Capital flow drill-down
Money in, revenue out: the full breakdown.
Issue 30 · Week 46 of 2026.
The four-category capital scorecard expanded with named transactions, burn-to-revenue context, and per-category trends across recent issues. Each row corresponds to a row in the summary table on the issue page.
Capital flow scorecard
Capital in, revenue out, burn ratio, and movement.
| Category | Capital in | Revenue out | Burn to revenue | Movement |
|---|---|---|---|---|
| Frontier LabsOpenAI, Anthropic, Google DeepMind, DeepSeek | None closed; the largest raise in the series is priced but not yet printed. Anthropic's S-1/A No. 2 (grade 5) sets a $70-78 range on 825M Class A shares (up to $64.35B gross; ~$1.4T fully diluted at the top) with a 123.75M-share option and a concurrent $26.0B Class N private placement at the IPO price from AMD, Broadcom, Google, Micron, Microsoft, NVIDIA and TSMC Partners; pricing is targeted for November 18 and trading November 19 as ANT, the book was reported about six times covered by November 12 (Bloomberg, Reuters; anonymous sources) and TSMC's board approved up to $500M (grade 5). OpenAI is reported at ~$15B raised so far at $1.25T in a round not yet closed (CNBC, grade 3; text only, as in W45), with TSMC's board approving up to $500M of OpenAI preferred. Nothing is charted until ANT prices; the big story carries the placement as the equity leg of the syndicated backstopPrior None charted. OpenAI is reported to have closed ~$11B of a round at $1.25T with up to ~$20B expected 'in the coming weeks' (Bloomberg, grade 3; text only, as last week's report was). Anthropic's S-1/A (grade 5) adds Q3 and discloses $77.0B of undrawn note-purchase capacity (Google $20B, Amazon $15B, Broadcom $42B at 9% PIK) and $146.6B of liquidity, with the price range still blank and the roadshow reported for the week of November 9; the Broadcom facility's implied cost is the house measurement and it is capacity, not capital in. Google's credit-support agreements for Anthropic's TeraWulf lease (8-K, grade 5) are a guarantee, carried in the big story, and move no cash to the lab. Fal's $1.3B Series F at $16B (grade 4) is an inference cloud and is charted under Neoclouds · Flat | ~$81B run-rate: Anthropic's S-1/A No. 2 discloses run-rate revenue of about $81B as of October 31, up $10B from September 30, on Q3 revenue of $16,956M (filed in W45); more than 11,900 customers above $100K produced $24.5B (74%) of nine-month revenue, with the five largest at ~25% (The Application Layer carries the customer mix). OpenAI's Q3 revenue is reported at ~$10B (+~50% QoQ, ~$50B annualised) with a $10.9B adjusted operating loss and enterprise at ~50% of the mix (WSJ, Bloomberg; grade 3). The series stays on Anthropic's filed number and moves from the W45 annualised Q3 figure (~$68B) to the October run-ratePrior ~$68B annualised: Anthropic's S-1/A files Q3 2026 revenue of $16,956M (+47% QoQ from $11,555M; +1,127% YoY), replacing last week's Reuters-reported $16.8B; nine-month revenue $33.2B; Q3 operating loss $567M, which includes a $2.0B non-cash charge for the donation-matching programme (about +$1.4B and an 8.5% margin without it, level with Q2) alongside compute and infrastructure expense of $9.86B (58% of revenue, from 48%); net loss $23.3B on a $21.9B non-cash fair-value charge. Consumption is 79% of nine-month revenue (The Application Layer carries the customer metrics). SpaceX's SI segment (~$30B annualised, W43) stays in text because the series is Anthropic's · Up | n/a (no chartable capital in this week). The filed stress number is not burn but the twelve-month commitment schedule: ~$70.7B of Anthropic's ~$900.9B of aggregate commitments falls due within twelve months of September 30, about 0.9x the October run-rate, which is the quantity the $26B placement and the IPO proceeds are sized against | Priced, not closed: the range, the placement and the roadshow are filed; the cash arrives November 18-19. The movement the series records is that a lab's equity is now the first link in the neocloud financing chain (big story), with OpenAI's reported round in text only |
| Hyperscaler-HostedAzure-OpenAI, AWS-Anthropic, Google Cloud-Gemini, Oracle-OCI | None closed in the graded packet this week; the category's cash moved the other way. Google, Microsoft, AMD, Broadcom, Micron, NVIDIA and TSMC are buyers in Anthropic's $26.0B Class N placement (charted under Frontier Labs when it prints), and Morgan Stanley counts $67B of hyperscaler bonds priced since September 30 (Microsoft, Meta, Oracle in the US; Alphabet in Europe) with spreads on most tightened since issue (Oracle the exception) and issuance expected quiet for the rest of 2026 (CNBC, grade 4). Goldman Sachs Research projects $420B of hyperscaler debt issuance in 2027 against ~$250B in 2026, and Morgan Stanley forecasts $1.66T of AI capex in 2027 (+75%); both are forecasts, not flowsPrior ~$35B closed: Oracle's $20B (86.2M shares at $116, an 11.2% discount to the last close, ~$9.9B net; and $10B of senior notes in five tranches at 7.000-8.650%, Treasuries +240 to +335 bp, Baa2 negative / BBB- stable), both settled November 4 and filling the $20B cap set at its October 28 Investor Day; and Meta's $15B of senior notes in five tranches at 5.850-7.500% (Treasuries +110 to +220 bp, Aa3 stable / AA- negative), settled November 5. Meta's underwriters also exercised the 6M-share option on the forward sale, lifting expected forward proceeds to ~$28.7B on settlement by November 19, 2027 (cash not yet received; not charted). Google's $17B Wyoming campus commitment with a 702 MW fuel-cell PPA is capex, not a raise, and is carried in text · Down | Indirect; no AI-segment revenue line, and no hyperscaler reported this week. The lines carried from W44 stand (Google Cloud $30.7B +103%, Azure +48%, AWS $46.0B +39%). The nearest new datapoints are from suppliers and customers: Cisco raised its FY27 hyperscaler AI-infrastructure revenue target to $9B from $7.5B (more than 2x FY26), which is hyperscaler spend seen from the vendor side; Anthropic's S-1/A says its revenue averages under 0.1% of its Fortune 500 customers' total expenses; Microsoft shares fell 2-3% on November 10, the day of Anthropic's Claude Meetings / Sheets / Whiteboards launch inside the $20 seat (The Application Layer; coincidence recorded, not cause). Azure's ND MI455X v7 Helios racks (72 MI455X, 18 EPYC Venice, 31 TB HBM4) are the hardware lens's availability evidence for AMDPrior Indirect; no AI-segment revenue line. No hyperscaler reported this week; the lines carried from W44 stand (Google Cloud $30.7B +103%, Azure +48%, AWS $46.0B +39%). The nearest new datapoint is the price of the capital: at matched tenors Oracle paid 165 bp more than Meta over Treasuries in 2031 (T+275 vs T+110), 160 bp more in 2033 and 135 bp more in 2046, the price gap between a BBB and an AA hyperscaler for the same use of proceeds (house measurement from the two FWPs) · Flat | n/a (no AI revenue line; cloud segments are not AI-attributable) | Quiet on issuance after $67B in six weeks, and net buyers of lab equity: the hyperscalers' capital this week went into Anthropic's Class N placement and into lobbying (Microsoft against the FCC transceiver ban), not into their own paper |
| NeocloudsCoreWeave, Nscale, Crusoe, Lambda, IREN, Zankore, NEXTDC | ~$2.3B of contract-secured bank paper charted, all at the asset or developer level and none unsecured: Rowan Digital Infrastructure's more than $1.5B of new corporate capacity (a $600M upsize of its Apterra-led holdco term loan to $900M, a $590M revolver and a $355M letter-of-credit facility led by Citibank, with accordions to $2B and $1B) for a 5.5 GW development pipeline (grade 5); Blue Owl Digital Infrastructure Trust's ~$512.6M floating-rate Citibank / Wells Fargo mortgage on a 54 MW Gainesville, Virginia building 100% leased to a hyperscaler, closed November 10 with the ~$779.5M acquisition (8-K, grade 5); and WhiteFiber's $300M Morgan Stanley-agented term loan at SOFR+600 with a 2% floor for the 40 MW NC-1 campus leased to Nscale, signed November 9 and unfunded pending customer acceptance (8-K, grade 5). Not charted: CoreWeave's ~$9B of expected customer prepayments over three months and its 'high investment-grade' asset-level facility expected within one to two weeks at ~7-7.5% (CEO on CNBC); Nebius's $17-20B of anticipated 2027 prepayments and its $3.5B Citi secured facility signed October 22 after shelving a high-yield bond (Reuters, November 9); Nscale's S-1/A with no price range. The unsecured market stayed shut: CoreWeave will not tap it through 2027 at current yields, Reuters reports banks reluctant to add exposure even to deals with marquee-tech guarantees and some offerings shelved, and Galaxy's Helios data-centre bonds fell to 89.25 to yield ~13% against a 9.875% July couponPrior ~$2.2B charted: Cipher's $850M 0% convertible due 2032 (closed, grade 4) and Fal's $1.3B Series F at a $16B valuation (grade 4), an inference cloud whose capital funds GPU capacity. IREN's $15B of remaining FY27 capex is to be funded by $9B of cash, committed GPU financing and prepayments plus $6B of targeted financing 'subject to market conditions', none closed this week. The category's price moved more than its volume: CoreWeave's 9.625% 2032 notes yielded above 16% and its five-year CDS spiked past 1,289 bp on November 5 (~1,155 on November 6); IFR reports five data-centre IPOs pulled or delayed and some developers offering tenants equity to keep projects alive; Bernstein sizes CoreWeave's remaining need at ~$85B to deliver its $104B backlog. Microsoft is consolidating its neocloud roster, declined to renew an Oracle capacity deal and passed on doubling its Nscale Texas contract (CNBC, Reuters; grade 4) · Flat | ~$20B annualised across the three operators that printed: CoreWeave Q3 revenue $3,661M (+168% YoY, more than $1B sequential; beat ~$3.55B consensus) with backlog ~$132.5B (from $104B), operating income $207M, adjusted operating income $379M, net interest expense $846M and a $709M net loss; 2026 revenue guide raised to $13.0-13.3B, capex to $37-39B, exit-2026 run-rate $20-21B, Q4 interest expense $1.05-1.15B, adjusted profitability targeted 1H 2028; short-term capacity pricing ~$40M/MW-year against $20-25M/MW-year on 4-5 year contracts. Nebius Q3 revenue $980.8M (+571%, including ~$325M non-recurring asset-light revenue), adjusted EBITDA $465.8M (47.5%), ARR $6.4B, 2026 guide trimmed to $3.0-3.2B on third-party capacity slipping into 1H 2027, 2027 guided to $12-13B on $40-45B of capex, pre-tax breakeven Q4 2027. Nscale Q3 revenue $307.5M (+1,344%) with a $1.69B net loss and $105.8B of active-and-contracted take-or-pay TCV ($6.4B active; ~5.7-year weighted life). The series moves from W45's first operating prints (IREN, Core Scientific, Cipher, TeraWulf) to the three largest operatorsPrior First operating prints in the series: IREN AI Cloud revenue $200M (+182% QoQ; backlog $20.3B, ~two-thirds investment-grade counterparties; >$4B ARR target slips from year-end to January-February 2027 after a Microsoft-requested network redesign, no penalty); Core Scientific colocation revenue $215.4M (+58% QoQ) with 485 MW billing of 590 MW contracted to CoreWeave; Cipher's first HPC lease revenue ($5.9M) with Barber Lake contracted revenue at ~$9.1B over 20 years; TeraWulf HPC lease revenue $63.0M. CoreWeave reports November 10 with Q3 interest expense guided at $860-940M against adjusted operating income of $200-260M · Up | CoreWeave's net interest expense of $846M was 2.2x its adjusted operating income of $379M, resolving last week's dated test (guided interest 3.3-4.7x guided adjusted operating income) at the favourable end because adjusted operating income beat its $200-260M guide; interest was 23% of revenue. Nebius lost $37.5M on $980.8M; Nscale lost $1.69B on $307.5M (5.5x revenue). The category-level burn is the capex: CoreWeave's call-stated $14.5B in the quarter against $3.66B of revenue (4.0x; the 10-Q cash-flow purchases figure is $9.44B, 2.6x, and the two differ on timing of accruals), Nebius ~$27B of 2026 capex against $3.0-3.2B of revenue | The funding model migrated from unsecured bonds to customer prepayments and contract-secured asset-level debt, a reading that is the consensus of the CoreWeave print (TechCrunch, TIKR, DCD, Reuters on Nebius; credited, not led with). The house adds the link: the lenders are underwriting lab contracts (Nscale's Anthropic agreement is the filed one; CoreWeave's and Nebius's prepaying customers are unnamed), so the category's credit now keys off the labs' twelve-month commitments, for which ANT's price is the public sentiment read from November 19 (big story) |
| On-Prem / HybridEnterprise GPU clusters, sovereign and national programs, open-weight and on-device deployment | ~$1.7B of sovereign-backed commitments charted, both signed and staged rather than drawn: SoftBank Group, Grab and Petroleum Sarawak's Joint Development Agreement for US$1B of initial seed capital (SoftBank $800M, Grab $100M, PETROS $100M) into a SoftBank-majority AI infrastructure vehicle in Sarawak, Malaysia, with a long-term 5 GW target and deployment staged to development and customer-contract milestones (grade 5); and Saudi Arabia's National Infrastructure Fund lending PIF-owned HUMAIN up to SAR 2.5B (~$667M) in two facilities, up to SAR 1.1B (~$293M) for GPUs representing ~25 MW of IT load and up to SAR 1.4B (~$373M) for substations, transmission and grid connections sized for up to 1 GW of future IT capacity, the first loans under Infra's AI Data Center Financing Program (grade 5). Enterprise deployment capital: none closed; the enterprise-side event was Anthropic's self-hosted Claude Managed Agents private beta, which moves the harness, sandboxes and session store into the customer's cloud or on-premises while routing model calls back to Anthropic or a hyperscaler (Agent Techniques Weekly), on-prem consumption without open weights. Open-weight deployment capacity arrived as artifacts, not financings: Inkling 2 Small's Apache 2.0 weights, llama.cpp v1.0.0, DeepSeek V4 on SageMaker JumpStart (The Model Pulse)Prior None charted. The one new sovereign-scale announcement, a Brookfield-Qai joint venture with Drax for a GW-scale AI campus at Drax Power Station (GBP 2.5B phase 1 for 100 MW by Q1 2028, potentially more than GBP 20B in total; grade 4), has no final investment decision and is carried in text. Enterprise deployment capital: none closed; the enterprise-side signals ran the other way, with Goldman Sachs restricting Claude Code and GitHub coding tools to engineers and quantitative strategists after months of banker and trader use (grade 4; The Application Layer carries the reading). Open-weight deployment capacity arrived as an artifact rather than a financing: Kimi K3.1 Pro's 1.56 TB of MXFP4 weights with an AMD day-zero MI355X path (The Model Pulse). Defence-adjacent rounds (Shield AI $2.5B at $28.8B to buy Echodyne; Zipline $1.5B at $21.5B) are not compute and are not charted · Up | Indirect. The enterprise-return datapoints of the week are the first central-bank staff review in the series and three surveys, all pointing the same way: the Federal Reserve's FEDS 2026-074 (a staff review of existing evidence, not a new measurement) finds aggregate productivity effects of AI 'appear to be small so far'; Genpact (2,000+ executives) has 79% with agents in production, 15% with AI investment fully tied to P&L and 92% having kept or grown headcount; Payscale and PA Consulting carry the worker-side reads (The Application Layer carries all three surveys in full). The two self-reported engineering deployments (Salesforce, Superhuman) are in Agent Techniques Weekly as proof of value; the vendor-side returns (Wix's Base44; monday.com's AI ARR) are in The Application LayerPrior Indirect. The enterprise-return datapoint of the week is the Bank of England's November Monetary Policy Report: firms say AI has raised productivity about 0.5% over three years (0.1-0.2 percentage points a year) while expecting ~0.9 points a year ahead, and AI supply rather than adoption drove about a fifth of 2026 UK growth (The Application Layer carries the reading) · Flat | n/a | Sovereign capital arrived in the load-pays-for-grid shape the big story describes in Kentucky: Infra's SAR 1.4B facility is specifically for substations and grid connections sized to 1 GW, and Sarawak's $1B is staged to customer-contract milestones. The second hypothesis is graded on its restated diffuse leg for the first time this week, with the Fed's staff review as the ruler (synthesis) |
Category
Frontier Labs.
OpenAI, Anthropic, Google DeepMind, DeepSeek
- Capital in
- None closed; the largest raise in the series is priced but not yet printed. Anthropic's S-1/A No. 2 (grade 5) sets a $70-78 range on 825M Class A shares (up to $64.35B gross; ~$1.4T fully diluted at the top) with a 123.75M-share option and a concurrent $26.0B Class N private placement at the IPO price from AMD, Broadcom, Google, Micron, Microsoft, NVIDIA and TSMC Partners; pricing is targeted for November 18 and trading November 19 as ANT, the book was reported about six times covered by November 12 (Bloomberg, Reuters; anonymous sources) and TSMC's board approved up to $500M (grade 5). OpenAI is reported at ~$15B raised so far at $1.25T in a round not yet closed (CNBC, grade 3; text only, as in W45), with TSMC's board approving up to $500M of OpenAI preferred. Nothing is charted until ANT prices; the big story carries the placement as the equity leg of the syndicated backstopvs None charted. OpenAI is reported to have closed ~$11B of a round at $1.25T with up to ~$20B expected 'in the coming weeks' (Bloomberg, grade 3; text only, as last week's report was). Anthropic's S-1/A (grade 5) adds Q3 and discloses $77.0B of undrawn note-purchase capacity (Google $20B, Amazon $15B, Broadcom $42B at 9% PIK) and $146.6B of liquidity, with the price range still blank and the roadshow reported for the week of November 9; the Broadcom facility's implied cost is the house measurement and it is capacity, not capital in. Google's credit-support agreements for Anthropic's TeraWulf lease (8-K, grade 5) are a guarantee, carried in the big story, and move no cash to the lab. Fal's $1.3B Series F at $16B (grade 4) is an inference cloud and is charted under Neoclouds · Flat
- Revenue out
- ~$81B run-rate: Anthropic's S-1/A No. 2 discloses run-rate revenue of about $81B as of October 31, up $10B from September 30, on Q3 revenue of $16,956M (filed in W45); more than 11,900 customers above $100K produced $24.5B (74%) of nine-month revenue, with the five largest at ~25% (The Application Layer carries the customer mix). OpenAI's Q3 revenue is reported at ~$10B (+~50% QoQ, ~$50B annualised) with a $10.9B adjusted operating loss and enterprise at ~50% of the mix (WSJ, Bloomberg; grade 3). The series stays on Anthropic's filed number and moves from the W45 annualised Q3 figure (~$68B) to the October run-ratevs ~$68B annualised: Anthropic's S-1/A files Q3 2026 revenue of $16,956M (+47% QoQ from $11,555M; +1,127% YoY), replacing last week's Reuters-reported $16.8B; nine-month revenue $33.2B; Q3 operating loss $567M, which includes a $2.0B non-cash charge for the donation-matching programme (about +$1.4B and an 8.5% margin without it, level with Q2) alongside compute and infrastructure expense of $9.86B (58% of revenue, from 48%); net loss $23.3B on a $21.9B non-cash fair-value charge. Consumption is 79% of nine-month revenue (The Application Layer carries the customer metrics). SpaceX's SI segment (~$30B annualised, W43) stays in text because the series is Anthropic's · Up
- Burn / rev
- n/a (no chartable capital in this week). The filed stress number is not burn but the twelve-month commitment schedule: ~$70.7B of Anthropic's ~$900.9B of aggregate commitments falls due within twelve months of September 30, about 0.9x the October run-rate, which is the quantity the $26B placement and the IPO proceeds are sized againstLower means more capital out than in.
The read
The lab layer's capital event is one week from printing and already priced the rest of the stack. Sell-side models anchor on 2028 revenue of ~$270-280B (Reuters, November 13; grade 3), bankers were reported weighing $1.5T or more (FT), and the house reads the $26B placement as Levine and Coben do (circular, drawn partly from existing commitments; credited) and then one step further: the placement is the equity leg of the syndicated backstop, since lab contracts are what neocloud lenders now underwrite (Anthropic's own filed counterparty is Nscale; CoreWeave's and Nebius's prepaying customers are unnamed). The counter-reading is the counterbrief's: prepayments relocate credit risk to a borrower with ~$70.7B due in twelve months whose funding is this IPO, so the first dated test is ANT's close against its $70 range low by December 31. The customer-side disclosures (revenue under 0.1% of Fortune 500 customers' expenses; >1,900 customers above $1M) live in The Application Layer.
This week's transactions
| Date | Transaction | Amount | Source |
|---|---|---|---|
| 2026-11-10 | Anthropic S-1/A No. 2: 825,000,000 Class A shares (801.35M primary, 23.65M secondary) at $70-78; 123.75M-share option; up to $64.35B gross at the top of the range; ~$1.4T fully diluted; $26.0B concurrent Class N placement at the IPO price from AMD, Broadcom, Google, Micron, Microsoft, NVIDIA and TSMC Partners; run-rate revenue ~$81B at October 31; ~$900.9B of commitments, ~$70.7B due in 12 months; Nasdaq: ANT, pricing targeted November 18 | Up to $64.35B IPO + $26.0B placement (priced, not closed) | SEC EDGAR (S-1/A No. 2); Anthropic; CNBC |
| 2026-11-12 | Anthropic's book reported about six times oversubscribed by the end of day three of the roadshow (Bloomberg, Reuters; anonymous sources); FT reports bankers weighing a valuation of $1.5T or more; ~20% retail allocation (CNBC) | ~6x covered (reported) | Bloomberg via Bloomberg Law; Reuters; CNBC |
| 2026-11-12 | OpenAI reported to have raised ~$15B so far at a $1.25T valuation in a round not yet closed, described as a bridge to a 2027 IPO (original ask up to $30B at $1.4T); Q3 revenue ~$10B with a $10.9B adjusted operating loss (WSJ / Bloomberg, November 10); Nvidia ~$30B and SoftBank ~$65B provided to date (grade 3; text only) | ~$15B (reported) | CNBC (Faber Report); Reuters citing WSJ via Investing.com |
| 2026-11-10 | TSMC's board approves up to $500M of Anthropic shares at the IPO price and up to $500M of OpenAI preferred stock, the foundry taking equity in both labs whose commitments fill its CoWoS book | Up to $1.0B (approved) | SEC EDGAR (Anthropic S-1/A No. 2); CNBC |
Trend across recent issues
- Capital in
- $0B, $10B, $0B, $0B, $0B, $40B, $0B, $0B
- Revenue out
- n/a, $4.6B, $4.6B, $67B, $67B, $67B, $68B, $81B
Category
Hyperscaler-Hosted.
Azure-OpenAI, AWS-Anthropic, Google Cloud-Gemini, Oracle-OCI
- Capital in
- None closed in the graded packet this week; the category's cash moved the other way. Google, Microsoft, AMD, Broadcom, Micron, NVIDIA and TSMC are buyers in Anthropic's $26.0B Class N placement (charted under Frontier Labs when it prints), and Morgan Stanley counts $67B of hyperscaler bonds priced since September 30 (Microsoft, Meta, Oracle in the US; Alphabet in Europe) with spreads on most tightened since issue (Oracle the exception) and issuance expected quiet for the rest of 2026 (CNBC, grade 4). Goldman Sachs Research projects $420B of hyperscaler debt issuance in 2027 against ~$250B in 2026, and Morgan Stanley forecasts $1.66T of AI capex in 2027 (+75%); both are forecasts, not flowsvs ~$35B closed: Oracle's $20B (86.2M shares at $116, an 11.2% discount to the last close, ~$9.9B net; and $10B of senior notes in five tranches at 7.000-8.650%, Treasuries +240 to +335 bp, Baa2 negative / BBB- stable), both settled November 4 and filling the $20B cap set at its October 28 Investor Day; and Meta's $15B of senior notes in five tranches at 5.850-7.500% (Treasuries +110 to +220 bp, Aa3 stable / AA- negative), settled November 5. Meta's underwriters also exercised the 6M-share option on the forward sale, lifting expected forward proceeds to ~$28.7B on settlement by November 19, 2027 (cash not yet received; not charted). Google's $17B Wyoming campus commitment with a 702 MW fuel-cell PPA is capex, not a raise, and is carried in text · Down
- Revenue out
- Indirect; no AI-segment revenue line, and no hyperscaler reported this week. The lines carried from W44 stand (Google Cloud $30.7B +103%, Azure +48%, AWS $46.0B +39%). The nearest new datapoints are from suppliers and customers: Cisco raised its FY27 hyperscaler AI-infrastructure revenue target to $9B from $7.5B (more than 2x FY26), which is hyperscaler spend seen from the vendor side; Anthropic's S-1/A says its revenue averages under 0.1% of its Fortune 500 customers' total expenses; Microsoft shares fell 2-3% on November 10, the day of Anthropic's Claude Meetings / Sheets / Whiteboards launch inside the $20 seat (The Application Layer; coincidence recorded, not cause). Azure's ND MI455X v7 Helios racks (72 MI455X, 18 EPYC Venice, 31 TB HBM4) are the hardware lens's availability evidence for AMDvs Indirect; no AI-segment revenue line. No hyperscaler reported this week; the lines carried from W44 stand (Google Cloud $30.7B +103%, Azure +48%, AWS $46.0B +39%). The nearest new datapoint is the price of the capital: at matched tenors Oracle paid 165 bp more than Meta over Treasuries in 2031 (T+275 vs T+110), 160 bp more in 2033 and 135 bp more in 2046, the price gap between a BBB and an AA hyperscaler for the same use of proceeds (house measurement from the two FWPs) · Flat
- Burn / rev
- n/a (no AI revenue line; cloud segments are not AI-attributable)Lower means more capital out than in.
The read
The hyperscaler layer spent the week as a supplier of capital rather than a raiser of it: three of the seven Class N buyers are hyperscalers, Morgan Stanley expects no further hyperscaler bond issuance in 2026, and the one hyperscaler policy action in the packet was Microsoft pressing the White House to drop an FCC ban on Chinese transceivers because 'many clusters need millions' and purchased GPUs would sit idle (Reuters, grade 3). Levine's structural point (credited) is the one for this row: a share of Anthropic is worth more to Google than to a public buyer because the money returns as TPU rent, so the placement is, from the hyperscaler's side, a prepayment on its own future revenue. The revenue line stays indirect; the supplier read (Cisco's $9B target, excluding neoclouds as 'relatively immaterial') is the week's best proxy for how much of the 2027 capex Morgan Stanley forecasts at $1.66T is already committed to network vendors.
This week's transactions
| Date | Transaction | Amount | Source |
|---|---|---|---|
| 2026-11-10 | Google, Microsoft, NVIDIA, AMD, Broadcom, Micron and TSMC Partners agree to buy $26.0B of Anthropic Class N stock at the IPO price (351,351,348 shares at the $74 midpoint); Google's and AMD's purchases reportedly draw on pre-existing commitments (Money Stuff, credited) | $26.0B (at IPO pricing) | SEC EDGAR (Anthropic S-1/A No. 2); Money Stuff via Milled |
| 2026-11-11 | Morgan Stanley: $67B of hyperscaler bonds priced since September 30, most spreads tighter since issue (Oracle the exception), issuance to stay quiet for the rest of 2026; 2027 AI capex forecast $1.66T (+75%), 2028 $1.77T; Goldman Sachs Research: $420B of hyperscaler debt issuance in 2027 vs ~$250B in 2026 (forecasts; grade 4) | $67B priced since Sep 30 (tally) | CNBC (Morgan Stanley note); Reuters via Investing.com |
| 2026-11-10 | Reuters: Microsoft and the Information Technology Industry Council press the White House to halt an FCC effort to bar new Chinese optical transceiver models; FCC reportedly weighing a 3.2 Tbps threshold; Microsoft has 35 US data centres under construction and 134 planned (Cleanview) (grade 3; policy, not a flow) | n/a | Reuters via Investing.com; Tom's Hardware |
Trend across recent issues
- Capital in
- $0B, $0B, $0B, $0B, $0B, $38B, $35B, $0B
- Revenue out
- n/a, n/a, n/a, n/a, n/a, n/a, n/a, n/a
Category
Neoclouds.
CoreWeave, Nscale, Crusoe, Lambda, IREN, Zankore, NEXTDC
- Capital in
- ~$2.3B of contract-secured bank paper charted, all at the asset or developer level and none unsecured: Rowan Digital Infrastructure's more than $1.5B of new corporate capacity (a $600M upsize of its Apterra-led holdco term loan to $900M, a $590M revolver and a $355M letter-of-credit facility led by Citibank, with accordions to $2B and $1B) for a 5.5 GW development pipeline (grade 5); Blue Owl Digital Infrastructure Trust's ~$512.6M floating-rate Citibank / Wells Fargo mortgage on a 54 MW Gainesville, Virginia building 100% leased to a hyperscaler, closed November 10 with the ~$779.5M acquisition (8-K, grade 5); and WhiteFiber's $300M Morgan Stanley-agented term loan at SOFR+600 with a 2% floor for the 40 MW NC-1 campus leased to Nscale, signed November 9 and unfunded pending customer acceptance (8-K, grade 5). Not charted: CoreWeave's ~$9B of expected customer prepayments over three months and its 'high investment-grade' asset-level facility expected within one to two weeks at ~7-7.5% (CEO on CNBC); Nebius's $17-20B of anticipated 2027 prepayments and its $3.5B Citi secured facility signed October 22 after shelving a high-yield bond (Reuters, November 9); Nscale's S-1/A with no price range. The unsecured market stayed shut: CoreWeave will not tap it through 2027 at current yields, Reuters reports banks reluctant to add exposure even to deals with marquee-tech guarantees and some offerings shelved, and Galaxy's Helios data-centre bonds fell to 89.25 to yield ~13% against a 9.875% July couponvs ~$2.2B charted: Cipher's $850M 0% convertible due 2032 (closed, grade 4) and Fal's $1.3B Series F at a $16B valuation (grade 4), an inference cloud whose capital funds GPU capacity. IREN's $15B of remaining FY27 capex is to be funded by $9B of cash, committed GPU financing and prepayments plus $6B of targeted financing 'subject to market conditions', none closed this week. The category's price moved more than its volume: CoreWeave's 9.625% 2032 notes yielded above 16% and its five-year CDS spiked past 1,289 bp on November 5 (~1,155 on November 6); IFR reports five data-centre IPOs pulled or delayed and some developers offering tenants equity to keep projects alive; Bernstein sizes CoreWeave's remaining need at ~$85B to deliver its $104B backlog. Microsoft is consolidating its neocloud roster, declined to renew an Oracle capacity deal and passed on doubling its Nscale Texas contract (CNBC, Reuters; grade 4) · Flat
- Revenue out
- ~$20B annualised across the three operators that printed: CoreWeave Q3 revenue $3,661M (+168% YoY, more than $1B sequential; beat ~$3.55B consensus) with backlog ~$132.5B (from $104B), operating income $207M, adjusted operating income $379M, net interest expense $846M and a $709M net loss; 2026 revenue guide raised to $13.0-13.3B, capex to $37-39B, exit-2026 run-rate $20-21B, Q4 interest expense $1.05-1.15B, adjusted profitability targeted 1H 2028; short-term capacity pricing ~$40M/MW-year against $20-25M/MW-year on 4-5 year contracts. Nebius Q3 revenue $980.8M (+571%, including ~$325M non-recurring asset-light revenue), adjusted EBITDA $465.8M (47.5%), ARR $6.4B, 2026 guide trimmed to $3.0-3.2B on third-party capacity slipping into 1H 2027, 2027 guided to $12-13B on $40-45B of capex, pre-tax breakeven Q4 2027. Nscale Q3 revenue $307.5M (+1,344%) with a $1.69B net loss and $105.8B of active-and-contracted take-or-pay TCV ($6.4B active; ~5.7-year weighted life). The series moves from W45's first operating prints (IREN, Core Scientific, Cipher, TeraWulf) to the three largest operatorsvs First operating prints in the series: IREN AI Cloud revenue $200M (+182% QoQ; backlog $20.3B, ~two-thirds investment-grade counterparties; >$4B ARR target slips from year-end to January-February 2027 after a Microsoft-requested network redesign, no penalty); Core Scientific colocation revenue $215.4M (+58% QoQ) with 485 MW billing of 590 MW contracted to CoreWeave; Cipher's first HPC lease revenue ($5.9M) with Barber Lake contracted revenue at ~$9.1B over 20 years; TeraWulf HPC lease revenue $63.0M. CoreWeave reports November 10 with Q3 interest expense guided at $860-940M against adjusted operating income of $200-260M · Up
- Burn / rev
- CoreWeave's net interest expense of $846M was 2.2x its adjusted operating income of $379M, resolving last week's dated test (guided interest 3.3-4.7x guided adjusted operating income) at the favourable end because adjusted operating income beat its $200-260M guide; interest was 23% of revenue. Nebius lost $37.5M on $980.8M; Nscale lost $1.69B on $307.5M (5.5x revenue). The category-level burn is the capex: CoreWeave's call-stated $14.5B in the quarter against $3.66B of revenue (4.0x; the 10-Q cash-flow purchases figure is $9.44B, 2.6x, and the two differ on timing of accruals), Nebius ~$27B of 2026 capex against $3.0-3.2B of revenueLower means more capital out than in.
The read
The neocloud layer had its best operating week and its clearest funding statement, and the two are one story. CoreWeave's revenue rose 168% with a $132.5B backlog on 4.6 GW of contracted power against 1.7 GW active (the gap the prepayments and the facility have to build), and its CEO said the company will not use the unsecured market at current yields, expects ~$9B of prepayments in three months and will print a high-investment-grade asset-level facility at ~7-7.5% within weeks; the stock rose ~25% and the 2031 notes rallied to 14.79% from 16.7% (CNBC; outside the graded packet). Nebius said the same thing in guidance form (75-85% of 2027 capex from prepayments and investment-grade asset-backed debt) and fell 16% because 2026 revenue slipped on third-party capacity delays; Nscale's S-1/A said it in a risk factor (financing for its Anthropic build 'not yet binding'; a $3.22B Microsoft work order terminated on a grid date; counterbrief reading). The three bank financings that did close (Rowan, Blue Owl, WhiteFiber) were each secured on a contract or a leased building, with WhiteFiber's SOFR+600 loan for an Nscale-leased campus funding only on customer acceptance, which is the lender pricing the tenant chain rather than the operator. TechCrunch names the cost (prepaid contracts lock pricing that looks worse if compute tightens; lenders discount the short contracts CoreWeave now favours at ~$40M/MW-year), and TIKR names the condition ('as long as they keep paying'), which is the condition the bond market was pricing at 16%. For an investor the test is dated: CoreWeave's facility at 7-7.5% by mid-December and the 2031 notes holding below their 16.7% pre-print yield through ANT's first week confirm the syndication reading; a prepayment-guidance cut at either operator's next print breaks it.
This week's transactions
| Date | Transaction | Amount | Source |
|---|---|---|---|
| 2026-11-10 | CoreWeave Q3: revenue $3,661M (+168%), backlog ~$132.5B, adjusted operating income $379M vs $200-260M guide, net interest expense $846M, net loss $709M; 2026 revenue $13.0-13.3B and capex $37-39B; ~$9B of customer prepayments expected in three months, ~70% of 2026 deals with prepayments; no unsecured issuance through 2027; asset-level facility expected within 1-2 weeks, high investment grade, ~7-7.5% (CEO on CNBC); weighted average cost of debt 7.9%; 10-Q filed November 12 | ~$9B prepayments expected (not closed) | CoreWeave IR; CNBC; Motley Fool (call transcript) |
| 2026-11-12 | Nebius Q3: revenue $980.8M (+571%), adjusted EBITDA $465.8M, ARR $6.4B; 2026 revenue guide cut to $3.0-3.2B on third-party capacity delays, 2026 capex raised to ~$27B; 2027: revenue $12-13B, capex $40-45B with 75-85% covered by $17-20B of prepayments plus investment-grade asset-backed financing; high-yield bond shelved in October for a $3.5B Citi senior secured facility signed October 22 (Reuters, November 9); shares -16% | $17-20B prepayments anticipated for 2027 (guidance) | Nebius (Q3 results release and shareholder letter); Reuters; CNBC |
| 2026-11-12 | Rowan Digital Infrastructure: $600M upsize of Apterra-led holdco term loan to $900M; new $590M revolver and $355M letter-of-credit facility led by Citibank (accordions to $2B and $1B); 5.5 GW US pipeline; $4.4B raised under its Green Finance Framework to date | $1.5B+ | Rowan Digital Infrastructure |
| 2026-11-10 | Blue Owl Digital Infrastructure Trust closes the ~$779.5M acquisition of a 54 MW Gainesville, Virginia data centre 100% leased to a hyperscaler, funded with offering cash and a ~$512.6M floating-rate mortgage from Citibank and Wells Fargo (8-K) | $512.6M mortgage | SEC EDGAR (Blue Owl Digital Infrastructure Trust 8-K) |
| 2026-11-09 | WhiteFiber (Enovum NC-1): $300M five-year term loan ($200M initial, $100M delayed-draw), SOFR+6.00% with a 2.00% floor, Morgan Stanley Senior Funding as agent, for the 40 MW NC-1 campus whose first 40 MW is leased to Nscale (~$865M over 10 years per prior disclosure); unfunded as of November 10, funding contingent on customer acceptance (8-K) | $300M (unfunded) | SEC EDGAR (WhiteFiber 8-K) |
| 2026-11-12 | Nscale S-1/A No. 1 (NYSE: NSCL; no price range): Q3 revenue $307.5M (+1,344%), net loss $1.69B (Q2: $1.02B on $140.6M); take-or-pay TCV $6.4B active and $105.8B active-and-contracted, ~5.7-year weighted life; $1.5B cash at June 30; $3.36B of pre-IPO convertible notes raised September 25 | IPO size blank | SEC EDGAR (Nscale S-1/A No. 1); CNBC |
Trend across recent issues
- Capital in
- $4.2B, $1.4B, $0.2B, $5.8B, $4.2B, $0B, $2.2B, $2.3B
- Revenue out
- n/a, n/a, n/a, n/a, n/a, n/a, n/a, $20B
Category
On-Prem / Hybrid.
Enterprise GPU clusters, sovereign and national programs, open-weight and on-device deployment
- Capital in
- ~$1.7B of sovereign-backed commitments charted, both signed and staged rather than drawn: SoftBank Group, Grab and Petroleum Sarawak's Joint Development Agreement for US$1B of initial seed capital (SoftBank $800M, Grab $100M, PETROS $100M) into a SoftBank-majority AI infrastructure vehicle in Sarawak, Malaysia, with a long-term 5 GW target and deployment staged to development and customer-contract milestones (grade 5); and Saudi Arabia's National Infrastructure Fund lending PIF-owned HUMAIN up to SAR 2.5B (~$667M) in two facilities, up to SAR 1.1B (~$293M) for GPUs representing ~25 MW of IT load and up to SAR 1.4B (~$373M) for substations, transmission and grid connections sized for up to 1 GW of future IT capacity, the first loans under Infra's AI Data Center Financing Program (grade 5). Enterprise deployment capital: none closed; the enterprise-side event was Anthropic's self-hosted Claude Managed Agents private beta, which moves the harness, sandboxes and session store into the customer's cloud or on-premises while routing model calls back to Anthropic or a hyperscaler (Agent Techniques Weekly), on-prem consumption without open weights. Open-weight deployment capacity arrived as artifacts, not financings: Inkling 2 Small's Apache 2.0 weights, llama.cpp v1.0.0, DeepSeek V4 on SageMaker JumpStart (The Model Pulse)vs None charted. The one new sovereign-scale announcement, a Brookfield-Qai joint venture with Drax for a GW-scale AI campus at Drax Power Station (GBP 2.5B phase 1 for 100 MW by Q1 2028, potentially more than GBP 20B in total; grade 4), has no final investment decision and is carried in text. Enterprise deployment capital: none closed; the enterprise-side signals ran the other way, with Goldman Sachs restricting Claude Code and GitHub coding tools to engineers and quantitative strategists after months of banker and trader use (grade 4; The Application Layer carries the reading). Open-weight deployment capacity arrived as an artifact rather than a financing: Kimi K3.1 Pro's 1.56 TB of MXFP4 weights with an AMD day-zero MI355X path (The Model Pulse). Defence-adjacent rounds (Shield AI $2.5B at $28.8B to buy Echodyne; Zipline $1.5B at $21.5B) are not compute and are not charted · Up
- Revenue out
- Indirect. The enterprise-return datapoints of the week are the first central-bank staff review in the series and three surveys, all pointing the same way: the Federal Reserve's FEDS 2026-074 (a staff review of existing evidence, not a new measurement) finds aggregate productivity effects of AI 'appear to be small so far'; Genpact (2,000+ executives) has 79% with agents in production, 15% with AI investment fully tied to P&L and 92% having kept or grown headcount; Payscale and PA Consulting carry the worker-side reads (The Application Layer carries all three surveys in full). The two self-reported engineering deployments (Salesforce, Superhuman) are in Agent Techniques Weekly as proof of value; the vendor-side returns (Wix's Base44; monday.com's AI ARR) are in The Application Layervs Indirect. The enterprise-return datapoint of the week is the Bank of England's November Monetary Policy Report: firms say AI has raised productivity about 0.5% over three years (0.1-0.2 percentage points a year) while expecting ~0.9 points a year ahead, and AI supply rather than adoption drove about a fifth of 2026 UK growth (The Application Layer carries the reading) · Flat
- Burn / rev
- n/aLower means more capital out than in.
The read
The on-prem and sovereign row carried the week's thesis-page consequence. The Federal Reserve's staff review is the ruler the W45 restatement of the second hypothesis asked for, and its first reading is 'small so far', which with Genpact's 15% and the Bank of England's 0.5% over three years (W45) is three independent reads of diffuse return near zero against a sixth week of returns printing one layer up the supply chain; the house applies the restatement (direct capture, tested on supplier and lab prints; diffuse capture, tested on central-bank and statistical-agency data), grades both legs supported in their first week, and retires the v1 diffuse wording unscored. The counter-evidence is real and is in the synthesis: Salesforce and Superhuman self-report halved cycle time and 60% more merged PRs, Wix's Base44 runs at a 58% gross margin, and more than half of the Fortune 500 pays Anthropic, which is diffusion of use if not yet of measured return. The sovereign capital is the other half of the row: both signed commitments this week finance power and grid ahead of compute (SAR 1.4B for substations sized to 1 GW; Sarawak staged to contracts), which is the same shape as the TeraWulf contract and the reason the fifth hypothesis is re-scoped to 2028 rather than dropped. For a CIO the practical datapoint is Anthropic's self-hosted Managed Agents: the harness can now sit inside the perimeter while the model stays remote, which is the demand route the fourth hypothesis did not anticipate and the one most enterprise buyers will take before they host weights.
This week's transactions
| Date | Transaction | Amount | Source |
|---|---|---|---|
| 2026-11-09 | SoftBank Group, Grab and Petroleum Sarawak Berhad sign a Joint Development Agreement for US$1B of initial seed capital (SoftBank $800M, Grab $100M, PETROS $100M) into a SoftBank-majority AI infrastructure company in Sarawak, Malaysia; long-term target 5 GW; deployment staged to development and customer-contract milestones; builds on the October 8 framework agreement | $1.0B seed (committed, staged) | SoftBank Group |
| 2026-11-09 | Saudi Arabia's National Infrastructure Fund (Infra) and HUMAIN announce up to SAR 2.5B (~$667M) in two facilities: up to SAR 1.1B (~$293M) for GPUs (~25 MW of IT load) and up to SAR 1.4B (~$373M) for substations, transmission and grid connections sized for up to 1 GW of IT capacity; first loans under Infra's AI Data Center Financing Program; HUMAIN targets 1.7 GW in 2027 and ~3 GW by 2030 | Up to SAR 2.5B (~$667M) | HUMAIN; EnterpriseAM |
| 2026-11-12 | Federal Reserve FEDS 2026-074: AI adoption growing rapidly but most firms use it lightly or not at all; labour-market effects 'small and somewhat mixed'; aggregate productivity effects 'appear to be small so far' (staff working paper; the diffuse-returns ruler) | n/a | Federal Reserve (FEDS 2026-074) |
| 2026-11-12 | Anthropic opens a private beta of self-hosted Claude Managed Agents: harness, sandboxes, credential vault, PostgreSQL session history and workers run in the customer's cloud or on-premises with model calls routed to Anthropic, Bedrock, Google Cloud, Microsoft Foundry or a customer gateway; pricing and GA not stated (Agent Techniques Weekly carries the controls) | n/a | Anthropic (claude.com); Help Net Security |
Trend across recent issues
- Capital in
- $0B, $0B, $0B, $0B, $0B, $9B, $0B, $1.7B
- Revenue out
- n/a, n/a, n/a, n/a, n/a, n/a, n/a, n/a