The AI Stack Weekly

Issue 13 · Week 29 of 2026.

July 18, 2026/Industry brief · ~7 min read/Public sources onlyDownload brief

The Bottom Line

The open frontier arrived priced like the closed one: Kimi K3 hit #4 on the independent index and #1 on a flagship arena at $3/$15 — while TSMC, PJM, and a first-ever state moratorium confirmed the constraints have moved fully downstream of silicon.

Flywheel arcAll three lenses

Moonshot AI's Kimi K3 (July 16) is the first model slated for open weights to reach the closed-frontier tier on independent evaluations — 57.1 on the Artificial Analysis Intelligence Index (#4 of 189, within three points of Claude Fable 5 and GPT-5.6 Sol) and #1 on LMArena's Frontend Code Arena at 1,679, ahead of both closed flagships. But it launched hosted-only at $0.30/$3/$15 per MTok — roughly 3x its predecessor and at parity with Claude Sonnet 5 — with weights and license promised by July 27. Add DeepSeek's first-of-its-kind peak-hour 2x surge pricing and the same-week Apache-2.0 release of Thinking Machines' 975B Inkling (weights actually live on day one), and the 'Chinese models are 10x cheaper' assumption in procurement models is over: open-weight labs now price like the frontier vendors they have become, and the differentiation is shifting to who actually ships weights. Meanwhile the physical layer told the other half of the story. TSMC posted a record $40.2B quarter with HPC at 66% of revenue, raised 2026 capex to $60-64B, added $100B to Arizona — and said advanced packaging capacity is 'so tight that now it's limiting my customers' growth.' PJM's 2028/29 capacity auction cleared at the FERC cap for the third straight year while coming up 6,831 MW short, and New York signed the first statewide data-center permit moratorium a day later. Compute itself started trading across former battle lines: Anthropic in reported talks to lease up to $10B of capacity from Meta, while GPU-backed credit institutionalized ($775M Nebius facility at SOFR+250, a Fitch-rated $2.6B CoreWeave DDTL, the first inference-ASIC-collateralized loan). Boards should re-price single-vendor model dependency against a credible self-host option arriving July 27; investors should note credit and equity now tell opposite stories about the same GPU assets; architects and operators should treat packaging, power, and market structure — not chips or capital — as the binding constraints for 2027 planning.

The three lenses

What moved this week, and what to do about it.

14 events across the flywheel — 5 software, 5 hardware, 4 networking.

Software.

Jul 13
OpenAI's GPT-5.6 family (Sol, Terra, Luna) reached GA on Amazon Bedrock via the bedrock-mantle Responses API at OpenAI first-party rates counting toward AWS commitments — with 272K context, 90% prompt-cache discounts, and Sol limited to two US East regions
AWS What's New; AWS Bedrock model cards
Jul 15
Thinking Machines Lab released Inkling — a 975B-parameter (41B active) Apache 2.0 multimodal MoE with 1M context, the largest US-origin open-weights model — with weights live on Hugging Face day one, BF16 plus calibrated NVFP4 checkpoints, and day-0 vLLM/SGLang/llama.cpp support
Hugging Face blog (co-published); Latent Space
Jul 15
OpenAI disclosed GPT-Red, an internal-only self-play RL red-teaming model that beat human red-teamers 84% vs 13% on indirect prompt-injection scenarios and hardened GPT-5.6 Sol to a 0.05% direct-injection failure rate — and said it will not be released due to offensive capability
OpenAI; The New Stack; Help Net Security
Jul 16
Moonshot AI launched Kimi K3 — a vendor-reported 2.8T-parameter MoE (16 of 896 experts active) with 1M context, native text/image/video input, and Kimi Delta Attention — hosted-only at $0.30/$3/$15 per MTok, with open weights promised by July 27
Moonshot/Kimi blog; Simon Willison; MarkTechPost
Jul 16
Kimi K3 debuted at 57.1 on the Artificial Analysis Intelligence Index (#4 of 189, vs Fable 5's 59.9 and GPT-5.6 Sol's 58.9) and took #1 on LMArena's Frontend Code Arena at 1,679 — the first open-weight-lab model to top an LMArena flagship board
Artificial Analysis; Arena.ai leaderboard

What this means

The Jevons arc ran hot this week: near-frontier capability got dramatically cheaper to acquire (K3 independently measured at $0.94 per Index task; Inkling free to self-host under Apache 2.0), and both flagship open releases ship 4-bit-native targeting very large self-hosted footprints — Moonshot recommends 64+ accelerator supernodes, Inkling needs ~600GB in NVFP4 — pulling inference demand toward private clusters and their memory and interconnect fabrics. Architects should hold procurement decisions on the K3 tier until the July 27 weights-and-license drop resolves, and see the Model Pulse for the full architecture read on why routing stability and prompt-cache hit rates are the new cost levers.

Hardware.

Jul 15
ASML Q2: EUR 9.3B net sales above the high end of guidance, system sales split 51% logic / 49% memory, ~65 Low-NA EUV shipments planned for 2026 (+45% YoY EUV system sales), 2027 Low-NA capacity 'close to fully covered' with +30% expansion planned
ASML Q2 2026 press release and investor call
Jul 15
Intel Foundry became the first to ship high-volume logic patterned with High-NA EUV — Panther Lake layers on Intel 18A, dual-qualified in Oregon at yields matching the standard NXE platform; TSMC has said it won't run High-NA in production before ~2029
ASML press release (joint with Intel Foundry)
Jul 15
NVIDIA's Huang, in Tokyo, denied the Rubin delay reports: 'Vera Rubin is already in production. Giant amounts of production incoming' — but named no customer-delivery date; Japan's FRONTia AI factory (27,500 Rubin GPUs) was announced the next day
Tom's Hardware; multiple outlets
Jul 16
TSMC Q2: record $40.2B revenue (+33.7% YoY), HPC at 66% of revenue, 2nm at 3% of wafer revenue in its first material quarter, FY26 capex raised to $60-64B, $100B added to Arizona — and CEO C.C. Wei: advanced packaging capacity is 'so tight that now it's limiting my customers' growth'
TSMC 2Q26 management report (SEC 6-K); earnings call
Jul 15
The Information: Google reportedly pitching TPUs to NVIDIA-centric neoclouds — offering to financially backstop TPU data centers and rent capacity back; Nebius, Lambda, and CoreWeave publicly demurred
The Information via WinBuzzer (secondary)

What this means

Both in-window prints confirmed the same structure: demand visibility is lengthening (TSMC guiding to slightly above 40% full-year growth; ASML nearly fully booked on 2027 EUV) while the binding constraint migrated downstream to advanced packaging — the CEO of the world's foundry said so on the record. Operators should treat Wei's packaging warning as the arbiter of the Rubin rack-delivery dispute, and watch Intel's High-NA production first as the first credible fork in leading-edge lithography cadence in a decade; if it converts to 14A customer wins, the Huang's-law slope gets a second supplier.

Networking.

Jul 14
Tower Semiconductor announced a ~$3B (net of $1B Japan government grants) dual-track 300mm silicon photonics and SiGe capacity expansion in Japan with METI support — targeting $3.6B revenue in 2028 on optical interconnect demand
Tower Semiconductor press release
Jul 15
3M and Microsoft announced a strategic partnership making Azure the first announced hyperscale cloud to deploy 3M Expanded Beam Optical fiber connectivity across AI data centers — attacking contamination-sensitive connector install and service time
Microsoft Source newsroom; 3M investor release
Jul 16
Japan's METI-backed FRONTia project launched national AI infrastructure — a 140 MW Vera Rubin AI factory (27,500 Rubin GPUs, 13,750 Vera CPUs) scaled with Spectrum-X Ethernet fabric, with ¥1T (~$6.2B) of government support over five years
NVIDIA Newsroom; Japan Times; NHK
Jul 16
Chelsio launched its seventh-generation AI Interconnect Platform with native 400Gb Ethernet and unified RDMA (RoCEv2 + iWARP) — widening the pool of standards-based NIC endpoints that can terminate open Ethernet AI fabrics
Chelsio Communications press release

What this means

The center of gravity this week was optics manufacturing capacity and deployment velocity rather than protocol wars: a $3B state-co-funded foundry bet on photonic ICs (Gilder — bandwidth supply industrializing ahead of demand) and a hyperscaler attacking the unglamorous fiber-connector bottleneck that throttles how fast bandwidth physically installs. A sovereign program standardizing on a single vendor's Ethernet fabric shows national fabrics becoming Metcalfe machines — every domestic lab and manufacturer that plugs in multiplies shared-fabric value, and deepens the lock-in debate at state scale. Watch the colo interconnect Q2 prints starting Jul 29 for the revenue confirmation.

Capital flow

Money in, revenue out.

4 categories tracked. Capital deployment up in 3 of 4; revenue follows at multiples of 0.21 to 0.6.

The four-category scorecard. Where capital is going in, where revenue is coming out, and how much of it is real. The one chart for the boardroom.

What’s real, what’s noise.

5 claims this week — 4 signal, 1 noise.

Each claim is scored 1–5 on source quality and triangulation. Anything 2 or below is flagged as noise. Where consensus is wrong, we say so.

5 / 5
PJM's 2028/29 capacity auction cleared at the $325/MW-day FERC cap for the third straight year — and still came up 6,831 MW short of the reliability requirement, with only 525 MW of new generation clearing.
Sources: PJM 2028/2029 BRA results report (primary); Utility Dive; Modo Energy
Real, and the headline understates it: the price fell 2.5% only because the cap itself fell, while PJM's own uncapped simulation cleared 71% higher ($554.72 RTO / $776.69 ComEd) — the gap between capped and simulated price has widened three auctions running. Anyone budgeting AI load in the largest US grid should watch PJM's FERC filings this month for the backstop auction and data-center connect-and-manage framework, not the headline price.
4 / 5
TSMC will spend another $100B in Arizona (total $265B) after a record Q2 — revenue $40.2B (+33.7% YoY), net income +77% YoY, FY26 capex raised to $60-64B.
Sources: TSMC SEC 6-K and earnings call (Q2 results grade 5); NYT, Reuters (Arizona commitment)
The financial capacity is audited and extraordinary; the commitment is real but the timeline is elastic — CEO C.C. Wei explicitly tied fab timing to 'how market conditions develop,' and the announcement's political utility is part of its purpose. Treat $265B as a ceiling with option value, not a schedule; groundbreaking dates and tool orders are the confirmation to watch.
4 / 5
Kimi K3 reached the closed-frontier tier on independent evaluations — 57.1 on the Artificial Analysis Index (#4 of 189) and #1 on LMArena's Frontend Code Arena — as an open-weights model.
Sources: Artificial Analysis (independent eval); Arena.ai leaderboard; Moonshot launch blog
The placements are genuinely independent and real; the 'open' framing is promissory — K3 launched hosted-only, the license text is unpublished, the HF repo 404'd, and Artificial Analysis classifies it proprietary until the July 27 weight drop lands. Buyers should also note effective cost exceeds sticker: always-on max-effort thinking consumed 130M tokens on the Index eval versus a 63M peer average. Hold the procurement conclusion two weeks.
3 / 5
New York became the first state to impose a data-center moratorium, halting AI buildout in the state.
Sources: EO 62 text (grade 5 for the event); Data Center Knowledge; law-firm client alerts
Directionally real, rhetorically inflated: the order pauses only discretionary state environmental permits for facilities ≥50 MW, exempting completed applications, ministerial permits, and local approvals. New York is a minor AI-capacity market — the significance is precedential, landing one day after PJM confirmed the shortage. Copycat orders in Virginia, Texas, or Georgia would change the read from symbolic to structural.
2 / 5 — noise
Meta and Anthropic have a $10B, two-year compute leasing deal.
Sources: NYT (three unnamed sources); CNBC; CNN — CNN's source calls the figures 'speculative'
Hype as priced: the talks are real (multiple outlets independently confirm talks), but the tradable 'fact' the market moved on — $10B — is explicitly disputed, terms are in flux, and no term sheet exists. A buyer-proposed monthly-payment lease with walk-away rights is an overflow-capacity option, not a strategic commitment. The direction it signals (labs renting rivals' GPUs, Meta seeking compute revenue) is real; the number is not yet.

Synthesis

The week, reasoned through.

3 cross-domain connections, 5 hypotheses tested (0 under pressure), 3 patterns tracked.

Reporting says what happened; this section says what it means when you put the pieces together. Every inference is labeled by type, linked to its evidence, and held against the working framework — so when the reasoning is wrong, you can see exactly where.

Connecting the dots

Deductive
77%
confidence
The 'Chinese models are 10x cheaper' assumption in enterprise procurement models ended this week — Chinese open-weight labs now price like the frontier vendors they have become, and the differentiation is shifting from price to who actually ships weights.
1. 01Moonshot priced Kimi K3 at $0.30/$3/$15 per MTok — roughly 3x its K2.6 flagship tier and at parity with Claude Sonnet 5's $3/$15 — the same day independent evaluations placed it #4 of 189 on the AA Intelligence Index.
2. 02DeepSeek introduced the first time-of-day surge pricing from a major model API (2x listed rates during Beijing peak hours, effective Jul 24 alongside its legacy-alias retirement) — importing electricity-market demand management into inference pricing.
3. 03The counter-move came from a US lab: Thinking Machines shipped Inkling under clean Apache 2.0 with weights live on day one — while K3's weights remained promissory (license unpublished, HF repo 404, AA classifying it proprietary until the Jul 27 drop).
Evidence: Kimi K3 pricing docs and launch analysis · DeepSeek surge pricing and alias retirement notice · AA Intelligence Index K3 placement · Inkling Apache 2.0 day-one release
Abductive
68%
confidence
Compute is becoming a traded commodity across former battle lines — the market structure is shifting from vertical exclusivity to a lease-and-credit market for GPU capacity, with credit and equity now telling opposite stories about the same assets.
1. 01Anthropic entered reported talks to lease up to ~$10B of capacity from Meta — a top-2 lab renting a rival's fleet — days after Meta doubled Hyperion to >$50B, meaning surplus hyperscaler capex is being productized as a fifth cloud.
2. 02The debt market institutionalized in the same week: Nebius closed an SEC-filed $775M GPU/contract-backed facility at SOFR+250, Fitch rated CoreWeave's $2.6B take-or-pay DDTL at BB+, and Upper90 wrote the first inference-ASIC-collateralized loan.
3. 03Meanwhile neocloud equity kept de-rating on exactly this development (CoreWeave down ~35% since the Meta Compute report) — lenders are underwriting contracted compute cash flows at institutional spreads while equity reprices the same assets' concentration risk downward.
Evidence: Anthropic-Meta compute lease talks · Nebius $775M secured facility (6-K) · Fitch CoreWeave DDTL rating note
Deductive
81%
confidence
The binding constraints on AI deployment have fully migrated downstream of silicon and capital — in one week, the world's foundry said packaging caps its customers' growth, the largest US grid cleared short at an administratively capped price, and a state added the first policy-layer pause — while capital itself flowed unimpeded.
1. 01TSMC posted a record quarter, raised capex to $60-64B, and added $100B to Arizona — capital is abundant — while CEO C.C. Wei stated advanced packaging capacity is 'so tight that now it's limiting my customers' growth.'
2. 02PJM's 2028/29 auction cleared at the FERC cap for the third straight year, 6,831 MW short of the reliability requirement, with PJM's own uncapped simulation 71% higher — administrative price suppression masking a worsening shortage.
3. 03New York signed EO 62 the next day — the first statewide data-center permit moratorium — converting local opposition into a governor-signed template exactly as the grid data confirmed the scarcity that fuels it.
Evidence: TSMC Q2 6-K and earnings call (packaging commentary) · PJM 2028/29 BRA results report · New York EO 62 text

Thesis test

The five standing hypotheses of the working framework, tested deductively against this week’s evidence. A framework that is never strained is not being tested.

Hypothesis 1
supported
The cycle is accelerating, not slowing.
Two trillion-parameter-class open MoEs landed in a single week (K3 at a vendor-reported 2.8T, Inkling at 975B), and an open-weights-slated model reached within ~3 points of the closed frontier on the independent index just five weeks after GPT-5.6's government preview began — a gap-closing cadence that took DeepSeek V4 months at the start of the year. On the hardware clock, TSMC's 2nm hit its first material revenue quarter (3% of wafer revenue) while FY26 growth guidance was raised to 'slightly above 40%', and Intel shipped the first high-volume High-NA EUV logic. Both the capability clock and the silicon clock ticked faster this week.
Evidence: K3 independent frontier-tier placement · TSMC Q2: 2nm at 3%, guidance raised · Intel High-NA EUV production milestone
Hypothesis 2
supported
Capital is concentrated, returns are diffuse.
The realized profits again landed at the physical layer while the model layer spent: TSMC booked a record $40.2B quarter with net income up 77% YoY and ASML beat the high end of guidance with 2027 EUV nearly fully booked — audited, banked returns — while frontier labs cut effective prices, rented rivals' compute, and Meta committed >$50B to a single site with no revenue attached. The concentration side sharpened too: two single-company sites (Hyperion, Tembo) now each approach the aggregate size of Japan's entire national AI program, and the clearest new revenue stream in the category was lease payments and debt coupons on contracted GPU capacity, not model margins.
Evidence: TSMC record Q2 (SEC-filed) · Meta Hyperion >$50B expansion
Hypothesis 3
supported
Networking is the durable layer.
The supply side voted with capital this week: Tower Semiconductor committed ~$3B (state-co-funded) to silicon photonics capacity explicitly for optical interconnect demand, Microsoft became the first hyperscaler to deploy 3M's expanded-beam optical connectivity across AI data centers (attacking deployment velocity, not just link speed), and Japan's sovereign program standardized on a single-vendor Ethernet fabric — a state-scale Metcalfe machine. The revenue-side test (interconnect growth outpacing compute growth at colo operators) comes with the Q2 prints starting Jul 29 and is now prediction p64.
Evidence: Tower $3B silicon photonics expansion · 3M-Microsoft EBO deployment partnership
Hypothesis 4
supported
Open weights pull the floor up.
The demand-catalyst mechanism operated on cue, with a new wrinkle: Inkling shipped Apache 2.0 weights at 975B scale with day-0 serving support and a calibrated NVFP4 checkpoint that fits ~600GB — explicitly positioned as a fine-tuning base for private deployment — and Moonshot's K3 launch materials recommend 64+ accelerator supernodes for self-hosting, a direct on-prem hardware pull at frontier scale. The wrinkle is pricing: the open-weight floor is rising in capability while its API price floor rises too (K3 at closed parity, DeepSeek surge pricing), which strengthens rather than weakens the self-host re-routing thesis — the cheaper path to open-model capability is increasingly your own cluster.
Evidence: Inkling Apache 2.0 release with NVFP4 checkpoint · K3 supernode self-hosting guidance
Hypothesis 5
supported
Power is the binding constraint for the next 24 months.
The strongest single-week confirmation since the hypothesis was authored: PJM cleared at the administrative cap for the third consecutive year while coming up 6,831 MW short — with the operator's own uncapped simulation 71% above the clearing price and only 525 MW of new generation clearing — and New York layered the first statewide permit moratorium on top of queue physics the very next day. The constraint now has three stacked layers: physical (generation shortfall), administrative (price caps masking scarcity), and political (state-level pauses). Japan's FRONTia sizing its national AI factory at 140 MW — a rounding error against a single Meta campus — shows even sovereign programs designing around power availability.
Evidence: PJM 2028/29 BRA results (at cap, 6,831 MW short) · New York EO 62 statewide moratorium

Pattern watch

Inductive3 weeks observed
Model-API pricing is converging from both directions onto a $1-3 input / $6-15 output frontier band — closed labs cutting down into it, Chinese open-weight labs raising up into it.
- W27: cost-per-task became the explicit competitive axis as closed-lab pricing moves anticipated GPT-5.6 tiering.
- W28: the closed frontier repriced downward — Terra GA at $2.50/$15 (half of GPT-5.5) and Grok 4.5 at $2/$6.
- W29: the open-weight side repriced upward — Kimi K3 at $3/$15 (roughly 3x its predecessor, parity with Sonnet 5) and DeepSeek adding 2x peak-hour surcharges from Jul 24.
Next week: DeepSeek V4's official GA pricing (due within days) lands at or above current V4-Pro rates rather than resetting the old ultra-cheap floor. If V4 GA undercuts to pre-K3 Chinese pricing levels, the convergence read fails and the price war resumes.
Inductive5 weeks observed
Government action is a standing gate on frontier-model availability — and the gate now includes labs gating themselves.
- W25: Fable 5 and Mythos 5 suspended under US export controls (Jun 12).
- W26: Mythos 5 restored only for ~100 'Annex A' critical-infrastructure organizations.
- W27: GPT-5.6 previewed to ~20 government-vetted partners at the US government's request.
- W28: GPT-5.6 GA'd only after a 12-day government-coordinated preview with CAISI evaluations; Beijing rationed H200 purchases and pushed the Manus unwind.
- W29: OpenAI announced GPT-Red will never be released due to offensive capability — the first explicit capability self-gating of a disclosed frontier-scale model — while Gemini 3.5 Pro slipped a third time.
Next week: The GPT-Red technical preprint ships within two weeks but the model does not — and no API access materializes. If OpenAI releases GPT-Red weights or endpoints in any form, the self-gating leg of this pattern breaks.
Inductive5 weeks observed
Physical-input scarcity (memory, power, now packaging) keeps marking itself to market with progressively harder money and harder words.
- W25-W26: all three HBM makers volume-shipping HBM4; Micron confirmed 2026 supply fully contracted.
- W27: SK hynix filed a ~$29.4B Nasdaq listing; price caps reportedly removed from long-term memory contracts.
- W28: the IPO closed at $26.5B with 7x demand; Samsung guided to a record quarter; Meta and MARA paid premiums for secured power.
- W29: PJM cleared at the cap and 6,831 MW short; TSMC's CEO said packaging is 'limiting my customers' growth' while raising capex to $60-64B; server DRAM contracts guided +13-18% QoQ for Q3; Tower committed $3B to optics capacity.
Next week: SK hynix (Jul 29) and Samsung (Jul 30) Q2 prints disclose HBM substantially sold out or committed into 2027 with capex raises attached. Prints showing softening HBM commitments or flat memory capex would be the first hard counter-evidence in five weeks.

Second-order effects

Trigger: Meta doubles Hyperion to >$50B while simultaneously negotiating to lease capacity to Anthropic, as GPU-backed credit institutionalizes at rated spreads.
A fifth cloud priced against neoclouds emerges from surplus hyperscaler capex, splitting the neocloud category by funding cost: operators with bankable take-or-pay contracts refinance into cheap secured debt while the rest face equity markets that just repriced their concentration risk. Expect consolidation — the weaker half of the neocloud category becomes acquisition inventory for hyperscalers and infrastructure funds within two or three quarters.
Horizon: Q4 2026 - H1 2027Who moves: Neocloud equity holders and lenders, frontier labs negotiating compute, infrastructure funds, and hyperscaler corp-dev teams
Trigger: New York signs the first statewide data-center permit moratorium one day after PJM clears at the cap with a 6,831 MW shortfall.
Data-center opposition movements in states with real AI load now have a governor-signed template, converting diffuse local resistance into replicable state policy — which raises the siting-risk premium on any campus without secured permits, accelerates the behind-the-meter generation and scale-across fabric workarounds, and pulls forward the value of already-permitted powered land as a distinct asset class.
Horizon: 2027Who moves: Data-center developers without generation balance sheets, state energy regulators, utilities, and holders of permitted powered-land inventory
Trigger: Kimi K3 reaches the closed-frontier tier on independent evaluations with open weights promised July 27 and self-hosting guidance targeting 64+ accelerator supernodes.
If the weights and a commercial-friendly license land, data-residency-constrained enterprises and sovereign programs get their first credible exit from frontier API dependency — which re-routes inference spend from closed-lab APIs toward private accelerator clusters, HBM, and interconnect, and forces closed labs to defend on harness, tooling, and injection-hardening (exactly where OpenAI positioned GPT-Red this week) rather than raw capability.
Horizon: 30-90 daysWho moves: Closed-lab API revenue, sovereign AI programs, enterprise infrastructure teams, and the accelerator/HBM supply chain

Strategic outlook

This week moved the 12-month posture on three fronts. First, procurement: the open-vs-closed decision is no longer about a capability gap (~3 Index points) or a price gap (K3 at closed parity) — it is about deployment control, and July 27 is the date that tells you whether frontier-class self-hosting is real; budget the evaluation now, because if the weights land clean, every closed-lab contract renewal after August negotiates against a credible internal alternative. Second, capital structure: compute is becoming a two-sided market — leased across rival lines, financed with rated secured debt — so treat GPU capacity commitments the way treasurers treat interest-rate exposure: as a portfolio of owned, leased, and optioned positions rather than a single vendor bet. Third, the constraint stack: with packaging capped at the foundry, power capped at the auction, and permits now pausable by executive order, the scarce inputs for 2027 are all downstream of money — which means the durable advantages are secured packaging allocation, energized land, and permitted sites, and the capital markets just started pricing all three accordingly.

Early warning panel

The levers we monitor.

10 metrics tracked — 2 rising, 1 falling, 7 steady.

Current vs prior period. Each metric has a threshold where the read materially changes — this panel flags the inflection before it lands in headlines. Click any metric for the methodology and this-week read.

Metric

Current

Prior

Dir

Threshold

Frontier lab cash position (avg months runway, top 3)
~34-37 mo; no new primary capital — the movement was compute-sourcing (Anthropic-Meta lease talks, reported ~$10B/2yr, disputed)
~34-37 mo; Anthropic at implied $1.2T on secondaries (broker-reported), IPO calendars unchanged
→
<18 mo triggers re-rating risk
What this measures
Top 3 frontier labs (OpenAI, Anthropic, Google DeepMind) by disclosed runway. No primary financing closed in-window; SoftBank's second $10B OpenAI tranche executed Jul 1 (pre-window). The structural shift is that labs are now negotiating compute from rivals' fleets — Anthropic-Meta talks on top of the earlier $45B Colossus 1 commitment — which converts capex exposure into opex flexibility. Boards should read this as runway-preserving behavior, not distress.
Hyperscaler capex / AI revenue ratio (top 4 weighted)
~5.0-5.5; Meta Hyperion doubled to >$50B/5 GW, Google behind 2.7 GW Wyoming 'Project Tembo' — numerator still hardening ahead of the Jul 22-30 earnings wave
~5.0-5.3; Meta targets 14 GW of compute in 2027 (2x 2026) on $125-145B capex guidance
↑
>6.0 invites investor pushback at next earnings
What this measures
Top 4 hyperscalers (MSFT, GOOG, META, AMZN) weighted aggregate of capex divided by AI-attributable revenue. Two single-site adds this week (Hyperion >$50B, Tembo ~$50B reported) push the committed numerator up without any new revenue disclosure — the ratio is unmeasurable to audit standard because AI revenue isn't segmented. The Jul 22-30 earnings wave (Alphabet, Microsoft, then Meta) is the first test of whether guides absorb these adds or raise.
CoreWeave revenue backlog
$99.4B as of Mar 31 (+284% YoY), restated in Fitch's Jul 16 DDTL note; equity down ~35% since the Meta Compute report
~$100B reported; Helios Phase I (133 MW) delivered on schedule — backlog now converting to lease revenue
→
Conversion velocity matters more than gross figure
What this measures
Booked but unrecognized revenue; official next print is Q2 in early August. This week the backlog got third-party validation of a different kind: Fitch rated a $2.6B contract-backed DDTL at BB+ against take-or-pay contracts, meaning a rating agency underwrote the backlog's bankability even as public equity de-rated it. Watch the Q2 print for anchor-customer mix and conversion velocity — credit and equity cannot both be right at current spreads.
NVIDIA Q-over-Q data center revenue
$75.2B Q1 FY27 (unchanged); Huang in Tokyo: 'Vera Rubin is already in production' — delay narrative rebutted, but no customer-delivery date named
$75.2B Q1 FY27; Q2 guide $91B (reports Aug 26); SemiAnalysis sees H2 ~20% above consensus despite Kyber dispute
→
Q2 FY27 guide $91B implies further +21% QoQ
What this measures
No earnings event in-window. Huang's Tokyo statement rebuts the early-July delay reporting for NVL72 racks, and Japan's FRONTia order (27,500 Rubin GPUs) puts a named sovereign customer behind the ramp — but TSMC's same-week warning that packaging capacity is 'limiting my customers' growth' is the supply-side caveat that keeps the delivery-date question open until the Aug 26 print.
Open vs closed gap on coding (SWE-Bench / agentic)
~3 pts on the AA Intelligence Index — Kimi K3 at 57.1 vs Fable 5's 59.9 — with K3 weights pending Jul 27; K3 already #1 on LMArena Frontend Code Arena
Effectively closed on cost-quality: GLM 5.2 statistically tied with Opus 4.8 at $1.28 vs $1.94/task (Databricks); Hy3 adds Apache-2.0 agentic-search lead
↓
Sustained open lead reshapes enterprise procurement
What this measures
The absolute-capability gap narrowed to ~3 Index points this week — the closest an open-weights-slated model has come to the closed frontier — and K3 topped a flagship human-preference board outright. Two caveats keep this from resolving: K3 is hosted-only until Jul 27 (AA classifies it proprietary in the interim), and its vendor-reported coding numbers ran on its own KimiCode harness. If the weights and license land clean, this lever's threshold — sustained open parity reshaping procurement — is effectively triggered.
Sovereign AI commitments (count / aggregate $)
~15 / ~$186B; Japan FRONTia adds ¥1T (~$6.2B/5yr) — the first national program explicitly scoped to physical AI and robotics
~14 / ~$180B+ (flat; Meta Alberta and MARA Texas are corporate capital on power-rich land, not sovereign programs)
↑
—
What this measures
Analyst-curated count of sovereign/national AI-compute commitments. FRONTia is the template evolution to watch: consortium structure (48 firms), national data, a hard chip order (27,500 Rubin GPUs at 140 MW), and a single-vendor fabric — sovereign programs are becoming reference-design purchases rather than bespoke buildouts. NVIDIA's SEC-filed commentary already attributes half its data-center revenue to non-hyperscale demand including sovereign; expect FRONTia clones in other industrial states within two quarters.
PJM 2026/27 capacity auction price ($/MW-day)
$325.00 — the 2028/29 BRA cleared at the (lowered) FERC cap Jul 14, 6,831 MW short of the reliability requirement; PJM's uncapped simulation: $554.72 RTO / $776.69 ComEd
$329.17; 2028/29 BRA bids closed Jul 7 — results post Jul 14 after 4 p.m. ET (prediction p54 resolves)
→
11x in 24 months — power is the new binding constraint
What this measures
Third consecutive auction pinned at the administrative cap; the nominal 2.5% decline is entirely the cap formula, not loosening scarcity. The information is in the gap: capped-vs-simulated spread widened from 18% to 59% to 71% across three auctions, only 525 MW of new generation cleared, and total cost still hit $16.4B. Budget at-cap through 2028; the September backstop auction and PJM's data-center connect-and-manage FERC filings are where the real price of AI load gets set.
Time-to-power, busiest US markets (months)
60-84 unchanged; New York adds the first statewide ≥50 MW discretionary-permit moratorium — state policy now stacks on top of queue physics
60-84; hyperscalers routing around queues — Meta fully funds its own generation in Alberta because the grid cannot host multiple large loads
→
—
What this measures
Months from new-load interconnection request to energization. No fresh queue data in-window, but the constraint acquired a policy layer: EO 62 pauses discretionary environmental permits for ≥50 MW facilities statewide while New York writes a generic environmental impact statement. Narrow in operational scope, precedential in kind — developers should now underwrite state-policy pause risk in any market with active data-center opposition, on top of the 60-84 month grid reality.
Cost-per-task, frontier reasoning model
DeepSeek's 2x Beijing-peak-hours surcharge (from Jul 24) imports electricity-market demand management into model APIs
~$0.06-$0.12 effective unchanged at the floor; new AA Index task costs: DeepSeek V4 Pro $0.04, GLM-5.2 $0.32, K3 $0.94, Sol $1.04 — and DeepSeek adds the first peak-hour surge pricing
~$0.06-$0.12 effective; GPT-5.6 GA tiering (Sol $5/$30 / Terra $2.50/$15 / Luna $1/$6), Grok 4.5 at $2/$6 — and harness choice swings per-task cost 2x+
→
—
What this measures
Median cost across frontier-tier reasoning models for a benchmark complex task. The floor held, but two structural changes landed: K3's always-on max-effort thinking shows sticker price diverging from effective cost in the expensive direction (130M tokens on the AA eval vs 63M peer average), and DeepSeek's time-of-day surge pricing means per-task cost now varies by clock. Cost models should carry cache-hit rates, reasoning-effort settings, and now time-of-day as explicit variables.
Custom silicon share of incremental AI compute
~34-37% unchanged; Google reportedly pitching TPUs to NVIDIA-centric neoclouds with financial backstops — custom silicon escalating from internal cost play to merchant distribution fight
~34-37%; Meta's Iris enters production in September, Broadcom-Apple extended through 2031 (8-K), AWS raising Trainium 3 orders 20-30%
→
>35% materially compresses merchant GPU pricing
What this measures
Estimated share of incremental AI compute capacity on hyperscaler custom silicon. No confirmed share shift this week — Nebius, Lambda, and CoreWeave all publicly demurred on the reported Google TPU pitch — but the pitch itself (backstopping TPU data centers and renting capacity back) is the clearest signal yet that custom silicon is contesting NVIDIA's own neocloud channel rather than just displacing internal purchases. NVIDIA reportedly counter-offering incentives to Nscale confirms the channel fight is live.

Frontier lab cash position (avg months runway, top 3)
~34-37 mo; no new primary capital — the movement was compute-sourcing (Anthropic-Meta lease talks, reported ~$10B/2yr, disputed)→vs ~34-37 mo; Anthropic at implied $1.2T on secondaries (broker-reported), IPO calendars unchanged
Threshold: <18 mo triggers re-rating risk
What this measures
Top 3 frontier labs (OpenAI, Anthropic, Google DeepMind) by disclosed runway. No primary financing closed in-window; SoftBank's second $10B OpenAI tranche executed Jul 1 (pre-window). The structural shift is that labs are now negotiating compute from rivals' fleets — Anthropic-Meta talks on top of the earlier $45B Colossus 1 commitment — which converts capex exposure into opex flexibility. Boards should read this as runway-preserving behavior, not distress.
Hyperscaler capex / AI revenue ratio (top 4 weighted)
~5.0-5.5; Meta Hyperion doubled to >$50B/5 GW, Google behind 2.7 GW Wyoming 'Project Tembo' — numerator still hardening ahead of the Jul 22-30 earnings wave↑vs ~5.0-5.3; Meta targets 14 GW of compute in 2027 (2x 2026) on $125-145B capex guidance
Threshold: >6.0 invites investor pushback at next earnings
What this measures
Top 4 hyperscalers (MSFT, GOOG, META, AMZN) weighted aggregate of capex divided by AI-attributable revenue. Two single-site adds this week (Hyperion >$50B, Tembo ~$50B reported) push the committed numerator up without any new revenue disclosure — the ratio is unmeasurable to audit standard because AI revenue isn't segmented. The Jul 22-30 earnings wave (Alphabet, Microsoft, then Meta) is the first test of whether guides absorb these adds or raise.
CoreWeave revenue backlog
$99.4B as of Mar 31 (+284% YoY), restated in Fitch's Jul 16 DDTL note; equity down ~35% since the Meta Compute report→vs ~$100B reported; Helios Phase I (133 MW) delivered on schedule — backlog now converting to lease revenue
Threshold: Conversion velocity matters more than gross figure
What this measures
Booked but unrecognized revenue; official next print is Q2 in early August. This week the backlog got third-party validation of a different kind: Fitch rated a $2.6B contract-backed DDTL at BB+ against take-or-pay contracts, meaning a rating agency underwrote the backlog's bankability even as public equity de-rated it. Watch the Q2 print for anchor-customer mix and conversion velocity — credit and equity cannot both be right at current spreads.
NVIDIA Q-over-Q data center revenue
$75.2B Q1 FY27 (unchanged); Huang in Tokyo: 'Vera Rubin is already in production' — delay narrative rebutted, but no customer-delivery date named→vs $75.2B Q1 FY27; Q2 guide $91B (reports Aug 26); SemiAnalysis sees H2 ~20% above consensus despite Kyber dispute
Threshold: Q2 FY27 guide $91B implies further +21% QoQ
What this measures
No earnings event in-window. Huang's Tokyo statement rebuts the early-July delay reporting for NVL72 racks, and Japan's FRONTia order (27,500 Rubin GPUs) puts a named sovereign customer behind the ramp — but TSMC's same-week warning that packaging capacity is 'limiting my customers' growth' is the supply-side caveat that keeps the delivery-date question open until the Aug 26 print.
Open vs closed gap on coding (SWE-Bench / agentic)
~3 pts on the AA Intelligence Index — Kimi K3 at 57.1 vs Fable 5's 59.9 — with K3 weights pending Jul 27; K3 already #1 on LMArena Frontend Code Arena↓vs Effectively closed on cost-quality: GLM 5.2 statistically tied with Opus 4.8 at $1.28 vs $1.94/task (Databricks); Hy3 adds Apache-2.0 agentic-search lead
Threshold: Sustained open lead reshapes enterprise procurement
What this measures
The absolute-capability gap narrowed to ~3 Index points this week — the closest an open-weights-slated model has come to the closed frontier — and K3 topped a flagship human-preference board outright. Two caveats keep this from resolving: K3 is hosted-only until Jul 27 (AA classifies it proprietary in the interim), and its vendor-reported coding numbers ran on its own KimiCode harness. If the weights and license land clean, this lever's threshold — sustained open parity reshaping procurement — is effectively triggered.
Sovereign AI commitments (count / aggregate $)
~15 / ~$186B; Japan FRONTia adds ¥1T (~$6.2B/5yr) — the first national program explicitly scoped to physical AI and robotics↑vs ~14 / ~$180B+ (flat; Meta Alberta and MARA Texas are corporate capital on power-rich land, not sovereign programs)
What this measures
Analyst-curated count of sovereign/national AI-compute commitments. FRONTia is the template evolution to watch: consortium structure (48 firms), national data, a hard chip order (27,500 Rubin GPUs at 140 MW), and a single-vendor fabric — sovereign programs are becoming reference-design purchases rather than bespoke buildouts. NVIDIA's SEC-filed commentary already attributes half its data-center revenue to non-hyperscale demand including sovereign; expect FRONTia clones in other industrial states within two quarters.
PJM 2026/27 capacity auction price ($/MW-day)
$325.00 — the 2028/29 BRA cleared at the (lowered) FERC cap Jul 14, 6,831 MW short of the reliability requirement; PJM's uncapped simulation: $554.72 RTO / $776.69 ComEd→vs $329.17; 2028/29 BRA bids closed Jul 7 — results post Jul 14 after 4 p.m. ET (prediction p54 resolves)
Threshold: 11x in 24 months — power is the new binding constraint
What this measures
Third consecutive auction pinned at the administrative cap; the nominal 2.5% decline is entirely the cap formula, not loosening scarcity. The information is in the gap: capped-vs-simulated spread widened from 18% to 59% to 71% across three auctions, only 525 MW of new generation cleared, and total cost still hit $16.4B. Budget at-cap through 2028; the September backstop auction and PJM's data-center connect-and-manage FERC filings are where the real price of AI load gets set.
Time-to-power, busiest US markets (months)
60-84 unchanged; New York adds the first statewide ≥50 MW discretionary-permit moratorium — state policy now stacks on top of queue physics→vs 60-84; hyperscalers routing around queues — Meta fully funds its own generation in Alberta because the grid cannot host multiple large loads
What this measures
Months from new-load interconnection request to energization. No fresh queue data in-window, but the constraint acquired a policy layer: EO 62 pauses discretionary environmental permits for ≥50 MW facilities statewide while New York writes a generic environmental impact statement. Narrow in operational scope, precedential in kind — developers should now underwrite state-policy pause risk in any market with active data-center opposition, on top of the 60-84 month grid reality.
Cost-per-task, frontier reasoning model
~$0.06-$0.12 effective unchanged at the floor; new AA Index task costs: DeepSeek V4 Pro $0.04, GLM-5.2 $0.32, K3 $0.94, Sol $1.04 — and DeepSeek adds the first peak-hour surge pricing→vs ~$0.06-$0.12 effective; GPT-5.6 GA tiering (Sol $5/$30 / Terra $2.50/$15 / Luna $1/$6), Grok 4.5 at $2/$6 — and harness choice swings per-task cost 2x+
DeepSeek's 2x Beijing-peak-hours surcharge (from Jul 24) imports electricity-market demand management into model APIs
What this measures
Median cost across frontier-tier reasoning models for a benchmark complex task. The floor held, but two structural changes landed: K3's always-on max-effort thinking shows sticker price diverging from effective cost in the expensive direction (130M tokens on the AA eval vs 63M peer average), and DeepSeek's time-of-day surge pricing means per-task cost now varies by clock. Cost models should carry cache-hit rates, reasoning-effort settings, and now time-of-day as explicit variables.
Custom silicon share of incremental AI compute
~34-37% unchanged; Google reportedly pitching TPUs to NVIDIA-centric neoclouds with financial backstops — custom silicon escalating from internal cost play to merchant distribution fight→vs ~34-37%; Meta's Iris enters production in September, Broadcom-Apple extended through 2031 (8-K), AWS raising Trainium 3 orders 20-30%
Threshold: >35% materially compresses merchant GPU pricing
What this measures
Estimated share of incremental AI compute capacity on hyperscaler custom silicon. No confirmed share shift this week — Nebius, Lambda, and CoreWeave all publicly demurred on the reported Google TPU pitch — but the pitch itself (backstopping TPU data centers and renting capacity back) is the clearest signal yet that custom silicon is contesting NVIDIA's own neocloud channel rather than just displacing internal purchases. NVIDIA reportedly counter-offering incentives to Nscale confirms the channel fight is live.

Predictions

What we expect next.

5 predictions for the next 30-90 days, confidence 57%-76%.

Each prediction is falsifiable, time-bounded, and tied to a specific signal we will watch. Future issues score these hit, miss, partial, or pending and build a public track record.

Prediction 01

72%

confidence

Software

Moonshot publishes Kimi K3 open weights on Hugging Face with a license permitting commercial self-hosting by August 10, 2026 (vendor-committed 'by July 27', with slippage buffer).

Deadline: By August 10, 2026

Trigger: Kimi K3 weights live on Hugging Face with published license text; Artificial Analysis reclassifies K3 from proprietary to open-weights.

Prediction 02

76%

confidence

Software

DeepSeek retires its legacy deepseek-chat and deepseek-reasoner aliases on July 24 as scheduled and ships DeepSeek V4 to official GA by July 31, 2026, with peak-hour surge pricing in effect.

Deadline: By July 31, 2026

Trigger: DeepSeek API docs showing V4 GA model IDs and the alias-retirement notice executed; surge pricing live in the rate card.

Prediction 03

68%

confidence

Hardware

SK hynix's July 29 Q2 earnings call discloses that 2027 HBM capacity is substantially sold out or committed under long-term agreements, extending the memory-scarcity trade into a second year.

Deadline: By July 29, 2026

Trigger: SK hynix Q2 2026 earnings call commentary on 2027 HBM capacity commitments and capex.

Prediction 04

71%

confidence

Networking

The largest colocation operators' Q2 prints (starting July 29) show interconnect/fabric revenue growth again outpacing overall revenue growth, sustaining the Metcalfe read for a second consecutive quarter.

Deadline: By August 15, 2026

Trigger: Q2 2026 colo earnings disclosures: interconnection and fabric revenue growth rates vs total revenue growth.

Prediction 05

57%

confidence

Power

At least one additional US state announces a moratorium, discretionary-permit pause, or equivalent statewide restriction on large data-center development by October 31, 2026, following New York's EO 62.

Deadline: By October 31, 2026

Trigger: A governor's executive order or enacted state legislation pausing or restricting ≥50 MW-class data-center permitting in a second state.

Track record

Scoring prior predictions.

4 prior predictions: 0 hit, 0 miss, 1 partial, 3 pending. Hit rate 0%.

4 predictions across issues so far. Hit rate: 0%. Hits 0, misses 0, partials 1, pending 3.

Prediction 01

58%

confidence

Software

Gemini 3.5 Pro reaches public general availability — a callable API model ID with published pricing — by July 31, 2026, after slipping past its June window and the reported July 17 target.

Deadline: By July 31, 2026

Trigger: Google Gemini API model list / pricing page showing a GA gemini-3.5-pro model ID.

pending— Missed the leaked Jul 17 target this week — its third slip — with reporting citing hallucination/reliability gaps and possible stopgap releases; the public API still lists only gemini-3.5-flash and gemini-3.1-pro-preview. Two weeks remain on the deadline; prediction markets price ~81% by Jul 31.

Prediction 02

64%

confidence

Software

At least one major agent platform (OpenAI, Anthropic, GitHub, or Cursor) ships product-level per-task or per-harness cost telemetry or routing controls — beyond session budget caps — by August 31, 2026.

Deadline: By August 31, 2026

Trigger: Product changelog or GA announcement exposing per-task cost measurement or harness-level cost controls.

pending— Adjacent evidence accumulated: Atlassian shipped DX AI cost-vs-output measurement (Jul 15) and Claude Code v2.1.212 shipped session-level search/subagent budgets — but Atlassian is outside the four named platforms and session caps are explicitly excluded by the prediction's own terms. The pattern is advancing; the specific trigger hasn't fired.

Prediction 03

62%

confidence

Hardware

TSMC's July 16 Q2 earnings raise or reiterate the top end of full-year 2026 capex guidance and report HPC/AI platform revenue up more than 50% year over year, confirming the packaging-constrained AI capex ramp.

Deadline: By July 16, 2026

Trigger: TSMC Q2 2026 earnings release and investor call (Jul 16; June revenue print Jul 13, typhoon-delayed).

partial— The capex leg hit decisively — TSMC raised the entire FY26 range to $60-64B (from $52-56B), well beyond reiterating the top end. The revenue leg fell just short: HPC at 66% of a $40.2B quarter implies ~47% YoY platform growth versus the >50% bar. The thesis behind the prediction (packaging-constrained ramp) was confirmed verbatim by the CEO: capacity 'so tight that now it's limiting my customers' growth.'

Prediction 04

61%

confidence

Networking

A second named vendor or operator announces a commercial cross-data-center scale-across AI fabric deployment or product launch — following DriveNets/WhiteFiber — by September 30, 2026.

Deadline: By September 30, 2026

Trigger: Vendor or operator press release for a commercial (not lab) multi-site training-fabric deployment; WhiteFiber's own Q3 commercial launch also qualifies if it lands with a named second customer.

pending— No second vendor announcement in-window; DriveNets/WhiteFiber drew continued trade analysis and their Q3 commercial launch remains on track. Hot Interconnects 2026 (Aug 19-21, themed 'Scale-Up, Scale-Out, Scale-Across') is the likely venue for the next entrant.

Watchlist

On the radar this week.

5 catalysts to watch, starting Jul 20-24.

Specific catalysts that would change the read materially. Watching these tells us whether the thesis is strengthening or weakening.

Jul 20-24
OpenAI's GPT-Red technical preprint
Promised 'later this week' — enables third-party scrutiny of the 84%-vs-13% red-teaming and 0.05% injection-failure claims that are currently vendor-reported only. If it substantiates, injection-hardening becomes a hard procurement criterion industry-wide.
Jul 22-23
Q2 earnings triple-header: ServiceNow and Alphabet (Jul 22), then SAP, Intel, and AMD's Advancing AI event (Jul 23)
First reads on agentic-AI monetization (ServiceNow), cloud AI revenue (Alphabet), the closed-gateway Joule strategy (SAP), 18A economics post-High-NA (Intel), and MI455X/Helios rack shipment clarity (AMD).
Jul 24
Two hard deadlines: M365 Copilot's OpenAI-subprocessor setting auto-enables, and DeepSeek retires its legacy API aliases
M365 tenants that don't opt out are auto-enrolled in a changed data-processing chain (silence is consent); DeepSeek's alias retirement plus surge pricing is a forced-migration event for one of the most-called model APIs.
Jul 27
Kimi K3 open-weights drop, license text, and technical report
The single highest-information event in view: if the weights and a commercial-friendly license land, frontier-class capability becomes self-hostable for the first time and prediction p61 resolves early; if they slip or the license restricts, the 'open frontier' narrative resets.
Jul 29-30
Memory and hyperscaler prints: SK hynix Q2 (Jul 29), Microsoft FY26 Q4 (Jul 29), Samsung divisional results (Jul 30), colo interconnect Q2 prints begin (Jul 29)
The supply-side read on 2027 HBM commitments (p63), whether hyperscaler capex guides absorb this week's Hyperion/Tembo adds, and the interconnect-revenue test of the networking hypothesis (p64).

Companion reads

The rest of the spine.

The AI Stack Weekly is the cross-stack flywheel read. Pair it with the model-and-tree spine and the working framework to get the full picture.