--- title: >- A four-week-old house frame finally got a denominator, and almost every other number this week resolved into an option publication: The AI Stack Weekly slug: 2026-W37 issueNumber: 21 isoYear: 2026 isoWeek: 37 publishedAt: '2026-09-12' canonicalUrl: https://brianletort.ai/industry/weekly/2026-W37 pdfUrl: https://brianletort.ai/downloads/ai-stack-weekly-2026-W37.pdf schemaVersion: 2026.05.02 flywheelArc: all-three capitalFlow: - category: Frontier Labs capitalIn: Unknown capitalInPrior: Unknown capitalInDirection: flat revenueOut: Unknown revenueOutPrior: Unknown revenueOutDirection: flat burnToRevenue: n/a - category: Hyperscaler-Hosted capitalIn: >- Unknown as a category total; two constituents disclosed $28.5B of quarterly capital expenditure and an announced plan of at least EUR 13B over two years respectively capitalInPrior: Unknown capitalInDirection: flat revenueOut: >- Unknown as a category total; one constituent disclosed triple-digit cloud infrastructure revenue growth and $664B of contracted backlog revenueOutPrior: Unknown revenueOutDirection: flat burnToRevenue: n/a - category: Neoclouds capitalIn: >- Unknown as a category total; in-window instruments comprise an up-to-$3.1B undrawn facility, A$1.1B of subordinated convertible notes, and $375M firm of an $875M headline capitalInPrior: Unknown capitalInDirection: flat revenueOut: Unknown revenueOutPrior: Unknown revenueOutDirection: flat burnToRevenue: n/a - category: On-Prem / Hybrid capitalIn: Unknown capitalInPrior: Unknown capitalInDirection: flat revenueOut: Indirect revenueOutPrior: Indirect revenueOutDirection: flat burnToRevenue: n/a levers: - metric: Frontier lab cash runway at current burn current: Unknown — no constituent disclosed cash, burn, or financing in the window prior: Unknown — W36 had no reproducible cash-and-burn input table either direction: flat threshold: Below 18 months for any disclosed-burn lab - metric: Hyperscaler AI capex to disclosed AI revenue ratio current: >- Unknown as a top-four ratio — one constituent disclosed $28.5B of quarterly capital expenditure and roughly $5B of negative free cash flow, with no AI-segment revenue line prior: Unknown — no reproducible top-four capex and AI-revenue input range direction: flat threshold: Above 6x sustained for two consecutive quarters - metric: CoreWeave contracted revenue backlog current: $104.2B as of June 30, unchanged — no CoreWeave filing landed in the window prior: $104.2B as of June 30 direction: flat threshold: Sequential decline, or conversion below 15% annually - metric: NVIDIA quarter-over-quarter data center revenue current: $89.0B for Q2 FY27, unchanged — no NVIDIA print in the window prior: $89.0B for Q2 FY27 direction: flat threshold: Two consecutive quarters of sequential decline - metric: Open-weight to closed-model capability gap on coding current: >- Not measurable this week — the reference index changed basis twice in four days, breaking comparability with the prior reading prior: Untested in W36 as well; no new independent open-frontier benchmark set direction: flat threshold: Open weights within 2 Index points of the closed leader - metric: Sovereign AI program commitments current: >- Unknown — no government-funded national compute programme in the window meets the ledger method prior: Unknown — W36 had no reproducible programme ledger direction: flat threshold: Above 20 programs or $250B committed - metric: PJM capacity auction clearing price current: $325.00 per MW-day for 2028/29, unchanged — no auction occurred in the window prior: $325.00 per MW-day for 2028/29 direction: flat threshold: An auction clearing below the cap, or a FERC-approved increase in the cap itself - metric: Time from interconnection request to energization current: >- Unknown as a queue duration — the window's only hard figure is 850 MW delivered within one quarter by a single operator, which measures delivery rather than energisation or queue time prior: Unknown — no reproducible multi-queue duration table direction: flat threshold: Below 48 months in two or more major queues - metric: Cost per task, frontier reasoning model current: >- $3.26 is the floor of the two leading models and $7.63 the other, per index task on the new v4.3 basis; no frontier-tier median is computable because the basis changed twice inside the window, and neither figure is commensurate with the prior reading prior: Prior reading withheld because the index basis changed twice inside the window direction: flat threshold: A frontier-tier reasoning model below $1 per million output tokens - metric: Custom silicon share of hyperscaler AI compute current: >- Unknown — a multi-generation custom inference agreement was signed but discloses a warrant-vesting ceiling rather than units, share, or a service date prior: Unknown — prior quarter's supplier disclosure combined custom accelerators with networking direction: flat threshold: Above 45% share with audited hyperscaler mix disclosure predictions: - id: p105-second-operator-delivered-capacity-mar31 lens: capital confidencePct: 62 deadline: By March 31, 2027 text: >- A publicly traded operator other than Oracle discloses, for a specific reporting period, both a megawatt capacity figure delivered or placed in service and a unit count of AI accelerators delivered, by March 31, 2027. - id: p106-loviisa-fid-jun30 lens: power confidencePct: 44 deadline: By June 30, 2027 text: >- Fortum announces an approved investment decision covering at least EUR 300 million of the EUR 700 million of Loviisa life-extension capital expenditure currently disclosed as pending, by June 30, 2027. - id: p107-runtime-manifest-mar31 lens: software confidencePct: 38 deadline: By March 31, 2027 text: >- A major model provider or evaluation publisher ships a machine-readable runtime or harness manifest that ties a published score to a reproducible configuration, by March 31, 2027. - id: p108-pjm-large-load-filing-dec31 lens: power confidencePct: 58 deadline: By December 31, 2026 text: >- PJM's Section 205 filing on large computational loads is docketed by December 31, 2026 and carries a telemetry or remote-disconnect requirement, not merely a ride-through envelope or a ramp-rate limit. - id: p109-1600zr-two-vendors-jun30 lens: networking confidencePct: 66 deadline: By June 30, 2027 text: >- At least two distinct vendors announce 1600ZR-conformant coherent pluggable optics products by June 30, 2027. - id: p110-agents-api-residency-mar31 lens: software confidencePct: 29 deadline: By March 31, 2027 text: >- OpenAI's managed Agents API supports zero data retention or a non-US data residency option by March 31, 2027. - id: p111-custom-inference-service-date-mar31 lens: hardware confidencePct: 33 deadline: By March 31, 2027 text: >- Qualcomm or its counterparty discloses a named service date, first-deployment date, or unit volume for the multi-generation custom AI inference agreement, by March 31, 2027. predictionsPrior: - id: p100-astra-neutral-memory-dec15 lens: software outcome: pending deadline: By December 15, 2026 text: >- ARC Prize or another provider-neutral evaluator publishes an Astra run without provider-private reasoning state that closes at least half of the 35.9-point matched-effort Standard-to-Provider-Adapter gap by December 15, 2026. - id: p101-fervo-expansion-firm-jun30 lens: power outcome: pending deadline: By June 30, 2027 text: >- Google accepts at least 500 MW of Fervo's conditional expansion and the parties execute a definitive agreement by June 30, 2027. - id: p102-second-queue-contract-dec31 lens: capital outcome: pending deadline: By December 31, 2026 text: >- A second AI infrastructure contract above $500M discloses both a future service date and an option, delayed-draw, earnout, or guarantee allocating schedule risk by December 31, 2026. - id: p103-fabric-one-control-schema-dec31 lens: networking outcome: pending deadline: By December 31, 2026 text: >- Equinix publishes Fabric One beta documentation that exposes approval, rollback, or auditable intent-history controls before December 31, 2026. - id: p104-broadcom-ai-split-dec31 lens: hardware outcome: pending deadline: By December 31, 2026 text: >- Broadcom discloses separate quarterly revenue figures for custom AI accelerators and AI networking by December 31, 2026. signalScores: - 5 - 5 - 4 - 3 - 2 - 2 keyTakeaways: - >- Oracle disclosed 850 MW of capacity and more than 300,000 accelerators delivered in one quarter, which gives this publication's four-week committed-versus-optioned frame its first dated denominator. Oracle's word is delivered, not energised, and the two are not the same quantity. - >- Almost every other large number this week is a ceiling conditioned on a decision the announcer does not control: unapproved utility capital, warrant-vesting tranches, an undrawn export-licence-contingent facility, eight operators' separate pipelines. - >- Approximately 70% of the capital expenditure behind Google's Finnish nuclear commitment is disclosed by the counterparty as pending an investment decision, and the deal contracts roughly 10.6 times more electricity than it adds. - >- Measurement instability, not capability, was the week's binding procurement risk: an index rebased twice in four days, a flagship endpoint rerouted to a smaller model, and a model deprecated inside agent mode on announcement day. - >- Networking was the best-evidenced lens: a 1.6-terabit coherent interface became a multi-vendor specification, high-end routing grew 25% year over year in the second quarter, and both major accelerator events turned on interconnect rather than compute. - >- Policy built enforcement machinery without attaching an obligation, with an auditor registry that mandates no audits and a grid-security rule whose substance is deferred to a later filing. byTheNumbers: - value: 850 MW label: Capacity delivered in one quarter - value: 300,000+ label: Accelerators delivered to customers in the quarter - value: $664B label: Contracted backlog, up $209B year over year - value: ~70% label: Finnish nuclear programme capital still pre-decision - value: +53.3% label: TSMC August revenue growth year over year - value: +94% label: Direct-to-cloud router sales growth, year over year --- # A four-week-old house frame finally got a denominator, and almost every other number this week resolved into an option *Issue 21 · Week 37 of 2026 · Published 2026-09-12* ## Executive summary - Oracle disclosed 850 MW of capacity and more than 300,000 accelerators delivered in one quarter, which gives this publication's four-week committed-versus-optioned frame its first dated denominator. Oracle's word is delivered, not energised, and the two are not the same quantity. - Almost every other large number this week is a ceiling conditioned on a decision the announcer does not control: unapproved utility capital, warrant-vesting tranches, an undrawn export-licence-contingent facility, eight operators' separate pipelines. - Approximately 70% of the capital expenditure behind Google's Finnish nuclear commitment is disclosed by the counterparty as pending an investment decision, and the deal contracts roughly 10.6 times more electricity than it adds. - Measurement instability, not capability, was the week's binding procurement risk: an index rebased twice in four days, a flagship endpoint rerouted to a smaller model, and a model deprecated inside agent mode on announcement day. - Networking was the best-evidenced lens: a 1.6-terabit coherent interface became a multi-vendor specification, high-end routing grew 25% year over year in the second quarter, and both major accelerator events turned on interconnect rather than compute. - Policy built enforcement machinery without attaching an obligation, with an auditor registry that mandates no audits and a grid-security rule whose substance is deferred to a later filing. **By the numbers.** - **850 MW** — Capacity delivered in one quarter (Oracle's Q1 FY27 disclosure for the quarter ended August 31, stated as almost three times the prior quarter and, on the same day's earnings call, 73% of all fiscal 2026 deliveries. The company's verb is delivered rather than energised.) - **300,000+** — Accelerators delivered to customers in the quarter (Company-reported operational fact from the same release; GPU models were not broken out.) - **$664B** — Contracted backlog, up $209B year over year (Remaining performance obligations, a contracted balance rather than revenue or cash. Management expects roughly half to convert over 36 months, which is guidance rather than commitment.) - **~70%** — Finnish nuclear programme capital still pre-decision (House computation: EUR 700M of an approximately EUR 1,000M life-extension programme, disclosed by Fortum as pending investment decisions.) - **+53.3%** — TSMC August revenue growth year over year (Record NT$514.8 billion monthly revenue, from the company's Form 6-K monthly report.) - **+94%** — Direct-to-cloud router sales growth, year over year (Dell'Oro measurement for 2Q 2026 against the same quarter a year earlier, alongside 25% growth in high-end routing overall. A market-research figure off an undisclosed base.) ## Big Story This publication has been separating committed capital from optioned capital since W34, and the distinction is not ours alone: the delivered-versus-announced frame was already built out in coverage that predates the window, and smaller operators had already filed dated in-service megawatt figures. What W37 added is the measurement. On September 10, Oracle reported that it had delivered 850 megawatts of data centre capacity and more than 300,000 accelerators to customers inside a single quarter, against $28.5 billion of quarterly capital expenditure and roughly $5 billion of negative free cash flow. The pairing is what is new: megawatts delivered next to accelerator units delivered, for one named period, at hyperscale. Nobody else published that this week, and it is what finally lets the optioned stack be sized. One caution the release imposes on us. Oracle's verb is delivered. It does not say energised, in service, or placed in service, and delivered capacity is not operating capacity. Run the week's other headline numbers through it and they resolve into something different from what they appear to be. The same filing put backlog at $664 billion and said the new contracts are structured so that customers prepay or bring their own hardware, requiring no incremental capital from Oracle... which is a real capital-efficiency innovation and also a transfer of the financing obligation to counterparties who did not disclose it. Google announced an investment plan of at least €13 billion in Finland and a twenty-two year claim on half a nuclear plant's output, and the price-protection claim attached to it depends on roughly €700 million of life-extension capital that the utility itself discloses as pending an investment decision. Qualcomm's multi-generation inference agreement carries a reported up-to-$60 billion ceiling from secondary reporting of a filing this publication could not open; it vests through milestone warrant tranches, with no service date disclosed. NVIDIA's two-gigawatt Australian target aggregates eight independent operators' pipelines. A two-month-old Indonesian platform's up-to-$3.1 billion facility is undrawn and collateralised against hardware awaiting export licences. And a $13.3 billion recurring-revenue figure is our arithmetic rather than the discloser's: $1.11 billion of monthly billing multiplied by twelve, on a contract whose term, counterparty and megawatts the executive declined to state. He gave a monthly rate and a separate company-wide target. He did not annualise this contract, and nothing in the public record lets anyone test the annualisation we just did. None of that is fraud and most of it is competent structuring. Staging exposure behind milestones is what careful parties do. But it means the honest unit of analysis is the conditionality rather than the headline, and the conditions are almost always disclosed by a counterparty rather than by the announcer. That is where diligence has to go, and almost nobody went there this week. The software layer moved the same way, from capability toward delivery terms. DeepSeek published an MIT-licensed checkpoint and simultaneously announced that requests to its existing flagship endpoint would begin serving a different, smaller model at different rates with no change required by the caller. Artificial Analysis revised its intelligence index twice in four days, replacing a benchmark, upgrading another by two major versions and raising private test-set weighting to 45% of the composite, which moved from v4.1.1's five-point gap between the two leading frontier models to v4.3's rounded tie without either model changing. OpenAI put its agent harness up for rent and documented that the managed service is US-residency-only with no zero data retention, including when the customer self-hosts the execution sandbox. A developer platform deprecated a model across agent mode on the day it announced the deprecation, in the same release that made unattended scheduled agent tasks available. Four independently reasonable decisions, one shared consequence: the system you evaluated is not the system you run. Policy built machinery and declined to attach an obligation to it. California enacted the first US statutory framework for third-party AI auditors, creating a registry and making unregistered covered audit work unlawful on reported 2028 and 2029 dates, while requiring no audits of anyone. The Department of Energy opened an information request implementing a bulk-power-system emergency authority and deferred the substance. A regional transmission organisation separately proposed ride-through, ramp-rate, telemetry and remote-disconnect requirements for gigawatt-scale computational loads, targeting a November filing, motivated by large-load disconnections that produced multi-gigawatt events. Ride-through means staying connected through a voltage or frequency disturbance instead of tripping off the grid, which is the opposite of what a data centre's own protection equipment is designed to do. That last item is the week's most consequential unwritten story: the largest campuses being financed on twenty-year leases may be legally curtailable before they are energised, and none of the leases appear to have priced it. The one lens where the evidence points cleanly in a single direction is networking. A 1.6-terabit coherent interface became a multi-vendor implementation agreement rather than a proprietary advantage, high-end routing grew 25% year over year in the second quarter with direct-to-cloud router sales up 94% on the same basis, a co-packaged optics company reached $650 million of 2026 capital, and both of the week's significant accelerator events turned on interconnect rather than compute: a memory-centric inference chip bought into the dominant ecosystem by joining its scale-up fabric, and the largest custom-silicon agreement explicitly bundled optical interconnect into scope. The counter-argument is right there in the first item, though. A multi-vendor specification is how an interconnect advantage becomes a commodity. Flywheel arc: `all-three`. ## Software lens - **Sep 10.** DeepSeek publishes an MIT-licensed V4.1-Flash and announces it will reroute its own flagship endpoint to a smaller model _([DeepSeek (model card)](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash))_ - **Sep 7.** Artificial Analysis revises its Intelligence Index twice in four days, breaking comparability with the prior basis _([Artificial Analysis](https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v4-3))_ - **Sep 10.** OpenAI makes GPT-Live-1 generally available and splits the voice layer from the reasoning model onto two meters _([OpenAI](https://openai.com/index/introducing-gpt-live-1-in-the-api/))_ - **Sep 8.** vLLM publishes a cache-residency policy that fits a full million-token context onto a single node _([vLLM](https://vllm.ai/blog/2026-09-08-glm53-part1-hybrid-sparse-offloading))_ - **Sep 9.** Anthropic discloses that its own agentic transcript scan missed cybersecurity evaluation records it should have found _([Anthropic](https://www.anthropic.com/news/alignment-assessment-cybersecurity-incidents))_ **What this means.** Nothing in the software lens this week was primarily about capability, and that is the finding. The two open releases are cheaper and smaller rather than stronger, and one of them arrived paired with a hosted retreat on the same vendor's flagship endpoint. The most-cited independent index changed its own measuring stick twice inside four days, which means no period-over-period capability comparison drawn from it is valid across the boundary, and v4.1.1's five-point gap between the two leading models became v4.3's rounded tie without either model changing. A voice product split one token meter into a per-minute meter plus a per-token meter, so the same workload now has two independently changeable rate cards. A serving project made long-context work fit on a single node, which is the week's clearest reduction in the on-premises hardware floor and the least-covered item in the lens. And a frontier lab disclosed that its own agentic scan of roughly 141,000 transcripts missed records, which is direct evidence against using agents for coverage claims in audit or discovery work. The operational lesson across all five is to pin the thing you tested and to monitor for the vendor changing it underneath you. ## Hardware lens - **Sep 10.** Oracle reports 850 MW of capacity and more than 300,000 GPUs actually delivered in one quarter _([Oracle (Q1 FY27 earnings release)](https://www.prnewswire.com/news-releases/oracle-announces-q1-results-driven-by-triple-digit-growth-in-cloud-infrastructure-revenues-302875728.html))_ - **Sep 8.** Qualcomm signs a multi-generation custom AI inference silicon and optical interconnect agreement with Amazon _([Qualcomm Technologies (release discloses no financial terms; up-to-$60-billion ceiling from secondary reporting of the related filing)](https://www.prnewswire.com/news-releases/qualcomm-announces-multi-generational-product-collaboration-with-amazon-to-build-next-generation-ai-data-center-infrastructure-302871895.html))_ - **Sep 10.** TSMC posts record August revenue of NT$514.8 billion, up 53.3% year over year _([TSMC Form 6-K (monthly revenue report)](https://www.stocktitan.net/sec-filings/TSM/6-k-taiwan-semiconductor-manufacturing-co-ltd-current-report-foreign--5794c1bfd249.html))_ - **Sep 8.** AMD puts MI450 production shipments at the end of the third quarter and defers its optical interconnect roadmap _([AMD at Citi 2026 Global TMT Conference (transcript)](https://stockanalysis.com/stocks/amd/transcripts/736068-citi-s-2026-global-tmt-conference/))_ - **Sep 10.** d-Matrix joins NVLink Fusion, putting a memory-centric inference accelerator inside NVIDIA's scale-up fabric _([d-Matrix](https://www.d-matrix.ai/announcements/d-matrix-rackscale-nvidia/))_ **What this means.** Hardware split cleanly into what was delivered and what was promised. Delivered: 850 megawatts of capacity and more than 300,000 accelerators delivered inside ninety days by one operator, on the company's own verb rather than an energisation claim, and a record foundry month up 53.3% year over year. Those are primary, dated, and fast, and together they are the strongest evidence this year that the physical layer is executing rather than merely announcing. Promised: a multi-generation custom inference and optical interconnect agreement whose up-to-$60-billion ceiling comes from secondary reporting of a filing this issue could not open, vests through milestone warrant tranches, and carries no disclosed service date, and a reported plan exceeding 38 gigawatts that rests on an unconfirmed internal document and carries the lowest source grade in this issue. In between sits the most honest disclosure of the week, a merchant challenger placing next-generation production shipments only at the end of the quarter and pushing its optical roadmap into the following year, which is a flattening in exactly the lens where the acceleration thesis needs a shortening. The structural signal is the memory-centric inference accelerator that joined the dominant vendor's scale-up fabric: the entry ticket to the ecosystem was interconnect compatibility, not silicon performance, which is a statement about where the moat actually sits. ## Networking lens - **Sep 9.** OIF publishes the 1600ZR Implementation Agreement, making 1.6T coherent interconnect a multi-vendor specification _([Optical Internetworking Forum](https://www.oiforum.com/oif-releases-critical-1600zr-coherent-interface-ia-doubling-capacity-per-wavelength-for-data-center-interconnects/))_ - **Sep 10.** Dell'Oro measures high-end routing up 25% and direct-to-cloud router sales up 94% year over year in the second quarter _([Dell'Oro Group](https://www.delloro.com/news/high-end-routing-and-aggregation-market-grew-25-percent-in-2q-2026/))_ - **Sep 10.** Ayar Labs adds $150 million, bringing 2026 capital to $650 million for co-packaged optics _([Reuters](https://www.reuters.com/world/asia-pacific/ayar-labs-backed-by-chip-giants-extends-funding-round-by-150-million-2026-09-10/))_ - **Sep 9.** Marvell details a 102-terabit Ethernet switch and a 256-lane scale-up switch _([Marvell Investor Relations](https://investor.marvell.com/news-events/press-releases/detail/1032/marvell-to-showcase-end-to-end-ai-data-center-connectivity-portfolio-at-ai-infra-summit-2026))_ - **Sep 10.** GlobalFoundries puts AI data centre optics and 800-volt power distribution on the same roadmap _([Converge Digest](https://convergedigest.com/globalfoundries-ai-optical-power-roadmap/))_ **What this means.** Networking is the week's best-evidenced lens and the only one where standards movement, revenue measurement, capital formation and silicon roadmaps all point the same way. A 1.6-terabit coherent interface became a multi-vendor implementation agreement, which is the precondition for volume deployment of datacenter interconnect at that capacity per wavelength. Measured revenue moved hard year over year in the second quarter, with high-end routing up 25% and direct-to-cloud sales up 94% against the same quarter a year earlier, though that second figure comes from a market-research house off an undisclosed base and carries a lower grade than the filings elsewhere in this issue. A co-packaged optics company reached $650 million of capital inside 2026, and a merchant vendor detailed a 102-terabit switch alongside a 256-lane scale-up part. The most interesting item is the least obvious: a foundry putting optics and 800-volt power distribution on one roadmap is an admission that at rack scale the interconnect problem and the power-delivery problem have become the same engineering problem. The honest tension in the lens is that the standards win cuts against the pricing-power thesis it appears to support, because a multi-vendor implementation agreement is precisely how a proprietary interconnect advantage becomes a commodity. ## Capital flow | Category | Capital in | Revenue out | Burn:Revenue | Movement | |---|---|---|---|---| | Frontier Labs (OpenAI, Anthropic, Google DeepMind, DeepSeek) | Unknown (was Unknown, flat) | Unknown (was Unknown, flat) | n/a | No financing or annual revenue disclosure from any constituent landed in the window, so no category point estimate is reproducible. What moved was delivery terms rather than capital: pricing structure, endpoint routing, and data-residency limits. | | Hyperscaler-Hosted (Azure-OpenAI, AWS-Anthropic, Google Cloud-Gemini, Oracle-OCI) | Unknown as a category total; two constituents disclosed $28.5B of quarterly capital expenditure and an announced plan of at least EUR 13B over two years respectively (was Unknown, flat) | Unknown as a category total; one constituent disclosed triple-digit cloud infrastructure revenue growth and $664B of contracted backlog (was Unknown, flat) | n/a | The only category with hard in-window primary disclosure, and the disclosure cuts both ways: 850 MW of capacity and more than 300,000 accelerators delivered against $28.5B of quarterly capital expenditure and roughly $5B of negative free cash flow. Both direction arrows are held flat because the category total and its prior are both Unknown, and a single company's quarter is not a category series. | | Neoclouds (CoreWeave, Nscale, Crusoe, Lambda, IREN, Zankore, NEXTDC) | Unknown as a category total; in-window instruments comprise an up-to-$3.1B undrawn facility, A$1.1B of subordinated convertible notes, and $375M firm of an $875M headline (was Unknown, flat) | Unknown (was Unknown, flat) | n/a | Every financing instrument in the category this week is structurally conditional. Nothing announced here is drawn, firm, or unconditioned in full, which is the pattern rather than an accident. | | On-Prem / Hybrid (Enterprise GPU clusters, sovereign and national programs, open-weight and on-device deployment) | Unknown (was Unknown, flat) | Indirect (was Indirect, flat) | n/a | The deployment floor fell measurably while the capital signal stayed unquantified. No government-funded national compute programme in the window meets this issue's ledger method, so no point estimate is published. | ### Frontier Labs — detail The week's frontier-lab news was commercial rather than financial. OpenAI put a full-duplex voice layer into general availability at a published per-minute rate and delegated reasoning to a separately billed text model, and rented out its agent harness as a managed service documented as US-residency-only with no zero data retention. DeepSeek published an MIT-licensed checkpoint and announced that its existing flagship endpoint would begin serving a different, smaller model. Anthropic disclosed a coverage failure in its own agentic transcript scan. None of that establishes a capital or revenue total, and prior rolling estimates are not reproduced because this issue's artifact set contains no constituent list, valuation date, or formula inputs. **Transactions:** - **Sep 10.** OpenAI makes GPT-Live-1 generally available and splits voice and reasoning onto two meters _([OpenAI](https://openai.com/index/introducing-gpt-live-1-in-the-api/))_ - **Sep 10.** OpenAI offers the Codex agent loop as a managed Agents API, without zero data retention _([OpenAI](https://openai.com/index/introducing-the-agents-api/))_ - **Sep 10.** DeepSeek publishes MIT-licensed V4.1-Flash and reroutes its flagship endpoint to a smaller model _([DeepSeek (model card)](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash))_ ### Hyperscaler-Hosted — detail Oracle's Q1 FY27 filing is the week's reference point because it separates delivered capacity from contracted capacity, which nobody else did. Backlog reached $664B, up $209B year over year and $26B sequentially, with management stating the new contracts are structured so that customers prepay or bring their own hardware and require no incremental Oracle capital, and expecting roughly half of the backlog to convert over 36 months as guidance rather than commitment. Google separately announced an at-least-EUR-13B Finnish investment plan for 2027 and 2028 alongside a twenty-two year power purchase agreement for up to half of an existing nuclear plant's output; the house measurement in this issue takes that transaction apart. Neither figure is a category total and neither vendor disclosed AI-segment operating margin. **Transactions:** - **Sep 10.** Oracle reports 850 MW and 300,000+ accelerators delivered, $28.5B quarterly capex, ~$5B negative free cash flow — $28.5B capex _([Oracle (Q1 FY27 earnings release)](https://www.prnewswire.com/news-releases/oracle-announces-q1-results-driven-by-triple-digit-growth-in-cloud-infrastructure-revenues-302875728.html))_ - **Sep 10.** Oracle discloses $664B of remaining performance obligations, up $209B year over year — $664B backlog _([Oracle Form 8-K](https://www.sec.gov/Archives/edgar/data/1341439/000119312526387905/orcl-20260910.htm))_ - **Sep 9.** Google announces an at-least-EUR-13B Finnish plan and a 22-year nuclear power purchase agreement — At least EUR 13B _([Google](https://www.googlecloudpresscorner.com/2026-09-09-Google-Deepens-Commitment-to-Finland-with-Two-Year-EUR13-Billion-investment-in-AI-Infrastructure))_ ### Neoclouds — detail A two-month-old Indonesian platform signed an up-to-$3.1B underwritten term loan that is undrawn and collateralised against hardware whose delivery depends on export licences that have not been granted. NEXTDC priced A$1.1B of subordinated convertible notes whose effective maturity is set by a 2029 investor put rather than by the stated term, with capped calls attached, meaning the issuer bought an option that limits how much dilution a conversion can cause. Positron AI's $875M headline is $375M firm plus up to $500M contingent. And a $1.11B monthly hosting rate annualises to roughly $13.3B, which is our arithmetic and not the discloser's: the executive stated no contract term, named no counterparty and disclosed no megawatts, so there is nothing in the public record against which to test an annualisation the company never made. Read together, the category is raising optionality rather than capital, which is prudent for the borrower and unpriceable from the public record for anyone underwriting the counterparty. **Transactions:** - **Sep 9.** Zankore signs an up-to-$3.1B underwritten term loan, undrawn, with export-licence-contingent collateral — Up to $3.1B _([Zankore](https://zankore.com/zankore-expands-southeast-asias-ai-compute-infrastructure-with-nvidia))_ - **Sep 10.** NEXTDC prices A$1.1B of subordinated convertible notes with a 2029 investor put — A$1.1B _([NEXTDC ASX release (launch)](https://announcements.asx.com.au/asxpdf/20260909/pdf/073xnbg99ksf0r.pdf))_ - **Sep 10.** Positron AI raises $875M headline, of which $375M is firm and up to $500M contingent — $375M firm _([Positron AI](https://www.prnewswire.com/news-releases/positron-ai-raises-875-million-at-a-5-billion-valuation-to-bring-its-next-generation-inference-silicon-to-market-302874601.html))_ - **Sep 10.** SpaceX discloses a $1.11B monthly hosting rate with no term, counterparty or megawatts; the roughly $13.3B annualisation is this issue's arithmetic — $1.11B per month _([Business Insider](https://www.businessinsider.com/spacex-strikes-billion-a-month-compute-deal-with-mystery-customer-2026-9))_ ### On-Prem / Hybrid — detail Two software events did more for on-premises economics than any financing announcement. A serving project published a cache-residency policy that fits a full million-token context onto a single node, which removes a multi-node requirement from long-context work, and two open-weight checkpoints were published, though the 1.6-trillion-parameter agentic model among them is reported as post-trained from the other lab's base rather than as an independent second lineage. NVIDIA's up-to-two-gigawatt Australian target is an aggregation of eight independent operators' pipelines rather than a government programme or a purchase, so it does not enter the sovereign ledger. The honest gap in this category is that nothing in the window attributes a sovereign or enterprise on-premises purchase to open-weight availability, which is the demand-routing mechanism the framework's fourth hypothesis actually claims. **Transactions:** - **Sep 9.** NVIDIA aggregates eight Australian operators into an up-to-2-gigawatt buildout target — Up to 2 GW _([NVIDIA Newsroom](https://nvidianews.nvidia.com/news/nvidia-expands-ai-infrastructure-capacity-in-partnership-with-australias-data-center-ecosystem))_ - **Sep 8.** vLLM fits a full million-token context onto a single node via a cache-residency policy _([vLLM](https://vllm.ai/blog/2026-09-08-glm53-part1-hybrid-sparse-offloading))_ - **Sep 8.** Nex-AGI publishes a three-tier agentic family including a 1.6T open-weight checkpoint _([Nex-AGI (model card)](https://huggingface.co/nex-agi/Nex-N2.5-Max))_ ## Signal vs noise - **Score 5/5 —** One operator delivered 850 MW of capacity and more than 300,000 accelerators inside the quarter ended August 31. - _Sources:_ Oracle Q1 FY27 earnings release and Form 8-K, September 10, with the 73%-of-deliveries and 36-month conversion figures from the same day's earnings call transcript rather than the release or the filing - _Read:_ The week's only dated delivered-capacity figure and the strongest signal in the issue. It is a company-reported operational fact in a filed document, stated as almost three times the prior quarter and 73% of all fiscal 2026 deliveries, and it gives the house frame the dated base it has been missing. One discipline: the company says delivered, so read it as delivery throughput and not as operating capacity. Read every other capacity number this week against it. - **Score 5/5 —** A multi-generation custom AI inference silicon and optical interconnect agreement was signed, with a ceiling reported at up to $60 billion. - _Sources:_ Qualcomm press release, September 8, which discloses no financial terms; the up-to-$60-billion ceiling and the warrant structure come from CNBC and Reuters reporting of the related SEC filing, which this issue could not open - _Read:_ The agreement is real, primary and filed. The figure is neither. It rests on secondary reporting of a filing nobody here has read, and even taken at face value it is a ceiling that vests through milestone warrant tranches tied to deployment decisions the supplier does not control, with no service date disclosed. Use the agreement as evidence of custom-silicon direction and refuse to put the headline number in a revenue model. - **Score 4/5 —** Seven pre-built job agents shipped, six generally available immediately, with no published price for any of them. - _Sources:_ Salesforce Newsroom, September 11 - _Read:_ Availability is verifiable and the packaging shift from platform to role is genuine. The economics are not verifiable at all, and six customer outcome percentages in the same release arrive without a base, a window, or a definition of the operative verb. Treat the availability as signal and the percentages as marketing until a contract defines what gets measured. - **Score 3/5 —** High-end routing grew 25% and direct-to-cloud router sales grew 94% year over year in the second quarter. - _Sources:_ Dell'Oro Group, September 10 - _Read:_ Directionally consistent with three other independent networking datapoints in the same week, which is why it survives at all. But it is a market-research measurement off an undisclosed base for a single quarter against the same quarter a year earlier, not sequential growth, so it belongs in a trend argument and not in a forecast. The 94% figure in particular should never be extrapolated. - **Score 2/5 —** An internal plan calls for more than 38 gigawatts of AI-focused data centre capacity. - _Sources:_ Bloomberg feature citing an unconfirmed internal document, September 10 - _Read:_ This is the largest number circulating this week and the weakest sourced item in the issue. It is a reported internal plan, unconfirmed by the company, with no dated commitment or filing behind it. It should not enter a capacity model, a competitive assessment, or a board slide, and its presence in secondary coverage as a peer to filed figures is exactly the error this section exists to catch. - **Score 2/5 —** A frontier lab is working with a memory manufacturer on next-generation chips. - _Sources:_ Reuters, September 9, with the framing subsequently walked back - _Read:_ The characterisation changed inside the window, which is itself the information. Treat it as an expression of intent and a deepening commercial relationship, not as a disclosed co-development programme with scope, silicon, or a date. Anyone building a supply-chain thesis on it is building on a quote that its own source revised. ## House measurement **Approximately 70% of the capital expenditure behind Google's Finnish nuclear commitment has not been approved, and the deal contracts roughly 10.6 times more electricity than it adds.** _[filing-derived]_ Method: House computation from two primary filings dated September 9, 2026: Fortum's inside-information release on the Loviisa power purchase agreement and Google's own energy blog post on the same transaction. The primary sources state, in their own words: (1) the Loviisa life-extension programme spanning 2023 to 2050 is approximately EUR 1 billion in total capital expenditure; (2) approximately 80% of the projects and EUR 700 million of that capital expenditure required to keep the plant operational through 2050 remain pending investment decisions; (3) the power purchase agreement enables a 10 MW uprate, on top of a separately planned 38 MW uprate; (4) the plant comprises two reactors of 507 MWe each, commissioned in 1977 and 1980; (5) Google contracts for up to 50% of the plant's output, beginning in 2028 at reduced volume and reaching the full contracted share for 2030 through 2049. Arithmetic, all figures as disclosed: un-sanctioned share of programme capital expenditure = EUR 700M / EUR 1,000M = approximately 70%, stated to two significant figures because the denominator is disclosed as approximate. Total added generation = 10 MW + 38 MW = 48 MW. Installed site capacity = 507 MWe x 2 = 1,014 MWe. Added generation as a share of the site = 48 MW / 1,014 MW = 4.7%. Power-purchase-enabled addition alone as a share of the site = 10 MW / 1,014 MW = 1.0%. Contracted volume expressed in capacity terms = 50% x 1,014 MW = 507 MW. Ratio of contracted volume to added generation = 507 MW / 48 MW = 10.6 times. No figure in this computation is derived from a secondary source, and no undisclosed input is assumed. - Programme capital expenditure still pending an investment decision: **Approximately 70% (EUR 700M of approximately EUR 1,000M)** (Fortum's own disclosure; the same sentence states that approximately 80% of the projects are also pending. This is the selected house measurement: the price-protection claim attached to the deal depends on money that has not been committed.) - Added generation as a share of installed site capacity: **4.7% (48 MW against 1,014 MWe)** (10 MW enabled by the power purchase agreement plus a separately planned 38 MW uprate, against two 507 MWe reactors. The deal is overwhelmingly a preservation transaction rather than an addition.) - Generation attributable to the agreement itself: **1.0% (10 MW against 1,014 MWe)** (Only the 10 MW uprate is stated as enabled by the agreement. The 38 MW uprate was already planned and would proceed without it, so attributing all 48 MW to the deal overstates its additionality by roughly a factor of five.) - Contracted volume relative to added generation: **10.6 times (507 MW contracted against 48 MW added)** (50% of a 1,014 MW site expressed in capacity terms, divided by total added generation. For every megawatt the transaction brings into existence, it removes roughly eleven megawatts from the pool available to other buyers.) - Share of Finnish electricity supplied by the plant: **Approximately 10%** (Fortum's stated figure for the existing plant, unchanged by the transaction. Half of that output is now under a 22-year claim by a single foreign counterparty, which is the sovereignty question the coverage treated as a sovereignty answer.) Implication: For a board or an investment committee evaluating a power-linked AI infrastructure announcement, the disclosed capital structure is the diligence item, not the headline capacity. Here the counterfactual runs the right way for the transaction and against the way it was reported: without the life-extension programme the plant cannot operate past 2030, so the agreement plausibly preserves roughly a tenth of a member state's electricity supply, which is a larger public benefit than 48 MW of uprates would suggest. But the mechanism that delivers it is 700 million euros of unapproved capital expenditure, and the buyer's published claim that its new demand does not raise household prices rests on that unapproved spending plus an unpublished model. An operator should therefore read this class of announcement as a conditional option on someone else's future investment decision, price the conditionality, and require the modelling before accepting the price-neutrality claim. The same test applied across this week's capital events separates the one dated delivered-capacity figure from a very large stack of commitments contingent on decisions nobody has taken. Caveats: This measurement does not show that the transaction is bad for Finnish consumers, nor that the price-neutrality claim is false. It shows only that the claim is not yet checkable from the disclosed record. Three specific limits apply. First, neither the contract price per megawatt-hour nor the contracted volume in terawatt-hours is disclosed, so no welfare conclusion of any kind follows. Second, the 507 MW contracted figure is a capacity-equivalent restatement of a 50% output share for arithmetic comparability; the agreement is an energy contract and the plant's actual output varies with availability, so the ratio is an order-of-magnitude comparison rather than a contractual quantity. Third, the approximately 70% share inherits the imprecision of an approximately one billion euro denominator and should not be read to more than two significant figures. The 80% figure in the same disclosure counts projects rather than euros and is not interchangeable with it. Sources: [Fortum inside information release, September 9, 2026](https://www.fortum.com/en/media/2026/09/inside-information-fortum-and-google-partner-drive-sustainable-growth-finland-sign-nuclear-power-purchase-agreement), [Google energy blog, September 9, 2026](https://blog.google/innovation-and-ai/infrastructure-and-cloud/global-network/clean-energy-finland/), [Google Cloud Press Corner, September 9, 2026](https://www.googlecloudpresscorner.com/2026-09-09-Google-Deepens-Commitment-to-Finland-with-Two-Year-EUR13-Billion-investment-in-AI-Infrastructure), [World Nuclear News, September 9, 2026](https://www.world-nuclear-news.org/articles/google-signs-up-for-electricity-from-finnish-nuclear-power-plant) ## Levers | Metric | Current | Prior | Direction | Threshold | |---|---|---|---|---| | Frontier lab cash runway at current burn | Unknown — no constituent disclosed cash, burn, or financing in the window | Unknown — W36 had no reproducible cash-and-burn input table either | flat | Below 18 months for any disclosed-burn lab | | Hyperscaler AI capex to disclosed AI revenue ratio | Unknown as a top-four ratio — one constituent disclosed $28.5B of quarterly capital expenditure and roughly $5B of negative free cash flow, with no AI-segment revenue line | Unknown — no reproducible top-four capex and AI-revenue input range | flat | Above 6x sustained for two consecutive quarters | | CoreWeave contracted revenue backlog | $104.2B as of June 30, unchanged — no CoreWeave filing landed in the window | $104.2B as of June 30 | flat | Sequential decline, or conversion below 15% annually | | NVIDIA quarter-over-quarter data center revenue | $89.0B for Q2 FY27, unchanged — no NVIDIA print in the window | $89.0B for Q2 FY27 | flat | Two consecutive quarters of sequential decline | | Open-weight to closed-model capability gap on coding | Not measurable this week — the reference index changed basis twice in four days, breaking comparability with the prior reading | Untested in W36 as well; no new independent open-frontier benchmark set | flat | Open weights within 2 Index points of the closed leader | | Sovereign AI program commitments | Unknown — no government-funded national compute programme in the window meets the ledger method | Unknown — W36 had no reproducible programme ledger | flat | Above 20 programs or $250B committed | | PJM capacity auction clearing price | $325.00 per MW-day for 2028/29, unchanged — no auction occurred in the window | $325.00 per MW-day for 2028/29 | flat | An auction clearing below the cap, or a FERC-approved increase in the cap itself | | Time from interconnection request to energization | Unknown as a queue duration — the window's only hard figure is 850 MW delivered within one quarter by a single operator, which measures delivery rather than energisation or queue time | Unknown — no reproducible multi-queue duration table | flat | Below 48 months in two or more major queues | | Cost per task, frontier reasoning model | $3.26 is the floor of the two leading models and $7.63 the other, per index task on the new v4.3 basis; no frontier-tier median is computable because the basis changed twice inside the window, and neither figure is commensurate with the prior reading | Prior reading withheld because the index basis changed twice inside the window | flat | A frontier-tier reasoning model below $1 per million output tokens | | Custom silicon share of hyperscaler AI compute | Unknown — a multi-generation custom inference agreement was signed but discloses a warrant-vesting ceiling rather than units, share, or a service date | Unknown — prior quarter's supplier disclosure combined custom accelerators with networking | flat | Above 45% share with audited hyperscaler mix disclosure | **Lever detail:** - **Frontier lab cash runway at current burn.** The required inputs remain unaudited and absent. GPT-Live-1 list pricing and the Agents API's managed-service terms are product and commercial signals, not financing inputs, and no lab published a raise or a revenue figure this week. - **Hyperscaler AI capex to disclosed AI revenue ratio.** Oracle's filing supplies a real capital-expenditure number and a real backlog number but no AI-segment revenue, which is the denominator the method requires. No hyperscaler disclosed AI-segment operating margin this week, so the ratio stays unpublished rather than estimated. - **CoreWeave contracted revenue backlog.** Backlog remains a filed stock value awaiting the next quarter. The week's neocloud financings, an undrawn term loan and a convertible note with an investor put, are category structure signals and cannot be substituted into this series. - **NVIDIA quarter-over-quarter data center revenue.** The in-window NVIDIA news is an Australian capacity aggregation across eight operators and a third-party accelerator joining NVLink Fusion. Neither is a revenue disclosure, and TSMC's record month is a supplier datapoint that cannot be substituted into this series. - **Open-weight to closed-model capability gap on coding.** This is a measurement failure rather than a capability result. Two open checkpoints shipped in the window, both explicitly cheaper and smaller rather than stronger, and one carries vendor-reported benchmarks with no independent evaluation. See Model Pulse for the lineage read. - **Sovereign AI program commitments.** The method counts government-funded national compute programmes only. This week's two candidates fail it for different reasons: an at-least-EUR-13B corporate investment plan with a power purchase agreement is private capital, and an up-to-two-gigawatt national target is an aggregation of eight commercial operators' pipelines. - **PJM capacity auction clearing price.** The prior threshold, a second consecutive auction clearing at the cap, is already satisfied three times over and can no longer move: 2026/27 cleared at $329.17, 2027/28 at $333.44 and 2028/29 at $325.00, each at the approved cap. PJM's own no-cap-or-floor simulation for 2028/29 reports $554.72 per MW-day for the RTO and $776.69 for ComEd, versus the $325 capped result, with $29.7 billion of simulated cleared value against $16.4 billion actual. The simulation does not isolate the cap as the sole cause, but it quantifies how much higher PJM's model clears without it. The in-window development is regulatory rather than price-setting: proposed large computational load ride-through, ramp-rate, telemetry and remote-disconnect requirements were presented on September 8, targeting a federal filing in November 2026. If mandatory curtailable status attaches to gigawatt-scale loads, it changes what capacity payments are actually buying. - **Time from interconnection request to energization.** No comparable queue-duration update was published. A Texas docket fight over forfeiture terms on a 75-megawatt interconnection rule is evidence of queue scarcity and of the cost of holding a position in one, but it is not a duration measurement. - **Cost per task, frontier reasoning model.** The cost spread is the most durable finding to survive the index revision and is a genuine procurement fact: 57% lower cost for the same rounded composite score. It is reported here as two point observations rather than as movement, because the composite they price changed underneath them. - **Custom silicon share of hyperscaler AI compute.** The week strengthens the directional case for custom inference silicon on two counts, a multi-generation agreement and a third-party accelerator entering the dominant scale-up fabric, while supplying no unit volumes, deployed utilisation, or mix disclosure from which a share could be computed. ## Predictions - **`p105-second-operator-delivered-capacity-mar31` _[capital]_ — A publicly traded operator other than Oracle discloses, for a specific reporting period, both a megawatt capacity figure delivered or placed in service and a unit count of AI accelerators delivered, by March 31, 2027.** - Confidence: 62%. Deadline: By March 31, 2027. - Trigger: Hit only if an earnings release, 10-Q, 10-K, or call transcript from an operator other than Oracle states both a megawatt capacity figure delivered or placed in service and a count of accelerators delivered or deployed, for a named period. A backlog, contracted-capacity, or announced-pipeline figure is a miss. - **`p106-loviisa-fid-jun30` _[power]_ — Fortum announces an approved investment decision covering at least EUR 300 million of the EUR 700 million of Loviisa life-extension capital expenditure currently disclosed as pending, by June 30, 2027.** - Confidence: 44%. Deadline: By June 30, 2027. - Trigger: Hit only if a Fortum release, interim report, or annual report states an approved investment decision of at least EUR 300 million for the Loviisa life-extension programme. Reaffirmation of the programme, or approval below EUR 300 million, is a miss. - **`p107-runtime-manifest-mar31` _[software]_ — A major model provider or evaluation publisher ships a machine-readable runtime or harness manifest that ties a published score to a reproducible configuration, by March 31, 2027.** - Confidence: 38%. Deadline: By March 31, 2027. - Trigger: Hit only if public documentation or a results artifact contains a versioned configuration identifier naming at least harness or adapter version, tool permissions, context persistence policy, and retry budget, which is the same four-element bar this issue's pattern watch sets for the same event. Prose disclosure of methodology changes, or publishing two harness results side by side without a configuration identifier, is a miss. - **`p108-pjm-large-load-filing-dec31` _[power]_ — PJM's Section 205 filing on large computational loads is docketed by December 31, 2026 and carries a telemetry or remote-disconnect requirement, not merely a ride-through envelope or a ramp-rate limit.** - Confidence: 58%. Deadline: By December 31, 2026. - Trigger: Hit only if a FERC filing by PJM docketed on or before December 31, 2026 includes a telemetry or remote-disconnect requirement applicable to large computational loads. A filing carrying only a voltage or frequency ride-through envelope or only a ramp-rate limitation is a miss, as is a further stakeholder presentation or any slip past year end. - **`p109-1600zr-two-vendors-jun30` _[networking]_ — At least two distinct vendors announce 1600ZR-conformant coherent pluggable optics products by June 30, 2027.** - Confidence: 66%. Deadline: By June 30, 2027. - Trigger: Hit only if product announcements or datasheets from two distinct vendors cite conformance or compliance with the OIF 1600ZR Implementation Agreement. Demonstrations, interoperability plugfests, and roadmap statements without a named product are a miss. - **`p110-agents-api-residency-mar31` _[software]_ — OpenAI's managed Agents API supports zero data retention or a non-US data residency option by March 31, 2027.** - Confidence: 29%. Deadline: By March 31, 2027. - Trigger: Hit only if OpenAI documentation states zero-data-retention eligibility or a non-US data residency option for the managed Agents API. A self-hosted sandbox, a roadmap statement, or residency support on other OpenAI surfaces but not the Agents API is a miss. - **`p111-custom-inference-service-date-mar31` _[hardware]_ — Qualcomm or its counterparty discloses a named service date, first-deployment date, or unit volume for the multi-generation custom AI inference agreement, by March 31, 2027.** - Confidence: 33%. Deadline: By March 31, 2027. - Trigger: Hit only if a filing, release, or call transcript from either party states a calendar service or first-deployment date, or a unit or megawatt volume, for the agreement. Restating the transaction value, the warrant structure, or a generational roadmap without a date or volume is a miss. ### Prior predictions scored - `p100-astra-neutral-memory-dec15` _[software]_ — **PENDING** — ARC Prize or another provider-neutral evaluator publishes an Astra run without provider-private reasoning state that closes at least half of the 35.9-point matched-effort Standard-to-Provider-Adapter gap by December 15, 2026. — No qualifying run appeared in the window and no third party has reproduced the provider-adapter result at all. The obstacle is structural rather than budgetary: the adapter is the vendor's own tooling and the preserved state is opaque even to the benchmark authors, who have committed to publishing both harness results side by side going forward. That fixes presentation, not verification. The week's relevant movement went the other way, with the most-cited independent index rebasing its own composite twice in four days, which does not address the harness question and makes the surrounding comparisons less stable rather than more. - `p101-fervo-expansion-firm-jun30` _[power]_ — **PENDING** — Google accepts at least 500 MW of Fervo's conditional expansion and the parties execute a definitive agreement by June 30, 2027. — No Fervo filing and no Google statement touching the expansion option appeared in the window. Google's in-window energy activity was a twenty-two year Finnish nuclear power purchase agreement, a different counterparty, technology and geography, and it says nothing about the geothermal option either way. The deadline is more than nine months out and the prediction remains live on its original terms. - `p102-second-queue-contract-dec31` _[capital]_ — **PENDING** — A second AI infrastructure contract above $500M discloses both a future service date and an option, delayed-draw, earnout, or guarantee allocating schedule risk by December 31, 2026. — This is the closest call in the set and it is being held open deliberately. Three in-window candidates were examined against all three conjuncts. The Qualcomm agreement clears the value test in a primary filing and clears the contingent-mechanism test through milestone warrant vesting, but discloses no service date, so it fails one conjunct outright. NEXTDC's A$1.1 billion note clears value and contingency through capped calls and a 2029 investor put, but its dates are financial maturities rather than service dates. Zankore's up-to-$3.1 billion facility is the only candidate that arguably satisfies all three, and it does so on two contestable readings: that the 'up to' ceiling constitutes a delayed-draw mechanism, and that the referenced first-phase timing and capacity constitute a named service date. Under the accountability rule that ambiguity resolves against the prediction, that is not a hit. What would resolve it is a single filing that states a calendar service date alongside an explicit contingent mechanism, which the next quarter of gigawatt-scale financings is likely to produce. - `p103-fabric-one-control-schema-dec31` _[networking]_ — **PENDING** — Equinix publishes Fabric One beta documentation that exposes approval, rollback, or auditable intent-history controls before December 31, 2026. — No Fabric One beta documentation was published in the window and nothing in this week's networking events touches it. The adjacent movement is worth noting without crediting it to this prediction: the strongest delegated-control artifact of the week came from a developer platform rather than a network operator, in the form of agent permissions whose managed restrictions cannot be weakened by user settings or a previously saved approval. That raises the bar this prediction's trigger is implicitly aiming at, since an approval workflow that an operator can override is weaker than what has now shipped elsewhere. - `p104-broadcom-ai-split-dec31` _[hardware]_ — **PENDING** — Broadcom discloses separate quarterly revenue figures for custom AI accelerators and AI networking by December 31, 2026. — No Broadcom disclosure landed in the window; the relevant quarter was reported on September 2, in the prior week, with a combined AI-semiconductor line. One opportunity remains before the deadline. The week's custom-silicon news, a multi-generation inference and optical interconnect agreement at a different supplier, cuts mildly against the prediction rather than for it, because bundling silicon and interconnect into one commercial scope is the opposite of the reporting split this prediction requires. ## Synthesis ### Connecting the dots - **A hyperscale operator's paired megawatt-and-accelerator delivery disclosure indicates that much of the week's announced capital will remain optioned rather than committed through the first half of 2027.** _[abductive, 74% confidence]_ 1. Oracle reported 850 MW of data centre capacity and more than 300,000 accelerators delivered inside a single quarter, against $28.5 billion of quarterly capital expenditure and roughly $5 billion of negative free cash flow. The company's verb is delivered rather than energised, and the paired megawatt-plus-unit disclosure for one named period at this scale is what no other operator published in the window. 2. The same filing put backlog at $664 billion and stated that the new contracts are structured so that customers prepay or bring their own hardware, requiring no incremental Oracle capital, which moves the financing obligation off the disclosing balance sheet and onto counterparties who did not disclose. 3. Read against that figure, the week's other headline figures are claims on decisions nobody has taken: a €13 billion Finnish programme whose €700 million of life-extension capital remains pre-decision, an up-to-$60 billion silicon ceiling from secondary reporting of a filing this publication could not open, a two-gigawatt national target that aggregates eight separate operators' pipelines, and an up-to-$3.1 billion facility that is undrawn with export-licence-contingent collateral. 4. The pattern repeats at the revenue line: a $13.3 billion annualised hosting figure is this publication's arithmetic rather than the discloser's, $1.11 billion of monthly billing multiplied by twelve, on a contract whose term, counterparty and megawatts the executive declined to state. Steel-man: One quarter's delivered-capacity figure from one operator is a thin base, and the comparison is not apples to apples: a prepay or bring-your-own-hardware structure is a genuine capital-efficiency innovation rather than an accounting dodge, and a customer willing to prepay is stronger evidence of demand than a vendor willing to spend. Several of the contingent instruments are also contingent in the buyer's favour, which is prudent rather than weak. The defensible version of the claim is narrower than the headline: the week established that delivered capacity and optioned capacity are now separately disclosable, and that almost nobody except Oracle disclosed the former. Falsified if, by 2027-06-30, at least three of the four named conditional instruments reach their originally announced scope without reduction or restructuring. That single conversion-rate test is consistent with the 26% uncertainty carried by this connection. Evidence: [hardware-02](https://www.prnewswire.com/news-releases/oracle-announces-q1-results-driven-by-triple-digit-growth-in-cloud-infrastructure-revenues-302875728.html), [capital-02](https://www.sec.gov/Archives/edgar/data/1341439/000119312526387905/orcl-20260910.htm), [capital-01](https://www.googlecloudpresscorner.com/2026-09-09-Google-Deepens-Commitment-to-Finland-with-Two-Year-EUR13-Billion-investment-in-AI-Infrastructure), [capital-03](https://zankore.com/zankore-expands-southeast-asias-ai-compute-infrastructure-with-nvidia), [capital-06](https://www.businessinsider.com/spacex-strikes-billion-a-month-compute-deal-with-mystery-customer-2026-9) - **The contingency in AI infrastructure contracts has migrated off the announcing party's balance sheet and onto a third party's undisclosed future decision, which makes the announcements unpriceable from the public record.** _[inductive, 71% confidence]_ 1. A hyperscaler's household-price guarantee in Finland depends on approximately 70% of a life-extension programme's capital expenditure, roughly €700 million, that the counterparty states is still pending investment decisions, and on a price and contracted volume neither party disclosed. 2. A multi-generation silicon and optical interconnect agreement carries a figure that vests through warrant tranches tied to milestones, so the ceiling is a function of someone else's deployment decisions rather than an order book. 3. A national two-gigawatt buildout target is assembled from eight independent operators' separate pipelines, each with its own siting, power and financing gates, none of which the aggregator controls. 4. A two-month-old platform's underwritten facility is undrawn and collateralised against hardware whose delivery depends on export licences that have not been granted, which means the credit is conditional on a government decision. Steel-man: Staging exposure behind milestones is ordinary commercial discipline and usually a sign of competent structuring rather than weakness. A warrant tranche that vests on deployment aligns a supplier with a customer's actual build; an undrawn facility costs almost nothing until it is needed; aggregating a national target across operators is how any country's capacity is in fact built. The claim should therefore not be read as an allegation of fragility. It is a claim about legibility: the conditions are real, they are held by parties who did not publish them, and the published number is the ceiling rather than the expectation. Falsified if, by 2027-06-30, two or more gigawatt-scale AI infrastructure announcements disclose the counterparty's governing condition in the announcing party's own filing at announcement, which would make the contingency legible from the announcer rather than only from whoever holds it. Evidence: [capital-01](https://www.googlecloudpresscorner.com/2026-09-09-Google-Deepens-Commitment-to-Finland-with-Two-Year-EUR13-Billion-investment-in-AI-Infrastructure), [hardware-01](https://www.prnewswire.com/news-releases/qualcomm-announces-multi-generational-product-collaboration-with-amazon-to-build-next-generation-ai-data-center-infrastructure-302871895.html), [hardware-04](https://nvidianews.nvidia.com/news/nvidia-expands-ai-infrastructure-capacity-in-partnership-with-australias-data-center-ecosystem), [capital-03](https://zankore.com/zankore-expands-southeast-asias-ai-compute-infrastructure-with-nvidia) - **After this week, measurement instability is more likely than capability to remain the binding procurement risk through the first quarter of 2027, because in three separate cases the thing an enterprise tested stopped being the thing it runs.** _[inductive, 78% confidence]_ 1. The most-cited independent intelligence index was revised twice in four days, the second revision replacing a benchmark, upgrading another by two major versions, and raising private test-set weighting to 45% of the composite, which moved from v4.1.1's five-point gap between the two leading frontier models to v4.3's rounded tie without either model changing. 2. A frontier vendor published a new open-weight checkpoint and simultaneously announced that requests to its existing flagship endpoint would begin serving a different, smaller model at different rates, with no change required by the caller. 3. A developer platform deprecated a model across chat, inline edits, ask and agent modes on the same day it announced the deprecation, with a replacement administrators may need to enable, while in the same release making unattended scheduled agent tasks available. 4. A frontier lab disclosed that its own agentic scan of roughly 141,000 evaluation transcripts had missed a subset, which is direct evidence that agent-based review of a large corpus does not establish coverage. Steel-man: Each of these is defensible on its own terms and two are actively good practice. The index publisher disclosed every change and explained that it was pulling forward planned work, which is more transparency than most evaluators offer. Rerouting an endpoint to a cheaper model is a price cut delivered without asking customers to migrate. The lab's disclosure of its own tooling failure is exactly the behaviour the field asks for. The claim is not that any actor behaved badly; it is that four independent, individually reasonable decisions produced the same operational consequence, and that consequence lands on whoever has to re-run an evaluation they already paid for. Falsified if, by 2027-03-31, a major provider or evaluator publishes a machine-readable configuration manifest and an independent enterprise rerun uses it to reproduce the tested score or behaviour after a vendor-side change. That single demonstrated replay would show measurement instability is manageable rather than binding. Evidence: [software-02](https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v4-3), [software-01](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash), [agents-06](https://github.blog/changelog/label/copilot/), [software-06](https://www.anthropic.com/news/alignment-assessment-cybersecurity-incidents) - **Policy machinery built this week is likely to keep front-loading compliance cost through year-end because operative obligations for large computational loads remain undefined.** _[abductive, 42% confidence]_ 1. California enacted the first US statutory framework for third-party AI auditors, creating a registry and making unregistered covered audit work unlawful on reported 2028 and 2029 dates, while mandating no audits of anyone. 2. The federal energy department opened an information request implementing a bulk-power-system emergency authority, deferring the substance of what large loads will actually be required to do to a later date. 3. A regional transmission organisation separately presented proposed ride-through, ramp-rate, telemetry and remote-disconnect requirements for gigawatt-scale computational loads, targeting a federal filing in November 2026, motivated by large-load disconnections that produced multi-gigawatt events. 4. A European cybersecurity agency was granted hands-on access to two frontier models, which creates state evaluation capacity without publishing an evaluation standard the vendors must meet. 5. A state commission took new large-load interconnection standards to a final adoption vote on September 11, applying to interconnections at or above 75 megawatts with financial security due on execution, and six intervenor filings landed in three days fighting over how much forfeiture risk a developer bears for abandoning a queue position. The vote outcome could not be confirmed from the docket, and the same filing states the per-megawatt security figure two different ways in its rule text and its preamble. Steel-man: Building capacity before imposing obligations is the normal and arguably correct sequence for technical regulation: you cannot mandate audits before a qualified auditor population exists, and you cannot specify load-flexibility requirements before you have measured what the loads actually do during faults. Read charitably, this is a regulator declining to write a rule it cannot yet write well. The cost of the charitable reading is borne asymmetrically, though. An enterprise signing a multi-year assurance arrangement or a twenty-year lease today has to guess at a requirement that will be defined after the ink dries, and the guess is not refundable. Falsified if PJM's targeted Section 205 filing is docketed by 2026-12-31 with an enforceable telemetry or remote-disconnect requirement for large computational loads. The 42% confidence is the complement of this issue's 58% probability for that same filing event. Evidence: [policy-01](https://www.gov.ca.gov/2026/09/09/governor-newsom-signs-first-in-the-nation-ai-safeguards-to-protect-californians-calls-on-the-federal-government-to-do-its-part/), [policy-02](https://www.federalregister.gov/documents/2026/09/09/2026-18370/securing-the-united-states-bulk-power-system), [policy-06](https://www.pjm.com/-/media/DotCom/committees-groups/committees/pc/2026/20260908/20260908-item-05---1-large-load-ride-through---presentation.pdf), [policy-03](https://www.reuters.com/technology/eus-cybersecurity-agency-granted-access-mythos-5-ai-model-commission-says-2026-09-10/), [policy-04](https://interchange.puc.texas.gov/Search/Filings?ControlNumber=58481) ### Thesis test - **Hypothesis 1 — The cycle is accelerating, not slowing.** — **STRAINED**. Prior verdict: strained in W36. It stays strained, and the objection is unchanged in kind. W37 supplies the strongest single delivery-cadence data point of the year and simultaneously undercuts the measurement that would let anyone verify a shortening doubling interval. One operator delivered 850 MW of capacity and more than 300,000 accelerators inside ninety days, and a foundry posted record monthly revenue up 53.3% year over year, both of which are hard, primary, and fast. But the most-cited capability index was rebased twice in four days, so period-over-period capability comparison is broken by construction, and the leading merchant challenger deferred its optical roadmap while placing next-generation production shipments only at the end of the quarter. Acceleration is measurable in delivery and in fabrication throughput; it is unmeasurable in capability this week, and it is flat to slower in one lens's silicon cadence. Nothing this week retires the prior objection, so nothing this week earns an upgrade. Against it: AMD placed MI450 production shipments at the end of the third quarter and deferred its optical interconnect roadmap into the following year, which is a flattening rather than a shortening in that lens. The week's genuinely new checkpoints lead on cost and footprint rather than capability, and one vendor rerouted its own flagship endpoint to a smaller model, which is a capability step sideways or down for anyone who had pinned it. The framework's refutation test is two consecutive quarters of flattening slope in any one lens, and the merchant-accelerator cadence now has one quarter on the board. Evidence: [hardware-02](https://www.prnewswire.com/news-releases/oracle-announces-q1-results-driven-by-triple-digit-growth-in-cloud-infrastructure-revenues-302875728.html), [hardware-03](https://www.stocktitan.net/sec-filings/TSM/6-k-taiwan-semiconductor-manufacturing-co-ltd-current-report-foreign--5794c1bfd249.html), [software-02](https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v4-3) - **Hypothesis 2 — Capital is concentrated, returns are diffuse.** — **STRAINED**. Prior verdict: strained in W36, and W37 does not retire the objection, so it stays strained. The concentration side is stark and primary: $28.5 billion of quarterly capital expenditure and roughly $5 billion of negative free cash flow at one operator, record foundry revenue, a warrant-backed silicon ceiling reported at up to $60 billion, and $550 million into a single legal-AI company at a $15.5 billion valuation. What changed is that the week's largest disclosure attacks the premise rather than the margins: the same filing states the new contracts are structured so customers prepay or bring their own hardware and require no incremental capital from the operator, which disperses the burden the hypothesis says a handful of balance sheets carry. The diffusion half did not firm up either. Value showed up as shipped product, seven generally available job agents, a globally available agentic security operations centre, a generally available procure-to-pay layer and contract extraction inside already-licensed document software, and not one vendor in the window published a price or an AI-segment operating margin. Both legs are weakened by in-window evidence, which is what strained means. Against it: The week's largest backlog disclosure partly contradicts the concentration mechanism: management stated the new contracts are structured so customers prepay or bring their own hardware and require no incremental capital from the operator, which disperses the capital burden onto counterparties rather than concentrating it. If that structure spreads, the hypothesis's premise that a handful of balance sheets carry the spend weakens even as the spend itself grows. Separately, none of the application-layer releases published a price, so the diffusion side remains an inference from shipped functionality rather than a measurement of realised return. Evidence: [capital-02](https://www.sec.gov/Archives/edgar/data/1341439/000119312526387905/orcl-20260910.htm), [capital-07](https://www.harvey.ai/blog/harvey-raises-dollar550m-at-a-dollar155b-valuation-to-help-legal-teams-own-their-intelligence), [applications-01](https://www.salesforce.com/news/stories/agentforce-job-ready-ai-agents/) - **Hypothesis 3 — Networking is the durable layer.** — **STRAINED**. Prior verdict: strained in W36, which asked for revenue evidence before this publication claims durable pricing power. W37 is the best architectural week of the year and it does not supply that evidence, so the verdict stays strained. On architecture the case is strong: an industry body published the 1600ZR implementation agreement, making 1.6-terabit coherent interconnect a multi-vendor specification; a co-packaged optics company reached $650 million of 2026 capital; a merchant silicon vendor detailed a 102-terabit Ethernet switch and a 256-lane scale-up switch; and the week's two most significant accelerator events both turned on interconnect rather than compute, one accelerator buying into the dominant ecosystem through its scale-up fabric and the largest custom-silicon agreement bundling optical interconnect into scope. On revenue there is one market-research quarter, high-end routing up 25% and direct-to-cloud routers up 94% year over year off an undisclosed base, which is not the pricing-power evidence W36 asked for and cannot clear that bar on its own. And the standards win cuts the wrong way for durability: a multi-vendor implementation agreement is how an interconnect advantage becomes a commodity. The claim W37 can actually defend is narrower than the hypothesis. Networking was the week's clearest architectural bottleneck; it was not shown to be its durable-margin layer. Against it: Standards publication is the opposite of pricing power: a multi-vendor implementation agreement for 1.6-terabit coherent optics is precisely the mechanism by which an interconnect advantage becomes a commodity, which cuts against the durability claim rather than for it. The 94% growth figure is one quarter from a market-research house off an unstated base and carries a lower source grade than the filings elsewhere in this issue. One merchant vendor also deferred its optical roadmap by a year, which is evidence that optical execution is hard rather than that optical margins are safe. Evidence: [networking-01](https://www.oiforum.com/oif-releases-critical-1600zr-coherent-interface-ia-doubling-capacity-per-wavelength-for-data-center-interconnects/), [networking-03](https://www.delloro.com/news/high-end-routing-and-aggregation-market-grew-25-percent-in-2q-2026/), [hardware-07](https://www.d-matrix.ai/announcements/d-matrix-rackscale-nvidia/) - **Hypothesis 4 — Open weights pull the floor up.** — **STRAINED**. Prior verdict: untested in W36. Strained here. The release half of the hypothesis held and the demand-routing half went untested, but the release half is narrower than it first appeared. A frontier Chinese lab published an MIT-licensed checkpoint. A separate organisation published a three-tier agentic family including a 1.6-trillion-parameter open-weight checkpoint, and independent coverage of that vendor's own materials reports the checkpoint is post-trained from the same lab's V4-Pro base, which its identical 1.6-trillion total and 49-billion active parameter shape corroborates. So the window produced two open-weight releases sitting on one lineage, not two labs raising the floor independently. The floor still moved, and the clearest movement came from neither: a serving project published a cache-residency policy that fits a full million-token context onto a single node, which materially lowers the hardware floor for on-premises deployment of long-context work. What no W37 event supplies is the mechanism the hypothesis actually claims: not one sovereign programme, national purchase, or enterprise on-premises deployment in the window was attributed to open-weight availability. The floor moved; the demand re-routing was not observed. On why the framework stands unamended after six weeks without support, prior verdicts being strained in W32 through W35, untested in W36, and strained again here: the hypothesis has two halves and only one of them is failing. Nothing in six weeks has shown demand routing away from open weights. What is missing is any published attribution in either direction, which is a measurement gap rather than contrary evidence, and narrowing a hypothesis because nobody measured it is how a framework gets quietly fitted to the record. The single observation that would settle it is a named sovereign programme or enterprise on-premises purchase whose own procurement document cites open-weight availability as a reason for the deployment. If two more quarters produce open-weight releases and still no such attribution anywhere, the honest move is to demote the demand-routing half out of the hypothesis rather than keep scoring it. Against it: The same vendor that published the open checkpoint simultaneously announced it would reroute its own hosted flagship endpoint to a different, smaller model, which is an open release paired with a hosted retreat rather than an advance on both fronts. The week's new open checkpoints are explicitly cheaper and smaller rather than more capable, so the frontier gap the hypothesis tracks did not close this week. One of the two open releases also rests on a grade-three source with vendor-reported benchmarks and no independent evaluation, so the claimed capability is not yet confirmable. Evidence: [software-01](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash), [software-04](https://huggingface.co/nex-agi/Nex-N2.5-Max), [software-05](https://vllm.ai/blog/2026-09-08-glm53-part1-hybrid-sparse-offloading) - **Hypothesis 5 — Power is the binding constraint for the next 24 months.** — **STRAINED**. Prior verdict: strained in W36, which held that power is a major site-timing constraint and not a proven unique bottleneck. W37 does not resolve that objection, so it stays strained, and the hypothesis's own word is where it fails: binding. What the week does support is that power has become a contractual constraint. The clearest tell is a unit change: the week's only operator to disclose delivered capacity reported it in megawatts and accelerators rather than in dollars, which is what a business does when its constraint is energisation rather than procurement. Around it, a hyperscaler announced an investment plan of at least €13 billion and contracted up to half of an existing nuclear plant's output for twenty-two years, an arrangement whose value is preservation of roughly a tenth of a member state's electricity rather than new generation. The federal energy department opened an information request implementing a bulk-power-system emergency authority. A state regulator's docket turned into a fight over forfeiture terms on a 75-megawatt interconnection rule, which is what scarcity looks like in a queue. What the week does not support is uniqueness. Financing, export licences, silicon timing and customer prepayment each gated a named in-window transaction, and this issue says so in four separate places. A constraint that shares the week with four others is not the binding one. Against it: The same quarter that makes the case also complicates it: 850 MW delivered in ninety days by a single operator is evidence that the constraint is being worked faster than a thirty-six to forty-eight month interconnection narrative implies. The largest capacity figure circulating this week, a reported plan exceeding 38 gigawatts, rests on an unconfirmed internal document and carries the issue's lowest source grade, so it should not be counted as evidence for anything. And the Finnish transaction adds roughly 4.7% of new generation at the site while contracting roughly ten times that volume, which means the marquee power event of the week was mostly a reallocation of existing supply rather than relief of the constraint. Evidence: [capital-01](https://www.googlecloudpresscorner.com/2026-09-09-Google-Deepens-Commitment-to-Finland-with-Two-Year-EUR13-Billion-investment-in-AI-Infrastructure), [policy-02](https://www.federalregister.gov/documents/2026/09/09/2026-18370/securing-the-united-states-bulk-power-system), [policy-04](https://interchange.puc.texas.gov/Search/Filings?ControlNumber=58481) ### Pattern watch - **Headline capital figures are increasingly ceilings conditioned on decisions the announcing party does not control, and the conditions are disclosed by the counterparty rather than by the announcer.** _[inductive, 4 weeks observed]_ - W34: a twenty-year land-and-power shell transaction used a chip vendor's residual-value guaranty, payable only after ready-for-service conditions, to make future accelerator-linked infrastructure financeable, with the guarantor disclosing the condition rather than the announcer. - W35: a hyperscaler reserved two million additional accelerators for 2027 and 2028, buying supply-chain schedule position rather than reporting installed capacity. - W36: the house measurement found 44.2% of a $2.60 billion acquisition consideration to be performance-contingent, with roughly 79 cents of earnout behind every upfront dollar; a robotics buyer committed $3.5 billion of compute against roughly $1.9 billion raised in its lifetime; and three separate buyers matched geothermal tranches, twenty-five-year capacity and delayed-draw facilities to deployment milestones years out. - W37: a €13 billion programme's price-neutrality claim depends on approximately €700 million of counterparty capital expenditure disclosed as pending investment decision; an up-to-$60 billion silicon ceiling comes from secondary reporting of a filing this publication could not open and vests through milestone warrant tranches; an up-to-$3.1 billion facility is undrawn with export-licence-contingent collateral; and a two-gigawatt national target aggregates eight operators' independent pipelines. Next week: Before 2026-12-31, at least one gigawatt-scale AI infrastructure announcement will be publicly revised downward or restructured when a third-party condition named in its own primary filing fails to clear. Falsified if the next four such announcements each reach a final investment decision or first drawdown on their originally announced scope and schedule. - **The object an enterprise evaluates and the object it subsequently runs are drifting apart, through benchmark revision, endpoint substitution, and same-day model deprecation.** _[inductive, 2 weeks observed]_ - W36: a matched-effort comparison on one agentic benchmark measured a 35.8-point spread between a provider-controlled harness and a provider-neutral one, and no third party has reproduced the provider-controlled result because the preserved state is opaque even to the benchmark authors. - W37: the most-cited intelligence index was revised twice in four days, moving from v4.1.1's five-point gap between the two leading models to v4.3's rounded tie without either model changing; one vendor announced its flagship endpoint would begin serving a different, smaller model with no caller change; and a developer platform deprecated a model inside agent mode on the day it announced the deprecation. Next week: Before 2027-03-31, at least one major model provider or evaluation publisher will ship a versioned runtime or harness manifest, naming adapter version, tool permissions, context persistence policy and retry budget, so that a score can be tied to a reproducible configuration. Falsified if the next three index or harness revisions ship without any machine-readable configuration identifier. ### Second-order effects - **Trigger:** An operator reported its quarter's AI delivery as 850 MW of capacity and more than 300,000 accelerators delivered, rather than as revenue or bookings. **Effect:** Delivered-capacity disclosure becomes a competitive norm, and the absence of it becomes informative. Expect analysts and investment committees to start asking every operator for megawatts delivered, megawatts energised and accelerators in service per period as three separate numbers, and to treat backlog-only disclosure as a weaker signal than it has been treated for the last two years. The reporting unit change also imports the energy sector's vocabulary into technology disclosure, which makes contingency easier to see because energy disclosure conventions already separate committed from pre-decision capital. _(Horizon: Q4 2026 to Q1 2027. Who moves: Public cloud and neocloud operators, equity analysts, investment committees, infrastructure lenders, enterprise capacity planners.)_ - **Trigger:** A state enacted an auditor registry that makes unregistered covered AI audit work unlawful on reported 2028 and 2029 dates while mandating no audits of anyone. **Effect:** Assurance arrangements signed today are being written against a requirement that will be defined after the contracts are in force, so evidence collected under informal arrangements may not be usable later. The rational response is to start generating registry-eligible artifacts now, which means insisting that current assurance work produce a documented scope, a named methodology, retained evidence and a versioned model inventory, whether or not the provider is registrable yet. The cost lands on enterprises rather than on regulators or vendors. _(Horizon: 2027 through 2029. Who moves: Enterprise AI governance and model-risk functions, internal audit, third-party assurance providers, AI vendors selling into regulated buyers.)_ - **Trigger:** Two separate vendor-side model changes landed inside one week: a flagship endpoint rerouted to a smaller model with no caller change, and a model deprecated across agent mode on the day of announcement, in the same release that made unattended scheduled agent tasks available. **Effect:** Model identity becomes a change-management artifact rather than a configuration detail. Teams running unattended agent loops will need pinned model versions, a regression suite that runs on vendor-side change, and a monitored signal for endpoint substitution, because agent behaviour can now change in production without the operator shipping anything. Procurement follows: expect deprecation notice periods and endpoint-stability commitments to appear as negotiated terms rather than as documentation. _(Horizon: Q4 2026. Who moves: Platform and AI engineering teams, change management, vendor management, anyone operating scheduled or unattended agent workloads.)_ ### Strategic outlook The useful question coming out of this week is not how much capacity the industry announced but how much of it is a commitment rather than an option. This publication has been asking that since W34, and W36 told readers plainly to separate operating megawatts from phased, optioned and contracted future capacity. What W37 added is a number. One operator's 850 MW and 300,000-plus accelerators delivered in ninety days, disclosed as a pair for a single named period, is the first dated denominator against which the optioned stack can be sized... and delivered is not energised, a distinction the filing itself respects and most of the coverage did not. Almost every other large figure in the window resolves into a ceiling conditioned on a decision somebody else has not taken: an unapproved €700 million of utility capital expenditure, a warrant-vesting schedule, an undrawn export-licence-contingent facility, eight operators' separate pipelines, and a monthly hosting rate that this publication multiplied by twelve because nobody else would state a term. None of that is fraudulent and most of it is competent structuring. It does mean the honest unit of analysis for the next two quarters is the conditionality rather than the headline, and that the conditions are usually disclosed by a counterparty rather than by the announcer. That is where the diligence has to go. Two practical consequences follow for anyone buying rather than selling. First, the measurement layer is now the binding procurement risk. An index rebased twice in four days, an endpoint rerouted to a smaller model, and a model deprecated inside agent mode on the day of announcement all produce the same operational result: the system you tested is not the system you run. Version pinning and regression-on-vendor-change stop being hygiene and become contract terms. Second, the controls are running ahead of the proof and the policy is running ahead of the obligation. Override-resistant agent permissions shipped generally available this week and are independently verifiable in your own tenant, which makes them worth more than any outcome percentage published in the window. Meanwhile a statutory auditor registry now exists that mandates no audits, and a grid-security rule defers its substance to a later filing. The durable move is to buy the controls you can verify, price the conditions you cannot, and refuse to treat an unbaselined vendor percentage as a measurement. ## Where we differ - **[EXTEND]** [SemiAnalysis](https://newsletter.semianalysis.com/p/nvidias-backstop-universe-heads-i): The largest guarantee structure of the cycle is circular financing, with a chip vendor manufacturing the demand that shows up in its own revenue. Our read: We agree the direction is right and think the mechanism matters more than the label: the instrument guarantees residual value on third-party land, power and shell leases rather than chip purchases, and pays only on a defined trigger after ready-for-service conditions, so commitment date and exposure date are deliberately separated. The per-gigawatt ordering in that analysis is the most useful number published on the topic all week. Attribution where it belongs: the circular-financing frame applied specifically to a residual value guaranty originates with Pip Theory on August 28, and the analysis we are extending is the in-window restatement of it. The number both of us should lead with is the gap between roughly $600 billion reported in July and roughly $105 billion actually signed. We extend it one step: the same structure appears in this week's neocloud financings at smaller scale, where an undrawn licence-contingent facility and a convertible note with an investor put do the same work of raising optionality rather than capital. - **[DIFFER]** [World Nuclear News, September 9](https://www.world-nuclear-news.org/articles/google-signs-up-for-electricity-from-finnish-nuclear-power-plant): Google's Finnish package is a landmark nuclear-for-AI commitment that brings new clean capacity to the grid and protects households from AI-driven price increases. Our read: The facts are solid and the framing is inflated. This issue's house measurement takes the transaction apart from the two primary filings: the additional generation disclosed is 48 MW against a 1,014 MWe site, 4.7%, of which only 10 MW is attributed to the agreement itself, while the contracted volume is roughly 10.6 times the addition. More decisively, the price-protection claim depends on approximately 70% of the life-extension programme's capital expenditure, EUR 700 million, that the counterparty discloses as pending investment decisions. No surveyed outlet set those two statements side by side, and both sit in the primary sources. - **[DIFFER]** [ModelCap snapshot, September 10, reproducing Intelligence Index v4.3; the index article itself is corroboration rather than the foil](https://modelcap.ai/benchmarks/aa-intelligence-index): Independent benchmarking now shows the two leading frontier models level, with the newer entrant having caught the incumbent. Our read: Our disagreement is with the downstream reproduction, not with the evaluator. The convergence is an artifact of instrument revision rather than model movement: Artificial Analysis v4.1.1 placed Astra five points behind Fable 5.1, while v4.3 revised the instrument and reported a rounded tie at 53 and 53 without either model changing. ModelCap's September 10 downstream snapshot reports 53.4 against 52.8 under a reasoning-unspecified configuration, which is not like for like with Artificial Analysis's max-with-fallback comparison and cannot be used to unround its tie. Artificial Analysis disclosed every change, which is why this is a lesson rather than a scandal. The finding that survives is the 57% cost spread for the same rounded score, which is a procurement fact. - **[DIFFER]** [A vendor quote in the Nscale and Figure releases of September 3, adopted across multiple outlets](https://www.nscale.com/press-releases/nscale-and-figure): Physical AI has arrived and robotics now justifies frontier-lab-scale compute, with the robotics flywheel activated by a multi-billion-dollar compute commitment. Our read: Credit first, because it is owed: Forbes is the outlet that caught the financing asymmetry, roughly $1.9 billion raised in the buyer's lifetime against $3.5 billion committed to rent computers, and most of the coverage did not. The bull frame itself originates in a vendor quote inside the two companies' own September 3 releases, and that is what we differ with. The underlying deal is dated September 3 and belongs to the prior week, so only the commentary is in this window, and the commentary did not read the buyer's own papers. The model the compute is meant to train was trained on roughly 500 hours of supervised data and deploys an 80-million-parameter policy onboard, while independent scaling work finds a power law in environment and object diversity and no clear power law from more demonstrations in the same setup. The 80-million-parameter onboard policy and the 100,000-accelerator order have not previously appeared in the same sentence, and they should. - **[OPEN]** [Counterbrief frame sweep of the surveyed coverage; the Federal Register notice of September 9 is the item that circulated](https://www.federalregister.gov/documents/2026/09/09/2026-18370/securing-the-united-states-bulk-power-system): The regulatory item the week's coverage picked up was the federal bulk-power-system notice; no outlet in the surveyed set elevated the September 8 large-load ride-through proposal at all. Our read: This is an unopposed house prioritisation rather than a disagreement with a named party, and we would rather say so than manufacture a foil out of a primary regulatory document that holds no editorial view. On September 8 a regional transmission organisation presented proposed ride-through, ramp-rate, telemetry and remote-disconnect requirements for large computational loads, motivated by load disconnections that produced multi-gigawatt events, targeting a November filing. No analyst in the surveyed set connected it to the gigawatt-scale campuses sitting inside that footprint. Whether it reaches them, and at what cost, is genuinely open, and it is the question we would most like answered before the next twenty-year lease is signed. ## Track record Cumulative ledger: **111 predictions made**, 57 resolved (23 hit / 15 partial / 19 miss), 54 pending, 0 overdue. Hit rate (partial = half): **54%**. Brier score: **0.200** (0 = perfect, 0.25 = coin-flip). Calibration by confidence band: - Bold (<55%): 1 resolved, hit rate 100% vs mean confidence 43% - Core (55-80%): 55 resolved, hit rate 52% vs mean confidence 66% - High-conviction (>80%): 1 resolved, hit rate 100% vs mean confidence 84% Recently resolved: - **HIT** (called at 72%): NVIDIA files exhibits with the 10-Q for the quarter ended July 26, 2026 that translate the SB Energy PORTS-Pike residual-value guaranty into a per-quarter contingent-obligation disclosure and identify the OpenAI affiliate as tenant, by October 31, 2026. — NVIDIA filed the Form 10-Q for the quarter ended July 26, 2026 on August 26, 2026 — inside the window. It satisfies all three trigger elements: guarantees 'capped at a total of $105 billion' with an exposure table of $3.5B AI-cloud guarantees plus $105.0B SB Energy for $108.5B total; effectiveness conditioned on SB Energy satisfying applicable ready-for-service conditions as each of nine phases is placed in service from fiscal 2029; and the tenant identified as 'an affiliate of OpenAI Group PBC' at the PORTS Technology Campus in Pike County, Ohio. Exhibit 10.1 is the Form of Residual Value Guaranty. - **PARTIAL** (called at 80%): Aggregate 2026 hyperscaler capex revises upward by 10% or more from the $700B baseline. — Q1 prints (MSFT $190B, GOOG $180-190B, META $125-145B, AMZN $200B reaffirmed) take 2026 aggregate to $695-725B (+77% YoY) vs the $700B W17 baseline. At/near baseline; +10% revision (~$770B) plausible by Q2 print. Score moves to hit if Q2 takes aggregate above $770B. - **HIT** (called at 43%): Z.ai publishes GLM-5.3 weights to Hugging Face by September 15, 2026, closing the two-week window promised at the model's August 14 announcement. — Z.ai published the full 753B-parameter GLM-5.3 weights to Hugging Face at zai-org/GLM-5.3 on August 27–28, 2026 — in-window and inside the trigger's September 15 window, distinct from GLM-5.2 — after GLM-5.3-Flash MIT weights landed Aug 26. The material nuance is licensing, not availability: GLM-5.3 ships under a bespoke GLM-5.3 license rather than MIT, requiring Z.AI security review before commercial use by any Model-as-a-Service operator whose group revenue exceeds $10B over any 12 consecutive months. - **HIT** (called at 66%): An independent benchmark finds Gemini 3.6 Flash at least 12% cheaper per completed agentic task than Gemini 3.5 Flash by August 31, 2026. — Artificial Analysis measured Gemini 3.6 Flash at $0.50 average cost per completed agentic task versus $0.59 for 3.5 Flash — a 15% reduction, above the 12% cheaper-per-task bar — before Aug 31. - **HIT** (called at 84%): DeepSeek V4's official GA pricing does not reset the ultra-cheap floor: off-peak deepseek-v4-pro output pricing stays at or above ¥6 (~$0.85) per MTok through August 31, 2026 — the kill-condition test for this issue's price-band-convergence claim. — DeepSeek's official API pricing page kept GA deepseek-v4-pro off-peak output at $1.98/MTok (~¥14+) through Aug 31 — well above the ¥6 (~$0.85)/MTok ultra-cheap floor the trigger set as the kill condition. - **HIT** (called at 64%): 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. — Cursor shipped Cursor Router in July 2026 with Auto Balance/Intelligence routing controls and published measured cost-per-commit figures ($4.63–$6.76) from live traffic — product-level harness routing and cost telemetry beyond session budget caps. ## Watchlist - **Sep 15-17 — AI Infra Summit disclosures on 102-terabit switching and scale-up fabrics.** Two of this week's networking items were pre-announcements staged for the conference, so the datasheets, availability windows and named customers all land immediately after this issue's cutoff and will determine whether the roadmap is deployable or aspirational. - **Sep 21-23 — OCP Summit and ECOC, where 1600ZR conformance and 800-volt rack power get their first public test.** The implementation agreement published this week is only worth what multi-vendor interoperability proves it is worth, and these are the venues where conformant product and power-delivery roadmaps are shown side by side rather than asserted separately. - **Nov 2026 — PJM's targeted Section 205 filing on large computational load requirements.** If ride-through, ramp-rate, telemetry or remote-disconnect obligations reach gigawatt-scale loads, twenty-year leases already signed cover capacity that can be curtailed involuntarily, which changes leasehold value, project bankability and the exposure profile of the guarantees written against them. - **Q4 2026 — The first quarter in which delivered capacity becomes a disclosure norm rather than an outlier.** One operator has now published megawatts delivered and accelerators delivered for the same period. Whether peers follow, or decline to, is the single most informative thing about the gap between contracted and delivered capacity across the whole sector. - **Dec 2026 — Substance of the bulk-power-system emergency authority deferred by this week's information request.** The rule as opened defines machinery without obligation, so what large loads are actually required to do arrives later. Anyone signing an interconnection or lease agreement before then is pricing a requirement that has not been written. - **2028-2029 — Commencement of the California AI auditor registry and verification-organisation requirements.** Reported statutory dates would make unregistered covered audit work unlawful, which would render today's informal assurance arrangements unusable as later evidence. The dates come from secondary analyses and should be re-checked against the enrolled statute before anyone plans against them. ## Changelog - W37 authored from the validated research, counterbrief, house measurement and synthesis artifacts, all dated to the September 12 cutoff. The research window is September 7 through 12; no event dated September 13 or later is referenced. - Three date traps were checked and excluded from the in-window event set: the Ohio campus guarantee announcement is dated August 17, the robotics compute agreement is dated September 3, and the relevant semiconductor quarter was reported September 2. Where this issue engages them it engages the in-window commentary and says so. - The house measurement is computed from two primary filings and reports approximately 70% to two significant figures, because the denominator is disclosed as approximate. The EUR 700 million figure and the 80%-of-projects figure in the same disclosure are not interchangeable, and the issue does not treat them as such. - No prediction resolution is recorded in the track record this week. Prediction p102 was the closest call: three candidates were examined and each fails or only arguably satisfies at least one conjunct of its own trigger, so under the accountability rule it stays pending with the contested conjunct named in its notes rather than being claimed as a hit. - The synthesis and house measurement blocks embedded here were serialised from content/industry/research/2026-W37/synthesis.json and primary-observation.json and then checked field by field against those artifacts, so the published issue and the research record are identical rather than merely consistent. Every evidence reference in the synthesis was separately reconciled against research.json, both that the URL appears in the verified research set and that a reference carrying an event id points at one of that event's own sources. Two defects were found and repaired this way: an initial draft contained links reconstructed from memory rather than copied, and a chain item on proposed large-load transmission requirements was backed by a different regulator's docket. The second is now cited to the presentation it describes. - Levers are held at Unknown or Not measurable wherever the window supplies no reproducible input, and the reason is stated per lever. The capability-gap lever is reported as not measurable rather than unchanged, because the reference index changed basis twice inside the window. - Issue art was generated after the editorial and deterministic gates: an abstract editorial scene separates diffuse option pathways from one measured delivery, carries no text or factual claims, and is published with descriptive alt text. - Review correction, SpaceX contract tenor. An earlier draft of this issue attached a 90-day commitment with a 90-day exit to the specific $1.11 billion monthly hosting contract, in the transaction label, the big-story body, a key point and the capital narrative. The disclosure does not support that. The executive characterised the company's compute contracts generally in those terms and explicitly did not state that the structure applies to this contract. Every such implication has been removed, and the published criticism now rests only on what was actually withheld: no term, no counterparty, no megawatts. The general characterisation is retained in the graded event record, where it is correctly scoped. - Review correction, capital-flow comparability. The Hyperscaler-Hosted category previously carried a numeric capital-in value of 28.5. That figure is one company's quarterly capital expenditure and is not comparable with an annual category series, so the numeric field is now empty and the disclosure survives only as a labelled single-company transaction and in the prose, where its basis is stated. - Review correction, tree-state reasoning. Hypothesis 1's reasoning and counter-evidence previously cited the absence of new rows in the evolutionary tree as evidence about industry cadence. The tree is a maintained reference artifact, so its row count measures maintenance state rather than the industry; the W37 tree edit has since been made and carries three additions and two updates, which does not change the point. Those clauses are struck. The substantive argument is unchanged and stands on its own evidence: capability is not measurable this week because the reference index changed basis twice inside the window, and the week's genuinely new checkpoints lead on cost and footprint rather than capability. - Restatement, matched-effort benchmark spread. Prediction p100 carries W36's wording of a 35.9-point matched-effort gap. Recomputing from the published figures, 62.71% against 98.55%, gives 35.8 points, and W37 uses 35.8 wherever it states the spread itself. The prediction text and its 80.7% trigger threshold are preserved verbatim, so the bar is not moved; the 0.1-point difference does not change which side of the threshold any candidate result falls on. - Seven events used in this week's publications but previously carried only in the raw research notes were consolidated into the graded event record, so every published claim now traces to a graded source: five application-layer items, one agent-layer item, and the transmission operator's large-load ride-through proposal, which is now graded as a policy event and cited as one in the synthesis rather than as a counterbrief note. The graded set is 49 events; three of the seven additions are grade 2, all three in the application layer, and each is flagged as such at every use. - Outlet attribution tightened. A comparison in Where We Differ previously credited three outlets while linking one. It now names only the outlet it links. - Review correction, novelty and attribution. The big story previously opened on the delivered-capacity denominator as though the committed-versus-optioned frame arrived with it. The frame is four weeks old in this publication, W36 told readers plainly to separate operating megawatts from phased and optioned capacity, and the delivered-versus-announced distinction was built out in coverage that predates the window. The opening now credits both and claims only what is new this week, which is the paired megawatt-and-unit disclosure for one named period. Two further attributions were repaired: the circular-financing frame applied to a residual-value guaranty is credited to its August 28 originator, and the physical-AI bull frame is attributed to a vendor quote inside the two companies' own releases rather than to the outlet that caught the financing asymmetry, which is credited for catching it. - Review correction, delivered is not in service. Oracle's filing says delivered. Earlier drafts of this issue used placed in service, in service and energised interchangeably with it across the big story, the synthesis and a second-order effect. Delivered capacity, energised capacity and operating capacity are three different quantities and the issue now keeps them apart, asks for all three separately where it asks operators for disclosure, and says so where the distinction changes the reading. - Review correction, Artificial Analysis. Earlier statements described the index revision as inverting the reported ordering and later replaced that error with an unsupported four-point gap and a misattributed 53.4-to-52.8 unrounding. The corrected comparison is versioned: v4.1.1 placed Astra five points behind Fable 5.1, while v4.3 reported a rounded tie at 53 and 53 without either model changing. The 53.4 and 52.8 figures are confined to ModelCap's September 10 downstream snapshot, whose reasoning-unspecified configuration is not like for like with Artificial Analysis's max-with-fallback comparison and is not used to resolve the tie. - Review correction, annualisation and Qualcomm attribution. The $13.3 billion hosting figure is this publication's arithmetic, one month's $1.11 billion billing rate multiplied by twelve, and earlier drafts attributed it to the discloser and rounded it to $13 billion. It is now labelled as house arithmetic and carried at $13.3 billion, with the same source limitations stated: no contract term, no counterparty, no megawatts. The Qualcomm agreement itself is primary and filed, but its reported up-to-$60 billion ceiling rests on CNBC and Reuters reporting of a filing this publication could not open. The two residual synthesis uses now restore both up to and the secondary-reporting qualification, and the public narrative uses the same scope. - Review correction, capital-flow direction arrows. Three arrows pointed up against an Unknown prior, which fabricates a direction off a non-comparable base. All three are now flat, and the single-company disclosures that motivated them survive as labelled transactions and in the prose, where their basis is stated. - Review correction, thesis verdicts. Hypotheses 2, 3 and 5 were scored supported over counter-evidence this issue itself raises at the mechanism level rather than at the margin: a prepay and bring-your-own-hardware structure disperses the capital burden the concentration hypothesis requires, a multi-vendor 1600ZR implementation agreement is how an interconnect advantage becomes a commodity, and four other constraints each gated a named in-window transaction, which is not what binding means. All three are now strained. Every hypothesis states its prior verdict and what did and did not change, and the fifth names W36's objection that power is a major site-timing constraint rather than a proven unique bottleneck, which this week does not retire. - Hypothesis 4 stands unamended in the living framework after six weeks without support, and this issue says why rather than editing the framework to fit the record. The release half of the hypothesis holds. The demand-routing half has never been measured in either direction, which is a measurement gap and not contrary evidence, and narrowing a hypothesis because nobody measured it is how a framework gets quietly fitted. The single settling observation is named in the hypothesis reasoning: a sovereign programme or enterprise on-premises purchase whose own procurement document cites open-weight availability as a reason for the deployment. - Synthesis history and falsifiers. The conditionality pattern now carries the four weeks it has actually been observed, W34 through W37, with a named instance for each, rather than the two weeks an earlier draft recorded. Connections zero and two now use one dated diagnostic outcome rather than low-probability multi-party conjunctions. Connection three is restated as a forward claim through year-end and repriced to 42%, the complement of p108's 58% probability for its single PJM filing falsifier. Connection one was already coherent and is unchanged. All synthesis fields were reapplied to content/industry/research/2026-W37/synthesis.json and the two artifacts verified field-identical again afterwards. - Calibration disclosure. The resolved record carries a persistent overconfidence gap in the Core band, 55 of 57 resolved predictions at mean confidence 66% against 52% realised, roughly fourteen points. Readers should shade this issue's three Core-band bets accordingly. One of them has already been repriced for a separate reason: the transmission-operator filing bet was restructured so the docketing is the precondition and a named telemetry or remote-disconnect requirement is the thing predicted, and its confidence was cut from 71% to 58% because the adjacent docket family has already slipped once. - Review correction, PJM no-cap simulation. An earlier lever falsely said PJM had not published a capped-versus-uncapped 2028/29 simulation. PJM's own results report gives the no-cap-or-floor case: $554.72 per MW-day for the RTO and $776.69 for ComEd versus the $325 capped result, with $29.7 billion of simulated cleared value against $16.4 billion actual. The lever now attributes those figures to PJM and treats them as a modelled comparison rather than proof that the cap alone caused the gap. - Review correction, Dell'Oro comparison basis. Earlier drafts described the reported 25% high-end-routing growth and 94% direct-to-cloud-router growth as sequential movement within one quarter. Dell'Oro reports both against the same quarter a year earlier. Every use now says year over year, and the issue does not extrapolate the market-research figure off its undisclosed base. - Review correction, Nex-N2.5 lineage. This issue originally presented Nex-AGI as an independent frontier lab whose 1.6-trillion-parameter open-weight checkpoint raised the deployment floor alongside DeepSeek's. A source recorded in the W37 research notes but not opened during the original pass has since been read directly. It reports that Nex-N2.5 Max is post-trained from DeepSeek-V4-Pro-Base and that the 35B sibling is post-trained from a Qwen3.5 variant, and the checkpoint's 1.6-trillion total and 49-billion active parameter counts match DeepSeek's V4-Pro exactly. Nex-AGI is therefore removed from the Frontier Labs constituent examples, and the fourth hypothesis and the sovereign-compute narrative now state that the week's two open-weight releases sit on one lineage rather than two. The attribution is independent coverage of the vendor's own materials, not a Nex-AGI page, and every use of it says so. The substantive finding is unchanged and better supported than before: the deployment floor fell this week, the clearest single reduction came from a serving project rather than from either checkpoint, and no event in the window attributed a sovereign or enterprise purchase to open-weight availability. --- Source of truth: `src/data/industry/weekly/2026-W37.ts`. Canonical HTML: . PDF: .