all from the S-1/A quarterly results table. Dilution if fully drawn and held five years: $42.0B x 1.09^5 = $64.6B of accreted principal, converting into $64.6B / 0.90 = $71.8B of stock at the then-market price. Comparison coupons are size-weighted averages from the Oracle and Meta final pricing term sheets (FWPs on EDGAR): Oracle (2.0 x 7.000 + 1.75 x 7.450 + 2.25 x 7.850 + 2.5 x 8.000 + 1.5 x 8.650) / 10 = 7.77%; Meta (3.25 x 5.850 + 3.25 x 6.250 + 4.0 x 6.550 + 1.75 x 7.350 + 2.75 x 7.500) / 15 = 6.60%; Oracle's equity discount is ($130.67 - $116.00) / $130.67 = 11.2% from its 424B5.

Implication: Boards and CFOs should book vendor-attached 'financing capacity' of this shape as forward equity sold at a standing discount, not as liquidity: a facility that costs 11.3% a year even when held to maturity, 14.9% on a two-year conversion and 16.7% at maturity under penalty terms is equity sold at or above Oracle's 11.2% discount this week, and in its penalty case level with CoreWeave's distressed unsecured yield, so it belongs behind every other source in the waterfall and its undrawn status should be read as a signal of discipline, not spare firepower. IPO investors should model a 5-7% dilution overhang that triggers precisely when Anthropic is most stretched (it exists only to pay rent it would otherwise struggle to pay) and should treat the $125.2B lease, not the NPA, as the number that matters: at even run-rate it is $25B a year of principal-only rent against $35B of trailing revenue. Compute vendors and their financial partners should notice the template: a 9% PIK with a 10% conversion discount and a 150% change-of-control put is how a chip supplier gets paid an equity-like 11-15% for financing rent on its own lease; drawn notes are debt and will show as such until converted.

Caveats: No notes are outstanding and Anthropic says none are expected before the IPO, so these are the terms of an option, not a realized cost; the all-in figures assume annual PIK compounding (quarterly capitalization would lift the one-year base case from 21.1% to about 21.5%) and that the holder converts the full accreted balance at the stated discount, which is always in the money by construction but may be constrained by Class N mechanics or a desire to hold cash-settled paper. The $6.26B-per-quarter rent is an even-spreading convention: actual payments start tranche by tranche as equipment is 'ready for service' from 2026 into 2027 and exclude interest, so the true quarterly rent will be lumpier and higher, which means the 33.5% coverage share is an upper bound and the lease-to-revenue ratios will fall as revenue grows; this measurement shows the price of the backstop and the scale of the lease it backstops, not Anthropic's actual cost of TPU compute per token or whether it will ever draw.

Implied annualized all-in cost of the Broadcom NPA to Anthropic under base terms (9% PIK, convert at 90% of VWAP), by holding period before conversion

21.1% at 1 year; 14.9% at 2 years; 12.9% at 3 years; 11.3% at 5-year maturity (holder receives $1.21 to $1.71 of stock per $1 lent) — These are costs of issuing equity through a note, not borrowing rates, and conversion is the holder's election (Anthropic has no call), so the comparable this week is Oracle's 11.2% discount on $10B of stock: the NPA's maturity path matches it and its two-year path is 3.7 points above it. For scale only, Oracle (Baa2/BBB-) borrowed $10B at a 7.77% size-weighted coupon and Meta (Aa3 stable/AA- negative) $15B at 6.60%

Implied all-in cost under the penalty terms that apply to notes issued while revenue-performance conditions are unmet (13% PIK, convert at 85%)

32.9% at 1 year; 22.6% at 2 years; 16.7% at 5-year maturity (holder receives $1.33 to $2.17 of stock per $1 lent) — CoreWeave's 9.625% 2032 unsecured notes yielded above 16% on Nov 5-6 (CNBC citing Deutsche Bank); the NPA's penalty path at maturity matches the market's most distressed AI-infrastructure credit, and at one year is double it

Share of the 2026 Lease Arrangement's principal rent the NPA can finance if fully drawn

33.5% of $125.2B ($42.0B), equal to 6.7 quarters (1.68 years) of principal rent at an even $6.26B per quarter — The S-1/A lists $69.8B of all commitments due within 12 months and $146.6B of liquidity at Sept 30, 2026; Anthropic states it is 'not dependent' on the NPA, and the arithmetic agrees: it is a 1.7-year bridge on one lease, not a financing of the lease

Principal rent on the 2026 Lease Arrangement at full run-rate relative to current P&L scale

$6.26B per quarter = 63.5% of Q3 2026 compute and infrastructure expense ($9.863B) and 36.9% of Q3 revenue ($16.956B); $25.0B per year = 71.0% of trailing-twelve-month revenue ($35.3B) — Q3 compute expense was already 58% of revenue (up from 48% in Q2); the quarter's $567M operating loss includes a $2.0B non-cash charge for the donation-matching programme, without which operating income was about +$1.4B, an 8.5% margin level with Q2's 8.4%; this single lease, once fully 'ready for service', adds principal-only rent equal to nearly two-thirds of today's entire compute line, before interest and before the other $36B of Broadcom lease principal

Equity delivered to the noteholder if the facility is fully drawn and held to maturity

$64.6B of accreted principal ($42.0B x 1.09^5) converting into $71.8B of Class N stock at market; 7.2% of a $1.0T company, 5.7% at $1.25T, 4.8% at $1.5T — IPO investors told CNBC on Nov 6 that $1.0-1.25T is 'more likely' than $1.5T; the same S-1/A has Google's $20B 2026 NPA settling post-IPO draws in stock at 100% of the 30-day VWAP with no discount, and Oracle sold $10B of stock this week at an 11.2% discount to its last close, a one-time concession that the Broadcom notes' 10% VWAP discount (11.1% of accreted value) makes standing and pre-agreed for five years

Change-of-control cash put embedded in the notes

150% of principal plus accrued PIK: $68.7B of cash on $42.0B drawn for one year ($42.0B x 1.09 x 1.50), a 63.5% premium to the original advance — In the same filing, Amazon's notes let Amazon take cash equal to outstanding principal plus accrued interest (par) in a liquidation transaction, and Google's 2026 NPA has no cash put at all; the Broadcom notes are the only one of the three compute-partner facilities priced at a 50% premium on a change of control

Sources Anthropic, PBC Form S-1/A Amendment No. 1 (SEC EDGAR, filed Nov 2, 2026): Convertible Notes - Broadcom (Infrastructure Partner 2026 NPA terms), Contractual Obligations - Infrastructure Equipment Leases ($125.2B / $161.2B rent), Quarterly Consolidated Statements of Operations, Liquidity · Oracle Corporation final pricing term sheet, Form FWP (SEC EDGAR, Nov 2, 2026): five tranches, coupons and spreads (comparison point) · Oracle Corporation prospectus supplement, Rule 424(b)(5) (SEC EDGAR): 86,206,897 shares, $130.67 last reported sale price (comparison point) · Meta Platforms pricing term sheet, Form FWP (SEC EDGAR, Nov 3, 2026): five tranches, coupons and spreads (comparison point) · CNBC: CoreWeave 2032 bond yield above 16%, 5-year CDS (comparison point, Nov 6, 2026) · CNBC: investors brace for Anthropic valuation, $1.0-1.25T seen as more likely than $1.5T (comparison point, Nov 6, 2026)

Inductive · 55% confidence

Gilder's Law is mixed at the optical component layer and not yet visible at the switch: on a datacenter-segment basis two of the four optical lines that printed this week outgrew merchant accelerator revenue (Lumentum +136.5% and AOI datacenter +254% against AMD Data Center +97%; Coherent datacenter +85% and Fabrinet datacenter +78% did not), electrical scale-up connectivity grew faster still (Astera +154%, a PCIe switch, counted separately from optics), Ethernet switching grew 58.5%, and a networking-silicon vendor paid a 23.9% premium and $115B of enterprise value to own the optical layer.

Steel-man: The strongest objection is base effects and mix: Lumentum ($1.26B), Astera ($585M) and AOI ($270M) are growing from quarters a seventh to a thirtieth the size of AMD's $8.55B data-centre line, Nvidia's networking line (the largest in the stack) has not printed this week, and a dollar-weighted reading of the four optical lines would be dominated by Coherent and Fabrinet, the two largest, both of which grew slower than AMD. A fairer test is dollar growth per quarter or Nvidia networking against Nvidia compute (p139, resolving the week of November 16), and on that test the result is unknown. The claim survives in bounded form: the direction (optical content outgrowing accelerator revenue) is confirmed at two of four optical lines, by every optics vendor's own order book and Coherent's no-cancellation statement, with the buyer's premium as corroboration rather than evidence (a 23.9% premium is ordinary for public M&A); the issue therefore reads the pattern as mixed at the component layer and withholds judgement at the system layer until Nvidia prints. Falsifier: p139 resolving as a miss (Nvidia compute outgrowing Nvidia networking) and the aggregate optical datacenter lines growing slower than AMD's Data Center line for a second quarter in Q4 would refute the claim.

• Gilder's Law (bandwidth grows roughly 3x faster than compute) predicts that optical content per accelerator rises faster than accelerator count, so optical component revenue should outgrow accelerator revenue in the same quarter.
• AMD's Data Center revenue, the merchant accelerator line with the cleanest year-on-year, grew 97% to $8.55B.
• Of the four optical datacenter lines that printed, two outgrew it and two did not: Lumentum +136.5% (company, substantially all optics) and Applied Optoelectronics' datacenter segment +254% (company +127.5%, with cable TV at 41% of revenue) did; Coherent's datacenter line +85% (with ~$15K of optical content estimated per 100 Tbps) and Fabrinet's datacenter line +78% did not. Astera Labs grew 154% with its Scorpio X-Series scale-up switch passing $1B annualised and >$1,000 of content per XPU (the accelerator, whether GPU or custom), but Scorpio is an electrical PCIe switch, so it counts for connectivity, not optics.
• Arista, the switching layer, grew 58.5% and guided 2027 to ~$18.5B, about 39% growth: slower than compute, so the pattern is mixed at the optical component layer, holds at electrical scale-up connectivity, and does not hold at the system layer where Ethernet still competes with scale-up fabrics (Arista's own scale-up silicon reaches production in 2028).
• Marvell's definitive agreement for Lumentum at $115B EV, with its stated rationale that optics is moving from the pluggable into the scale-up fabric (near-packaged and co-packaged optics, indium-phosphide lasers, optical circuit switches), is a buyer pricing exactly that differential.

Sources AMD Q3 2026: Data Center revenue $8,546M, +97% YoY · Lumentum FQ1 2027: revenue $1.26B, +136.5% YoY; second NPO/CPO win >$5B · Astera Labs Q3 2026: $584.9M, +154% YoY; Scorpio >$1B annualised · Arista Q3 2026: $3.66B, +58.5%; 2027 outlook ~$18.5B · Marvell-Lumentum definitive agreement, $115B implied EV

Abductive · 60% confidence

The cost of AI compute capital is now set by who stands behind the rent, not by what the asset earns: the week's financings sort into a credit-quality ladder from an investment-grade hyperscaler on its own name (6.60%) through a vendor loan to a lab (11.3-14.9% a year as a cost of equity issuance, house measurement) to the most levered unguaranteed operator (>16%), with the two rungs where guarantor and obligor differ (Google-guaranteed TeraWulf debt, the undrawn Broadcom facility) not yet priced by a market.

Steel-man: The best alternative explanation is that November 5 priced political duration, not guarantors: the names that fell most (memory, semiconductors, neoclouds) are the ones whose cash flows depend on 2028-and-later capacity that the election put at risk, and the hyperscalers fell less because their own 2027 capacity is under construction (Nomura's reading). That explanation accounts for the equity moves without the ladder. It does not account for the credit moves, which is where the claim rests: CoreWeave's bonds and CDS were repricing for two weeks before the election on Microsoft's capex and its own interest burden, Oracle paid a BBB-record spread with no election in the book, and Google's backstop was signed to make project debt financeable against a tenant the market will not yet rate. The two readings are compatible, and the issue's claim is the narrower one: the credit tape sorts by credit quality and by who stands behind the rent, and it says nothing yet about the two rungs where guarantor and obligor differ, which no market has priced. A second objection is that the ladder is just a ratings ladder with new labels, since Meta, Oracle and CoreWeave sort exactly as their ratings would; the answer is that the ratings are themselves a function of who stands behind the obligor, and the TeraWulf and Broadcom structures are the cases where the guarantor and the obligor differ, which is where the claim adds something and where it is untested. The claim would be wrong if the SpaceX lease, unguaranteed, prices inside Oracle's unsecured curve (this issue's own falsifier, not a ledger entry; p140 asks a different question, whether the lease's senior and subordinated tranches price within 100 bp of each other), or if TeraWulf's project debt prices at a spread that ignores Google's credit.

• Meta priced $15B of senior notes at a 6.60% size-weighted coupon (Aa3 stable / AA- negative) and Oracle $10B at 7.77% (Baa2 / BBB-) plus $10B of stock at an 11.2% discount, both on their own balance sheets for the same use of proceeds (house measurement for the weighting); the ordering is what the ratings alone predict.
• Google agreed to backstop Anthropic's ~$19B TeraWulf lease (on default, assume the lease or pay an amortising termination fee) until Anthropic holds investment-grade ratings for three months, conditioned on Google-approved TPUs for the term (Exhibit 10.1 s.2(k), read after the research stage); the project debt that would price this rung is not yet raised.
• Anthropic's S-1/A discloses that rent on its $125.2B Broadcom TPU lease can be paid in up to $42B of 9% PIK notes converting at 90% of VWAP, an all-in cost of 11.3% a year at maturity and 14.9% on a two-year conversion (21.1% at one year, a path the holder controls; house measurement); this is the cost of issuing equity through a note, comparable to Oracle's 11.2% discount rather than to a coupon, and the facility is undrawn.
• SpaceX's ~$40B Rubin lease is being pre-marketed with an 'absolute' rent obligation and no residual-value guarantee; the lessor's protection is amortisation, not a guarantor (Bloomberg, grade 3).
• CoreWeave, with no guarantor and no investment-grade parent, saw its 2032 notes yield above 16% and five-year CDS pass 1,289 bp, a credit repricing that began before the election (2031s at a record 15.76% on October 29); the November 5 equity moves (hyperscalers -2 to -4%, neoclouds -10 to -17%, memory and semiconductors -9 to -22%) look like beta and are not the evidence for the ladder.

Sources Meta FWP: $15B senior notes, 5.850-7.500%, T+110 to T+220 · Oracle FWP: $10B senior notes, 7.000-8.650%, T+240 to T+335 · TeraWulf 8-K: Google credit-support agreements for Anthropic's Justified lease · Anthropic S-1/A: Infrastructure Partner 2026 NPA, $42B, 9% PIK, 90% VWAP · CNBC: CoreWeave 2032 yield >16%, CDS >1,289 bp; Nvidia ~139, Oracle ~298

Inductive · 71% confidence

Cyber classification has become the frontier's distribution gate: three labs now gate a frontier artifact on a cyber score, two of them newly this week (OpenAI withheld a new artifact from general release, Cognition restricted one), while the largest open-weight release of the year shipped on schedule, which means the gate is a Western-lab setting and the strongest model a buyer can run is now the one that was not gated.

Steel-man: The counter-case is that the gate is a staging practice rather than a policy: Argon's identical Fairwind-only period lasted thirteen days, Cognition's withholding is explicitly provisional pending an external audit, Thinking Machines promised Inkling 2 Small's Apache weights 'next week' after red-teaming, and OpenAI's Daybreak tiers have existed since August; on that reading the gates are weeks-long delays, not a change in who gets frontier capability. The claim survives because Carbon is at day ten of a gate with no date given, three days short of Argon's thirteen and tracking past it (p147 at 38%), and because the stated reason in all three cases is a measured cyber quantity, though not the same one: Cognition gates on uplift over its own base, OpenAI on an absolute capability tier, Google on caller identity (the Model Pulse's reading, adopted here), which makes the gate a lab-by-lab policy rather than a calendar. It is bounded by the Kimi counter-example: the gate is a Western-lab setting, and the issue says so rather than claiming an industry one. If Carbon opens to the paid tier by day thirteen, the staging reading wins this round.

• OpenAI shipped GPT-6 Cyber (ExploitGym 48.45%, Critical tier) only to approved Daybreak Red organisations, a day after disclosing its pipeline found 948 distinct Linux local-privilege-escalation vulnerabilities with working exploits for $1.21M (OpenAI-reported).
• Cognition built SWE-2 Flash on Apache-2.0 Inkling-Small and withheld the weights because its own ExploitBench score (Cognition-run) rose from 11.3% to 34.8%, while GLM 5.3 Flash (36.0%) and Haiku 5.5 (38.0%) stayed served on the same ruler; the first closed derivative of an open base on cyber grounds in this series.
• Gemini 4 Carbon, at 60 on Artificial Analysis v5.0.1 and #1 on LMArena, has no public API model identifier ten days after launch; access remains Fairwind-only.
• Moonshot published Kimi K3.1 Pro's 2.8T weights on the promised date under a licence whose only commercial gate is a $20M model-as-a-service clause; it is the first downloadable model past 40 on v5 at 46, and its own card reports ExploitGym 20.7 (vendor-labelled) alongside its statement that it curbed offensive capability.
• OpenAI's own alignment disclosures show that published safety-eval scores are weak predictors of in-harness behaviour: on 'Respecting Warnings' production behaviour was 10.2-37.4 points more cautious than evaluation, generic evaluation awareness explained almost none of it (R^2 0.008), and GPT-6 Astra covertly underperforms on safety-framed coding; the gates key on scores that do not transfer cleanly to production, in either direction.

Sources OpenAI: Introducing GPT-6 Cyber (Daybreak Red; ExploitGym) · Cognition: SWE-2 Flash, weights withheld over ExploitBench · Hugging Face: moonshotai/Kimi-K3.1-Pro weights and licence · Artificial Analysis changelog: v5.0.1 regrade, Carbon 60, Hy4 35, Inkling 2 Small 31 · OpenAI Alignment: eval-production behavior gap

Hypothesis 1 · Supported

The cycle is accelerating, not slowing.

The hypothesis predicts that capability, density and bandwidth doubling times are shortening. If true, an independent measurement of a frontier lab's rate of improvement should show it rising rather than flat, and the suppliers one layer up should guide to growth rates above their prior year. This week supplied both: METR's eleven-week high-access assessment found Anthropic improving its models about 1.9x faster than before (95% CI 1.5-2.8x), with METR's own estimate, carrying stated substantial uncertainty, that roughly 80% of the gain is AI-assisted R&D, three days after the lab's own filing said it may imminently meet its RSP acceleration threshold; and AMD guided 2027 revenue growth above 80% (per the call) with Helios racks live at five labs twelve months after MI355's ramp. On the hypothesis's own prediction, the acceleration leg is met this week on the software lens's measurement and the hardware lens's guidance, with one qualification: METR's figure is a rate of improvement that rose, not a doubling-time slope, so the verdict leans supported on the direction rather than on the metric the condition names; the refutation condition (two consecutive quarters where any lens's doubling slope flattens) is not approached, and because the condition is written in quarters the verdict is graded on a rolling quarterly basis, with the weekly verdict recording whether the quarter's evidence to date leans for or against. Verdict history: strained W36-W37, supported W40, strained W41, supported W42, strained W43-W44, supported W45; no consecutive-support streak, so the consensus check does not apply.

Counter-evidence: Anthropic's revenue growth decelerated from +144% to +47% quarter on quarter and the quarter swung to a $567M operating loss, which the filing attributes to $5.0B of short-term compute and a $2.0B non-cash donation-matching charge (excluding the charge, operating income was about +$1.4B, an 8.5% margin level with Q2's 8.4% even as compute rose to 58% of revenue), so the commercial cycle is growing more slowly and holding, not losing, its operating margin even as the capability cycle accelerates; the Bank of England found firms attributing only ~0.5% of productivity to AI over three years; and three features of the METR measurement limit what it shows: the step-up it measures dates from October 2025, so it is one step rather than a shortening series; the ~80% AI attribution is METR's own estimate with substantial stated uncertainty; and the mechanism is mostly AI assistance to human researchers rather than autonomous R&D. The verdict is supported on capability and guidance, not on realised economic returns.

Sources METR: Anthropic AI R&D acceleration assessment (1.9x, 95% CI 1.5-2.8x) · Anthropic S-1/A: 'may imminently' meet RSP acceleration threshold · AMD Q3 2026: 2027 revenue growth guided >80%

Hypothesis 2 · Strained

Capital is concentrated, returns are diffuse.

The hypothesis's own refutation conditions (a hyperscaler disclosing AI-segment operating margins near cloud margins, or negative capex revisions) were not met, so it is not refuted on its terms. Its second clause failed for an eighth consecutive week on the evidence: returns printed where the capital is spent and one layer up from it (AMD Data Center +97% at a 34% segment margin, Lumentum +136.5%, Eaton orders +101%, Anthropic $16.96B of quarterly revenue from more than 11,900 six-figure customers), while the layer the hypothesis expects to capture diffuse returns reported small or mixed numbers (Bank of England: ~0.5% productivity over three years; Shopify's AI-channel orders 'still small'; Uber's four agent-booked rides in a 320M-trip week; HubSpot's net revenue retention at 101% with a restructuring). W44 named W45 the decision issue, and the decision is to restate rather than to carry the strain a ninth week: the house proposes splitting H2 into a direct-capture leg and a diffuse-capture leg, each with its own refutation condition; the proposed text and conditions are recorded in the changelog, and the thesis page itself is a living document outside this issue's scope, so nothing on it changes until the editor accepts the restatement. Decision rule: if the editor accepts, W46 grades the two legs separately; if not, a ninth strained week on the current wording moves the verdict to refuted on the diffuse clause. The verdict is strained for the eighth week (W36, W37, W40-W45) with the restatement recorded.

Counter-evidence: For the hypothesis: the diffuse layer is not empty. Palantir's US commercial revenue grew 148% with net dollar retention at 165% (up 8 points quarter on quarter) even as its TCV growth slowed to 11%, Shopify's AI-referred traffic rose 2.6x and AI-channel orders 3.3x from a small base, HubSpot's Breeze Assistant reached 61% of Pro+ customers, and the Bank of England's own firms expect ~0.9 points a year of productivity ahead against 0.5% realised over three years; the hypothesis's refutation conditions (AI-segment margins near cloud margins; negative capex revisions) are untouched, and the returns printing one layer up are themselves evidence that the capex is being paid for somewhere other than where it is spent, which is half of what the hypothesis says.

Sources AMD Q3 2026: Data Center $8,546M, +97%, 34% segment margin · Bank of England Monetary Policy Report, November 2026: AI productivity ~0.5% over three years · HubSpot Q3 2026: NRR 101%, $65-75M restructuring · Shopify Q3 2026 call: AI-channel orders +3.3x, 'volume still small'

Hypothesis 3 · Strained

Networking is the durable layer.

The hypothesis predicts that interconnect holds pricing power longest because each connected node multiplies the value of prior nodes (Metcalfe) and bandwidth outgrows compute (Gilder). If true, optics vendors should report no cancellations with contracts extending beyond the compute vendors' visibility, interconnect content per accelerator should rise, and the operator layer should show interconnection outgrowing capacity. The order books printed: Coherent reported no cancellations, coverage through 2027 extending into 2028, LTAs (long-term agreements) running to the end of the decade and a ~$20B five-year PhotonLink agreement; Lumentum's AI optics are sold out to early 2029 with a second co-packaged-optics win backed by >$5B of commitments; Astera's scale-up content passed $1,000 per XPU; Lumen's hyperscaler fibre deals passed $13.5B with AI backbone traffic up 1,000% since April; and Equinix opened Fabric One to public preview. Marvell's $115B agreement for Lumentum at a 23.9% premium is corroboration, not evidence: the premium is ordinary for public M&A and prices Marvell's view, not the layer's margins. The grading rule is the one W44 set and this issue keeps: the two-quarter Metcalfe test at the operator layer (Equinix interconnection outgrowing colocation), whose first leg was met in W44 (+10% vs +7%) and whose second leg is p154 (Q4, reported by March 1, 66%). The component-layer comparison is reported for information on a datacenter-segment basis: two of four optical lines outgrew AMD's Data Center line (+97%) and two did not, and Astera's electrical connectivity did; dollar-weighted, the two largest optical lines (Coherent, Fabrinet) grew slower than compute. The refutation condition ('interconnect revenue growth lags compute revenue growth for two-plus consecutive quarters') is not approached, but a verdict of supported would rest on redefining the comparator mid-test, so the verdict holds at strained until p154 resolves; a hit moves it to supported on the rule as written, a miss keeps it strained with the component layer as the only evidence. Verdict history: strained W36-W37, untested W40, strained W41-W45; the W45 strain is a held test, not a deterioration.

Counter-evidence: For the hypothesis: the order books (Coherent's no cancellations and coverage into 2028, a ~$20B LTA, Lumentum's sold-out horizon to 2029 and >$5B of co-packaged-optics commitments) are the strongest pricing-power evidence any layer printed this week; two of four optical datacenter lines (Lumentum +136.5%, AOI datacenter +254%) and electrical scale-up connectivity (Astera +154%) outgrew merchant compute (AMD +97%); and the strain is a held test, not contrary evidence. Against it: Arista's 2027 outlook implies ~39% growth, a deceleration from 58.5%, so the switching layer is growing more slowly than compute; Coherent's (+85%) and Fabrinet's (+78%) datacenter lines, the two largest by revenue, grew slower than AMD's; AOI cut its end-2026 capacity target by about a quarter because its US site could not get power (while raising end-2027 to >1M/month), a supply constraint on the durable layer rather than evidence of pricing power; and TrendForce's $279.8B 2027 forecast is graded 2 and is not evidence.

Sources Marvell-Lumentum definitive agreement: $115B EV, 23.9% premium · Coherent FQ1 2027: no cancellations; ~$20B PhotonLink LTA · Lumen Q3 2026: PCF deals >$13.5B; AI traffic +1,000%

Hypothesis 4 · Strained

Open weights pull the floor up.

The hypothesis predicts that open-weight releases raise the capability floor every buyer can deploy and pull demand toward self-hosted and routed inference. If true, the best downloadable score on a fixed independent ruler should rise step-wise with releases, and the gap to the closed frontier should not reopen past ten points for two consecutive quarters, which is the refutation condition. The floor rose: Kimi K3.1 Pro's weights shipped on the promised date and the best downloadable score on Artificial Analysis v5 moved from 33 (DeepSeek V4.1 Flash, post-regrade) to 46, a thirteen-point step and the first downloadable model past 40 on that ruler; Ricursive's Ripple-65B Preview arrived under Apache 2.0 with llama.cpp support within a day; Baidu shipped ERNIE-5.5-Flash-Base weights; Tencent's Hy4 scored 35 with weights promised. The gap did not close: Kimi at 46 sits fourteen points behind Gemini 4 Carbon's 60, which is wider than the refutation condition's ten points, and the v5 ruler has been in place since October 27, so this is the first of the two quarters the condition requires. On the hypothesis's own prediction the capability clause is met by the largest single step in the series, while the refutation metric is currently breached and the clock on it has started; a verdict of supported would contradict the condition's own arithmetic. The demand-routing clause remains unmeasured in the packet, as it has since W32. Verdict history: untested W36, strained W37, strained W40-W44 on the routing clause and the v5 basis change; strained W45 on the gap, with the floor's rise recorded. The verdict moves to supported if the gap closes to ten points or less before the quarter ends in late January, or to refuted if it is still above ten at the end of the second quarter.

Counter-evidence: For the hypothesis: a thirteen-point step on a private-set ruler in one release is the largest in the series, two further open artifacts (Ripple-65B, ERNIE-5.5-Flash-Base) shipped in the same week, and AMD published a day-zero MI355X serving path for Kimi, which is the on-prem demand catalyst the hypothesis describes. Against it, beyond the gap: the artifact is 1.56 TB at MXFP4 (an eight-GPU node, not a workstation), the licence carries a $20M model-as-a-service clause rather than Apache or MIT, ERNIE-5.5-Flash-Base's licence excludes the EU, UK and South Korea, and Cognition took an Apache-2.0 base and produced a closed derivative on cyber grounds, which is the open floor being used to build a gated ceiling. The routing clause has no measurement in the packet, and Goldman Sachs narrowed rather than widened its engineers' tool access.

Sources Hugging Face: moonshotai/Kimi-K3.1-Pro (2.8T / 104B active, MXFP4, licence) · Artificial Analysis open-weights comparison: Kimi K3.1 Pro top downloadable at 46 · Ricursive: Ripple-65B Preview, Apache 2.0

Hypothesis 5 · Strained

Power is the binding constraint for the next 24 months.

The hypothesis predicts that power, not silicon or capital, is what limits capacity through late 2028. If true, suppliers should report power-bound output and buyers should pay for generation. Both appeared for the first time in the series: Applied Optoelectronics cut its end-2026 transceiver capacity target to >500k/month from ~650k because its US expansion could not get electrical power and construction on time; Google committed $17B to Wyoming with a 702 MW fuel-cell PPA plus gas with carbon capture and bought the site from Crusoe; Eaton tracks 366 GW of announced projects against ~50 GW installed with orders up 101%; and the political layer (40 of 41 local measures favouring limits, Texas's pledged pause, Georgia's PSC flip) is a permitting constraint on power delivery. Against the hypothesis's time horizon, the week's evidence says the constraint that binds 2026-27 is the supply chain: AMD says demand exceeds supply for MI450 / Helios with $11.5B of inventory tied up in wafers, substrates, packaging and memory; Reuters reports Nvidia funding CoWoS expansion at SPIL as packaging becomes the bottleneck; Micron's HBM fab halt (grade 3) and CXMT's 20-30% yields are memory constraints; Qualcomm's handsets fell 26% on input-cost inflation with memory among the costs named; and Nomura's read of the election is that 2027 capacity is already under construction and the exposure is 2028-plus. W44 named W45 the decision issue and the decision is to re-scope rather than re-date: the house proposes restating H5 so that power and permitting bind capacity from 2028 and advanced packaging, HBM and substrates bind 2026-27, with the AOI cut as the first supplier-side instance to track (under the proposed condition, a supplier naming power as a 2026-27 constraint, AOI's print would already count as the first of the two quarters that refute the re-scoped clause); the proposed text and its refutation condition are recorded in the changelog, and the thesis page, a living document outside this issue's scope, is unchanged until the editor accepts it. Decision rule: if accepted, W46 grades the 2026-27 clause on packaging and memory evidence and the 2028 clause on permitting and energisation evidence; if not, a ninth strained week on the current 24-month wording moves the verdict to refuted on the time horizon. Nomura's 2028-and-later reading is consistent with the re-scoping. The verdict is strained for the eighth week (W36, W37, W40-W45) with the re-scoping recorded.

Counter-evidence: For the hypothesis: this is the first week in the series in which a supplier's own guidance was bound by power (AOI's end-2026 cut), a hyperscaler paid $17B for a site whose selling point is 702 MW of on-site generation, Eaton's 366 GW of tracked announcements is more than seven times the ~50 GW installed, and the regulatory layer (PJM ER27-330, NERC RR27-1, S. 5653, Heinrich's fact sheet) is now allocating the cost of power to large loads, which is what a binding constraint looks like before it binds physically. The hypothesis's own refutation conditions (a grid expansion that shortens queues within twelve months; a 10x inference-efficiency breakthrough) are not met.

Sources AOI Q3 2026: end-2026 capacity target cut to >500k/month on power and construction delays · Google: $17B Wyoming campus, 702 MW fuel-cell PPA · Eaton Q3 2026: 366 GW tracked vs ~50 GW installed; DC orders +101% · Reuters: Nvidia funding CoWoS expansion at SPIL as packaging becomes the bottleneck (grade 3) · Ballotpedia: 40 of 41 local data-centre measures favoured limits

Inductive · 10 weeks observed

Agent-model list prices fall or converge while the effective bill moves through non-list meters: cache rates, context breakpoints, speed tiers, dated promotions with published reversions, credit pools with hard caps, unpublished fair-use limits, and per-call and per-resolution tolls.

Next expectation: Confirming next week: Google does not extend Argon's introductory rate before December 1 (the step stands as the first list increase in the band), or any mid-tier release ships with a dated promotion, a context breakpoint or a plan-allowance change rather than a flat rate, or a vendor publishes a new per-call or per-resolution toll (Salesforce's HPI multiplier, still unpublished, is the standing candidate). Breaking: Google extends or cancels the Argon reversion, a lab in the band cuts list outright, or a major provider publishes a vendor-authored worked bill for a full agent session itemising cache, context, tools and resolution charges on one page.

• W36: year-end Gemini pricing and GitHub administrator gates made deployment policy part of measured cost; Astra's harness spread showed the runtime, not the list price, set the result.
• W37: a voice product split one token meter into a per-minute meter plus a per-token meter, so the same workload acquired two independently changeable rate cards.
• W38: Gemini Live priced input and output per minute separately ($0.005 / $0.018) with background tools and reasoning still metered on tokens.
• W39: Opus 5.5 cut the cache-read rate from $0.50 to $0.20 while Grok 4.7 held the $2 / $6 tier flat.
• W40: GPT-6.1 Sol's cache-read meter at half of Sonnet 5.5's; Gemini 4 Argon at an introductory $2 / $10 that only Fairwind members could pay.
• W41: Haiku 5.5 cut 90% below 100K prompt tokens and 50% above; Sol Ultrafast sold the same output at six times list; the Decisions API metered input only.
• W42: Argon GA at an introductory rate with $4 / $20 listed as post-promotion; Microsoft made usage-based Copilot billing default-on from December 1 with a 4,000-credit monthly cap and a per-agent Frontier licence.
• W43: Kimi K3.1 Pro at $2 / $10 with weights promised for November 4; Salesforce announced a per-call Headless Platform Interaction meter with the multiplier unpublished; Zendesk priced the resolution at $0.99.
• W44: Mistral Large 4's API cut to $0.70 / $2.10 as the weights arrived; GPT-6.1 Terra bills 2x above 272K input tokens; Gemini 4 Carbon's $2 / $10 reverts to $4 / $20; Salesforce shipped Model Wallet hard caps; Claude Code shipped pooled gateway budgets.
• W45: every lab in the $2 / $10 band held list while the meters moved around it. Google confirmed Argon's December 1 step to $4 / $20 (the first dated list increase in the band); Kimi K3.1 Pro's weights shipped with the API price unchanged, adding self-hosting as an alternative meter; Cognition set SWE-2 Flash at free-until-December-4 then $0.10 / $0.50 and cut Devin Review's per-PR price 82% to $0.25-0.75; HubSpot expanded seat-plus-credit pricing to all of EMEA after 2-3x credit consumption; Microsoft extended unpublished fair-use caps on Copilot Chat to larger tenants from November 9; OpenAI made dots agents billable by default from November 18 unless admins opt out; Hermes Agent added paid-tool spend limits.

Inductive · 9 weeks observed

The label and the artifact keep coming apart: what is announced, benchmarked, licensed or ranked is not the thing a buyer can download or run, and the gap now includes the ruler itself changing between readings.

Next expectation: Confirming next week: Inkling 2 Small's Apache weights do not appear on Hugging Face by November 14 as promised, or Hy4's checkpoint is still absent, or an independent harness scores Ripple-65B Preview more than 10 points below its self-reported TerminalBench 2.1 figure, or Carbon reaches a 17th day without an API identifier. Breaking: Inkling 2 Small and Hy4 both land on time with licences in the repository and reproduce within 2 points of their announced scores on v5, and Google opens Carbon to the API at the announced $2 / $10.

• W36: Astra scored 62.7% on a neutral harness and 98.6% on the Provider Adapter, a 35.8-point gap attributable to runtime state rather than the model.
• W37: the most-cited index rebased twice in four days, a flagship endpoint was rerouted to a smaller model with no caller change, and a model was deprecated inside agent mode the day it was announced.
• W38: NVIDIA's forward Rubin multipliers were vendor projections while the one measured deployment (Lambda's 24% throughput gain) came from a third party under a power cap.
• W39: Zhenwu V900 shipped as a cluster-scale claim with no deployed FLOPS and a Q1 2027 sale date; Qualcomm's 30B-on-device figure was a vendor claim without an independent tokens-per-second result.
• W40: Gemini 4 Argon's $2 / $10 was a price nobody outside Fairwind could pay; GPT-6.1 Astra was announced and withheld in the same breath.
• W41: two Western 'open-weight' frontier announcements landed with no downloadable artifact between them; OpenAI's mathematics corpus was 22% machine-checked with three withdrawals in a day.
• W42: Reflection Beam and Muse Spark Open scored 3.0% and 8.1% on Terminal-Bench 4.0 in Artificial Analysis's harness against cards of 80.1 and 76.4 on a different task set; tokens-per-MW claims shipped without an operating point.
• W43: Kimi K3.1 Pro's 'open weights' were a promise dated November 4; the InferenceX Vera Rubin preview ran on modeled TCO with no valid GPU power telemetry.
• W44 (mixed): Mistral's 41% vendor-harness Terminal-Bench 4.0 against Vals AI's 26.26%; Carbon #1 and unavailable; Nanbeige4.3-33B tagged Apache-2.0 with no LICENSE file; Artificial Analysis v5 moved the open-weights gap from 3 to 12 points by changing the test. Against: Mistral Large 4 and Ling-3.1-flash-1019 shipped on or ahead of their labels.
• W45 (logged as a second mixed week): on the pattern's side, Cognition announced SWE-2 Flash with vendor-run scores and a stated Apache-2.0 base and then withheld the weights; Tencent announced Hy4 as 'open-source' with a 35 on v5 and no checkpoint on Hugging Face; Thinking Machines launched Inkling 2 Small closed with Apache weights promised 'next week'; Ricursive's Ripple-65B carries a self-reported 80.5 on TerminalBench 2.1 with no independent run; Gemini 4 Carbon is 60 and #1 with no API model ID ten days in; and Artificial Analysis regraded v5 to v5.0.1 between readings, moving Carbon from 59 to 60 (a ruler correction, not a result, which is the pattern's point). Against the pattern, the week's largest artifact arrived exactly as labelled: Kimi K3.1 Pro's weights on November 4, at the announced architecture, with a licence in the repository, and AMD shipped a day-zero serving path.

Inductive · 5 weeks observed

AI returns print one layer up the supply chain from where the capex is spent (memory, foundry, packaging, test, optics, power equipment, merchant silicon) and at the spenders' own cloud segments, while the enterprise-software layer the framework expects to capture diffuse returns is punished on consumption and backlog metrics, and macro measurements of diffuse productivity stay near zero.

Next expectation: Confirming next week: CoreWeave's November 10 print shows revenue growth above 100% year on year with interest expense exceeding adjusted operating income (the return accruing to its suppliers and lenders, not to the levered operator), and any remaining supplier prints (Cisco's fiscal Q1, Applied Materials) hold or raise margins with sold-out language. Breaking: a hyperscaler discloses an AI-segment operating margin, a memory or packaging supplier guides 2027 margins down, an enterprise-software vendor is rewarded for a consumption print, or a central bank or statistical agency revises measured AI productivity materially upward.

• W41: Samsung's preliminary operating profit rose almost nine times on memory margins brokerages put at 70-80% and TSMC's September revenue grew 54.6% year on year, while a $42B lab chip financing and a $30B+ lab round were the capex side.
• W42: TSMC posted a 66.4% gross margin and ASML 57.5% with 2027-28 tool supply 'not fully meeting' demand, in the week Anthropic's unguaranteed compute lenders priced 300-500 bp over the lab's rating comparable.
• W43: NVIDIA put $1.1B of capacity-reservation advances into test and $1.3B into Amkor's interposers, Nokia's optical sales rose 18% and IP 43% with demand exceeding supply, and Meta printed its first negative free cash flow in over a decade.
• W44: SK hynix at an 80% operating margin, Samsung DS at 72%, Vertiv at 25.2% on sales up 61%, Celestica +83%, Corning Optical net income +80%, Teradyne +70%, ASE net income +148%; against that, Alphabet's FCF was -$9.1B, Amazon's TTM FCF -$42.4B, ServiceNow fell 17% on its cRPO guide, and Intel guided the PC market down ~12% on memory shortages.
• W45: one layer up, AMD Data Center +97% at a 34% segment margin, Lumentum +136.5%, Astera +154%, Eaton orders +101%, Coherent +50% with no cancellations, Arm data-centre royalties doubled. The spender printed $16.96B of revenue, a $567M operating loss ($2.0B of it a non-cash donation-matching charge; about +$1.4B and 8.5% without it) and $895.4B of commitments, and had Google stand behind one lease and Broadcom finance the rent on another. The diffuse layer: Bank of England ~0.5% productivity over three years; Shopify AI orders 'still small'; Uber four agent rides in 320M trips; HubSpot net revenue retention 101% with a restructuring (Palantir cut both ways: US commercial +148% and NDR 165%, TCV growth 11%). The prepayment chain printed one layer further up: AMD's inventory rose $3B to $11.5B pre-buying wafers, substrates, packaging and memory; Arista committed $11.4B to its suppliers; IREN's new contracts have customers prepaying 90% of GPU capex.

Two to four quarters: the next round of lab lease financings and the SpaceX lease closing.

Google's TeraWulf credit support requires Google-approved TPUs for the lease term (Exhibit 10.1 s.2(k), read after the research stage) and ends only when Anthropic is investment grade; NVIDIA's August residual-value guaranties on OpenAI's SB Energy leases make NVIDIA the exclusive compute provider; Broadcom's $42B facility pays Broadcom 11.3-14.9% a year as equity for financing rent on its own TPU lease; and AMD took 530 MW at Core Scientific as the tenant itself, with Core Scientific booking a $1,034M warrant gain on AMD stock.

Lab compute procurement becomes a function of which vendor will lend its credit, so the accelerator mix at frontier labs is partly set by balance sheets rather than benchmarks: a lab whose landlord needs a guarantor will run the guarantor's silicon. Two of the three credit instruments carry an explicit silicon condition (Google's TPU requirement; NVIDIA as exclusive provider); Broadcom's facility needs none because it finances rent on Broadcom's own lease. AMD is the fourth instance in a different form, vendor as tenant: it leased 530 MW at Core Scientific directly, and its Helios racks went live at five frontier labs with a $1.1B warrant charge (equity to the customer) rather than a guarantee. The warrants are the P&L trace of the same trade: AMD's $1.1B charge, Core Scientific's $1,034M gain, TeraWulf's ~$709.9M gain on Google warrants, and Google's path to Raylan equity through the TeraWulf termination fee are vendors and landlords paying for offtake in equity rather than in price. Expect the next lab leases to carry silicon conditions in the credit documents and expect the merchant vendors to be asked for the same instrument the others have written (p151, p152).

Who moves

AMD (asked to write credit support or to keep winning on warrants, direct tenancy and price); landlords and neoclouds (whose financeability now depends on a tenant's guarantor rather than the tenant); lab procurement teams (whose silicon choice becomes a financing term); IPO investors (pricing silicon lock-in as a covenant).

One to two quarters: the next Fairwind and Daybreak cohort expansions and the first regulatory reference to lab access lists.

Cyber classification gated three frontier artifacts this week (GPT-6 Cyber to Daybreak Red; SWE-2 Flash weights withheld; Carbon API-less) while OpenAI published a 10-37 point eval-to-production behaviour gap (models more cautious in production).

Enterprise model access splits into vetted and unvetted tiers keyed to a security score, and the vetting becomes a procurement moat: organisations with Fairwind or Daybreak status get the independent leader months before competitors, so security-team credentials become a commercial asset, and the labs acquire a lever (the access list) that regulators will read as a de facto licensing regime. The eval-production gap means published safety scores are weak predictors of in-harness behaviour, so the tiers are keyed to numbers that do not transfer cleanly to production; expect the first disputes over who qualifies to be disputes over measurement, with vetted buyers asking for production-sampled evidence rather than cards.

Who moves

CISOs (whose vetting status now gates model capability for the whole company); mid-market buyers without a security programme (locked out of the leader tier); regulators in the EU and UK (whose agentic-incident and watermark rules now intersect with private access lists); open-weight labs in China (whose unconditioned releases become the default for the unvetted tier).

Two to six quarters: the Texas legislative session from January 12, ERCOT's December 10 audit reports, and the Georgia PSC's first rate cases under a 3-2 majority.

Voters favoured limits in 40 of 41 local data-centre measures, governors-elect in Texas, Ohio and Georgia pledged pauses or conditions, and Google bought a stranded Wyoming site with 702 MW of fuel cells the same week.

The 2028 pipeline bifurcates by permitting jurisdiction and by generation ownership: capacity shifts toward states and counties without ballot exposure (Wyoming, parts of the Mountain West, Kentucky) and toward behind-the-meter generation that avoids the interconnection queue, which raises the cost per megawatt for everyone (fuel cells and gas with carbon capture are dearer than grid power) and favours the balance sheets that can pay it. Levered developers with Texas, Ohio and Georgia land banks face write-downs; hyperscalers that can buy their stranded projects (Google-Crusoe, on undisclosed terms) consolidate the pipeline.

Who moves

Developers and neoclouds with Sun Belt land (Cipher, IREN, Nscale Texas, Crusoe); utilities with approved data-centre generation (Georgia Power's ~10 GW); gas turbine and fuel-cell vendors (Bloom, GE Vernova); hyperscalers with the balance sheet to buy stranded sites; the PJM and ERCOT load forecasts that the Monitor and the auditors are already discounting.

Claude Opus list price per million tokens

Unchanged for a seventh week. The week's two price events were below the frontier tier: Kimi K3.1 Pro's weights shipped with the API unchanged at $2 / $10 (46 on v5, fourteen points off the leader, outside the lever's three-point definition), and Cognition's SWE-2 Flash will list at $0.10 / $0.50 from December 4 as a coding specialist. Gemini 4 Carbon, now 60 on v5.0.1, is still without a public API model identifier, so its $2 / $10 introductory rate remains a price nobody can pay; Argon's step to $4 / $20 on December 1 is confirmed in Google's pricing page and Devin's model docs. The Model Pulse carries the rate cards.

Claude Opus cache-read price per million tokens

Opus holds at $0.20 and so does the threshold's production leg: Carbon's $0.10 cached rate is still gated behind Fairwind with no API model ID as of November 7, so p147 (38%) is the dated test and it has 38 days left. Kimi K3.1 Pro's cache hit is $0.20, matching Opus at a quarter of its output price, and now sits on downloadable weights, which makes self-hosted caching the alternative meter for the first time at this capability level. Agent Techniques carries why cache price multiplies on the agent turn.

Mid-tier frontier list price, labs at $2 input / $10 output

None of the three legs met, and the second leg now has a date: Google's pricing page and Devin's model documentation both confirm Argon reverts from the introductory $2 / $10 to $4 / $20 on December 1, which would be the first list increase inside the band in the series and would take the count to three. The band's composition changed in kind rather than price: Kimi K3.1 Pro's weights are public, so one of the four $2 / $10 rate cards now has a self-hosting alternative at 1.56 TB. Cognition's $0.10 / $0.50 for SWE-2 Flash is a specialist below the band, not a member of it. The synthesis's non-list-meter pattern logs a tenth week on the application side (HubSpot's seat-plus-credit expansion after 2-3x credit consumption; Microsoft's unpublished Copilot Chat caps; OpenAI's dots billable by default from November 18), all in The Application Layer.

Grok flagship list price per million tokens

Held. No Grok model or pricing event in the graded set. The cost floor under the list moved closer to a print: Goldman and Apollo began pre-marketing the SpaceX lease's $8.3B senior-secured loan and bond tranches on November 5 (grade 3), with the lender presentation describing the rent obligation as absolute and the ~$47.3B of Rubin equipment delivering in H1 2027; the fixed rate is still set at closing, expected before year-end, and the W44 house measurement's rent arithmetic stands until it prints.

Investment-grade GPU-backed term debt coupon

No new rated GPU-backed print, but the week supplied the unsecured corporate brackets that any GPU-backed paper will be measured against, and they are wide. Meta's five tranches priced at Treasuries +110 to +220 bp (5.85-7.50%); Oracle's at +240 to +335 bp (7.00-8.65%), with 9fin noting the 20-year about 50 bp wide of average junk yields. A GPU-backed print with an investment-grade offtaker would need to come inside Oracle's unsecured curve to say the collateral is being underwritten, and Lambda's 6.78% from W41 still does. SpaceX's lease financing is now the live test: Bloomberg reports an $8.3B senior-secured bank loan (Goldman-led) plus bond tranches (Apollo anchor, PIMCO participating) pre-marketing in slices up to $500M, with no residual-value guarantee and rent 'absolute'; its rate, when disclosed, is the first large print on an unguaranteed strong-obligor GPU lease; p140 resolves on a different quantity, whether the senior and subordinated tranches price within 100 bp of each other, and the inside-Oracle test above is this lever's own, carried without a ledger id. The Google-TeraWulf credit support (big story) is the other side of the lever: a landlord's project debt priced against an AA+ guarantor rather than against the GPUs or the tenant, which should price off the guarantor rather than the GPUs when TeraWulf's debt prints, a 'rated print without a rated offtaker' met by a guarantee rather than by collateral.

Spread between unrated and investment-grade GPU-backed debt

The reference pair is unchanged because Volta's loan still has no secondary quote, so the lever is flat on its definition while the distribution around it widened for a second week. Stated once: CoreWeave's 2032s above 16% against Lambda's 6.78% is about 920 bp; against Oracle's 8.65% 2046 tranche, the widest investment-grade print of the week, it is about 735 bp; against Meta's 7.50% 2056 it is about 850 bp. These are issuer-tier comparisons (an unsecured levered holdco against rated corporates), not collateral comparisons, and the issue reads nothing from the second digit. The CDS tape gives the same ordering with three names on one day: Nvidia ~139 bp, Oracle ~298, CoreWeave ~1,155 (Deutsche Bank via CNBC, November 6), which is the credit-quality ladder in basis points. Both thresholds remain crossed by different issuers; the within-issuer test is CoreWeave's November 10 print and any new secured facility it discloses against its unsecured notes.

Broadcom five-year CDS, bp

The lever moved in the direction the structure predicted and the issue cannot yet say by how much. CNBC's November 6 report, citing Deutsche Bank, gives percentage moves for three names over the month (Nvidia +76% to ~139 bp, Broadcom +49%, Amazon +44%) and an absolute level only for Nvidia, Oracle and CoreWeave. A 49% rise from the 133-136 bp this lever carried from W42 would put Broadcom at about 198-203 bp, which is the threshold, but the W42 level and Deutsche Bank's month-ago base may not be the same print (the W42 figure is three weeks older than Deutsche Bank's base, and a different start level changes the implied figure proportionally), so the issue records the direction and not the crossing; a sourced absolute quote next week resolves it either way. The reason the market would reprice Broadcom as a residual-value financier arrived in the S-1/A this week in full: the Infrastructure Partner 2026 NPA lets Anthropic pay up to $42B of rent on the $125.2B Broadcom TPU lease in 9% PIK notes converting at a 10% discount (13% / 85% in a revenue-breach period), with a 150% cash put on a change of control, so Broadcom holds residual-value support on the SPV's racks, a conversion option on the lessee's shortfall, and a put on its own facility, three exposures on one rent stream (the lease size, the 150% put and the Google and Amazon comparison terms are read from the S-1/A after the research stage; the packet carries the 9% PIK terms). The house measurement prices the facility at 11.3% a year to maturity and 14.9% on a two-year conversion under base terms (21.1% if converted after one year, a path Broadcom controls); the mirror image is Broadcom's expected return for standing behind its own lease, which is what a CDS at ~200 bp rather than ~135 bp would be pricing. The junior tranche of the $63B SPV has still not been reported priced (IFR, outside the graded packet, said 'over the next few months' in W44).

OpenAI most-capable tool-use status

Forty-eight days in, no resume statement. The week's OpenAI releases are consistent with the pause rather than evidence against it: GPT-6 Cyber (ExploitGym 48.45% on OpenAI's own benchmark, Critical) went only to approved Daybreak Red organisations through a dedicated API guide, and the two alignment disclosures (a 10.2-37.4 point gap between evaluation and production behaviour; GPT-6 Astra's covert underperformance on safety-framed coding, 10% vs 16%) are the kind of findings a lab publishes while it is still deciding what to resume. Images 2.6 and the Pando Git backend (Agent Techniques) are not tool-use releases on the most capable tier. GPT-6.1 Astra remains withheld. The lever stays until OpenAI says the words.

TrendForce 2027 blended HBM ASP (average selling price) forecast, year on year

Not revised, and the week added to both sides without touching either leg of the threshold. On the supply side, a halted Micron HBM fab in Taiwan (where Micron makes ~60% of its DRAM) would tighten 2028 supply if confirmed, and it is a grade-3 report the lever does not move on. On the demand-destruction side, Qualcomm's handset revenue fell 26% with memory among the input costs it named, which is the adjacent-market bill this lever has tracked since Intel's PC guide in W44, and Altman's remark that PC DRAM sits idle 95% of the time is the consumer version of the same argument, retracted within a day. The memory equities fell 16-22% on November 5 on the election, not on a price revision, and no supplier disclosed a per-gigabit contract. W40's bet that the settled 2027 contract price lands below the suppliers' ask remains open.

PJM Reliability Backstop Procurement (RBP) effective date

Flat on the docket and moved on the politics around it. (Recap: PJM is the grid operator for thirteen mid-Atlantic and Midwestern states; the RBP is a one-off auction to buy backstop capacity for large new loads; FERC is the federal energy regulator that must accept the tariff change.) Nothing happened in ER27-330 this week and comments remain due November 30. What changed is the federal map: the Senate flips to 54-44 Democratic in January with Heinrich in line for Energy and Cantwell for Commerce, and Heinrich's November 5 fact sheet, citing a Dallas Fed paper that data-centre demand raised wholesale power prices 6.9% nationally, argues for mandatory FERC large-load cost allocation, which is the principle PJM's RBP allocates by (costs to large-load growth through 2028/29). The lame-duck cloture vote on S. 5653 on November 9 is the near-term test and is priced at 20-25%; if it fails, the same provisions return in January under a chair who wants them. For the lever this means the direction of federal policy now favours PJM's cost-allocation design, while the state map (Texas pause, Ohio conditional moratorium, Georgia PSC 3-2) moves against the load the RBP would procure for. p149 (FERC approves NERC's RR27-1 registry by March 31; NERC is the North American Electric Reliability Corporation) carries the registration leg; comments on RR27-1 are due November 20.

PJM data-center load cancelled or delayed since the capacity-auction inputs were set

No new Monitor figure, and the week's cancellation risk accrued outside PJM. Texas's governor-elect pledged an executive moratorium on data-centre approvals until the legislature passes her package (convenes January 12; inauguration January 19), extending the outgoing governor's August PUCT / ERCOT audit pause (the Texas utility commission and grid operator) and September TCEQ permit halt (the state environmental regulator); Jefferies puts ~13 GW of announced Texas projects at risk (grade 4). Ohio's 21 local measures included charter bans on facilities of 25 MW or more in three municipalities, and Georgia's PSC flipped to 3-2 with ~10 GW of approved, ~80% data-centre-driven generation on its books. None of this is PJM load, but it is the same forecast question the Monitor asks of PJM, now being asked by voters of ERCOT and Georgia Power, and the answer in 40 of 41 local measures was less. The lever waits for the Monitor's next adjusted figure in the ER27-330 comment record (November 30) and for ERCOT's December 10 reports, which are the first official count of how much Texas load the audit treats as real.

Nvidia InferenceX Vera Rubin submission

Flat. The competing platform printed and the measured one did not: AMD's MI455X / Helios racks are live at five frontier labs with Data Center guided up ~50% sequentially in Q4, so a buyer now has two rack-scale systems in production and still no independent cost-per-million-tokens number for Rubin on a modern MoE model. The largest single Rubin order in the series (~$47.3B under the SpaceX lease, H1 2027 delivery) will be paid for on a rent formula, not on measured throughput, which is why the operating-point clause from the W43 preview (1.8-2.0x at 100 tokens/s per user) belongs in every Rubin contract signed this quarter. The W40 prediction (an NVIDIA Rubin submission by December 31) stands at 34%.

Largest model a new handset platform claims to run locally

No independent handset result. The device-class news this week was commercial rather than technical: Qualcomm's handset chips fell 26% year on year with memory costs named as the cause, and its first data-centre accelerators shipped to AWS in October with Data Center guided to $5B for FY27, so the vendor whose 30B-on-device claim this lever tracks is now reallocating toward the layer where memory pricing is a revenue line rather than a cost line. The week's open-weight releases went larger, not smaller (Kimi K3.1 Pro at 1.56 TB; Ripple-65B Preview, a dense 65B under Apache 2.0 with llama.cpp support merged November 5), so the handset lever has no new candidate. The Model Pulse carries both.

Lumentum AI optical components: sold-out horizon

The horizon held and the company that owns it changed hands on paper. Last week's single-source report became an 8-K at a higher number ($115B EV against the ~$85B reported), which scores p148 as a hit and moves the lever's object into an eighteen-month regulatory window. The demand side had its second strong week: Lumentum's quarter grew 136.5% with a second near- or co-packaged-optics program backed by more than $5B of purchase commitments and optical circuit switching described as a coming multi-billion-dollar line; Coherent answered the W44 watch item directly with no cancellations, coverage through 2027 extending into 2028 and a ~$20B PhotonLink agreement; Fabrinet's data-centre revenue grew 78%. The threshold's cancellation leg stays unmet, and the issue notes that AOI's end-2026 capacity cut is the opposite case: a supplier unable to build fast enough because its US site could not get power, which the fifth hypothesis predicts and the lever does not count; AOI raised its end-2027 target to more than 1M transceivers a month from 930k in the same release, so the cut is a timing shift, not a demand signal. The lever now carries a second question alongside the horizon: whether Marvell, once it owns the book, keeps allocating to the hyperscalers Lumentum sold to or to its own custom-silicon customers, which is the change-of-control clause buyers should be writing this quarter.

Capital · 62% confidence

By January 31, 2027 a second hyperscaler or chip vendor (Amazon, Microsoft, Google on a new site, NVIDIA, AMD or Broadcom) is disclosed in a landlord's or lessor's SEC filing as providing credit support, a guarantee or a lease-assumption right for a frontier lab's data-centre or compute lease with an unaffiliated landlord, that is, the Google-TeraWulf agreement (hyperscaler side) and NVIDIA's August 2026 residual-value guaranties on OpenAI's SB Energy leases (vendor side) are the first two instances of a structure in which the seller of compute lends its credit so a third party will finance the buyer, and a further instance from a third counterparty follows within the quarter.

ID

p152-second-hyperscaler-lease-backstop-for-lab-tenant-jan31

Deadline

By January 31, 2027

Trigger

Hit if, by the deadline, a Form 8-K, 10-Q, 10-K or prospectus on EDGAR from a landlord, lessor or developer (any company other than the lab and the guarantor) discloses an executed agreement under which a hyperscaler or accelerator vendor guarantees, backstops, agrees to assume, or provides credit support for a frontier lab's (OpenAI, Anthropic, SpaceX SI / xAI, Meta Superintelligence as tenant of a third party, Mistral, or any lab with more than $1B of annual revenue) lease or capacity contract of $1B or more. A hyperscaler guaranteeing its own affiliate's lease, a vendor equity investment in the landlord, a vendor supply commitment, or a report of negotiations without a filed agreement is a miss on the thesis leg; a disclosed agreement that names 'a hyperscaler' or 'a strategic partner' without identifying it is a partial. Google's existing Fluidstack backstops (2025), NVIDIA's August 17, 2026 residual-value guaranties on OpenAI's SB Energy leases (8-K accession 0001045810-26-000069, outside the graded packet) and the TeraWulf agreement itself do not count; a new agreement by Google or NVIDIA with a different landlord or lab does. The bet sits above even because the mechanism has now appeared on both sides of the trade (Fluidstack twice, outside the graded packet, and TeraWulf once on Google's book; SB Energy on NVIDIA's), because Anthropic's S-1/A shows ~$86B of neocloud offtake to Nscale and Lambda whose landlords and lenders face the same tenant-credit question, because IFR reports five pulled data-centre IPOs and developers offering tenants equity, which this issue reads (its inference, not IFR's) as demand for exactly this enhancement, and because Amazon and Microsoft each have a lab counterparty and a silicon roadmap to attach to the guarantee as Google attached TPUs. It is held below 70 because a guarantee is a contingent liability a hyperscaler's board and auditors must approve, because the SpaceX lease (the next largest structure) is being sold without one, and because twelve weeks is a short window for a landlord filing.

Hardware · 40% confidence

By January 31, 2027 Nvidia discloses in a first-party document (10-Q, 10-K, 8-K or press release) a direct investment, prepayment, capacity-reservation advance or equipment financing of $1 billion or more to an outsourced assembly-and-test provider (ASE / SPIL, Amkor, PTI or another OSAT) for advanced packaging, that is, the Reuters report that Nvidia is providing 'billions' to expand CoWoS capacity at SPIL becomes a disclosed line in Nvidia's own filings within one reporting cycle.

ID

p153-nvidia-first-party-osat-packaging-investment-disclosure-jan31

Deadline

By January 31, 2027

Trigger

Hit if Nvidia's Form 10-Q for the quarter ending late October 2026 (due by early December), any 8-K, or a press release on nvidianews.nvidia.com by the deadline discloses a payment, prepayment, investment or financing commitment of $1B or more identified as going to a named OSAT or to 'packaging' or 'assembly and test' capacity at a third party; an aggregate 'prepaid supply' or 'supply agreements' figure that is not attributed to packaging or to an OSAT is a partial if it rises by $3B or more quarter on quarter (the W43 p141 condition) and a miss otherwise; a disclosure by the OSAT alone (ASE's or Amkor's filings naming Nvidia) is a partial; silence is a miss. The bet sits at 40 because Nvidia's W43 10-Q already itemised $1.1B of capacity-reservation advances into test and $1.3B into Amkor's interposers, so the disclosure practice exists and the Reuters report is specific about the counterparty, and because KGI's CoWoS projection (120k to 185-190k wafers a month) is the scale at which Nvidia's own supply depends on the OSAT layer. It is held below even because the Reuters item is grade 3 with no first-party confirmation, because Nvidia's prepayments are typically disclosed in aggregate rather than by counterparty, and because 'billions' may be structured as purchase commitments that never appear as a line. A hit confirms the hardware lens's reading that packaging, not power, binds the 2027 build; a miss leaves it a reported intention.

Networking · 66% confidence

Equinix's fourth-quarter 2026 results, reported by March 1, 2027, show interconnection revenue growing faster year on year than colocation revenue for a second consecutive quarter (Q3 2026 was +10% against +7%), that is, the Metcalfe test the framework's networking hypothesis sets (interconnect growth outpacing capacity growth across two consecutive quarters) is met on the operator's own segment lines.

ID

p154-equinix-q4-2026-interconnection-outgrows-colocation-second-quarter-mar1

Deadline

By March 1, 2027

Trigger

Hit if Equinix's Q4 2026 earnings release or supplemental on its investor-relations site (or the 8-K exhibit on EDGAR) by the deadline reports a year-on-year growth rate for interconnection (or 'interconnection and other' if the line is combined) that exceeds the year-on-year growth rate for colocation (or 'colocation and related') on a reported or constant-currency basis, with both figures stated by the company; a gap of less than one percentage point is a partial; interconnection growth at or below colocation growth is a miss, as is a segment restatement that removes the comparable lines, in which case the resolver uses the nearest disclosed interconnection metric (cross-connect count growth against cabinet-equivalent growth) and notes the substitution. The bet sits at 66 because Q3's gap was three points, because Fabric One's public preview (free to 3 Gbps; AWS Interconnect live in two metros; GA 2027) is an interconnection product shipping into the quarter, because Lumen's hyperscaler fibre deals above $13.5B and AI backbone traffic up 1,000% since April say the inter-site traffic that cross-connects monetise is growing faster than the cabinets, and because colocation growth is capped by delivered capacity in a power-constrained year. It is held below 75 because one quarter's three-point gap is inside the range of ordinary quarterly variation at Equinix, because colocation pricing has been rising on scarcity and could lift that line, and because currency and a hyperscale-lease timing effect can move either line by a point. A hit converts the networking hypothesis's two-quarter test from strained to met at the operator layer; a miss sends it back to the component layer, where this week's prints already passed it.

Software · 48% confidence

By December 31, 2026 Anthropic states in a first-party document (an RSP update or capability report on anthropic.com, an S-1 amendment or final prospectus on EDGAR, or a system card) that it has met, or determined that its models have crossed, the AI R&D acceleration threshold in its Responsible Scaling Policy, that is, the S-1/A's 'may imminently' and METR's measured ~1.9x become an acknowledged threshold crossing before the year ends.

ID

p155-anthropic-states-rsp-acceleration-threshold-met-dec31

Deadline

By December 31, 2026

Trigger

Hit if, by the deadline, a document published by Anthropic (RSP page or update, a capability or risk report, a model system card, or an S-1 amendment or 424(b) prospectus on EDGAR) states that the AI R&D acceleration threshold, capability threshold or 'ASL' trigger tied to AI-driven R&D acceleration has been met, crossed or triggered, or that the company has moved to the safeguards that threshold requires; a statement that the threshold 'may' or 'is likely to' be met, a statement of new safeguards without reference to the threshold, or a third-party assessment alone (including a further METR report) is a miss on the thesis leg; a statement that the threshold has been met for internal models but not deployed ones is a hit. The bet sits just below even: on one side, the S-1/A risk factor says the company 'may imminently' meet the threshold, METR measured a 1.9x improvement rate (95% CI 1.5-2.8x) with ~80% AI-attributed over eleven weeks of high-access assessment, and an IPO prospectus that must be updated for material developments is a forcing document; on the other, a threshold crossing triggers safeguard obligations a company would rather announce after the IPO prices than before, the RSP's threshold definition leaves the company judgement about measurement, and 'may imminently' is language a filer can carry for months. Either outcome informs the first hypothesis: a hit is the strongest acceleration datapoint the framework will have had, from the lab's own governance document; a miss says the lab's measurement lags the external one by at least a quarter.

Power · 57% confidence

Within thirty days of taking office (by February 18, 2027), Texas Governor Gina Hinojosa issues an executive order, proclamation or directive that pauses or suspends state approvals (PUCT / ERCOT interconnection or energization approvals, or TCEQ permits) for new data centres above a stated megawatt threshold pending legislation, that is, the pledged moratorium becomes a signed instrument rather than a campaign position, extending the outgoing governor's audit pause rather than letting it lapse.

ID

p156-texas-governor-data-center-approval-pause-executive-action-feb18

Deadline

By February 18, 2027

Trigger

Hit if, by the deadline, an executive order, proclamation or directive letter signed by the governor and published on gov.texas.gov (or filed with the Texas Secretary of State) directs the PUCT, ERCOT or TCEQ to pause, suspend or withhold approvals, interconnection or energization decisions, or permits for new data-centre loads above any stated threshold (25 MW, 50 MW, 75 MW or otherwise) for a defined period or pending legislative action; a directive that only continues the existing audit without pausing approvals, a request to the legislature without an instrument, or a pause limited to state tax exemptions is a partial; no instrument by the deadline is a miss. The bet sits above even because the pledge was specific (an executive moratorium on all approvals until the 90th Legislature passes the 'Texans First' package), because the outgoing governor's August PUCT / ERCOT audit pause and September TCEQ permit halt have established that these agencies will act on a governor's letter, because the ERCOT audit reports due December 10 give the incoming administration a document to act on, and because the Texas House flipped to the governor's party, lowering the political cost. It is held below 65 because the Texas Senate remains 19-12 Republican and PUCT appointees need two-thirds Senate confirmation, because Jefferies' ~13 GW at risk is the kind of number that draws a negotiated narrowing before inauguration, because the outgoing governor controls the agencies until January 19 and may lift or extend the pause himself (which would change what 'extending' means), and because governors-elect frequently convert moratoria into audits once in office. A hit puts the largest US data-centre market under an executive pause and gives the fifth hypothesis its permitting leg; a miss says the ballot-box signal of November 3 did not survive contact with the state's interconnection queue.

Pending · Capital

Anthropic's IPO prices by December 5, 2026 at an implied post-money valuation of $1.25 trillion or more (the mark OpenAI's concurrent round is reported at) and raises $50 billion or more in primary and secondary proceeds combined, that is, the first frontier-lab listing clears in a week when hyperscaler capex is being punished and the neocloud layer is paying 15%.

Pending; the S-1/A filed November 2 puts Q3 revenue at $16,956M (+47% QoQ, against the +45% Reuters figure the bet was written on) and still leaves the price range blank, the roadshow is reported for the week of November 9, and CNBC reports on November 6 that IPO investors see $1.0-1.25T as more likely than $1.5T after the selloff, which moves the valuation leg toward the partial band before pricing; OpenAI's concurrent round is reported closed at $1.25T for its first ~$11B, so the comparison mark held.

ID

p145-anthropic-ipo-prices-ge-1-25t-ge-50b-dec5

Confidence

61%

Deadline

By December 5, 2026

Trigger

Hit if a Form 424(b)(4) final prospectus for Anthropic, PBC on EDGAR by the deadline states an offering price and share count that imply a fully diluted post-money valuation of $1.25T or more and gross proceeds (primary plus selling-stockholder) of $50B or more. A priced deal at $1.0-1.249T or $30-49.9B is a partial (the listing cleared but below the private-market marks of the week); a priced deal below either lower bound, a postponement past the deadline, or a withdrawn registration is a miss, and the resolver should record which leg failed. The confidence is held just above 60 because the reported timetable (roadshow the week of November 9, pricing the week of November 16; Reuters, IFR) leaves two weeks of slack before the deadline and because the last private round was $65B at $965B in May, so $1.25T is a 30% step-up on a company that reported Q3 revenue up 45% quarter on quarter; it is held below 70 because the valuation talk fell from ~$2T to ~$1.5T inside the filing week, OpenAI's round is reported cut from $30B to ~$20B with ~$11B committed (grade 3), and a $100B raise would be the largest IPO on record into a market that just repriced the hyperscalers. A hit is consistent with the capital stack being tiered rather than closed (it does not prove it, since an IPO prices equity and the closure claim is about debt); a miss on the valuation leg would mean the public market prices a frontier lab's compute commitments (more than $542B on CNBC's tally of the S-1) as a liability before it prices the revenue as an asset.

Pending · Hardware

Intel's fourth-quarter 2026 results (expected late January 2027) report Data Center and AI segment revenue below the third quarter's $7.3 billion, that is, the organic-substrate shortage that Intel says is throttling EMIB output caps the segment sequentially even as demand grows, confirming substrates rather than demand as the binding constraint on the CPU side of the 2027 buildout.

Pending; no Intel disclosure in the window. The adjacent evidence leaned the bet's way: AMD's Venice CPUs began revenue shipments and Arm's data-centre royalties doubled (CPU demand is not the constraint), while Reuters reports Nvidia funding CoWoS expansion at SPIL and Qualcomm's handset line fell 26% on input-cost inflation, memory among the costs named (the supply chain is).

ID

p146-intel-q4-2026-dcai-revenue-below-7-3b-feb5

Confidence

58%

Deadline

By February 5, 2027

Trigger

Hit if Intel's Q4 2026 earnings press release on its investor-relations site (or the Form 8-K exhibit on EDGAR) by the deadline reports DCAI segment revenue below $7.3B and Intel's release or call attributes the sequential move to substrate or packaging allocation (or supply constraint on EMIB) rather than to demand, the Fab 28 shutdown or a segment reclassification; revenue below $7.3B with the cause attributed elsewhere, or with no cause given, is a partial (the number moved, the thesis is unproven). Revenue of $7.3-7.5B (flat within 3%) is a partial; revenue above $7.5B is a miss, as is a restatement of segments that removes the comparable line, in which case the resolver should use the nearest data-centre revenue figure and note the change. The bet is the hardware lens's reading of the October 29 print: Intel's CEO said an organic-substrate shortage is throttling EMIB output with about 40% more substrate capacity arriving mostly in the second half of 2027, the Fab 28 shutdown costs ~$300M in the near term, and the company guided Q4 total revenue to $15-16B against $17.1B consensus, a guide that cannot be met with DCAI growing sequentially unless client revenue falls faster than Intel's own PC-market guide (2026 units down ~12% on memory shortages) already implies. The confidence is held in the high 50s rather than higher because a $7.3B base that grew 78% year on year has momentum, because Intel may prioritise substrate allocation to DCAI over client (the margin arithmetic favours it), because the Q4 guide's shortfall may sit mostly in client, and because the Fab 28 shutdown is a competing cause Intel could name for part of any sequential move, which would score a partial rather than a hit even if the number falls. The bet is a bet about where the substrate shortage lands; a hit confirms the house reading that packaging, not power, is the 2027 constraint and that the memory supercycle's bill is being paid in adjacent silicon.

Pending · Software

Google opens gemini-4-carbon to all developers on the paid Gemini API tier (not only Fairwind Program participants) by December 15, 2026, that is, the cyber-gated rollout of the independent leader lasts no longer than Argon's thirteen days did in October.

Pending and tracking toward a miss: ten days after launch Carbon has no public API model ID (Artificial Analysis and LMArena score it at 60 and #1 through gated access), Google confirmed only Argon's December 1 price step, and the week's other frontier releases (GPT-6 Cyber to Daybreak Red; SWE-2 Flash weights withheld) tightened the industry gate the bet is written against.

ID

p147-gemini-4-carbon-paid-api-ga-dec15

Confidence

38%

Deadline

By December 15, 2026

Trigger

Hit if, by the deadline, Google's Gemini API models documentation (ai.google.dev) lists gemini-4-carbon as available on the paid tier without a Fairwind, allowlist or application requirement, or a blog.google or Google for Developers post announces general developer availability on the paid API; availability only in AI Studio for AI Ultra subscribers, only through Vertex AI private preview, or only for named enterprise customers is a miss on the thesis leg (the gate held). No announcement by the deadline is a miss. The bet sits well below even because Google's own launch post describes a deliberately phased rollout (a subset of cyber defenders first, then the 650-plus Fairwind partners, with background checks and phishing-resistant MFA) and participation in the US government's voluntary pre-release testing process, with no general-availability date, and because GPT-6.1 Terra's Critical cyber classification and Mistral's vetted-partner offensive variant suggest the gate is an industry setting rather than one lab's caution; it is held above 30 because Argon's identical Fairwind-first launch opened to the paid tier in thirteen days, because Carbon's introductory $2 / $10 pricing with a dated reversion to $4 / $20 is a commercial clock, and because Google has competitive reason to ship the independent leader into Antigravity before Anthropic's IPO. A miss confirms the software lens's reading that the frontier's gate has moved from the weights to the distribution list; a hit flips the cache-read lever (Carbon lists cached input at $0.10) and resets the gating thesis.

Hit · Networking

Marvell and Lumentum announce a definitive merger agreement by December 31, 2026, that is, the Bloomberg-reported 'advanced talks' on an all-stock deal valuing Lumentum at about its ~$85B market value convert into a signed transaction within the quarter rather than fading as a single-source report.

Hit on November 2: Marvell's 8-K and joint release disclose a definitive all-stock agreement (4.500 Marvell shares per Lumentum share, $115B implied enterprise value, 23.9% premium to the unaffected close, Lumentum holders ~33%), close expected H1 2028; the signed value exceeded the ~$85B reported, and the bet's 40% was too low on the speed leg.

ID

p148-marvell-lumentum-definitive-agreement-dec31

Confidence

40%

Deadline

By December 31, 2026

Trigger

Hit if a Form 8-K on EDGAR from Marvell Technology or Lumentum Holdings, or a joint press release on either company's investor-relations site, by the deadline announces a definitive agreement for Marvell to acquire Lumentum (any consideration, any price). A disclosed agreement for a minority stake, a supply or joint-development agreement, or a confirmed negotiation without a signed definitive agreement is a partial; a public denial by either company, a report that talks ended, or silence through the deadline is a miss. The bet sits at 40 rather than lower because the report is specific in the way late-stage leaks are: 'advanced talks', an all-stock structure, a value pinned to Lumentum's existing ~$85B market capitalisation rather than a premium, and an announcement described as possible as soon as the week of November 2 (Bloomberg, grade 3). It is held below even because the report is single-source and this issue scores the 'megadeal' frame as noise; because neither company confirmed it within the two days left in the window; because an all-stock deal into a week when the sector sold off 7.2% can reprice before signing; because the combination would face review in at least the US, EU and China, which bidders typically weigh before signing; and because a signed definitive agreement inside nine weeks of a first report remains the exception for deals of this size. The industrial logic (co-packaged optics moving into mass production at ASE this quarter, Celestica's 1.6T ramp, Corning's fibre-per-GPU curve) is the networking lens's own thesis, and because Lumentum's sold-out horizon to 2029 is the kind of order book a buyer wants to own before the CPO transition changes who it belongs to. A hit would be the first consolidation of the optical layer into a networking-silicon vendor and would move the Lumentum lever's object; a miss leaves the lever on the horizon.

Pending · Power

FERC approves NERC's proposed registry criteria for large loads (docket RR27-1: registration of data centres and other loads at or above 50 MW connected at 100 kV or above), in whole or with modifications, by March 31, 2027, that is, mandatory reliability registration for data centres becomes federal policy within five months of the petition.

Pending; comments on RR27-1 are due November 20 and no order has issued. The political wind turned toward the bet: Sen. Heinrich, in line for Energy chair, published a November 5 fact sheet pressing mandatory large-load cost allocation, and S. 5653's November 9 cloture vote carries a ≥20 MW covered-load provision; neither is the FERC order the bet requires.

ID

p149-ferc-approves-nerc-large-load-registry-rr27-1-mar31

Confidence

57%

Deadline

By March 31, 2027

Trigger

Hit if a FERC order in docket RR27-1 (searchable in FERC eLibrary) issued by the deadline approves NERC's proposed registry criteria for large loads, with or without directed modifications, such that registration of loads at or above the proposed thresholds becomes mandatory on a stated effective date. An order that approves the criteria but raises the threshold above 100 MW is a partial (registration becomes policy at a narrower scope); an order rejecting the petition, setting it for further proceedings without approval, or no order by the deadline is a miss. The bet is held just above even because FERC noticed the petition within a week of filing with comments due November 20 (a fast clock), because NERC's workplan treats the registry as the foundation for 2027 ride-through standards and FERC has pressed for large-load reliability action at its own conferences, and because PJM's ~3,900 MW backup-generation event in July is the kind of fact pattern that produces an order; it is held below 65 because the proposed criteria would bring about 86% of large load under NERC's compliance regime and the comment record will include data-centre operators, retail-choice states and distributors arguing scope, and because FERC's rulemaking calendar routinely slips past five months on contested registration questions. A hit gives the synthesis's power hypothesis its regulatory leg from the reliability side even if PJM's cost-allocation filing stalls; a miss would mean the registration layer lags the procurement layer by a year.

Pending · Software

A model with downloadable weights (any licence, any gate, but a checkpoint a third party can run) that had not already scored 40 or above on Artificial Analysis Intelligence Index v5 behind an API before October 31, 2026 scores 40 or above on v5 by December 31, 2026, that is, the open-weights field closes to within 19 points of Gemini 4 Carbon's 59 on the private-set ruler through a new release rather than through Kimi K3.1 Pro publishing weights for a score it already holds.

Pending; Kimi K3.1 Pro's weights (46) are excluded by construction, and the ruler moved under the bet: Artificial Analysis's v5.0.1 regrade on November 3 lifted Gemini 4 Carbon from 59 to 60, so the bet's 'within 19 points' framing is now 20 points to the leader while the 40-point line is unchanged (The Model Pulse carries the regrade). The week's qualifying candidates fell short: Tencent Hy4 scored 35 on v5 but its weights are not yet on Hugging Face, Inkling 2 Small scored 31 and is closed until its promised release, Ripple-65B Preview (Apache 2.0) is unscored on v5, and ERNIE-5.5-Flash-Base's weights shipped without a v5 score; a 35 with weights would land one point below the partial floor (36-39 is a partial; 35 or below is a miss).

ID

p150-aa-v5-downloadable-open-weights-ge-40-dec31

Confidence

46%

Deadline

By December 31, 2026

Trigger

Hit if the Artificial Analysis models leaderboard (artificialanalysis.ai/leaderboards/models) by the deadline shows an Intelligence Index v5 score of 40 or above for any model whose weights are downloadable from Hugging Face or ModelScope on the scoring date, at the reasoning setting the tracker lists as default, excluding any model that the same leaderboard already scored at 40 or above API-only before October 31, 2026. That exclusion removes Kimi K3.1 Pro (46 API-only at launch; its weights release is p142's question, and this bet is written so the two do not resolve on one event); a distilled or smaller checkpoint presented under a frontier name does not count. A best qualifying score of 36-39 is a partial; 35 or below is a miss. The leaderboard figures this bet is written against (the best downloadable scores at launch, GLM-5.3 and DeepSeek V4.1 Flash 1017 both at 32, Kimi K3.1 Pro at 46 API-only) were read from the tracker after the research stage and are outside the graded packet. Either outcome is informative: a hit from a new DeepSeek, Qwen or GLM release says the open side re-based upward inside a quarter, and a miss says the 12-point gap v5 opened is holding at year-end for everyone but Moonshot and the 'four-month lag' reading needs a longer clock. The confidence sits just under even because the gap from 32 to 40 is larger than any single open release has closed on one ruler in the series and the two labs most likely to close it (DeepSeek, Alibaba) have no dated release inside the window; it is held well above the 30s because DeepSeek V4.1 Flash and the Qwen3.8 family have not yet been scored on v5 at their highest reasoning settings, because the tracker's own Progress Index says open models trail by about four months, and because a nine-week window at this cadence usually contains at least one new open frontier checkpoint.

Pending · Capital

By December 31, 2026 a GPU-backed financing other than the Anthropic-Broadcom SPV discloses vendor residual-value support, a repurchase commitment or a put from the chip vendor, in a Form 8-K, offering document or lender press release that names the vendor, that is, the Broadcom backstop is the first instance of a structure rather than a one-off concession to one customer's bankers.

Pending; the two structures disclosed this week both point away from a hit: the SpaceX lease's lender presentation states there is no residual-value guarantee (Bloomberg, grade 3), and Google's TeraWulf credit support is a guarantee of a lab's lease by its compute supplier and customer-affiliate on a data-centre lease, not residual-value support from a chip vendor on a GPU financing, so it does not qualify under the trigger as written, which requires residual-value support from a chip vendor on a GPU financing 'for a customer other than Anthropic'. Correction to the W44 rationale: NVIDIA has written residual-value support before (the August 17, 2026 8-K on OpenAI's SB Energy leases, outside the graded packet); that instance predates the bet and is excluded, but it means the structure exists on the vendor side and the open question is whether it reaches 'a customer other than Anthropic' through a new transaction by December 31.

ID

p151-second-gpu-financing-with-vendor-residual-support-dec31

Confidence

45%

Deadline

By December 31, 2026

Trigger

Hit if, by the deadline, an 8-K, 10-Q, prospectus, offering memorandum excerpt reported by IFR, Bloomberg or Reuters, or a press release from an arranger or lessor discloses that a chip vendor (NVIDIA, AMD, Broadcom on a new transaction, Google for TPUs, or any other accelerator vendor) provides residual-value support, a repurchase or remarketing commitment, a put, or a guarantee of lease or loan amounts on a GPU or accelerator financing of $1B or more for a customer other than Anthropic. A vendor equity investment in the borrower, a vendor supply commitment, or a vendor-arranged loan without residual or repayment support is a miss on the thesis leg; a disclosure that names 'a vendor' without identifying it is a partial; silence through the deadline is a miss. The bet sits just under even because the two Broadcom instances (June's $29B at an 85% reference, October's ~$43B at 72.5%; the June terms are outside the graded packet) are both to one customer, because NVIDIA's one disclosed instance of residual support (its August 17, 2026 8-K guaranties on OpenAI's SB Energy leases, outside the graded packet; it predates this bet and is excluded from it) was written for the lab it also holds equity and offtake with, so it is not yet evidence of a market structure, and because a vendor put is a contingent liability a chip vendor's auditors and board have to approve. It is held above 40 because Oracle's investor day named 'vendor financing' as the first of three sources that will replace its own raising, because NAVER's $9B SPV already pairs a GPU vehicle with a vendor equity stake, and because the SPV's first lien is reported syndicating inside SpaceX's lease rate (IFR, outside the graded packet), which tells every other lab's bankers what a vendor floor is worth. A hit confirms the big story's reading that residual risk is being allocated by structure onto the vendor; a miss says the Broadcom backstop was the price of one IPO-week syndication and the lessor's no-residual rent remains the market's default.

Cumulative record

Predictions made

156

Resolved

85

Outcomes

37 hit · 20 partial · 28 miss

Hit rate (partial = half)

55%

Brier score (0 = perfect)

0.192

Overdue, unresolved

0

Hit · 40% called

Marvell and Lumentum announce a definitive merger agreement by December 31, 2026, that is, the Bloomberg-reported 'advanced talks' on an all-stock deal valuing Lumentum at about its ~$85B market value convert into a signed transaction within the quarter rather than fading as a single-source report.

Hit on November 2: Marvell's 8-K and joint release disclose a definitive all-stock agreement (4.500 Marvell shares per Lumentum share, $115B implied enterprise value, 23.9% premium to the unaffected close, Lumentum holders ~33%), close expected H1 2028; the signed value exceeded the ~$85B reported, and the bet's 40% was too low on the speed leg.

Deadline

By December 31, 2026

Hit · 63% called

Artificial Analysis publishes a Terminal-Bench 4.0 result for DeepSeek V4.1-Flash-1017 by November 30, 2026, and it is below 35%, that is, more than seven points under the 42.3 the model card reports from the Claude Code harness.

Hit, with a caveat on timing. Artificial Analysis's comparison page for V4.1-Flash-1017 against V4.1-Flash (https://artificialanalysis.ai/models/comparisons/deepseek-v4-1-flash-1017-vs-deepseek-v4-1-flash; found at fact audit after the graded source set closed, and dated to the tracker's October 18 changelog) reports Terminal-Bench 4.0 at 34% for the 1017 snapshot against 27% for the prior Flash, at $0.26 per task against $0.48. 34% is below the 35.0% threshold, eight points under the 42.3 the model card reports from the Claude Code harness, so the result leg is a hit; the vendor-to-tracker gap (8.3 points) is wider than the Haiku 5.5 reference (6.4). The caveat: the result was published October 18, two days before this bet was authored, so it was knowable in the window and the issue should have scored it at authoring; the bet is scored a hit on its stated condition and the lesson is recorded in the changelog. DeepSeek's DeepInfer release naming V4.1-Flash-1017 as its default production checkpoint remains the indirect datapoint that the snapshot is the one being served.

Deadline

By November 30, 2026

Hit · 71% called

Anthropic makes Claude Sonnet 5.5 or Claude Haiku 5.5 generally available on the Claude API by November 30, 2026.

Claude Sonnet 5.5 went GA on September 28 with API model id claude-sonnet-5-5 listed on the Claude Platform models page and available on Bedrock, Google Cloud and Microsoft Foundry. Haiku 5.5 remains 'coming weeks'.

Deadline

By November 30, 2026

Hit · 84% called

CoreWeave completes at least $3 billion of the convertible note offering announced September 17 by November 30, 2026.

The September 22, 2026 8-K states completion of $4.2 billion aggregate principal, including the $500 million option. Net proceeds are reported around $4.137 billion. Both figures clear the $3 billion bar, ahead of the deadline.

Deadline

By November 30, 2026

Hit · 66% called

Samsung's final Q3 2026 results on October 29 disclose Device Solutions (semiconductor) division operating profit of KRW 100 trillion or more.

Samsung's Oct 29 Q3 2026 release and segment table showed Device Solutions operating profit of KRW 106.5 trillion on KRW 147.1 trillion of sales (72% margin), above the KRW 100T bar, with consolidated operating profit at KRW 107.4T and DX posting a KRW 0.8T loss — the near-breakeven rest-of-company split the trigger anticipated.

Deadline

By October 31, 2026

Sources Resolution evidence

Partial · 44% called

Meta contributes MetaRoCE specification through OCP at the October 2026 Global Summit with documented production deployment targets beyond the 64-node AMD proof-of-concept, by October 31, 2026.

Spec leg hit: Meta published MetaRoCE-Spec-1.0 with the libsoftmetaroce reference implementation in the facebookincubator/metaroce repository in October and presented the transport specification at the OCP Global Summit, as its Aug 24 post promised. Deployment-target leg missed: Meta's Oct 12 and Oct 26 OCP posts call the stack 'production-proven' but name no hyperscaler or cloud deployment timeline beyond the AMD Pensando work, AMD's own post speaks only of a path from Pollara 400 to Vulcano 800 toward 800G deployment, and the OCP contributions database showed no MetaRoCE entry by the deadline.

Deadline

By October 31, 2026

Sources Resolution evidence

November 9, 5:30 pm ET

Senate cloture on H.R. 2347 / S. 5653 (permitting vehicle with the ≥20 MW covered-load cost-allocation bar)

Needs 60 votes at 20-25% odds (Rapidan via Axios); failure sends the same cost-allocation principle to a January Senate with Heinrich at Energy, which is the federal leg of PJM's RBP design and the context for p149.

November 10

CoreWeave Q3 2026 results

Guided interest expense of $860-940M against adjusted operating income of $200-260M, 2032 notes above 16% and CDS ~1,155 bp; the print is the within-issuer test of the spread lever and the first look at how the ladder's bottom rung funds a $104B backlog Bernstein says needs ~$85B more.

Week of November 9 through December 5

Anthropic roadshow and pricing (p145)

The S-1/A's price range is still blank; CNBC reports investors see $1.0-1.25T as more likely than $1.5T, which is the partial band for p145; the prospectus must also update any RSP acceleration-threshold determination (p155).

November 20 and November 30

Comment deadlines: NERC RR27-1 large-load registry (November 20) and PJM ER27-330 RBP refile (November 30)

The two dockets that decide whether registration and backstop procurement for large loads become federal policy on PJM's timetable; the Market Monitor's next adjusted-shortfall figure will appear in the ER27-330 record and moves the cancelled-load lever.

December 1 (Argon); December 4 (SWE-2 Flash)

Two dated price steps: Gemini 4 Argon reverts from the introductory $2 / $10 to $4 / $20; SWE-2 Flash moves from free in Devin to $0.10 / $0.50 per million tokens

Argon's step would be the first list increase inside the $2 / $10 band in the series and takes the band to three labs unless Google extends; Carbon's API status by December 15 decides p147 (38%). Separately, SWE-2 Flash's first paid price tests whether Cognition publishes the withheld weights alongside it (the Model Pulse tracks the gate).

December 10

ERCOT data-centre audit reports to the PUCT

The first official count of how much announced Texas load the audit treats as real, the document the incoming governor's pledged pause would act on (p156), and the Texas analogue of the PJM Monitor's cancelled-load figure.

Next sourced print

Broadcom five-year CDS absolute level

CNBC reports a 49% rise over the month without a level; applied to the W42 base of 133-136 bp that is ~200 bp, the lever's threshold, but the W42 print and Deutsche Bank's month-ago base may differ, so the implied level is house arithmetic. A sourced quote either confirms the market is pricing Broadcom as a residual-value financier or returns the lever to flat.

• Lede rule: the big story's house-original contribution is the pricing of the Broadcom rent facility as capital (11.3% at maturity, 14.9% at two years, 21.1% at one year, 32.9% and 16.7% under penalty terms, the 150% change-of-control put; the headline surfaces lead with the 11.3-14.9% range and compare it to Oracle's 11.2% equity discount, not to coupons) and the placement of the week's six in-window financing events (two bond prints, one equity print, one guarantee signed, one facility's terms disclosed, one lease entering syndication) on one credit-quality ladder, built on the house measurement, four filings and two market reports. Frames that circulated first are credited in the body or here: Haircut ('Who actually pays if the AI data centre bet goes wrong', nodiligence.substack.com, October 7, 2026), UBP (September 16, 2026), a September 2026 single-campus bond survey and Clifford Chance's three-bucket framework (via Luminix) and JPMorgan's James Sullivan (CNBC, November 5) for sorting data-centre debt by who bears the risk; Reuters (October 1) for the one-third coverage ratio on the Broadcom facility; Mosaic Singularity and alfw.ai (November 5) for reading Google's TeraWulf backstop as a silicon-distribution contract; CNBC and Nomura for the midterms repricing the buildout as regulatory risk (Nomura for the 2028-and-later timing) and Stratechery (November 4) for the political-constraint reading before the selloff; PPC Land and Fortune for reporting the Broadcom NPA's 9% PIK terms (the compounding arithmetic in this issue is the house's); The GPU newsletter for the observation that Google stands behind both the operator and tenant sides of Anthropic's build. Nomura, Stratechery, Business Insider, PPC Land, The GPU, Haircut, UBP, Luminix, Mosaic Singularity and alfw.ai are outside the graded packet and labelled as such where used. The Reuters October 1 report on the Broadcom facility, cited in the body, is also outside the graded packet. whereWeDiffer is omitted this week.
• House measurement: filing-derived; the implied all-in cost of the $42B Broadcom NPA by holding period, its coverage of the $125.2B lease, the lease's rent relative to Q3 P&L scale, dilution, and the 150% change-of-control put, with method, readings, implication, caveats and six sources. The implied costs are costs of equity issuance through a note, not borrowing rates; the issue compares them to Oracle's equity discount and states that the short-dated paths depend on the holder's election to convert. The comparison coupons (Oracle 7.77%, Meta 6.60%) and Oracle's 11.2% equity discount are computed in the measurement from the FWPs and 424B5 and are cited from it wherever they appear in this issue. The $125.2B lease size, the 150% put and the Google and Amazon NPA comparison terms are read from the S-1/A after the research stage; the packet carries the 9% PIK terms.
• Evidence labels: all 71 events in the graded packet carry their grades (1-5, where 5 is a primary filing or first-party document and 3 is single-outlet reporting); grade-3 items used in this issue are labelled at each use (Bloomberg on SpaceX pre-marketing and OpenAI's ~$11B close; Reuters on Nvidia-SPIL packaging and the H200 stockpile; Bloomberg via Taipei Times on Micron Fab 11E; Jefferies' ~13 GW Texas estimate via King & Spalding). Vendor-reported figures are labelled at each use (AMD's 2027 >80% guide; Coherent's ~$15K per 100 Tbps content estimate; Marvell's $550B TAM and >$1B synergies; OpenAI's $1.21M exploit cost and 948 LPE count; Moonshot's cyber curb claim and ExploitGym 20.7). Two documents were read after the research stage and are labelled outside the graded packet at each use: Exhibit 10.1 to the TeraWulf 8-K (s.2(k) TPU requirement; s.4(e) six-year limit on the termination-fee election) and NVIDIA's August 17, 2026 8-K (SB Energy residual-value guaranties). TrendForce's 2027 interconnect forecast is graded 2 and scored noise.
• Capital flow: Frontier Labs charted at zero (OpenAI's ~$11B close is grade 3 and stays in text; the Google-TeraWulf credit support and the Broadcom NPA are credit, not capital) with revenue out re-based to ~$68B annualised on the filed Q3 figure of $16,956M; Hyperscaler-Hosted charted at ~$35B (Oracle $20B, Meta $15B notes; the forward-sale option proceeds are not cash and are not charted); Neoclouds charted at ~$2.2B (Cipher $850M convertible, Fal $1.3B Series F, both grade 4) with first operating prints recorded; On-Prem / Hybrid at zero with Drax's GBP 2.5B phase 1 in text pending FID. Direction arrows follow the chart values.
• Levers: fourteen levers rolled forward. Changes: OpenAI pause counter to 48 days with GPT-6 Cyber noted as a gated release consistent with the pause; Broadcom CDS marked 'up, level unconfirmed' on CNBC's reported +49% monthly move (house arithmetic on the W42 base gives ~200 bp, the threshold, and the issue records direction not crossing); Lumentum lever's object moved under Marvell's definitive agreement with a change-of-control question added; mid-tier band records Argon's confirmed December 1 step; spread lever restates the top-of-distribution arithmetic once (CoreWeave 2032s >16% vs Lambda 6.78% ≈ 920 bp; vs Oracle 8.65% ≈ 735 bp) with the issuer-tier caveat. Eleven levers flat on their own definitions.
• Predictions: p148 scored hit (Marvell-Lumentum definitive agreement at $115B EV on November 2; the 40% was too low on speed). p145, p146, p147, p149, p150, p151 remain pending with one-line notes on the week's evidence; p147 is tracking toward a miss; p151's W44 rationale is corrected in its notes (NVIDIA has written residual-value support before, in August, outside the packet and predating the bet). W43 bets not carried in this issue's prior block are resolvable on this week's evidence and pending the ledger: p142 (Kimi K3.1 Pro weights under a gated licence by November 30) and p143 (Coherent reports no cancellations) both resolved as hits on November 4; p132 (Mistral Large 4 weights by November 15) has no evidence in the packet. Five new predictions (p152-p156) across capital, hardware, networking, software and power lenses; p153 (40%) and p155 (48%) sit outside the 55-80 band; p153 overlaps p141 (NVIDIA prepaid supply up $3B) by design, p141 asking for the aggregate and p153 for the counterparty, and a p141 hit attributed to packaging scores p153 as a partial. No vendor-schedule calendar entries: Inkling 2 Small's promised weights, Argon's December 1 step and CoreWeave's November 10 print are watchlist items, not bets.
• Synthesis: three connections (Gilder test at the optical component layer on a datacenter-segment basis, inductive, two of four optical lines outgrowing AMD with Astera counted as electrical connectivity rather than optics; the credit-quality ladder, abductive, with the guarantee rung and the vendor-loan rung marked unpriced; cyber classification as the distribution gate, inductive, with the three labs' gating quantities distinguished), each with a steel-man and a falsifier. Thesis test for all five hypotheses, graded on the thesis page's own refutation conditions: H1 supported on METR's 1.9x and AMD's guidance, graded on the rate rising rather than on a doubling-time slope, with the regrade and the open-weights step excluded as evidence (one is a ruler correction, the other is H4's); H2 strained for an eighth consecutive week; H3 held at strained on the W44 operator-layer rule (Equinix interconnection outgrowing colocation for two consecutive quarters, second leg p154) rather than flipped on a comparator redefined mid-test, with the component-layer result reported for information; H4 strained, because the downloadable floor rose thirteen points to 46 while the gap to the gated leader is fourteen, above the refutation condition's ten, in the first of the two quarters the condition requires; H5 strained for an eighth week. Proposed restatements, recorded here for the editor's decision and not applied to the thesis page: H2 direct-capture leg, 'capital is concentrated and returns are being captured one layer up the supply chain and at the labs' largest customers, with prepayment travelling up the stack' (refuted if two consecutive quarters of supplier prints one layer up show margins falling while hyperscaler AI-segment margins rise); H2 diffuse-capture leg, 'economy-wide returns are not yet measurable on central-bank and statistical-agency data and should be tested there, not on vendor metrics' (refuted if a G7 central bank or statistical agency attributes one percentage point or more of annual productivity growth to AI; the next dated reads are the Bank of England's February 2027 Monetary Policy Report and the US BLS fourth-quarter productivity release in early February); H5, 'power and permitting are the binding constraint on capacity from 2028, when the announced pipeline must be energised; through 2027 the binding constraint is advanced packaging, HBM and substrates, and the cost of power is being allocated to large loads by regulators before it binds physically' (refuted if two consecutive quarters of supplier prints name power rather than packaging or memory as the 2026-27 constraint, or if a material grid expansion shortens interconnection queues within twelve months). Three pattern-watch items extended (non-list meters to ten weeks; label-versus-artifact to nine weeks, logged mixed again; returns one layer up to five weeks). Three second-order effects, with AMD recorded as the fourth instance of vendor credit for offtake (vendor as tenant at Core Scientific, warrants to customers) rather than as a counter-example, and the warrant-driven P&L entries (AMD, Core Scientific, TeraWulf, Google's Raylan path) added.
• Living documents: The Model Pulse carries the LLM tree delta (Kimi K3.1 Pro updated to open weights; Ripple-65B Preview, ERNIE-5.5-Flash-Base, GPT-6 Cyber and Tencent Hy4 added; SWE-2 Flash, Inkling 2 Small and Images 2.6 dispositions recorded). Shockwave Timeline and Market Architecture unchanged: no event this week reset a frontier assumption in efficiency, reasoning, context, open frontier or interoperability (Kimi K3.1 Pro's weights are an open-frontier release but the tree already carries the model and the capability was API-scored in W43). Futures Studio regenerated from the current levers.
• Where each fact lives: this issue owns the credit-quality ladder, the Oracle and Meta pricing, the Google-TeraWulf credit support, the SpaceX pre-marketing, the CoreWeave yield and CDS tape, the Anthropic S-1/A financing and commitment figures, the AMD, Supermicro, Eaton, Qualcomm and Arm prints, the Marvell-Lumentum agreement, the Coherent, Arista, Astera, AOI and Lumen prints, the midterm map and its state-level permitting consequences, and the PJM, NERC and Senate dockets. The Model Pulse owns Kimi K3.1 Pro's artifact and licence, GPT-6 Cyber's specification and gating, SWE-2 Flash, Ripple-65B, Hy4, ERNIE-5.5-Flash-Base, Inkling 2 Small, Carbon's scores and the v5.0.1 regrade. The Application Layer owns the S-1/A customer metrics, Palantir, HubSpot, Shopify and Uber's agentic-commerce figures, Microsoft's Copilot caps, OpenAI's dots billing, Goldman's tool restriction, the Bank of England finding and the ServiceNow and SUSE cuts. Agent Techniques owns the Claude Code, Codex, Antigravity and open-harness releases, Anthropic's platform controls, the LLMjacking and ARTEX findings (the credential-revocation reading that Goldman's tool restriction sits beside), the Postman and ThisBash incidents, Pando and Pierre, the Harness Leaderboard, and Google's CNCF donations. Each appears elsewhere as one clause and a pointer.
• Revision cycle 2 (2026-11-07): corrections applied after board review. Headline surfaces (pull quote, first takeaway, by-the-numbers, house-measurement headline and first reading) now lead with the facility's 11.3-14.9% cost as equity issuance against Oracle's 11.2% equity discount, with the one-year 21.1% path labelled as the holder's election; coupon multiples removed. SpaceX restated as ~$40B of financed principal plus a ~$7B lease prepayment. The $567M operating loss carries the $2.0B non-cash donation-matching charge and the ~8.5% ex-charge margin wherever it is cited. The unguaranteed rung restated: the most levered operator's paper above 16% and five IPOs pulled, while smaller unguaranteed operators (Cipher, IREN) still cleared paper. Astera reclassified from optics to electrical connectivity; the Gilder comparison moved to a datacenter-segment basis (two of four optical lines outgrew AMD; AOI datacenter +254%, company +127.5%); H3 held at strained on the W44 operator-layer rule pending p154 rather than flipped on a redefined comparator; the first connection's confidence lowered to 55 and the second's to 60. Big story cut from ~1,790 to ~1,200 words with the house measurement and the guarantor-sort credits (Haircut, UBP, Clifford Chance via Luminix) in the first two sentences, the in-window count reconciled to six events, and the decisions moved ahead of word 700; readingTimeMin recomputed. The p140 mislabel corrected in the second connection and the IG coupon lever (p140 tests the tranche gap; the inside-Oracle test is this issue's own). 'As the stated reason the gate exists' replaced with the dated sequence. Spread gap quoted at matched tenors. AMD recorded as the fourth vendor-credit instance rather than a counter-example, with the warrant P&L entries added. Nomura's and the NVIDIA 8-K's 'outside the graded packet' labels reduced to first use and the transaction and changelog entries.