--- title: The frontier became a sovereign and capital-markets object, not just a benchmark race. publication: The AI Stack Weekly slug: 2026-W24 issueNumber: 8 isoYear: 2026 isoWeek: 24 publishedAt: '2026-06-13' canonicalUrl: https://brianletort.ai/industry/weekly/2026-W24 pdfUrl: https://brianletort.ai/downloads/ai-stack-weekly-2026-W24.pdf schemaVersion: 2026.05.02 flywheelArc: all-three capitalFlow: - category: Frontier Labs capitalIn: ~$95B capitalInPrior: ~$90B capitalInDirection: up revenueOut: ~$21B revenueOutPrior: ~$20B revenueOutDirection: up burnToRevenue: ~1.3x - category: Hyperscaler-Hosted capitalIn: ~$185B capitalInPrior: ~$181B capitalInDirection: up revenueOut: ~$62B revenueOutPrior: ~$60B revenueOutDirection: up burnToRevenue: ~0.3x - category: Neoclouds capitalIn: ~$12B capitalInPrior: ~$12B capitalInDirection: flat revenueOut: ~$5B revenueOutPrior: ~$5B revenueOutDirection: flat burnToRevenue: ~3x - category: On-Prem / Hybrid capitalIn: ~$93B capitalInPrior: ~$91B capitalInDirection: up revenueOut: ~$36B revenueOutPrior: ~$35B revenueOutDirection: up burnToRevenue: ~2x levers: - metric: Frontier lab cash position (avg months runway, top 3) current: ~34-37 mo (xAI now public via SpaceX IPO) prior: ~33-36 mo direction: up threshold: <18 mo triggers re-rating risk - metric: Hyperscaler capex / AI revenue ratio (top 4 weighted) current: ~5.0-5.3 (Oracle FY27 ~$70B guide; top-4 flat) prior: ~5.0-5.2 direction: up threshold: '>6.0 invites investor pushback at next earnings' - metric: CoreWeave revenue backlog current: $99.4B (next print Q2, early Aug) prior: $99.4B direction: flat threshold: Conversion velocity matters more than gross figure - metric: NVIDIA Q-over-Q data center revenue current: $75.2B (Q1 FY27); Q2 guide $91B, reports Aug 26 prior: $75.2B (Q1 FY27); Rubin production ramp confirmed direction: flat threshold: Q2 FY27 guide $91B implies further +21% QoQ - metric: Open vs closed gap on SWE-Bench Pro (coding) current: Closed lead widened on paper (Fable 5 80.3% vendor scaffold) vs open ~58.6%; contested prior: Closed +~19pp (no new Pro challenger yet) direction: up threshold: Sustained open lead reshapes enterprise procurement - metric: Sovereign AI commitments (count / aggregate $) current: ~14 / ~$180B+ (NAVER+NVIDIA 1GW added) prior: ~13 / ~$160B+; power-first gating rising direction: up threshold: null - metric: PJM 2026/27 capacity auction price ($/MW-day) current: $329.17 (2028/29 BRA clears Jun 30-Jul 7) prior: $329.17 direction: flat threshold: 11x in 24 months — power is the new binding constraint - metric: Time-to-power, busiest US markets (months) current: 60-84 (new PJM); 2027/28 auction cleared short of target prior: 60-84 (new PJM); power-first campuses rising direction: flat threshold: null - metric: Cost-per-task, frontier reasoning model current: ~$0.10-$0.15 (effective); Fable 5 sets a 2x-Opus top tier ($10/$50 per MTok) prior: ~$0.10-$0.15 (effective; unchanged) direction: flat threshold: null - metric: Custom silicon share of incremental AI compute current: ~33-36%; d-Matrix Corsair inference ASIC in volume production prior: ~33-36%; Broadcom AI revenue +143% YoY direction: up threshold: '>35% materially compresses merchant GPU pricing' predictions: - id: p37-gemini-3-5-pro-below-fable5 lens: software confidencePct: 58 deadline: By June 30, 2026 text: >- Gemini 3.5 Pro reaches public GA by June 30, 2026 but debuts below Claude Fable 5 on the Artificial Analysis Intelligence Index in its first independent pass. - id: p38-fable5-access-restored lens: software confidencePct: 60 deadline: By July 31, 2026 text: >- Anthropic restores public (non-government) access to a Fable-5-class model — with geo-gating or KYC controls — by July 31, 2026. - id: p39-second-inference-asic-milestone lens: hardware confidencePct: 65 deadline: By August 31, 2026 text: >- At least one more merchant inference-ASIC vendor (Groq, Cerebras, Tenstorrent, or d-Matrix) announces a volume-production or named-hyperscaler milestone before August 31, 2026. - id: p40-second-1-6t-design-win lens: networking confidencePct: 60 deadline: By August 31, 2026 text: >- A second of (Marvell, Credo) cites a 1.6T or co-packaged-optics production design win by August 31, 2026, joining Broadcom's Tomahawk 6. - id: p41-no-public-s1 lens: capital confidencePct: 62 deadline: By August 31, 2026 text: >- Neither OpenAI nor Anthropic converts its confidential draft to a publicly visible S-1 on SEC EDGAR before August 31, 2026. - id: p42-hyperscaler-power-first lens: power confidencePct: 60 deadline: By September 30, 2026 text: >- A hyperscaler announces a >500MW power-first AI campus or behind-the-meter generation deal by September 30, 2026. predictionsPrior: - id: p32-gemini-3-5-pro-ga lens: software outcome: pending deadline: By June 30, 2026 text: >- Gemini 3.5 Pro reaches public GA by June 30, 2026, but does not exceed Claude Opus 4.8 on SWE-Bench Pro in its first independent Artificial Analysis run. - id: p33-vera-rubin-first-shipments lens: hardware outcome: pending deadline: By September 30, 2026 text: >- At least one major OEM announces customer shipment or formal order availability for Vera Rubin NVL72-class systems before September 30, 2026. - id: p34-hbm4-allocation-tightness lens: hardware outcome: pending deadline: By August 31, 2026 text: >- Before August 31, 2026, at least one memory supplier or supply-chain analyst reports HBM4 allocation tightness despite three-supplier qualification. - id: p35-cpo-design-win lens: networking outcome: partial deadline: By August 31, 2026 text: >- Broadcom, Marvell, or NVIDIA announces a new CPO/1.6T production design win or revenue guide uplift tied to AI networking before August 31, 2026. - id: p36-power-first-followthrough lens: power outcome: pending deadline: By September 30, 2026 text: >- A hyperscaler announces another >500MW power-first AI campus or behind-the-meter generation deal by September 30, 2026. signalScores: - 5 - 4 - 5 - 2 - 1 keyTakeaways: - >- A U.S. government export-control directive forced Anthropic to disable Claude Fable 5 and Mythos 5 for every customer within three days of a #1 debut — the first known government-forced takedown of a deployed frontier model, converting 'which model is best' into 'which model can be switched off by a third party.' - >- The frontier-lab capital story went public: SpaceX, with xAI embedded, priced the largest IPO ever (~$75B raised) and closed near a $2.1T valuation, while OpenAI's confidential S-1 kept the next two labs pre-public. - >- Demand concentrated visibly: Oracle booked $638B of remaining performance obligations (+363% YoY) and TSMC printed a record month with CoWoS sold out through 2027 — but Oracle's roughly -$24B FY26 free cash flow sharpened the can-it-be-funded debate. - >- The inference layer began diversifying: d-Matrix moved its Corsair ASIC into volume production explicitly architecting around the HBM chokepoint, and 1.6T Ethernet fabric started shipping in production (Arista on Broadcom's Tomahawk 6). - >- Signal-vs-noise: Fable 5's 80.3% SWE-Bench Pro figure is a vendor-harness result — noise until replicated; GPT-5.6 imminence is rumor with no primary artifact. - >- Watch whether and how Anthropic restores Fable-class access — the template for sovereign access risk — and PJM's 2028/29 capacity auction (Jun 30-Jul 7), the quarter's most important power-cost signal. byTheNumbers: - value: '64.9' label: 'Claude Fable 5''s #1 debut on the AA Intelligence Index' - value: ~$2.1T label: SpaceX close after the largest IPO ever, with xAI embedded - value: $638B label: Oracle remaining performance obligations, +363% YoY - value: $3.25B label: CoreWeave senior notes priced at 8.5-9.625% - value: $10/$50 label: Fable 5 per MTok — a new top tier at ~2x Opus 4.8 - value: 5.6GW label: Gas-paired data-center pipeline Stark Power moved to acquire --- # The frontier became a sovereign and capital-markets object, not just a benchmark race. *Issue 08 · Week 24 of 2026 · Published 2026-06-13* ## Executive summary - A U.S. government export-control directive forced Anthropic to disable Claude Fable 5 and Mythos 5 for every customer within three days of a #1 debut — the first known government-forced takedown of a deployed frontier model, converting 'which model is best' into 'which model can be switched off by a third party.' - The frontier-lab capital story went public: SpaceX, with xAI embedded, priced the largest IPO ever (~$75B raised) and closed near a $2.1T valuation, while OpenAI's confidential S-1 kept the next two labs pre-public. - Demand concentrated visibly: Oracle booked $638B of remaining performance obligations (+363% YoY) and TSMC printed a record month with CoWoS sold out through 2027 — but Oracle's roughly -$24B FY26 free cash flow sharpened the can-it-be-funded debate. - The inference layer began diversifying: d-Matrix moved its Corsair ASIC into volume production explicitly architecting around the HBM chokepoint, and 1.6T Ethernet fabric started shipping in production (Arista on Broadcom's Tomahawk 6). - Signal-vs-noise: Fable 5's 80.3% SWE-Bench Pro figure is a vendor-harness result — noise until replicated; GPT-5.6 imminence is rumor with no primary artifact. - Watch whether and how Anthropic restores Fable-class access — the template for sovereign access risk — and PJM's 2028/29 capacity auction (Jun 30-Jul 7), the quarter's most important power-cost signal. **By the numbers.** - **64.9** — Claude Fable 5's #1 debut on the AA Intelligence Index (Disabled by a U.S. government directive three days later) - **~$2.1T** — SpaceX close after the largest IPO ever, with xAI embedded (~$75B raised; the first traded read on a frontier lab) - **$638B** — Oracle remaining performance obligations, +363% YoY (Only ~12% of RPO converts within 12 months) - **$3.25B** — CoreWeave senior notes priced at 8.5-9.625% (vs ~5.9% on its investment-grade, GPU-backed facilities) - **$10/$50** — Fable 5 per MTok — a new top tier at ~2x Opus 4.8 (Priced as an async heavy-lift tool, not a default router target) - **5.6GW** — Gas-paired data-center pipeline Stark Power moved to acquire (Firm power and land now gate strategic compute) ## Big Story W24 was the week the model layer stopped behaving like a product cycle and started behaving like critical infrastructure with a kill switch. Anthropic shipped Claude Fable 5, a new "Mythos-class" tier above Opus, and it debuted #1 on the independent Artificial Analysis Intelligence Index at 64.9 — then, three days later (Jun 12), a U.S. government export-control directive forced Anthropic to disable Fable 5 and Mythos 5 for every customer because it could not filter foreign-national access in real time; queries now fall back to Opus 4.8. That is the first known government-forced takedown of a deployed frontier model, and it converts "which model is best" into "which model can be switched off by a third party." The same week, the capital story went public: SpaceX, with xAI embedded, priced the largest IPO ever (~$75B raised) and closed near a $2.1T valuation, giving the market its first traded read on a frontier lab, while OpenAI's confidential S-1 (Jun 8) kept the next two labs at the pre-public stage. Underneath, demand concentrated visibly — Oracle booked an RPO of $638B (+363% YoY), TSMC printed a record month with CoWoS sold out through 2027, and d-Matrix put a merchant inference ASIC into volume production explicitly to route around the HBM chokepoint. Net/net: boards should treat top-tier closed models as single points of failure and mandate hard fallbacks; investors now have a public frontier-lab proxy to price against; architects should design multi-model routing and second-source compute; operators should assume the binding constraints are regulatory access, power, and packaging, not raw intelligence. Flywheel arc: `all-three`. ## Software lens - **Jun 9.** Anthropic shipped Claude Fable 5 (and Mythos 5), a new Mythos-class tier above Opus, GA across the Claude API, Bedrock, Vertex AI, and Microsoft Foundry; it debuted #1 on the Artificial Analysis Intelligence Index at 64.9 _([Anthropic launch; Artificial Analysis](https://www.anthropic.com/news/claude-fable-5-mythos-5))_ - **Jun 12.** A U.S. government export-control directive forced Anthropic to disable Claude Fable 5 and Mythos 5 globally for all customers; unable to filter foreign-national access in real time, Anthropic routed queries back to Opus 4.8 (still suspended at week's end) _([VentureBeat, MarkTechPost (Reuters/BBC/Axios cited secondhand)](https://venturebeat.com/technology/anthropic-blocks-all-public-access-to-claude-fable-5-mythos-5-following-us-government-order-what-enterprises-should-do))_ - **Jun 10.** Google DeepMind released DiffusionGemma, an Apache-2.0 open text-diffusion model (26B / ~3.8B active MoE, multimodal-in, 256K context) generating ~1,000+ tok/s on an H100 with native vLLM support _([Google blog; Hugging Face model card](https://blog.google/innovation-and-ai/technology/developers-tools/diffusion-gemma-faster-text-generation/))_ - **Jun 13.** Gemini 3.5 Pro remained not-GA at W24 close: still a limited Vertex enterprise preview with no public API identifier, price row, or independent benchmark, despite a Google-stated June target _([TechTimes; Presenc launch brief](https://www.techtimes.com/articles/317919/20260606/google-gemini-35-pro-nears-june-launch-2-million-token-context-deep-think-reasoning.htm))_ **What this means.** The model layer's news was binary: a new closed leader arrived and then was switched off by a government within 72 hours, while the durable open progress was an efficiency path (text diffusion) rather than another frontier crown. Architects should standardize an evaluation harness and a fallback router now so a top-tier model can be slotted — or revoked — without freezing deployments; see The Model Pulse for the full architecture read. ## Hardware lens - **Jun 10.** Oracle reported Q4 FY26 with remaining performance obligations of $638B (+363% YoY, +$85B QoQ), OCI/IaaS +93% to $5.8B, ~1GW of new AI capacity targeted in Q1 FY27, and a ~$70B FY27 net capex guide _([Oracle Investor Relations; CNBC](https://investor.oracle.com/investor-news/news-details/2026/Oracle-Announces-Record-Q4-and-FY-2026-Results-Driven-by-Cloud-Infrastructure--Cloud-Applications/default.aspx))_ - **Jun 10.** TSMC reported record May revenue of NT$416.98B (+30.1% YoY), confirming sustained AI/HPC pull-through at the foundry chokepoint, with CoWoS advanced packaging sold out through 2027 _([TSMC press release; SEC Form 6-K](https://pr.tsmc.com/english/news/3320))_ - **Jun 9.** d-Matrix moved its Corsair inference accelerator into volume production — a digital in-memory-compute ASIC (TSMC N6) that the company says runs inference ~10x faster and ~3x cheaper than a standalone GPU while explicitly architecting around the HBM/DRAM chokepoint _([CNBC; d-Matrix](https://www.cnbc.com/2026/06/09/nvidia-d-matrix-chip-production-microsoft.html))_ **What this means.** Hardware's signal this week was demand and supply, not silicon launches: Oracle proves AI compute is being booked as long-duration backlog, TSMC proves it is foundry- and packaging-gated, and d-Matrix proves the inference layer is diversifying away from merchant GPUs and HBM. Investors with concentrated GPU exposure should track contracted-backlog conversion and ASIC ramps; operators should secure packaging-bound capacity early. ## Networking lens - **Jun 9.** Arista launched its first 1.6T rack-scale switch portfolio (7060XE7) built on Broadcom's Tomahawk 6 — a fresh Broadcom 1.6T production design win for AI back-end fabric _([Arista; Broadcom](https://www.arista.com/en/products/7060x-series))_ - **Jun 8.** Broadcom disclosed a scale-up Ethernet design win with inference-ASIC startup FuriosaAI, extending AI-Ethernet beyond hyperscaler back ends into merchant accelerator racks _([Broadcom; FuriosaAI](https://www.broadcom.com/products/ethernet-connectivity))_ - **Jun 9.** Marvell's pending S&P 500 inclusion (replacing Campbell Soup, ~late June) marked capital-markets validation of the custom-silicon and AI-networking designer thesis _([The Next Platform](https://www.nextplatform.com/compute/2026/06/09/ai-chip-shepherds-broadcom-and-marvell-have-skinned-the-golden-fleece/5253011))_ **What this means.** 1.6T fabric is now shipping in production, but the in-window proof points all ran through Broadcom — the marquee CPO/optics catalysts from Marvell, Credo, Coherent, and Lumentum fell just outside the week. Network architects should treat 1.6T Ethernet as deployable today while pressing for a second silicon source before committing a fabric roadmap to one vendor. ## Capital flow | Category | Capital in | Revenue out | Burn:Revenue | Movement | |---|---|---|---|---| | Frontier Labs (OpenAI, Anthropic, Google DeepMind, xAI) | ~$95B (was ~$90B, up) | ~$21B (was ~$20B, up) | ~1.3x | The frontier-lab capital story went public: SpaceX, with xAI embedded, priced the largest IPO ever and closed near a $2.1T valuation, the first traded read on a frontier lab. | | Hyperscaler-Hosted (Azure-OpenAI, AWS-Anthropic, Google Cloud-Gemini, Oracle-OCI) | ~$185B (was ~$181B, up) | ~$62B (was ~$60B, up) | ~0.3x | Oracle's $638B RPO and ~$70B FY27 capex guide made hosted AI demand visible as contracted backlog, but negative free cash flow sharpened the can-it-be-funded debate. | | Neoclouds (CoreWeave, Nscale, Crusoe, Lambda, Fluidstack, IREN) | ~$12B (was ~$12B, flat) | ~$5B (was ~$5B, flat) | ~3x | CoreWeave priced $3.25B of senior notes at 8.5-9.625%, putting a hard number on the neocloud cost of unsecured capital. | | On-Prem / Hybrid (Enterprise GPU clusters, sovereign and national programs, Cisco / Dell / HPE) | ~$93B (was ~$91B, up) | ~$36B (was ~$35B, up) | ~2x | Sovereign and power-first capacity advanced: NAVER and NVIDIA committed to a gigawatt-scale Korean AI factory, and Stark Power moved to acquire a 5.6GW gas-paired data-center pipeline. | ### Frontier Labs — detail W24 turned private frontier-lab marks into a public-market signal. SpaceX (xAI merged in February) priced ~$75B and popped ~19% to ~$2.1T on its Nasdaq debut, embedding xAI's value in a traded security, while OpenAI's confidential S-1 (Jun 8) and Anthropic's confidential DRS keep the next two labs pre-public. Investors finally have a frontier-lab proxy to price against; the first audited public S-1 remains the regime change to watch. **Transactions:** - **Jun 11-12.** SpaceX (with xAI embedded) prices largest-ever IPO; closes ~$2.1T (Nasdaq: SPCX) — ~$75B raised _([CNBC; SEC filing](https://www.cnbc.com/2026/06/12/spacex-ipo-spcx-live-updates.html))_ - **Jun 8.** OpenAI submits confidential draft S-1 (not publicly visible on EDGAR) — Undisclosed _([OpenAI](https://openai.com/index/openai-submits-confidential-s-1/))_ ### Hyperscaler-Hosted — detail Oracle's Q4 print is the clearest in-window evidence that AI infrastructure demand is being booked as long-duration backlog (RPO $638B, +363% YoY), with ~1GW more capacity targeted in Q1 FY27. But FY26 free cash flow was about -$24B and the shares fell on a planned debt/equity raise, so the hyperscaler-hosted row is now a financing-risk story as much as a demand story. Boards should read contracted backlog as real but watch conversion velocity (only ~12% of RPO converts within 12 months). **Transactions:** - **Jun 10.** Oracle Q4 FY26: RPO $638B (+363% YoY); OCI/IaaS +93%; ~$70B FY27 net capex guide — $638B RPO _([Oracle Investor Relations; CNBC](https://investor.oracle.com/investor-news/news-details/2026/Oracle-Announces-Record-Q4-and-FY-2026-Results-Driven-by-Cloud-Infrastructure--Cloud-Applications/default.aspx))_ ### Neoclouds — detail The neocloud row held flat on capacity but repriced on funding cost: CoreWeave's $3.25B senior notes ($1.25B at 9.625% plus EUR 2B at 8.5%, due 2032) carry roughly 8.5-9.6% coupons versus ~5.9% on its investment-grade, GPU-backed facilities. That spread quantifies how the market prices balance-sheet risk against contracted-asset risk. Operators evaluating neocloud counterparties should weigh cost of capital alongside backlog. **Transactions:** - **Jun 11.** CoreWeave prices $1.25B + EUR 2B senior notes (8.5-9.625%, due Jul 2032) — $3.25B _([CoreWeave Investor Relations](https://investors.coreweave.com/news/news-details/2026/CoreWeave-Announces-Pricing-of-1-25-Billion-of-Senior-Notes-and-2-Billion-of-Senior-Notes/default.aspx))_ ### On-Prem / Hybrid — detail The sovereign and power-first build-out kept compounding. NAVER and NVIDIA committed to a gigawatt-scale Korean sovereign AI factory (55MW in H1 2027 scaling toward 1GW; ~$18B revenue goal by 2030), and Stark Power agreed to acquire Sagebrush's 5.6GW co-located, gas-paired pipeline. Neither is a hyperscaler behind-the-meter deal, but both confirm that firm power and land now gate strategic compute. Sovereign and on-prem buyers should assemble power and permitting before announcing GPU counts. **Transactions:** - **Jun 7-8.** NAVER + NVIDIA gigawatt-scale sovereign AI factory (DSX; 55MW -> 1GW roadmap) — ~$18B revenue goal by 2030 _([NVIDIA Newsroom; NAVER](https://nvidianews.nvidia.com/news/naver-ai-infrastructure))_ - **Jun 8.** Stark Power to acquire Sagebrush Infrastructure (5.6GW gas-paired DC pipeline) — 5.6GW pipeline _([PR Newswire](https://finance.yahoo.com/sectors/energy/articles/stark-power-acquire-sagebrush-infrastructure-123000856.html))_ ## Signal vs noise - **Score 5/5 —** Anthropic shipped Claude Fable 5, a new Mythos-class tier, and it debuted #1 on the independent Artificial Analysis Intelligence Index at 64.9. - _Sources:_ Anthropic launch announcement; Artificial Analysis (independent pre-release access); arena.ai/LMArena board adds. Caveat: vendor pricing, independent index. - _Read:_ SIGNAL. A genuine new top of the closed frontier — but priced as an async heavy-lift tool ($10/$50 per MTok, ~2x Opus 4.8). Architects should route Fable 5 to high-value reasoning, not default every call to it. - **Score 4/5 —** A U.S. government export-control directive forced Anthropic to disable Fable 5 and Mythos 5 for all customers within three days of launch. - _Sources:_ VentureBeat and MarkTechPost reporting; Reuters/BBC/Time/Axios cited secondhand. Caveat: the directive letter itself is not public. - _Read:_ SIGNAL. The first known government-forced takedown of a deployed frontier model. This is a concrete sovereign/regulatory single point of failure — boards should mandate hard model fallbacks and avoid single-sourcing the top tier. - **Score 5/5 —** SpaceX, with xAI embedded, priced the largest IPO ever and closed near a $2.1T valuation, putting a frontier lab inside a public security. - _Sources:_ SEC registration; CNBC live IPO coverage. Multi-source, audited filing. - _Read:_ SIGNAL. Investors now have a public-market frontier-lab proxy (SPCX) that OpenAI and Anthropic roadshows will be measured against. The private-mark-only era for frontier labs is ending. - **Score 2/5 —** Claude Fable 5 is the definitive coding leader at 80.3% on SWE-Bench Pro. - _Sources:_ Vendor-scaffold figure; independent vals.ai shows 95% on SWE-Bench Verified, but Scale SEAL has no Fable entry and an Epoch replication is pending. - _Read:_ NOISE until replicated. The AA Index lead is real and independent; the headline SWE-Bench Pro number is a vendor-harness result. Do not anchor a migration decision on it before third-party replication. - **Score 1/5 —** GPT-5.6 is imminent and will reset the frontier this month. - _Sources:_ Backend codenames and prediction-market odds only; no system card, model page, or API contract exists. - _Read:_ NOISE. A rumor with no primary artifact. Keep current GPT-5.5 baselines and do not delay procurement on an unreleased SKU. ## Levers | Metric | Current | Prior | Direction | Threshold | |---|---|---|---|---| | Frontier lab cash position (avg months runway, top 3) | ~34-37 mo (xAI now public via SpaceX IPO) | ~33-36 mo | up | <18 mo triggers re-rating risk | | Hyperscaler capex / AI revenue ratio (top 4 weighted) | ~5.0-5.3 (Oracle FY27 ~$70B guide; top-4 flat) | ~5.0-5.2 | up | >6.0 invites investor pushback at next earnings | | CoreWeave revenue backlog | $99.4B (next print Q2, early Aug) | $99.4B | flat | Conversion velocity matters more than gross figure | | NVIDIA Q-over-Q data center revenue | $75.2B (Q1 FY27); Q2 guide $91B, reports Aug 26 | $75.2B (Q1 FY27); Rubin production ramp confirmed | flat | Q2 FY27 guide $91B implies further +21% QoQ | | Open vs closed gap on SWE-Bench Pro (coding) | Closed lead widened on paper (Fable 5 80.3% vendor scaffold) vs open ~58.6%; contested | Closed +~19pp (no new Pro challenger yet) | up | Sustained open lead reshapes enterprise procurement | | Sovereign AI commitments (count / aggregate $) | ~14 / ~$180B+ (NAVER+NVIDIA 1GW added) | ~13 / ~$160B+; power-first gating rising | up | — | | PJM 2026/27 capacity auction price ($/MW-day) | $329.17 (2028/29 BRA clears Jun 30-Jul 7) | $329.17 | flat | 11x in 24 months — power is the new binding constraint | | Time-to-power, busiest US markets (months) | 60-84 (new PJM); 2027/28 auction cleared short of target | 60-84 (new PJM); power-first campuses rising | flat | — | | Cost-per-task, frontier reasoning model | ~$0.10-$0.15 (effective); Fable 5 sets a 2x-Opus top tier ($10/$50 per MTok) | ~$0.10-$0.15 (effective; unchanged) | flat | — | | Custom silicon share of incremental AI compute | ~33-36%; d-Matrix Corsair inference ASIC in volume production | ~33-36%; Broadcom AI revenue +143% YoY | up | >35% materially compresses merchant GPU pricing | **Lever detail:** - **Frontier lab cash position (avg months runway, top 3).** Top 3 frontier labs (OpenAI, Anthropic, Google DeepMind) by disclosed runway, with xAI now a public-market entity inside SpaceX (SPCX). The SpaceX/xAI IPO and OpenAI's confidential S-1 extend the category's access to capital. Boards should not assume frontier-lab funding pressure as a forcing function for short-term commercial concessions — capital access just widened. - **Hyperscaler capex / AI revenue ratio (top 4 weighted).** Top 4 hyperscalers (MSFT, GOOG, META, AMZN) weighted aggregate of total capex divided by AI-attributable revenue. No new top-4 prints in-window, but Oracle's ~$70B FY27 net capex guide and -$24B FY26 free cash flow push the broader hosted-capex ratio up and re-open the funding-vs-demand debate. Investors should keep the bubble hypothesis on funding/conversion, not demand. - **CoreWeave revenue backlog.** Booked but unrecognized revenue. The $99.4B audited figure (Mar 31) is unchanged; the in-window event was financing, not backlog — CoreWeave priced $3.25B of senior notes at 8.5-9.625%. Operators should watch the Q2 backlog print in early August and the cost-of-capital spread versus GPU-backed facilities. - **NVIDIA Q-over-Q data center revenue.** Q1 FY27 Data Center revenue of $75.2B was reported May 20; the Q2 guide is $91B and the next earnings print is Aug 26. No within-window change. Adjacent demand (Oracle RPO $638B, TSMC record month, CoWoS sold out through 2027) supports the guide; packaging and power remain the binding constraints, not demand. - **Open vs closed gap on SWE-Bench Pro (coding).** Claude Fable 5's vendor-scaffold SWE-Bench Pro figure (80.3%) widens the headline closed-over-open gap, but it is not yet replicated by Scale SEAL or Epoch; independent vals.ai shows 95% on SWE-Bench Verified. The strongest open item, MiniMax-M3 (AA Index 55), has not released weights. Architects should treat the widened gap as provisional and keep open self-host options in pilot. - **Sovereign AI commitments (count / aggregate $).** Analyst-curated count of sovereign/national AI-compute commitments. NAVER + NVIDIA's gigawatt-scale Korean AI factory (roadmap to 1GW; ~$18B revenue goal by 2030) was added in-window. Operators with APAC workloads should treat Korea as a credible sovereign landing zone while pricing in multi-year build timelines. - **PJM 2026/27 capacity auction price ($/MW-day).** The 2026/27 BRA cleared at the FERC cap ($329.17) and took effect June 1, 2026; the 2027/28 BRA cleared at $333.44. The next signal is the 2028/29 BRA, scheduled Jun 30-Jul 7, 2026, with a FERC collar capping it near $325. Architects should not assume near-term price relief; budget capacity at-cap through 2028. - **Time-to-power, busiest US markets (months).** Months from new-load interconnection request to energization. PJM's 2027/28 base auction cleared ~6,600MW short of its reliability requirement — the first RTO-wide shortfall, with a large share driven by data-center load. Architects should pre-commit power and long-lead grid equipment before pre-committing GPU SKUs. - **Cost-per-task, frontier reasoning model.** Median cost across frontier-tier reasoning models for a benchmark complex task. Claude Fable 5 lists at $10 in / $50 out per MTok (~2x Opus 4.8), pushing the ceiling up, but effective cost-per-task depends on how much work is routed to the top tier versus cheaper models. Operators running agents at scale should re-benchmark on cost-per-task, not list price. - **Custom silicon share of incremental AI compute.** d-Matrix's Corsair inference ASIC entered volume production in-window — a concrete merchant non-GPU data point, not a projection — and Marvell's pending S&P 500 inclusion validates the custom-silicon designer thesis. Investors with concentrated NVIDIA exposure should diversify into ASIC co-design (Broadcom, Marvell) and inference-specialist silicon, and into advanced packaging / power. ## Predictions - **`p37-gemini-3-5-pro-below-fable5` _[software]_ — Gemini 3.5 Pro reaches public GA by June 30, 2026 but debuts below Claude Fable 5 on the Artificial Analysis Intelligence Index in its first independent pass.** - Confidence: 58%. Deadline: By June 30, 2026. - Trigger: Google AI Studio / Vertex GA announcement plus Artificial Analysis leaderboard update. - **`p38-fable5-access-restored` _[software]_ — Anthropic restores public (non-government) access to a Fable-5-class model — with geo-gating or KYC controls — by July 31, 2026.** - Confidence: 60%. Deadline: By July 31, 2026. - Trigger: Anthropic status page, model documentation, or press confirming restored availability. - **`p39-second-inference-asic-milestone` _[hardware]_ — At least one more merchant inference-ASIC vendor (Groq, Cerebras, Tenstorrent, or d-Matrix) announces a volume-production or named-hyperscaler milestone before August 31, 2026.** - Confidence: 65%. Deadline: By August 31, 2026. - Trigger: Vendor press release, earnings disclosure, or customer design-win announcement. - **`p40-second-1-6t-design-win` _[networking]_ — A second of (Marvell, Credo) cites a 1.6T or co-packaged-optics production design win by August 31, 2026, joining Broadcom's Tomahawk 6.** - Confidence: 60%. Deadline: By August 31, 2026. - Trigger: Earnings call, product release, or customer design-win disclosure. - **`p41-no-public-s1` _[capital]_ — Neither OpenAI nor Anthropic converts its confidential draft to a publicly visible S-1 on SEC EDGAR before August 31, 2026.** - Confidence: 62%. Deadline: By August 31, 2026. - Trigger: SEC EDGAR public filings; Reuters / Bloomberg / The Information confirmation of public S-1 vs confidential DRS. - **`p42-hyperscaler-power-first` _[power]_ — A hyperscaler announces a >500MW power-first AI campus or behind-the-meter generation deal by September 30, 2026.** - Confidence: 60%. Deadline: By September 30, 2026. - Trigger: Hyperscaler energy/data-center announcement; utility or developer disclosure. ### Prior predictions scored - `p32-gemini-3-5-pro-ga` _[software]_ — **PENDING** — Gemini 3.5 Pro reaches public GA by June 30, 2026, but does not exceed Claude Opus 4.8 on SWE-Bench Pro in its first independent Artificial Analysis run. — As of W24 close (Jun 13), Gemini 3.5 Pro is still not GA — limited Vertex preview, no public benchmark. Deadline June 30 not reached; the GA leg is now at risk and the bar has moved (Fable 5, not Opus 4.8, is the new top). - `p33-vera-rubin-first-shipments` _[hardware]_ — **PENDING** — At least one major OEM announces customer shipment or formal order availability for Vera Rubin NVL72-class systems before September 30, 2026. — Tracking ahead: the first operational Vera Rubin rack (Dell PowerEdge XE9812 -> CoreWeave) shipped Jun 1, with standalone Vera CPU order availability opening ~August and full systems at fall GA. Deadline Sep 30 not reached. - `p34-hbm4-allocation-tightness` _[hardware]_ — **PENDING** — Before August 31, 2026, at least one memory supplier or supply-chain analyst reports HBM4 allocation tightness despite three-supplier qualification. — No new in-window allocation print, but Samsung's reported Gwangju advanced-packaging review and pending 2027 HBM4 price hikes reinforce the thesis. Deadline Aug 31 open. - `p35-cpo-design-win` _[networking]_ — **PARTIAL** — Broadcom, Marvell, or NVIDIA announces a new CPO/1.6T production design win or revenue guide uplift tied to AI networking before August 31, 2026. — Arista's 1.6T 7060XE7 portfolio on Broadcom's Tomahawk 6 (Jun 9) is a fresh Broadcom 1.6T production design win, satisfying the 1.6T leg; no co-packaged-optics production win or vendor revenue-guide uplift yet. Tracking to a full hit by deadline. - `p36-power-first-followthrough` _[power]_ — **PENDING** — A hyperscaler announces another >500MW power-first AI campus or behind-the-meter generation deal by September 30, 2026. — In-window power-first activity (Stark/Sagebrush 5.6GW gas-paired pipeline; NAVER+NVIDIA 1GW sovereign factory) was not a hyperscaler behind-the-meter deal. Deadline Sep 30 open. ## Watchlist - **Jun 14-30 — Anthropic Fable 5 / Mythos 5 access restoration.** Whether and how Anthropic restores a top-tier model after the government takedown will set the template for sovereign access risk. Watch for geo-gating, KYC, or a permanent capability split between civilian and government tiers. - **Jun 14-30 — Gemini 3.5 Pro GA and first independent benchmark.** Google's Pro has slipped to the edge of its June commitment. A GA with an independent Artificial Analysis pass would show whether June is a real frontier reset or a third-place launch behind Fable 5 and GPT-5.5. - **Jun 30-Jul 7 — PJM 2028/29 base residual capacity auction.** After two near-cap clears and a 2027/28 reliability shortfall, the 2028/29 auction (collar near $325/MW-day) will show whether power price pressure persists. It is the single most important power-cost signal of the quarter. - **Jun-Aug — Second 1.6T / CPO design win and inference-ASIC adoption.** The networking and custom-silicon theses need a second source. Marvell, Credo, Groq, Cerebras, or Tenstorrent commentary will show whether 1.6T fabric and merchant inference ASICs broaden beyond Broadcom and d-Matrix. - **Jun-Jul — SPCX as frontier-lab proxy; OpenAI/Anthropic public S-1.** SpaceX/xAI's post-IPO trading is now the public benchmark for frontier-lab value. A move to a publicly visible S-1 by OpenAI or Anthropic would be the next regime change for AI capital markets. ## Changelog - Reframed W24 around concentration and sovereign/regulatory risk at the model layer: Claude Fable 5's launch and the first government-forced frontier takedown, the SpaceX/xAI IPO, and Oracle's $638B RPO. --- Source of truth: `src/data/industry/weekly/2026-W24.ts`. Canonical HTML: . PDF: .