AI's scarce asset is now a governed path to service, not an announced model or megawatt

Key takeaways

• ARC Prize's matched-effort Astra comparison shows a 35.9-point runtime-associated spread; version the model, memory policy, context compression, and harness together.
• The week's infrastructure pattern is capital committed to dated future service, not operating capacity: a power-purchase agreement starts in 2028, a 150-terabit-per-second cable targets Q4 2028, and delayed-draw funding releases against milestones.
• Vertiv put 44.2% of UtilityInnovation Group's maximum $2.60B consideration behind EBITDA delivery; interpreting that as seller risk transfer is a house reading, not a disclosed fact.
• Broadcom's $16.7B AI-semiconductor quarter proves supplier monetization, not utilization or application returns; custom accelerators and networking remain combined.
• PJM's comment window closed while Texas staff proposed $50,000/MW security above 75 MW, keeping power obligations inside operating contracts.

By the numbers

Astra matched-effort harness-associated spread

35.9 points — ARC Prize: 62.7% Standard versus 98.6% Provider Adapter, both at maximum reasoning.

Vertiv-UIG maximum price that is contingent

44.2% — $1.15B earnout divided by $2.60B maximum consideration.

Broadcom Q3 AI-semiconductor revenue

$16.7B — Filed unaudited quarterly; 56.4% of $29.6B total revenue.

Firm geothermal capacity under Google-Fervo power-purchase agreement

396 MW — Four 99 MW tranches beginning Q3 2028; conditional expansion excluded.

Nscale delayed-draw commitments

$3.05B — $1.85B Texas plus $1.20B North Carolina maximum commitments.

Flywheel arc · all-three

The asset being bought is a controlled path from intent to completed work, not a standalone model, chip, cable, or megawatt.

• Credit ARC Prize's finding and use its matched-effort comparison: Astra's Provider Adapter is associated with a 35.9-point gain at maximum reasoning.
• Separate operating supply from dated future-service commitments across geothermal power, subsea fiber, and milestone-based deployment finance.
• Use milestone, option, earnout, and rollback terms to price delivery risk rather than quoting maximum headline capacity.
• Vertiv left a material share of UIG's maximum consideration contingent on post-close EBITDA delivery; see the house measurement.

What this means

Architects should stop procuring model names as if the surrounding runtime were interchangeable. Astra's neutral-versus-provider harness spread, GitHub's administrator gates, and year-end Gemini pricing all make deployment policy part of measured capability and cost. See The Model Pulse for the full model and tree read.

• Version the model, harness, memory policy, and retry budget together; Astra's headline result is not model-only.
• Treat Fable and Gemini availability, retention, administrator enablement, and promotional pricing as stock-keeping unit (SKU) properties.
• See Model Pulse for tree additions, benchmark comparability, and the quiet open-weight week.

Sep 3

OpenAI begins phased GPT-6 Astra rollout with 1M context and $10/$50 per million-token Standard API pricing

Sources OpenAI, ARC Prize, Artificial Analysis

Sep 1

Claude Fable 5.1 becomes generally available across GitHub Copilot; enterprise administrators must enable it

Sources GitHub

Sep 3

Gemini 3.8 Flash enters GitHub Copilot under introductory provider pricing through year-end

Sources GitHub

Sep 1

NVIDIA releases Apache-2.0 Muse-Glimmer-30B NVFP4 (4-bit floating-point) checkpoint, reducing storage from 60 GB to 24.7 GB

Sources NVIDIA on Hugging Face

What this means

Infrastructure buyers should separate production systems, contracted phases, and architecture announcements. The HUMAIN cluster is live but unquantified; Cerebras has only its first Finnish phase under construction; MediaTek-NVLink has no delivery milestone. Broadcom's filed unaudited quarterly disclosure is the hard signal, but its combined custom-accelerator-and-networking line cannot resolve which layer holds margin.

• Count HUMAIN as production evidence, but not as proof of the 250 MW or 1 GW roadmap ceilings.
• Keep custom-accelerator and NVLink dependence in the same architecture model; MediaTek disclosed an ecosystem commitment, not shipping silicon.
• Broadcom's AI line proves supplier revenue conversion while withholding the split needed to rank compute versus networking economics.

Aug 31

AMD, Cisco, and HUMAIN put an MI355X and 800G Ethernet cluster into production; live scale and utilization undisclosed

Sources AMD, Cisco

Aug 31

MediaTek joins NVIDIA NVLink Fusion for custom cloud accelerators (XPUs) without a named customer, tape-out, or shipment date

Sources NVIDIA

Sep 1

Cerebras contracts phased 50 MW, 80 MW, and 165 MW Finnish campus capacity under seven-year service orders

Sources Cerebras announcement mirror, Converge Digest

Sep 2

Broadcom reports AI-semiconductor revenue at 56.4% of Q3 company revenue, without splitting accelerators from networking

Sources Broadcom earnings release

What this means

Network architects should design delegated changes around simulation, approval, evidence, and rollback, not just easier invocation. AWS-Azure and Fabric One make intent-driven multicloud connectivity concrete, while the Firmus contract shows long-duration queue hedging. The durable-layer thesis remains strained until revenue growth proves pricing power.

• Buy route diversity and policy simulation before natural-language network changes reach production.
• Treat Fabric One as architecture direction until pricing, service levels, provisioning time, and rollback evidence publish.
• Firmus bought future schedule position; it did not demonstrate a present bandwidth shortage or lit capacity.

Aug 31

AWS-Azure managed private multicloud interconnect enters preview in four paired regions

Sources AWS

Sep 1

DE-CIX deploys dual-homed Azure ExpressRoute Metro across five global metros

Sources DE-CIX

Sep 2

Equinix announces pre-beta Fabric One for intent-driven routing, encryption, resilience, and failover

Sources Equinix

Sep 3

Firmus commits approximately $300M for up to 150 terabits per second on APX East for 25 years, targeting Q4 2028 service

Sources Firmus and SUBCO

Frontier Labs detail

Prior rolling estimates are not reproduced because this issue's artifact set lacks the constituent list, valuation date, and formula inputs. Astra pricing and Anthropic's safeguards remain product signals, not category capital or revenue totals.

• Sep 3 · OpenAI launches GPT-6 Astra Standard API (list pricing in software events)
• Sep 1 · Anthropic announces customer-controlled Enterprise Frontier Safeguards

Sources OpenAI · Anthropic

Hyperscaler-Hosted detail

Google contracted the firm Fervo geothermal capacity noted above and received an option for approximately 600 MW of conditional expansion. If accepted under a definitive agreement, total contracted capacity would be not less than approximately 950 MW, with June 2030 as the expansion's guaranteed commercial-operation date. None belongs in a capital total without disclosed consideration.

• Sep 1 · Google Energy signs 396 MW, 15-year Fervo power-purchase agreement; initial delivery targeted Q3 2028
• Aug 31 · AWS-Azure private multicloud interconnect enters four-region preview

Sources Fervo filing mirror · AWS

Neoclouds detail

Nscale closed $1.85B for Texas and $1.20B for North Carolina. These are maximum commitments available against future deployment milestones, not cash funded on August 31. They are shown as a transaction, not added to a fabricated rolling ledger.

• Aug 31 · Nscale closes two investment-grade delayed-draw facilities · Up to $3.05B
• Sep 3 · Firmus reserves future APX East capacity for its Australian AI-factory program · ~$300M

Sources Nscale · Firmus and SUBCO

On-Prem / Hybrid detail

CPP Investments and Equinix completed the atNorth acquisition, and AMD-Cisco-HUMAIN announced a live Saudi cluster. Neither supplies a reproducible aggregate for this category.

• Sep 2 · CPP Investments and Equinix complete atNorth acquisition · ~$4.0B
• Aug 31 · AMD-Cisco-HUMAIN production MI355X and 800G cluster goes live; scale undisclosed

Sources Equinix and CPP Investments · AMD

Signal score 4/5

Broadcom reported Q3 FY2026 revenue of $29.6B and AI-semiconductor revenue of $16.7B, 56.4% of company revenue.

The reported line supports supplier monetization. It does not separate custom accelerators from networking or show customer utilization and application returns.

Sources

Broadcom filed unaudited quarterly earnings release, September 2, 2026.

Signal score 5/5

Google Energy signed a 396 MW, 15-year Fervo power-purchase agreement in four 99 MW tranches beginning in Q3 2028.

Count 396 MW as firm future capacity. Exclude conditional expansion until Google accepts it and a definitive agreement is executed; do not treat either amount as energized today.

Sources

Fervo filing mirror and Utility Dive, September 1, 2026.

Signal score 4/5

Astra scored 99.9% on ARC-AGI-3 and therefore nearly saturates general intelligence.

The number is real but the conclusion is not. ARC Prize's Standard harness landed far below the headline adapter score; the best-observed adapter result used different reasoning effort, provider-private state, and context compression. Model Pulse carries the matched-effort scores. ARC says the closed-ended benchmark is not proof of general intelligence.

Sources

OpenAI and ARC Prize, September 3, 2026.

Signal score 2/5

DeepSeek will deploy at least 160,000 Huawei Ascend accelerators at a roughly 1 GW Inner Mongolia inference site.

Potentially material domestic-inference substitution, but no purchase contract, delivered quantity, rack design, or schedule is public. Do not put it into capacity forecasts yet.

Sources

Bloomberg and The Decoder, September 4, 2026; no primary confirmation.

Signal score 1/5

Fabric One has already made the network an autonomous production control plane for enterprise AI.

Architecture direction, not operating proof. The product is pre-beta with no price, service level, provisioning benchmark, rollback evidence, or production customer result.

Sources

Equinix announcement and launch coverage, September 2-4, 2026.

Filing-Derived

Vertiv made 44.2% of UtilityInnovation Group's maximum $2.60 billion consideration performance-contingent, putting 79.3 cents of earnout behind every upfront dollar.

Method: House computation from Vertiv's September 2, 2026 acquisition disclosure. The primary release states: (1) approximately $1.45 billion cash consideration at closing; (2) up to $1.15 billion additional cash consideration tied to 12- and 24-month EBITDA targets; and (3) the upfront price is approximately 13 times expected 2027 EBITDA. Arithmetic, all in US dollars: maximum consideration = $1.45B + $1.15B = $2.60B; contingent share of maximum = $1.15B ÷ $2.60B = 44.23%; upfront share = $1.45B ÷ $2.60B = 55.77%; contingent dollars per upfront dollar = $1.15B ÷ $1.45B = $0.793; current implied expected 2027 EBITDA = $1.45B ÷ 13 = approximately $111.54M; maximum earnout ÷ current implied expected EBITDA = $1.15B ÷ $111.54M = 10.31x. The last ratio is a scale comparison, not the deal's full-price EBITDA multiple, because paying the earnout requires undisclosed higher EBITDA targets.

Implication: The disclosed structure makes a material share of potential consideration contingent rather than fixed. The house interpretation is that this transfers some delivery risk to sellers, but the disclosure does not quantify that transfer. Boards should compare the contingent share of maximum consideration and the earnout-to-upfront ratio before comparing headline transaction values.

Caveats: Seller risk transfer is an interpretation, not a disclosed fact: Vertiv does not publish seller obligations, the EBITDA thresholds, their weighting, payout curve, revenue concentration, or baseline. The approximate inputs limit precision, the $2.60B maximum is not the expected purchase price, and the 10.31x scale comparison is not a transaction multiple. Closing remains expected in Q4 2026.

Performance-contingent share of maximum consideration

44.2% ($1.15B of $2.60B) — $1.15B maximum earnout ÷ ($1.45B upfront + $1.15B maximum earnout); this is the selected house measurement

Earnout dollars per upfront dollar

$0.793 per $1.00 upfront — $1.15B maximum earnout ÷ $1.45B cash at close; the structure places nearly eighty cents of additional price behind each upfront dollar

Upfront share of maximum consideration

55.8% ($1.45B of $2.60B) — $1.45B cash at close ÷ $2.60B maximum; closing cash is still the majority of potential consideration

Implied expected 2027 EBITDA

Approximately $111.5M — $1.45B upfront price ÷ the disclosed approximately 13x expected 2027 EBITDA multiple; approximate because both inputs are rounded

Maximum earnout relative to current implied EBITDA

10.31x current implied expected 2027 EBITDA — $1.15B ÷ $111.5M; this sizes the contingent pool only and is not an acquisition multiple because the earnout depends on higher undisclosed EBITDA targets

Sources Vertiv acquisition announcement and financial terms

Abductive · 76% confidence

Enterprise AI is re-layering around a delegated-control boundary: policy, evidence custody, approval, and rollback are becoming separately owned infrastructure around model execution.

Steel-man: Three of the four controls are new, previewed, or rolling out later: Anthropic's Enterprise Frontier Safeguards starts in phases, Copilot approval is public preview, and Fabric One is not yet in beta. These could remain vendor-specific features rather than a durable independent layer. The strongest part of the claim is narrower: across code, model traffic, and networking, vendors are converging on explicit scopes, retained evidence, and revocation. Falsified if production releases fold those controls back into opaque model behavior or customers cannot export and audit the resulting policy state by Q2 2027.

• Astra's launch joined persistent model state with action-level classifiers and automatic review, making runtime control part of the released capability rather than an external compliance attachment.
• Anthropic's Enterprise Frontier Safeguards separates detector operation from data custody: monitoring runs across sessions, while logs, keys, flagged-event review, and human access stay in the customer's cloud account.
• GitHub now lets a Copilot approval satisfy branch-protection rules, but scopes the authority by repository and path and dismisses the approval after a new commit, converting AI judgment into revocable policy state.
• Equinix plans to translate portal, application programming interface, agent, and natural-language intent into routing, encryption, resilience, and failover, moving delegated control from software changes into network state.

Sources agents-01 · agents-02 · agents-03 · networking-04

Inductive · 74% confidence

An accumulating procurement pattern is visible across AI infrastructure: buyers are committing capital to dated future-service paths in power, fiber, and deployment before those assets operate.

Steel-man: Power agreements, subsea capacity, delayed-draw loans, and acquisitions are different instruments with different risks; grouping them can obscure ordinary long-lead infrastructure planning. None proves that queue positions are independently tradable or earn excess returns. The inference is a continuing procurement pattern, not a new asset class or measured directional shift. Falsified if comparable power, fiber, and compute capacity becomes repeatedly available on short notice without precommitment or scarcity-linked terms by mid-2027.

• Google's 15-year Fervo power-purchase agreement contracts 396 MW in four tranches beginning in Q3 2028 and includes approximately 600 MW of conditional expansion, reserving a future delivery path rather than consuming present generation.
• Firmus committed approximately $300 million for up to 150 terabits per second over 25 years on an Australia-US cable targeted for service in Q4 2028, paying for route position before the system is built.
• Nscale's $3.05 billion facilities are delayed-draw commitments, meaning funds become available against future milestones rather than all at closing, for two deployments spanning compute, network, cooling, and site work.
• Vertiv and Flex agreed to acquire power-control and conversion capabilities that influence whether a permitted site can move from interconnect application to energized rack.

Sources capital-02 · networking-05 · capital-01 · capital-05

Abductive · 67% confidence

Agent procurement complexity is accumulating in bundles, credits, workflow entitlements, retention routes, and runtime choices, making utilization forecasting a core buying competency.

Steel-man: Salesforce is one vendor, its credits are not directly convertible to model tokens, and introductory Gemini pricing may expire without changing Copilot seat economics. Customers can still capture value primarily from superior capability on difficult tasks. This evidence supports procurement complexity, not margin migration. Falsified if major agent suites converge on transparent per-outcome prices and publish stable task-level usage denominators by Q2 2027.

• Salesforce embedded Agentforce, collaboration, analytics, security, and support into three seat editions while coupling access rights to different consumption pools.
• Astra launched as a premium application-programming-interface tier while Gemini 3.8 Flash entered Copilot under temporary introductory provider pricing, widening the model-price range inside comparable agent surfaces.
• GitHub administrators can choose whether Fable 5.1 is available across nine surfaces, showing that enterprise entitlement and retention policy can matter as much as public model availability.
• AWS's migration pattern deliberately separates tools, loop, runtime, memory, and observability so model choice can change without rebuilding every operating layer.

Sources applications-01 · software-01 · software-02 · agents-04

Hypothesis 1 · Strained

The cycle is accelerating, not slowing.

W36 shows rapid software packaging: Astra arrived with persistent memory, asynchronous clarification, and action monitoring, while Anthropic, GitHub, and AWS announced adjacent distribution and control changes in the same week. It does not measure a shorter prior-to-current doubling interval across software, hardware, and networking, and therefore cannot support the hypothesis's flywheel-cadence test.

Counter-evidence: W36 had no production-volume next-generation silicon release: MediaTek-NVLink is a collaboration, Tensordyne is modeled pre-silicon, and Cerebras's 165 MW is a phased campus target with only the first phase under construction. The acceleration is clearest in software distribution and control-layer iteration, not uniformly across the hardware lens; two quarters of stalled production would still trigger the framework's refutation test.

Sources software-01 · agents-02 · software-03

Hypothesis 2 · Strained

Capital is concentrated, returns are diffuse.

Large infrastructure commitments show capital concentration, including Nscale's delayed-draw facilities and the atNorth, UtilityInnovation Group, and EPC Power transactions. Salesforce packaging, Wonderful's valuation, and Atira's reported deployments do not measure diffuse returns or margins. The week supports the concentration half of the hypothesis but leaves the return-distribution half untested.

Counter-evidence: Broadcom captured $16.7 billion of quarterly AI-semiconductor revenue, 56.4% of company revenue, showing returns can concentrate with an upstream custom-silicon and networking supplier rather than diffuse downstream. Wonderful and Atira disclose funding, customer counts, and selected outcomes but not audited margins or retention, so application-layer return diffusion remains more plausible than measured.

Sources capital-01 · capital-05 · applications-03

Hypothesis 3 · Strained

Networking is the durable layer.

W36 strengthened networking's strategic role: AWS-Azure private multicloud interconnect entered preview, DE-CIX expanded dual-homed ExpressRoute Metro, Equinix announced intent-driven Fabric One, and Firmus committed approximately $300 million to 25 years of APX East capacity. These events show orchestration, route diversity, and long-duration capacity value. They do not satisfy the framework's economic test because none discloses cross-connect or fabric revenue growth relative to compute revenue.

Counter-evidence: Fabric One is pre-beta, the AWS-Azure preview currently reports up to 1 Gbps, and APX East targets Q4 2028. Meanwhile Broadcom's AI-semiconductor line combines networking with custom accelerators, preventing a clean margin comparison. The architectural evidence is favorable, but the week's public data cannot show networking pricing power holding longer than compute pricing power.

Sources networking-04 · networking-05 · networking-01

Hypothesis 4 · Untested

Open weights pull the floor up.

NVIDIA's Apache-2.0 low-precision Muse-Glimmer checkpoint reduces deployment footprint while preserving published capabilities. But it is an optimization of an existing W33 model, not a new open-weight frontier release, and W36's dominant application and agent launches were closed or metered. One optimized checkpoint cannot test whether open weights narrowed the closed-model capability gap or rerouted meaningful new enterprise compute demand this week.

Counter-evidence: Astra's closed launch produced the week's strongest benchmark movement, and Salesforce, GitHub, and Anthropic all reinforced governed hosted access. The open floor may still be rising over a multiweek horizon, especially after W35's model wave, but W36 adds no comparable fresh benchmark set and should not be scored as support from absence of refutation.

Sources software-04 · applications-01

Hypothesis 5 · Strained

Power is the binding constraint for the next 24 months.

Capital and policy converged on energization: Vertiv and Flex agreed to power-control and conversion acquisitions; Google contracted 396 MW of geothermal with delivery beginning in 2028; Texas staff retained $50,000 per MW security for large-load interconnection; PJM's comment window closed on a bring-capacity-or-curtail proposal; and the Department of Energy temporarily authorized customer standby generation during emergency conditions. This supports power as a major site-timing constraint, but does not establish it as the unique or dominant constraint across the industry.

Counter-evidence: The Nscale facilities fund GPUs, networking, storage, cooling, and site work, not power alone, while Cerebras's Finnish build and the HUMAIN production cluster show projects can still secure viable sites. The evidence supports power as a site-timing constraint, not the only industry constraint; accelerator supply, financing, permits, and customer readiness still determine how fast announced capacity becomes revenue.

Sources capital-05 · capital-02 · policy-02 · policy-03

Inductive · 3 weeks observed

Agent deployment is shifting from capability access to explicit containment, delegated authority, and evidence custody.

Next expectation: Before 2026-12-31, at least one major enterprise agent platform will publish an exportable authorization and evidence schema covering tool scope, approvals, retained traces, and revocation. Falsified if the next two major agent releases provide capability GA without new authority or audit controls.

• W34: Anthropic's versioned-skills interface, Salesforce's externally callable data and agent services, and UiPath Maestro converged on identity-inherited agent orchestration.
• W35: The METR/OpenAI incident postmortems made evaluation-network isolation, scorer integrity, and default-deny tool boundaries central deployment requirements.
• W36: OpenAI attached action monitoring to Astra, Anthropic placed monitoring logs under customer keys, and GitHub made AI approval a scoped and revocable branch-protection state.

Inductive · 4 weeks observed

Power constraints are migrating from site-selection assumptions into enforceable contracts, financial security, curtailment priority, and emergency dispatch.

Next expectation: By 2026-10-31, either FERC will act on PJM ER26-3515 or another US state or grid operator will publish a large-load rule that assigns measurable security, curtailment, or capacity obligations. Falsified if PJM's proposal is withdrawn and no comparable rule advances during that period.

• W30: Georgia Power's 3.2 GW Project Camellia contract paired service with up to 1 GW of curtailment during grid stress.
• W34: Pennsylvania Executive Order 2026-05 applied new grid-readiness requirements above 25 MW, and a long-duration data-center transaction tied value to power availability.
• W35: Georgia formalized ratepayer safeguards, PJM proposed curtailment-first treatment for large loads without qualifying capacity, and Virginia began data-center electricity-tax collection.
• W36: PJM's comment deadline passed, Texas staff retained $50,000 per MW security above 75 MW, and DOE temporarily authorized customer standby assets as last-resort grid resources.

Inductive · 3 weeks observed

AI infrastructure buyers are reserving future schedule position through guarantees, capacity contracts, and milestone-timed capital rather than waiting for operating supply.

Next expectation: Before year-end 2026, another AI infrastructure contract above $500 million will disclose a future service date plus an option, earnout, guarantee, or delayed-draw mechanism that allocates schedule risk. Falsified if the next three comparable commitments are fully funded against currently operating capacity without milestone conditions.

• W34: A 20-year land-and-power shell transaction used an NVIDIA residual-value guaranty to make future accelerator-linked infrastructure financeable.
• W35: AWS reserved 2 million additional NVIDIA GPUs for 2027-2028, buying supply-chain position rather than reporting installed capacity.
• W36: Google contracted four 99 MW Fervo tranches from Q3 2028, Firmus committed to 25-year APX East capacity targeted for Q4 2028, and Nscale matched delayed-draw facilities to future deployment milestones.

Q4 2026

ARC Prize's matched maximum-reasoning comparison measured Astra at 62.7% on the Standard harness and 98.6% on the Provider Adapter, a 35.9-point spread associated with runtime state and context compression.

Enterprise evaluations will have to version the model, harness, memory policy, tool surface, and retry budget as one tested system. Model-only scorecards will become procurement-incomplete, and portability reviews will ask which gains disappear when provider-private state is removed. ARC Prize, not this publication, established the core harness finding.

Who moves

Enterprise AI architecture teams, model-evaluation vendors, frontier API providers, procurement and model-risk functions

Two procurement cycles

GitHub allowed Copilot approvals to count toward protected-branch requirements while Anthropic placed long-window monitoring evidence in customer-controlled accounts.

Security and audit teams will define a new separation-of-duties rule: an AI may execute or approve, but the same provider-controlled evidence path cannot be the sole basis for both. Expect independent trace retention, human escalation thresholds, and policy-engine attestations in regulated software delivery.

Who moves

Software engineering leaders, internal audit, regulated-industry security teams, GitHub administrators, agent-platform vendors

2027 planning cycle

Vertiv and Flex committed up to $7.0 billion to grid-to-chip controls and power conversion while large-load rules attached security and curtailment obligations to interconnection.

Data-center design authority will move earlier toward firms that can model utility, onsite generation, storage, power conversion, and rack loads as one permitted system. Cooling and compute vendors without an upstream power-control partner will face acquisition, partnership, or specification risk before equipment selection.

Who moves

Data-center developers, electrical equipment vendors, utilities, engineering firms, hyperscalers, neoclouds, and infrastructure investors

Differ

GPT-6 Astra's 99.9% ARC-AGI-3 result is near-saturation evidence of broadly human-level agentic intelligence.

The 99.9% adapter result is best-observed and descriptive, but it uses a different reasoning effort from the matched Standard-harness maximum. Model Pulse carries the matched-effort Standard versus Provider Adapter scores. Astra plus its runtime nearly saturates a closed-ended benchmark; model-only saturation and general intelligence do not follow.

Sources OpenAI and ARC Prize, Sep 3, 2026

Open

Broadcom's 56.4% AI revenue share proves custom-silicon returns have already concentrated upstream.

The reported revenue is material, but it comes from a filed unaudited quarterly disclosure, combines custom accelerators and networking, and omits customer utilization, component margins, and downstream returns. It supports shipment monetization, not where durable return pools settle.

Sources Broadcom earnings release; RCR Tech

Differ

Vertiv and Flex spent $7.0B to buy one integrated grid-to-chip control plane.

The headline sum collapses different assets. Vertiv buys site architecture and microgrid orchestration with a material contingent share of maximum price (see house measurement); Flex buys conversion hardware and grid-forming controls at 5.5x forecast 2026 revenue. Both are pending acquisitions, not operating capacity.

Sources Vertiv and Flex announcements, Sep 2-3, 2026

Extend

Equinix Fabric One has turned the network into the autonomous control plane for distributed AI.

Intent-driven connectivity is the direction, but Fabric One is pre-beta. The control-plane product is not natural-language provisioning alone; it is policy simulation, scoped approval, retained intent history, rollback, and production service evidence that have not yet published.

Sources Equinix and SiliconANGLE, Sep 2-4, 2026

Frontier lab cash runway at current burn

The required inputs are unaudited and absent from this issue's artifact set. Astra pricing and Enterprise Frontier Safeguards cloud-storage costs do not establish a financing input.

Hyperscaler AI capex to disclosed AI revenue ratio

The methodology requires a range and a top-four input table. Neither is present in W36, and the Fervo power-purchase agreement has no disclosed consideration.

CoreWeave contracted revenue backlog

Backlog remains a filed stock value awaiting the next quarter. Nscale delayed-draw financing is a category signal, not a CoreWeave backlog input.

NVIDIA quarter-over-quarter data center revenue

No NVIDIA print landed this week. Broadcom's filed unaudited quarterly disclosure broadens supplier evidence but cannot be substituted into NVIDIA's series.

Open-weight to closed-model capability gap on coding

The quantized checkpoint improves deployment footprint, not the measured frontier gap. See Model Pulse for the benchmark and lineage read.

Sovereign AI program commitments

The method includes government-funded national compute programs only. Corporate acquisitions, power agreements, and deployments stay outside, but no constituent ledger is available here to support a point estimate.

PJM capacity auction clearing price

No auction occurred. No capped-versus-uncapped 2028/29 simulation was identified in the evidence reviewed for W36; the policy signal is a pending capacity-or-curtail rule.

Time from interconnection request to energization

W36 makes compliance costs more explicit but does not publish a comparable queue-duration update. At the 75 MW Texas threshold, proposed security equals $3.75M.

Cost per task, frontier reasoning model

Methodology v2 requires a named-basis floor and median. ARC Prize's suite totals are retained in Model Pulse but cannot move this lever against the prior Artificial Analysis basis.

Custom silicon share of hyperscaler AI compute

The revenue line confirms scale but cannot produce a custom-compute share because networking is included and deployed utilization is absent.

Software · 32% confidence

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

p100-astra-neutral-memory-dec15

Deadline

By December 15, 2026

Trigger

Public ARC-AGI-3 result with a reproducible visible-memory and context-compression configuration, no provider-private reasoning state, and Astra at 80.7% or higher.

Power · 41% confidence

Google accepts at least 500 MW of Fervo's conditional expansion and the parties execute a definitive agreement by June 30, 2027.

ID

p101-fervo-expansion-firm-jun30

Deadline

By June 30, 2027

Trigger

Fervo filing or Google announcement stating that at least 500 MW of the approximately 600 MW expansion option has become firm under a definitive agreement.

Capital · 68% confidence

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

p102-second-queue-contract-dec31

Deadline

By December 31, 2026

Trigger

Primary filing or announcement with transaction value above $500M, named service date, and explicit contingent or milestone mechanism.

Networking · 57% confidence

Equinix publishes Fabric One beta documentation that exposes approval, rollback, or auditable intent-history controls before December 31, 2026.

ID

p103-fabric-one-control-schema-dec31

Deadline

By December 31, 2026

Trigger

Public Fabric One beta documentation naming at least one of policy simulation, approval workflow, rollback, or exportable intent history.

Hardware · 24% confidence

Broadcom discloses separate quarterly revenue figures for custom AI accelerators and AI networking by December 31, 2026.

ID

p104-broadcom-ai-split-dec31

Deadline

By December 31, 2026

Trigger

Hit only if a Broadcom earnings release, 10-Q, or call transcript reports distinct dollar revenue for both custom AI accelerators and AI networking; a combined AI-semiconductor line is a miss.

Pending · Hardware

SemiAnalysis publishes AgentX v3 multi-turn benchmark results for OpenAI Jalapeño on production-representative agentic traces, with methodology comparable to Vera Rubin NVL72 AgentX runs cited by NVIDIA, by October 31, 2026.

Resolution window remains open; no comparable AgentX publication in the validated W36 evidence.

ID

p95-jalapeno-agentx-oct31

Confidence

36%

Deadline

By October 31, 2026

Trigger

SemiAnalysis newsletter or InferenceX page listing Jalapeño AgentX throughput-per-megawatt and cost-per-million-tokens on multi-turn traces, not solely 8k/1k InferenceX STP runs.

Pending · Software

The highest single ISO week of OpenRouter aggregate token volume in September 2026 exceeds the Ox Alpha stealth-week peak (week of August 20–26, 2026) by at least 15%, by September 30, 2026.

September is incomplete at the September 5 cutoff; a highest-week comparison cannot yet be scored.

ID

p96-openrouter-volume-sep30

Confidence

40%

Deadline

By September 30, 2026

Trigger

OpenRouter public stats page or Requesty/OpenRouter blog post reporting weekly tokens processed for each September 2026 ISO week against the Ox Alpha peak week — week versus week, same unit. The baseline peak week ran at zero list price and is community-reported (grade 3 on our 1–5 source scale), so the comparison inherits that grade.

Pending · Power

Commerce BIS publishes a Federal Register notice of proposed rulemaking on remote access to advanced US AI compute by Chinese end users, by November 30, 2026.

Resolution window remains open; no qualifying Federal Register notice appears in the validated W36 evidence.

ID

p97-bis-remote-gpu-nprm-nov30

Confidence

27%

Deadline

By November 30, 2026

Trigger

Federal Register NPRM from Commerce/BIS with docket number and comment period addressing remote GPU access via third-country data centers.

Pending · Hardware

NVIDIA Q3 FY2027 earnings disclosure states Vera Rubin contributed more than 25% of datacenter revenue for the quarter ended October 26, 2026.

The predicted quarter has not ended and the earnings disclosure has not occurred.

ID

p98-nvidia-rubin-mix-q3-earnings

Confidence

74%

Deadline

By NVIDIA Q3 FY2027 earnings release (expected November 2026)

Trigger

NVIDIA Form 10-Q or earnings call transcript for quarter ended October 26, 2026 stating Vera Rubin datacenter revenue mix above 25%.

Pending · Networking

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

The October summit and deadline remain ahead; no qualifying production target appears in W36 evidence.

ID

p99-metaroce-ocp-spec-oct31

Confidence

44%

Deadline

By October 31, 2026

Trigger

OCP Global Summit 2026 materials or Meta Engineering blog publishing MetaRoCE spec with named hyperscaler or cloud deployment timeline distinct from the August 64-node lab cluster.

Cumulative record

Predictions made

104

Resolved

57

Outcomes

23 hit · 15 partial · 19 miss

Hit rate (partial = half)

54%

Brier score (0 = perfect)

0.200

Overdue, unresolved

0

Hit · 72% called

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.

Deadline

By October 31, 2026

Partial · 80% called

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.

Deadline

By October 31, 2026

Hit · 43% called

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.

Deadline

By September 15, 2026

Hit · 66% called

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.

Deadline

By August 31, 2026

Sources Resolution evidence

Hit · 84% called

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.

Deadline

By August 31, 2026

Sources Resolution evidence

Hit · 64% called

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.

Deadline

By August 31, 2026

Sources Resolution evidence

Sep 8

DOE emergency-order expiry

Order 202-26-43 expires after temporarily authorizing customer standby generation as a last-resort grid resource.

Sep 9

GLM-5.3-Flash promotion ends

Reprice W35 open-weight agent pilots on steady-state list rates rather than the launch discount.

Sep 11

Texas large-load rule vote

Tests whether the 75 MW threshold and $50,000/MW security move from staff recommendation into final policy.

Later in 2026

Fabric One beta documentation

Look for policy simulation, approval, rollback, service levels, and exportable intent evidence before production use.

Q4 2026

Vertiv-UIG and Flex-EPC closings

Regulatory clearance and disclosed integration terms will test whether grid-to-chip consolidation converts from agreement to operating capability.

• W36 authored from validated research, counterbrief, house measurement, and synthesis artifacts through the September 5 cutoff.
• Three model-tree rows added: GPT-6 Astra, Claude Fable 5.1, and Gemini 3.8 Flash.
• All five W35 predictions remain pending because their resolution windows have not matured.