brianletort.ai
All issues

The AI Stack Weekly

Issue 18 · Week 34 of 2026.

/Industry brief · ~7 min read/Public sources onlyDownload brief

This week in 30 seconds

Full read ~7 min

  • Last week's house-original thesis just moved into a filing: NVIDIA's own 8-K discloses a residual-value guaranty on a 20-year OpenAI lease at PORTS-Pike, Ohio, cumulatively capped at $105B on roughly 4.25 IT-GW of initial commitment — the GPU vendor is now the underwriter of land-power-shell, not just a supplier into it.
  • Marvell sold Google a warrant on up to ~58.97M shares at $206.58 (~$12.2B fully exercised) that vests only as Custom Products revenue arrives — 240 tranches of $500M each for a $120B ceiling, and Google has no purchase obligation. The residual-value bet is no longer just neocloud credit; it is how silicon vendors underwrite demand.
  • Anthropic promoted Skills API, Files API, and computer use to GA on August 20 with HIPAA/BAA availability, 1TB/org file capacity, and progressive-disclosure loading — the first agent extensibility surface that treats skills as versioned artifacts under inherited identity. Full technique read is in the Agent Techniques Weekly.
  • Salesforce, UiPath, and Google converged on identity-inheritance for enterprise agents inside the same week (Headless 360 MCP surfaces, Maestro Flow, Antigravity bundled into Gemini Enterprise). Service-account patterns are now the exception rather than the pattern. See the Application Layer.
  • OpenAI cut GPT-5.6 Sol pricing by more than 20% on August 21 for a dated three-month promotional window and published the reversal date at launch — the frontier is now pricing on scheduled windows rather than steady state. Full architecture read is in the Model Pulse.
  • Pennsylvania's August 18 executive order binds GRID requirements to DCs above 25 MW, removes them from Permit Fast Track, and conditions the tax exemption on GRID — regulatory friction lengthening the interconnect queue rather than shortening it, and the first state-level rule to explicitly rope hyperscale into a slower permitting lane.

By the numbers

$105B / 4.25 GW
Cumulative cap and initial IT-GW covered by NVIDIA's residual-value guaranty for the SB Energy PORTS-Pike lease
$120B / 240 tranches
Implied ceiling and structure of Marvell's warrant to Google, one $500M tranche of Custom Products revenue at a time
$4.94B net
Nebius upsized convertible notes offering priced August 19 ($3.0B 0.50% due 2030 + $2.0B 4.50% due 2034)
>20%
OpenAI's dated GPT-5.6 Sol price cut, three-month promotional window with published reversal
5x / 1TB
Anthropic Files API rate-limit multiple and per-organization capacity at GA (August 20)
>25 MW
Pennsylvania Executive Order 2026-05 GRID threshold binding on data centers effective August 18

The Bottom Line

The duration-mismatch thesis just moved into the GPU vendor's own 8-K — NVIDIA is now the residual-value underwriter of a 20-year OpenAI lease, and Marvell is teaching Google how to buy silicon on the revenue that has not yet arrived

Flywheel arcAll three lenses

The short version

  • NVIDIA's own 8-K puts the residual-value bet on the GPU vendor's contingent-obligation line: $105B cumulatively capped, 4.25 IT-GW initial, 20-year OpenAI lease at Ohio PORTS-Pike, triggered by tenant default and beginning 2028 subject to ready-for-service conditions — ~$24.7B of credit support per IT-GW as a filing-derived house measurement.
  • Marvell issued Google a warrant on up to ~58.97M shares at $206.58 that vests one $500M-revenue tranche at a time to a $120B ceiling over Q3 FY2027–FY2033, with Google under no purchase obligation — a different structural form of the same residual-value logic.
  • Nebius priced $5B of upsized convertibles, Cerebras claimed 750 PFLOPS on the CS-4, Etched doubled its mark to $21B on thin corroboration, and Pennsylvania added GRID as a binding regulatory stage for DCs above 25 MW — the week did four things to the duration stack at once.
  • Anthropic Skills API, Files API and computer use are GA with progressive-disclosure loading, HIPAA/BAA availability, and 1TB per organization — the shortest-tenor layer of the stack just re-based its extensibility model in one release, treated in full in the Agent Techniques Weekly.
  • Do: model the 10-Q exhibits when they arrive with NVIDIA's quarter-ended-July-26 filing; require warrant-style vesting mechanics before the industry standardizes on one; add PA EO 2026-05 to any PJM-region site plan above 25 MW; put OpenAI's dated Sol reversal into any November agent business case.
The residual-value bet is no longer just neocloud credit — it is now how the silicon vendors themselves underwrite demand.

The full story

Last week this publication argued that the AI infrastructure trade was systematically financing itself on a longer duration than the contracts underneath it, and that the mismatch was structural rather than specific to any one operator. This week that thesis moved into a filing. NVIDIA's Form 8-K dated August 17, 2026 discloses residual value guaranties supporting SB Energy's PORTS-Pike technology campus in Pike County, Ohio, at approximately 4.25 gigawatts of IT load, with an OpenAI affiliate as the tenant on 20-year leases. NVIDIA can optionally support an additional roughly 3.8 GW at the same site. The aggregate payment obligation is cumulatively capped at $105B for the initial commitment, disclosed under Item 2.03 as an off-balance-sheet arrangement, subject to ready-for-service conditions expected beginning 2028, and triggered by tenant insolvency or default under the lease. In a companion transaction NVIDIA committed $1.5B of equity investment into SB Energy, with the site contracted exclusively for NVIDIA compute for 20 years. This publication's filing-derived house measurement is that $105B / 4.25 IT-GW is approximately $24.7B of residual-value credit support per IT-GW of initial commitment — reproducible from the 8-K text alone, and material to how the category should now be modelled.

The first thing to say about the number is what it is not. It is not $105B of cash out the door. It is a cumulative cap on a contingent obligation that turns on a tenant default across a 20-year lease that has not yet commenced service. NVIDIA's July press narrative talked about ~$250B; the filing is $105B, conditional, and phased. Every revenue-opportunity figure the vendor puts alongside this — the ~1.5 million GPUs per generation, the ~$150-200B revenue opportunity language on the NVIDIA blog — is vendor-stated and belongs behind that label. What the filing does say precisely is who is now on the hook if the tenant fails to pay: not the developer, not the neocloud, not the bond insurance market, but the accelerator vendor itself. Once the residual-value assumption on 20-year land-power-shell sits inside a public GPU vendor's off-balance-sheet disclosure, the industry has stopped debating whether GPUs are financeable long assets and started underwriting them like them. Two competing narratives are running through coverage this week — 'NVIDIA becomes a financier' and 'a $105B guarantee' — and both are correct in outline and misleading in emphasis. The tighter reading is that NVIDIA has taken the residual-value bet the neoclouds had been carrying and rehypothecated it onto its own balance sheet's contingent line, which is a structural change in how AI capital gets underwritten and a rating question that will not be resolved by the next earnings call.

The same week produced the mirror image on the silicon side. On August 19, Marvell filed an 8-K disclosing a warrant issued to Google covering up to 58,970,907 shares of Marvell common stock at an exercise price of $206.58 per share — roughly $12.2B if fully exercised and approximately 7% of the company. The commercial agreement underneath is dated July 29 and covers custom silicon around the TPU ecosystem. 1,360,867 shares vest quarterly through the first year; the remaining shares vest in 240 equal tranches, one tranche released for each $500M of Custom Products revenue Marvell attributes to Google under the agreement between Q3 FY2027 and FY2033. Google has no purchase obligation. Multiply the tranche count by the revenue trigger and the implied ceiling is $120B of Custom Products revenue over roughly six fiscal years for Google's silicon relationship with Marvell alone — a ceiling the market has already begun quoting as a headline number even though the vesting is optional and back-loaded. What matters for this issue is the shape of the instrument. Marvell has effectively sold Google the option to acquire ~7% of the company at $206.58 against an August 19 close of $237.27 — modestly in-the-money at disclosure, with nearly all shares still unvested and funded by revenue that has not yet arrived and that only appears if Google actually chooses to source at that scale. The silicon vendor is not underwriting demand in the same way NVIDIA is at PORTS-Pike, but it is issuing equity as a function of a customer's future purchasing decision — which is a different and quieter form of the same residual-value logic.

Around those two structural filings, the rest of the week's capital moved in ways that either confirmed the pattern or made the promised timeline slower. Nebius priced $5B of convertibles on August 19 — $3B of 0.50% notes due 2030 and $2B of 4.50% notes due 2034, upsized from $4.5B, with roughly $4.94B of net proceeds — which is a second public neocloud rolling long paper against shorter customer contracts inside the same quarter that CoreWeave did it. Cerebras announced the CS-4 SUPERNOVA on August 18 with three WSE-3 Turbo processors at 750 vendor-claimed PFLOPS and first shipments at the end of Q3 2026, which is a genuine architectural step and a claim measured on Cerebras's own harness. Etched raised $700M at a $21B valuation, roughly double its June mark, after Jane Street led the round following a rack deployment — an eye-catching valuation move on thin corroboration and best treated as noise-adjacent. Pennsylvania Executive Order 2026-05 dated August 18 bound GRID requirements to data centers above 25 MW, removed those DCs from Permit Fast Track, and conditioned the state tax exemption on GRID compliance while requiring local approvals before DEP permits — the first state-level rule to explicitly rope hyperscale into a slower permitting lane. Read together, the week did four things at once: filed the duration mismatch into a public balance sheet's contingent line, invented a silicon-issuance vehicle indexed to a customer's future purchasing, added a $5B convertible print to the neocloud tenor stack, and lengthened the regulatory path for Pennsylvania sites inside PJM.

The practical instruction is narrower and sharper than 'watch NVIDIA.' If you underwrite AI infrastructure, the residual-value line is no longer a private credit assumption inside a neocloud model. It is now a public off-balance-sheet disclosure at a $4T market-cap company, and the rating implication compounds every time another operator's lease falls under the same guaranty envelope — so read the 10-Q exhibits when they arrive alongside the quarter ended July 26 print. If you buy silicon at scale, the Marvell warrant is the template you will be offered next; require the vesting mechanics in the term sheet before the industry standardizes on a version that only vests when convenient to the vendor. If you plan capacity in PJM territory, PA EO 2026-05 has just added a permitting stage to every site above 25 MW, and this compounds with an interconnect queue already at 60-84 months rather than shortens it. And if you architect agents, the layer with the shortest tenor just moved: Anthropic's Skills API, Files API and computer use went GA on August 20, Salesforce, UiPath and Google converged on identity-inheritance in the same week, and OpenAI cut Sol pricing for a dated three-month window while publishing the reversal date. Each of those is a full home in another publication this week — the important thing to see at the weekly level is that all three are now compressing capital durations and agent extensibility durations against each other faster than any board paper cycles.

JevonsMetcalfeGilderSoftwareJevonsHardwareHuangNetworkingMetcalfe + Gilder

The three lenses

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

15 events across the flywheel — 5 software, 5 hardware, 5 networking.

Software.

  • NVIDIA published TensorRT Model Connect as a public preview under Apache 2.0 — a Hugging Face to TensorRT path with no ONNX intermediate and aarch64 wheels, and stated on the release page that the tool was built using Codex agents

    NVIDIA TensorRT Model Connect on GitHub

  • OpenAI announced ChatGPT advertising expanding to 31 EU countries with rollout beginning August 24 and separately launched ChatGPT for Teens with new safeguards — consumer surfaces monetizing and segmenting on the same day

    OpenAI, coverage via TechCrunch and Reuters

  • Ornith AI released Ornith-1.5 as an MIT-licensed open-weights family — a 397B/35B-active mixture-of-experts flagship with 35B dense and 9B distilled companions, positioned against DeepSeek V4 Pro and GLM-5.3 on vendor-run coding and terminal benchmarks

    RuntimeWire; independent trackers unranked at publish

  • DeepSeek launched V4-Flash-Vision-Exp as an experimental API-only multimodal model with no downloadable weights and no committed graduation date — a second week of served-versus-downloadable divergence at the same vendor

    DeepSeek platform announcement

  • OpenAI cut GPT-5.6 Sol token pricing by more than 20% for a dated three-month promotional window, publishing the reversal date at launch and preserving the same price on Sol Max — see the Model Pulse for the architecture read

    OpenAI GPT-5.6 announcement

What this means

  • Cost per task moved down again — but on a dated window with the reversal published, not on a durable rate. Any November agent business case must be modelled against the step-up.
  • The served-versus-downloadable gap widened in a new direction: DeepSeek shipped an API-only experimental multimodal model in the same week Ornith released open weights it had already benchmarked against.
  • Consumer surfaces started monetizing and segmenting in the same day — ChatGPT ads to 31 EU countries and ChatGPT for Teens — a sign that inference cost pressure is being answered on the demand side rather than the model side. See The Model Pulse for full model architecture treatment.
Full reasoning +

The software layer's week was a dated cut and a widening definition of what 'open' means, both of which push the same procurement conclusion: agent contracts should be short, indexed, and written with the assumption that both the price and the benchmarked artifact can change without notice. OpenAI publishing a reversal date at launch is now the second frontier vendor to price on a schedule rather than a spot rate in the last six weeks, which turns inference cost from a market to a calendar. DeepSeek serving a vision endpoint whose weights are not obtainable, in the same week Ornith released weights already benchmarked against DeepSeek on the vendor's own harness, means the industry's headline measure of the open-versus-closed gap keeps drifting further from measuring substitutability. Architects should treat inference contracts and skill artifacts as separate durations — one repricing in months, the other in weeks — and stop letting either one anchor the other. The Model Pulse carries the full architecture read on Ornith, the Sol price cut, and DeepSeek-V4-Flash-Vision-Exp.

Hardware.

  • NVIDIA filed Form 8-K disclosing residual value guaranties supporting SB Energy's PORTS-Pike, Ohio campus for a 20-year OpenAI-affiliate lease at approximately 4.25 IT-GW, cumulatively capped at $105B for the initial commitment and subject to ready-for-service beginning 2028 (see Big Story for full treatment)

    NVIDIA Form 8-K, August 17, 2026

  • Cerebras announced the CS-4 SUPERNOVA — three WSE-3 Turbo processors delivering a vendor-claimed 750 PFLOPS with first shipments targeted at the end of Q3 2026

    Cerebras Systems

  • Qualcomm and Modular unveiled Dragonfly AI200 at ModCon with Modular's Mojo 1.0 shipping under Apache 2.0 — no TOPS figure was published for AI200 at launch and the software stack is where the news lives

    Qualcomm and Modular ModCon 2026

  • Etched raised $700M at a roughly $21B valuation — approximately doubling its June $10B mark 26 days later — with Jane Street leading after deploying a rack for evaluation, third-party benchmark comparability still limited

    TechCrunch

  • Anthropic hired former Google TPU founder Amir Salek to lead in-house silicon work, per Bloomberg reporting relayed through secondary coverage — a spend-side commitment against custom accelerators rather than a primary raise

    Bloomberg via secondary coverage

What this means

  • The hardware lens's headline is a public filing rather than a spec sheet. NVIDIA's 8-K puts the residual-value bet on the vendor's contingent-obligation line — capital risk now rides with the accelerator, not just the accelerator's owner.
  • Cerebras and Qualcomm/Modular shipped genuine architectural moves without independent verification of performance. Treat 750 PFLOPS and AI200's absent TOPS as vendor claims, and price capacity against measured throughput on the same workload.
  • Etched's valuation doubling on thin corroboration is the week's cleanest noise call — Jane Street's rack deployment is a datapoint, not a benchmark, and reasoning about ASIC economics from a private mark that moved 100% in 26 days is procuring on a signal that has not landed.
Full reasoning +

The hardware lens produced two structural filings and one noise call in the same seven days. NVIDIA's disclosure changes how the industry should model residual-value risk: what used to be a private credit assumption inside a neocloud DSCR test is now an off-balance-sheet contingent obligation on a mega-cap public semiconductor company, and it will move rating agency posture on both sides of that trade. Cerebras and Qualcomm/Modular each shipped a real architectural step, but the numbers around them are vendor harnesses without independent third-party reproduction — and Etched raising at $21B on a rack demo is the week's reminder that private-mark inflation on custom accelerators is running ahead of any measured procurement case. The honest read on custom silicon this week is that the demand side (Marvell/Google warrant, Anthropic hiring Salek, Microsoft Maia 300 talks in market) is moving faster than the measured performance side, and the risk is a category priced on optionality rather than on shipped throughput per dollar.

Networking.

  • Marvell filed an 8-K disclosing a warrant issued to Google covering up to 58,970,907 shares at $206.58 (~$12.2B if fully exercised, ~7% of the company), with 1,360,867 vesting quarterly year one and the remainder in 240 tranches of one per $500M of Custom Products revenue from Q3 FY2027–FY2033 — an implied $120B ceiling, no purchase obligation (see Big Story)

    Marvell Form 8-K, August 19, 2026

  • Hot Interconnects 2026 confirmed NVIDIA co-packaged-optics Ethernet switches shipping in production, with vendors and independent commentators putting optical scale-out at approximately 10% of total compute power in advanced deployments

    Converge Digest coverage of Hot Interconnects 2026

  • Broadcom detailed Ethernet scale-up, scale-out and scale-across positioning at Hot Interconnects with new switch and optics roadmaps, positioning against NVIDIA CPO on a merchant fabric

    Broadcom Hot Interconnects sessions, Converge Digest

  • Ciena launched Nitro 2004 targeting approximately 4 meter copper reach at roughly 1.5 pJ/bit for in-rack and near-rack AI interconnects — the strongest continuing case for copper's role as the near-rack default before optics

    Ciena press and Hot Interconnects sessions

  • Pennsylvania Executive Order 2026-05 imposed binding GRID requirements on data centers above 25 MW, removed qualifying DCs from Permit Fast Track and conditioned the tax exemption on GRID — regulatory friction added to a PJM interconnect queue already at 60-84 months

    Pennsylvania Governor's Office Executive Order 2026-05

What this means

  • The networking lens is where the residual-value logic met the silicon issuance: Marvell selling Google a warrant that only vests as custom-silicon revenue arrives is a new instrument that ties fabric vendor equity to hyperscaler purchasing behaviour.
  • CPO is real, single-source and single-week — NVIDIA claiming production while Broadcom stakes the merchant-fabric position and Ciena defends near-rack copper together define the shape of the 2027-28 fabric choice, not a resolved winner.
  • PA EO 2026-05 is regulatory friction not queue shortening. Any PJM-region site above 25 MW now inherits a GRID compliance stage before DEP permits and a local-approval gate on top — model the schedule impact before the site plan, not after.
Full reasoning +

Networking produced this week's second structural filing and its most consequential policy move on the same three-day window. The Marvell warrant is a different form of the residual-value logic that appeared in NVIDIA's 8-K: NVIDIA is guaranteeing land-power-shell against a tenant default, and Marvell is issuing equity against a customer's future purchasing decision, but both are the industry moving underwriting risk onto public balance sheets that used to sit inside private neocloud credit assumptions. On the physical layer, Hot Interconnects 2026 confirmed that CPO is now a shipping technology from one vendor rather than a coalition-standard technology from many, and the OCP CPO specification's stated Q4 2026 submission window remains the schedule input that determines whether a second source exists for 2027-28 refresh cycles. Pennsylvania's EO 2026-05 is the reminder that fabric planning still sits inside a physical constraint stack, and that constraint just got heavier at the largest US market for hyperscale demand. Architects should now separate three fabric decisions explicitly rather than roll them together: single-vendor CPO now, multi-vendor CPO later, and the near-rack copper layer where Ciena's Nitro 2004 keeps the pluggable window open through the transition.

Capital flow

Money in, revenue out.

4 categories tracked. Capital deployment up in 1 of 4; revenue follows at multiples of 0.21 to 0.6.

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

  • Frontier Labs

    OpenAI, Anthropic, Google DeepMind, xAI

    Capital In

    ~$95B

    vs ~$95B

    Revenue Out

    ~$21B

    vs ~$21B

    Burn / Rev

    ~4.5x

    Movement

    No lab closed primary financing in-window. Anthropic hired former Google TPU founder Amir Salek to lead in-house silicon work — a spend-side commitment against custom accelerators rather than a raise. OpenAI's Sol price cut and EU ads rollout are product-and-monetization moves that do not change the lab's aggregate capital or disclosed revenue in-window.

  • Hyperscaler-Hosted

    Azure-OpenAI, AWS-Anthropic, Google Cloud-Gemini, Oracle-OCI

    Capital In

    ~$250B

    vs ~$250B

    Revenue Out

    ~$70B

    vs ~$70B

    Burn / Rev

    ~3.6x

    Movement

    No hyperscaler reported or revised capex guidance in-window. The category's motion was structural rather than budgetary: NVIDIA's residual-value guaranty on the SB Energy PORTS-Pike site is a contingent obligation on the accelerator vendor tied to an OpenAI affiliate's 20-year lease, and Marvell's warrant to Google indexes silicon vendor equity to custom-product revenue. Google also bundled Antigravity into Gemini Enterprise on August 20 — see the Application Layer.

  • Neoclouds

    CoreWeave, Nscale, Crusoe, Lambda, Fluidstack, IREN

    Capital In

    ~$24.6B

    vs ~$19.6B

    Revenue Out

    ~$8B

    vs ~$8B

    Burn / Rev

    ~3.1x

    Movement

    Nebius priced $5B of convertible notes on August 19 — $3B of 0.50% notes due 2030 and $2B of 4.50% notes due 2034, upsized from $4.5B, with roughly $4.94B in net proceeds — the largest AI-infrastructure convertible print of the year. No neocloud reported earnings in-window.

  • On-Prem / Hybrid

    Enterprise GPU clusters, sovereign and national programs, Cisco / Dell / HPE

    Capital In

    ~$103B

    vs ~$103B

    Revenue Out

    ~$38B

    vs ~$38B

    Burn / Rev

    ~2.7x

    Movement

    Pennsylvania Executive Order 2026-05 on August 18 bound GRID requirements to data centers above 25 MW, removed qualifying DCs from Permit Fast Track, and conditioned the state tax exemption on GRID compliance — regulatory friction added to the largest US hyperscale market's interconnect queue. Cerebras announced CS-4 SUPERNOVA with vendor-claimed 750 PFLOPS and first shipments at end of Q3 2026.

Burn-to-Revenue is revenue divided by committed capital. Lower means more capital is going out than coming in.

Signal vs noise

What’s real, what’s noise.

7 claims this week — 4 signal, 3 noise.

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

  • 5 / 5

    NVIDIA's Form 8-K discloses residual value guaranties supporting SB Energy's PORTS-Pike, Ohio campus at approximately 4.25 IT-GW for a 20-year OpenAI-affiliate lease, cumulatively capped at $105B for the initial commitment, disclosed under Item 2.03 as an off-balance-sheet arrangement and triggered by tenant insolvency or default.

    Sources: NVIDIA Form 8-K filed August 17, 2026; NVIDIA newsroom announcement; NVIDIA corporate blog. Exhibits pending with the 10-Q for the quarter ended July 26, 2026.

    Structural change in how AI capital gets underwritten, disclosed in a primary filing. The residual-value assumption that used to live inside neocloud credit models is now on a mega-cap public semiconductor company's contingent-obligation line, and the rating implication compounds every time another site falls under the same guaranty envelope. Model the 10-Q exhibits when they arrive. The $105B figure is a cumulative cap on a contingent obligation, not cash out the door, and every revenue-opportunity figure NVIDIA has attached to the transaction is vendor-stated — treat the label accordingly.

  • 5 / 5

    Marvell issued Google a warrant on up to 58,970,907 shares at $206.58 (~$12.2B if fully exercised, ~7% of the company) vesting one $500M-revenue tranche at a time over 240 tranches from Q3 FY2027 through FY2033, implying a $120B ceiling of Google Custom Products revenue with no purchase obligation.

    Sources: Marvell Form 8-K filed August 19, 2026; commercial agreement dated July 29 covering custom silicon around the TPU ecosystem.

    The other structural filing of the week and a different form of the residual-value logic. Silicon vendor equity is now indexed to a hyperscaler's future purchasing decision at 240 discrete revenue triggers, funded by revenue that has not yet arrived. Enterprises buying at scale should require warrant-style vesting mechanics in their term sheets before the industry standardizes on a version that only vests when convenient to the vendor. The $120B ceiling is the multiplication of tranche count and trigger; do not carry it as a revenue forecast.

  • 4 / 5

    Nebius priced $5B of upsized convertible notes on August 19 — $3B of 0.50% notes due 2030 and $2B of 4.50% notes due 2034, with roughly $4.94B in net proceeds — the largest AI-infrastructure convertible print of the year.

    Sources: Nebius Group Form 6-K filed August 19, 2026.

    Public disclosure at scale by an audited issuer, so directionally reliable. It continues the pattern this publication named in W33 — long paper against shorter contracts — with a second neocloud rolling the same tenor stack inside the same quarter as CoreWeave. Rates term structure (0.50% near, 4.50% far) is the market rewarding the growth story on the near tranche and demanding a step-up on the further one, which is a rational and unremarkable answer to a still-forming private-mark environment. The number itself is signal; the structural implication for GPU residual value beyond 2030 remains a hypothesis rather than a settled question.

  • 3 / 5

    Cerebras CS-4 SUPERNOVA delivers 750 PFLOPS using three WSE-3 Turbo processors, with first shipments at the end of Q3 2026 — a step change in wafer-scale performance for large model training and inference.

    Sources: Cerebras Systems product announcement dated August 18, 2026. Number is a vendor claim measured on Cerebras's own harness; no independent third-party reproduction available at publish.

    Architectural progress is real and the shipping timeline is credible. The PFLOPS number is a vendor claim without independent verification, and Cerebras's methodology on prior generations has diverged from Artificial Analysis measurement on comparable workloads. Treat the announcement as directionally reliable for planning and as vendor-stated for procurement. On-prem operators evaluating wafer-scale should schedule a measured comparison against a Blackwell-Ultra reference workload rather than rely on the harness delta.

  • 2 / 5 — noise

    Etched's transformer-ASIC valuation doubled to $21B in 26 days on the basis of a Jane Street rack deployment — a signal that dedicated inference silicon is being repriced by the market against merchant GPUs.

    Sources: TechCrunch reporting on the round dated August 18, 2026, citing anonymous sources on the pricing and Jane Street's participation.

    Noise-adjacent. Private-mark inflation of 100% in a month is not evidence of category revaluation; it is evidence of a specific investor's willingness to lead at that mark. Jane Street's rack deployment is a datapoint, not a benchmark, and dedicated-inference ASIC economics still depend on measured throughput per dollar on a stable workload that has not been published. The category may well reprice but nothing in this transaction supports the claim as it stands.

  • 2 / 5 — noise

    HBM/DRAM Korean press reports on August 20 indicate a further tightening at Samsung and SK Hynix that will require accelerator vendors to de-spec memory configurations across the next generation, with material shipment implications for 2027.

    Sources: Grade-3 Korean press aggregation dated August 20, 2026, unsourced back to a manufacturer disclosure.

    Directionally consistent with what Micron said publicly in W33 and with the pattern this publication has been tracking, but this specific report is aggregation without a manufacturer citation and should not be carried as a fifth-grade signal. The underlying constraint is real; the specific de-spec framing is a rumour continuation without primary confirmation. Wait for a manufacturer disclosure or an accelerator-vendor product page rather than move a share estimate on a Korean press item.

  • 1 / 5 — noise

    SB Energy is preparing a $50B initial public offering on the strength of its PORTS-Pike hyperscale contracts, per unnamed sources cited across secondary coverage of the NVIDIA guaranty announcement.

    Sources: Secondary coverage tying the NVIDIA guaranty announcement to a rumoured SB Energy IPO valuation; no primary source, no filing, no company confirmation.

    Lowest-confidence claim of the week and the one most likely to end up in a slide. Every part of the number is inference: the valuation, the timing, and even that a listing is being prepared. Anchoring a $50B private mark to the same-day announcement of a contingent guaranty conflates two very different balance-sheet events, and no company involved has confirmed any listing preparation. Do not quote the number in any comparison table.

Synthesis

The week, reasoned through.

3 cross-domain connections, 5 hypotheses tested (1 under pressure), 2 patterns tracked.

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

Connecting the dots

  • Abductive

    78%

    confidence

    The residual-value bet on AI infrastructure has moved from private neocloud credit assumptions into two different public-issuer disclosures in the same week — NVIDIA guaranteeing land-power-shell on a tenant default and Marvell indexing equity to a hyperscaler's future silicon purchasing — which means the underwriting question is no longer whether GPUs are a financeable long asset, but who publicly carries the residual-value bet once they are.

    1. 01NVIDIA's Form 8-K dated August 17 discloses residual-value guaranties supporting SB Energy PORTS-Pike at approximately 4.25 IT-GW for a 20-year OpenAI-affiliate lease, cumulatively capped at $105B, Item 2.03 off-balance-sheet, triggered by tenant insolvency or default.
    2. 02Marvell's Form 8-K dated August 19 discloses a warrant to Google covering up to 58,970,907 shares at $206.58 with the remainder vesting one $500M-revenue tranche at a time across 240 tranches from Q3 FY2027 through FY2033, implying a $120B ceiling with no purchase obligation.
    3. 03Nebius priced $5B of upsized convertibles on August 19, rolling long paper (0.50% 2030 + 4.50% 2034) against still-shorter customer contracts — a second neocloud in the same quarter as CoreWeave doing the same duration trade.
    4. 04The 8-K exhibits from NVIDIA are still to come with the 10-Q for the quarter ended July 26, which means the auditability of the residual-value line grows this quarter regardless of whether NVIDIA reports a revenue surprise.

    Steel-man

    The strongest counter is that this reading conflates three genuinely different instruments — a contingent obligation, a warrant, and a convertible — into a single narrative when they answer different capital questions. NVIDIA guaranteeing a tenant default is not the same instrument as Marvell issuing Google an equity option indexed to future purchasing, and Nebius issuing convertibles is normal growth-cap-stack behaviour rather than novel duration engineering. There is real risk of pattern-matching separate filings into a shared story. The claim still stands in a narrower form: what all three do share is that the residual value of AI infrastructure — either the site behind the accelerator, the accelerator vendor's equity, or the neocloud debt — is being underwritten longer than the contracts underneath it, and that is a common structural feature rather than a coincidence.

    Evidence: NVIDIA Form 8-K 2026-08-17: $105B cumulative cap on residual-value guaranty, 4.25 IT-GW initial · Marvell Form 8-K 2026-08-19: warrant to Google, 58.97M shares at $206.58, 240 tranches · Nebius Form 6-K 2026-08-19: $5.0B upsized convertibles priced across 2030/2034 · NVIDIA newsroom: 20-year exclusive NVIDIA-compute lease, $1.5B NVIDIA equity investment in SB Energy

  • Deductive

    66%

    confidence

    The framework's hypothesis that Metcalfe's law and the power constraint reinforce each other predicted the fabric layer's pricing behaviour, and this week's PA EO 2026-05 confirms the mechanism from the regulatory side by lengthening the interconnect queue for the largest US hyperscale market — which means the two-vendor CPO decision (single-source now versus multi-vendor Q4 2026 spec later) becomes a scheduling question against a queue that is measurably getting longer rather than shorter.

    1. 01The working framework's Metcalfe law predicts that interconnect value scales as the square of connected nodes, and hypothesis 5 holds power as the binding constraint on deployment — together these predict that campuses fragment under power limits and drive optics demand, which W33's Cisco call commentary directly supported.
    2. 02Pennsylvania Executive Order 2026-05 on August 18 binds GRID requirements to DCs above 25 MW, removes qualifying DCs from Permit Fast Track and conditions the tax exemption on GRID compliance — the first state-level rule to explicitly rope hyperscale into a slower permitting lane in the largest US market.
    3. 03Hot Interconnects 2026 confirmed NVIDIA CPO Ethernet switches shipping in production with vendors and independent commentators putting optical scale-out at approximately 10% of total compute power, and Ciena launched Nitro 2004 at approximately 4m copper reach at 1.5 pJ/bit as the near-rack default before optics.
    4. 04The OCP Open Silicon Photonics workstream's stated Q4 2026 specification submission target remains the fork in the road for whether a multi-vendor path exists for 2027-28 fabric refresh cycles.

    Steel-man

    The strongest counter is that PA EO 2026-05 is one state's regulatory posture rather than a national trend, and that fabric-refresh scheduling decisions run off national accelerator supply and hyperscaler capex plans that PJM-region permitting does not gate. If operators respond by routing new sites out of PJM into ERCOT, MISO, or Nordic markets — as several have already signalled they can — the queue-length effect is a Pennsylvania problem rather than a national fabric-choice problem. The deduction still holds because it does not depend on PA being unique; it depends on state-level friction lengthening some part of the effective queue somewhere, and PA EO 2026-05 is direct evidence of that mechanism arriving in the market that matters most for hyperscale demand. If a comparable rule appears in a second major-market state within the quarter, the claim gets stronger; if PA reverses, it gets weaker.

    Evidence: Pennsylvania Executive Order 2026-05, August 18, 2026 (GRID binding on DCs above 25 MW) · Hot Interconnects 2026: NVIDIA CPO in production, optical scale-out ~10% of compute power · Ciena Nitro 2004 launch: ~4m copper reach at ~1.5 pJ/bit for in-rack AI interconnect

  • Inductive

    72%

    confidence

    Anthropic's Skills API and computer use GA on August 20 and Salesforce, UiPath, and Google's identity-inheritance releases in the same week are converging on the same architectural pattern — versioned skill artifacts running in a hosted sandbox under the caller's identity — without any coordinating body having asked for it, which suggests the industry has decided identity inheritance is the enterprise-viable default and that service-account patterns have become an exception carrying a written justification rather than a routine choice.

    1. 01Anthropic Skills API, Files API and computer use graduated to GA on August 20 with progressive-disclosure loading, HIPAA/BAA availability, 5x file rate limits and 1TB per organization — skills executing in Anthropic's sandbox with outbound calls under the caller's identity.
    2. 02Salesforce expanded Headless 360 on August 19 with MCP servers, skills and identity/permission inheritance across Data Cloud, Agentforce, MuleSoft and Tableau — agents calling Salesforce data under caller entitlements rather than service credentials.
    3. 03UiPath launched Maestro Flow on August 19 for identity-scoped agent orchestration, and Google bundled Antigravity into Gemini Enterprise on August 20 for identity-inherited agent workflows on the Google identity plane.
    4. 04Alipay's agentic commerce release on August 17 uses MCP as commerce infrastructure with per-user consent scoping rather than aggregate merchant credentials — the same architectural choice on a different vendor's substrate.

    Steel-man

    The strongest counter is that four announcements in a single week are a coincidence of release calendars rather than a convergent architectural decision, and that identity inheritance is what these vendors have been building for six to twelve months regardless of what any of them shipped this week. The 'convergence' language may also be reading a shared pattern into products that solve the identity problem in materially different ways (Salesforce's is a data platform choice, Anthropic's is a lab-side sandbox, Google's is a hyperscaler identity plane). The claim survives because the useful thing to say about the pattern is not who moved first but that the enterprise architecture question — service account or inherited identity — now has one visible answer across the three surfaces most enterprises will encounter, and buyers who assume the older default in a term sheet drafted today are the exception rather than the pattern. See the Agent Techniques Weekly for the full technique treatment and the Application Layer for the platform-side read.

    Evidence: Anthropic Skills API GA (Aug 20): versioned skills, hosted sandbox, inherited identity · Salesforce Headless 360 expansion (Aug 19): MCP surfaces, identity/permission inheritance · UiPath Maestro Flow launch (Aug 19): identity-scoped agent orchestration · Google Antigravity bundled into Gemini Enterprise (Aug 20)

Thesis test

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

  • Hypothesis 1

    supported

    The cycle is accelerating, not slowing.

    The frontier lab release cadence continues to compress: OpenAI cut GPT-5.6 Sol pricing more than 20% inside a dated three-month window, Anthropic promoted Skills API and computer use to GA, DeepSeek launched V4-Flash-Vision-Exp as an experimental API model, and Ornith AI shipped a 397B/35B-active MoE flagship with 35B and 9B companions in one release. Agent tooling GA'd inside the same seven-day window as consumer-tier product surfaces monetized in 31 EU countries. The framework's law that pricing on a schedule rather than a spot rate is now the observable behaviour of the top tier is holding.

    Against it: The main honest counter is that this week's biggest lab-side events were about extensibility and pricing rather than capability advances. Anthropic did not ship a new base model; OpenAI did not raise Sol's ceiling capability; Ornith's benchmarks are vendor-run. Release cadence accelerating while pretraining cadence does not is the weaker form of the hypothesis, and this week's evidence sits on that side of the line.

    Evidence: OpenAI GPT-5.6 Sol dated promotional cut (Aug 21) · Anthropic Skills API GA (Aug 20) · Ornith-1.5 MIT-licensed release (Aug 19)

  • Hypothesis 2

    supported

    Capital is concentrated, returns are diffuse.

    The two structural filings of the week are direct evidence: NVIDIA carrying a $105B cumulative cap of contingent obligation for a single OpenAI-affiliate site and Marvell issuing Google an equity option indexed to $120B implied Custom Products revenue. Concentration is compounding. Nebius adding $5B of convertible paper against still-shorter contracts continues the same shape. Returns remain diffuse on the enterprise side — the Application Layer documents Salesforce, UiPath, Google and Alipay all shipping identity-inheritance features in the same week, spreading agent value across horizontal and vertical platforms rather than concentrating it back to the lab.

    Against it: The honest counter is that both filings are contingent instruments, not booked capital. NVIDIA's $105B is a cap on a payment obligation triggered by tenant default with obligations only beginning 2028; Marvell's warrant only vests as Google chooses to purchase. Neither counts as capex concentration in the near-term reporting sense. The hypothesis still holds because the underwriting concentration — who has to carry the residual value if things go wrong — is now on public balance sheets in a way it was not a month ago.

    Evidence: NVIDIA Form 8-K $105B cumulative cap contingent obligation · Marvell Form 8-K warrant to Google ($120B implied ceiling) · Salesforce Headless 360 MCP surfaces (Aug 19)

  • Hypothesis 3

    supported

    Networking is the durable layer.

    Hot Interconnects 2026 confirmed NVIDIA CPO Ethernet switches shipping in production, Broadcom staked scale-up/out/across positioning on merchant fabric, and Ciena launched Nitro 2004 for the near-rack copper window. Marvell's warrant to Google indexes the silicon vendor's equity value to hyperscaler purchasing behaviour on custom silicon, which is the closest thing to a valuation-fed pricing-power confirmation the networking layer has produced. Optical scale-out at approximately 10% of compute power in advanced deployments is the durable-layer share showing up as a fabric-power datapoint rather than an ARR one.

    Against it: Single-vendor CPO shipping while OCP's multi-vendor specification is still on a Q4 2026 target and ASE says the test ecosystem is not ready means the durable pricing power may capture to one vertically integrated silicon vendor rather than accrue to networking suppliers as a class. If the merchant fabric layer loses share to NVIDIA CPO across the next refresh cycle, the durability applies to a smaller category than the hypothesis assumes.

    Evidence: Hot Interconnects 2026 NVIDIA CPO production, Broadcom scale-across, Ciena Nitro 2004 · Marvell Form 8-K warrant to Google indexed to Custom Products revenue

  • Hypothesis 4

    strained

    Open weights pull the floor up.

    Ornith-1.5 shipped as MIT-licensed at 397B/35B-active with vendor-run benchmarks against DeepSeek V4 Pro and GLM-5.3, and no independent tracker had ranked it at publish. DeepSeek in parallel launched V4-Flash-Vision-Exp as API-only with no downloadable weights. The mechanism the hypothesis depends on — open weights re-routing compute demand to on-prem and sovereign deployment — requires that the weights be obtainable in the benchmarked form, and this week produced one open release that has not been externally benchmarked and one served build that is not downloadable at all.

    Evidence: Ornith-1.5 MIT-licensed release with vendor-run benchmarks · DeepSeek V4-Flash-Vision-Exp API-only experimental release

  • Hypothesis 5

    supported

    Power is the binding constraint for the next 24 months.

    NVIDIA guaranteeing a 20-year land-power-shell arrangement at 4.25 IT-GW in Ohio is the strongest single week of evidence for power (and grid interconnect) as the binding physical constraint the framework has produced. The deal exists because power at that scale is the scarce resource the tenant is contracting for, and NVIDIA writing a residual-value guaranty against tenant default is the accelerator vendor absorbing what would otherwise be an unfinanceable siting risk. Pennsylvania Executive Order 2026-05 in the same week added a permitting-side friction to sites above 25 MW in the largest US hyperscale market — a state directly acting on the constraint. The honest ambiguity is whether NVIDIA's absorption of the residual-value bet should be read as power constraint (evidence for the hypothesis) or as capital reroute (evidence for hypothesis 2 instead). This publication reads it as power-first because the guaranty is contingent on the physical site rather than on the accelerator.

    Against it: The strongest counter is that NVIDIA writing a residual-value guaranty on a 20-year lease could be read as demand routing capital around a power constraint that stops binding once the site is in service in 2028 — in which case this week's evidence is capital solving power on the ground rather than power binding capital. Under that reading the hypothesis is supported for the near term and structurally weakening for the medium term. The verdict remains supported because the guaranty exists at all only because power at scale is scarce, but the counter is honest and worth naming.

    Evidence: NVIDIA 8-K residual-value guaranty on 4.25 IT-GW at PORTS-Pike, 20-year lease · Pennsylvania Executive Order 2026-05 binding GRID on DCs above 25 MW

Pattern watch

  • Abductive2 weeks observed

    Silicon vendors are moving residual-value and demand-underwriting risk from private customer credit assumptions onto their own public disclosure lines — first as commercial narrative, then as filed contingent obligations and warrants.

    • W33: CoreWeave's $2.6B DDTL at SOFR+550 explicitly lengthened financing beyond three-year customer contracts and required lenders to underwrite GPU residual value past the contracted period.
    • W34: NVIDIA filed an Item 2.03 off-balance-sheet residual-value guaranty for a 20-year OpenAI-affiliate lease at approximately 4.25 IT-GW, cumulatively capped at $105B, and Marvell filed a warrant to Google indexed to a customer's forward Custom Products revenue with a $120B implied ceiling and no purchase obligation.

    Next week: At least one additional silicon vendor or major hyperscaler counterparty discloses a residual-value guaranty, warrant, or take-or-pay-linked equity instrument tied to future purchasing behaviour in an SEC or equivalent filing by October 31, 2026. Falsified if the next eight weeks produce no such filing and no vendor commentary confirming the trend at a scheduled earnings event.

  • Inductive2 weeks observed

    Enterprise agent surfaces are converging on identity inheritance across the agent boundary as the default authentication pattern, deprecating service-account-based orchestration as an exception rather than a routine choice.

    • W33: Salesforce previewed Headless 360 architecture ahead of Dreamforce and Grok Bot demonstrated the front-door integration pattern with per-user consent scoping.
    • W34: Anthropic Skills API GA runs skills in the hosted sandbox under inherited identity; Salesforce, UiPath and Google shipped identity-inherited MCP surfaces on August 19-20; Alipay released agentic commerce with per-user consent scoping.

    Next week: At least one major enterprise vendor publishes a deprecation notice or migration path for service-account-based agent orchestration by December 31, 2026, or two additional platforms ship inherited-identity MCP surfaces in Q4. Falsified if two consecutive quarterly release cycles produce new agent surfaces built primarily on service-principal patterns without an inherited-identity alternative.

Second-order effects

  • Trigger: NVIDIA disclosed a $105B cumulatively capped residual-value guaranty as an Item 2.03 off-balance-sheet arrangement with exhibits pending in the 10-Q for the quarter ended July 26, 2026.

    The rating agencies now have to model contingent-obligation exposure on the GPU vendor whose credit was previously priced almost entirely off product margins. Even without a call event, the disclosure changes the ratings math on future issuance and on any hyperscaler counterparty that would sit under the same guaranty envelope — which will show up as tighter documentation on the next comparable lease before it shows up as a downgrade.

    Horizon: 3-9 monthsWho moves: NVIDIA rating agencies, SB Energy investors, hyperscaler tenants negotiating similar leases, insurance underwriters writing residual-value cover, and neocloud borrowers whose next facility now has a public comparable for the guaranty market
  • Trigger: Marvell's warrant to Google indexes silicon vendor equity to a customer's future purchasing decision with 240 discrete revenue triggers and no purchase obligation.

    The warrant becomes a template. Broadcom, Astera Labs, Alchip and every custom-silicon vendor selling into a hyperscaler will be offered — or asked to accept — a similar instrument on the next major supply agreement. Enterprises buying at scale will either standardize the vesting mechanics in their term sheets before the industry converges on a vendor-friendly version, or inherit whatever the first two or three major deals set as the norm.

    Horizon: 6-18 monthsWho moves: Custom-silicon vendors, hyperscaler procurement organizations, buy-side analysts pricing merchant-versus-custom share, and legal/tax teams designing the next generation of long-term supply contracts
  • Trigger: Pennsylvania Executive Order 2026-05 binds GRID requirements to DCs above 25 MW and conditions the state tax exemption on GRID compliance, adding a permitting-side stage to the largest US hyperscale market's site pipeline.

    Site selection for above-25-MW campuses shifts materially against Pennsylvania and toward ERCOT, MISO and Nordic markets. The demand does not disappear; it routes around the state, which lengthens effective queue times in less-friction markets by adding demand and puts fabric planning under further pressure to design for multi-site training runs that cross ISO boundaries.

    Horizon: 6-24 monthsWho moves: PJM-region developers, hyperscaler site-selection teams, ERCOT and MISO grid operators anticipating displaced demand, and network operators being asked to design more inter-DC fabric than the current physical layer supports

Strategic outlook

The twelve-month posture this week argues for is to treat the residual-value bet on AI infrastructure as a public disclosure item rather than a private credit assumption, and to write it into vendor term sheets before the market standardizes on the version that is most convenient to the issuer. NVIDIA has moved the residual-value line inside its own contingent-obligation footnote, Marvell has moved it into a warrant vesting mechanism, and Nebius has extended the neocloud tenor stack another $5B against still-shorter contracts — three different vehicles, one shared structural feature. For a board setting AI infrastructure strategy into 2027, the useful diagnostic is not capacity or capability but disclosure surface: which counterparties in your stack have written the residual-value bet down publicly, in what form, on what schedule of exhibits. If yours has not, it is either not thinking about it or it is not telling you. On the shorter-tenor side, the agent-tooling layer just re-based on Anthropic's Skills API and computer use GA — a decision that will look either premature or obviously right depending on whether OpenAI ships a comparable versioned-skill surface inside the quarter — and the OpenAI Sol promotional window has published its own reversal. Contracts in that layer should be dated, indexed, and written with the assumption that both the price and the benchmarked artifact can change without notice. Pennsylvania's EO 2026-05 is a reminder that the physical layer still owns the schedule, and the last twelve months of net site displacement out of PJM will be measurable well before any of the near-tenor layers rebase.

Where we differ

Our read against the field.

4 top-tier positions engaged, 3 disagreements on the record.

The best analysts covered this week too. Here is what they said, what we borrow with credit, and where our read genuinely departs from theirs — on the record, so you can score us later.

  • Their take: NVIDIA has effectively pledged $105B to guarantee the SB Energy PORTS-Pike campus, becoming the financier behind OpenAI's next 20-year home.

    Our read: The framing collapses two important distinctions. The $105B is a cumulative cap on a contingent obligation triggered by tenant default across a 20-year lease with obligations subject to ready-for-service starting 2028 — it is not a $105B pledge and it is not cash out the door. And 'becoming the financier' overstates the shift: the developer is SB Energy, the tenant is an OpenAI affiliate, and NVIDIA is the residual-value underwriter on the lease alongside a $1.5B equity investment in SB Energy. The structural point is real and larger than the framing: the accelerator vendor now sits inside the contingent-obligation footnote for land-power-shell, which is a rating question. But the number and the identity of the underwriter both need the primary-filing precision.

  • Their take: The July NVIDIA-OpenAI announcement was worth $250B; the filing confirms the deal at that scale.

    Our read: The July $250B figure was a press narrative that the filed 8-K does not confirm. The filed number is a $105B cumulative cap on a contingent obligation for the initial commitment, with an optional additional ~3.8 GW capacity NVIDIA could support but has not committed to. Anyone quoting $250B this week is quoting a July press figure that the primary filing revised downward and conditioned, and the revision matters because the ratings implication runs off the filed number rather than the press one.

  • Their take: Cerebras CS-4 at 750 vendor-claimed PFLOPS is roughly a 30x lift over Blackwell-Ultra rack performance on comparable workloads.

    Our read: The 30x framing is a vendor comparison against a specific reference configuration and specific workload, not an independently reproduced measurement. Cerebras's architectural step is genuine and the shipping timeline is credible; the specific multiple over Blackwell-Ultra depends on assumptions about model, precision, batch size, and utilization that are not visible from the marketing page. We are not calling this either way until independent measurement lands. On-prem operators evaluating wafer-scale should schedule a measured comparison rather than accept the vendor multiple.

  • Their take: Marvell's warrant to Google is a $120B silicon supply agreement.

    Our read: The $120B figure is the implied ceiling of the vesting mechanism (240 tranches × $500M revenue trigger), not a booked supply agreement and not a Google purchase commitment. Google has no obligation to source at that scale, and the tranches only vest as attributed Custom Products revenue actually arrives across Q3 FY2027 through FY2033. Calling this a $120B supply deal reads the maximum of a forward option as an executed contract. The right framing is that Marvell issued Google an optionality vehicle indexed to a customer's future purchasing — a real instrument, worth watching, but not a booked commitment.

Early warning panel

The levers we monitor.

10 metrics tracked — 0 rising, 1 falling, 9 steady.

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

  • Frontier lab cash runway at current burn

    ~30-40 months, unchanged — no lab closed primary financing in-window. Anthropic's hire of former Google TPU founder Amir Salek for in-house silicon is a spend-side commitment that turns forward cash needs upward without disclosing a delta; OpenAI's Sol promotional cut and EU ads rollout are demand-side moves that do not change disclosed runwayvs ~30-40 months, unchanged — no lab closed primary financing in-window. OpenAI's ~$7B transaction was an employee tender at a flat $852B mark, which supplies liquidity to shareholders rather than capital to the business

    Threshold: Below 18 months for any top-four lab

    What this measures

    Measures how long the frontier labs can sustain current burn without new capital. The category continues to ship without needing to price new equity, which argues it is not currently capital-constrained. The Salek hire is the highest-signal cost commitment of the week and will feed into future runway estimates when a program is disclosed; for now it is a spend-side variable without a number attached.

  • Hyperscaler AI capex to disclosed AI revenue ratio

    ~3.6x, unchanged — no hyperscaler reported or revised guidance in-window. NVIDIA's residual-value guaranty on the SB Energy PORTS-Pike site is a contingent obligation at the accelerator vendor, not hyperscaler capex, and does not enter the numerator until it is called; Google's Marvell warrant is silicon vendor equity indexed to future purchasing rather than a capex disclosurevs ~3.6x, unchanged — no hyperscaler reported or revised guidance in-window; the last moves were the late-July earnings cycle and nothing this week touched either side of the ratio

    Threshold: Above 6x sustained for two consecutive quarters

    What this measures

    No new denominator disclosure this week. The ratio remains an estimate built on committed capital against a revenue proxy since no hyperscaler breaks out AI-attributable revenue. Two structural filings changed how the residual-value assumption behind the numerator should be modelled without moving the ratio itself — hyperscaler capex is now cross-supported by a GPU vendor contingent line and a silicon vendor optional-equity vehicle.

  • CoreWeave contracted revenue backlog

    $104.2B as of June 30, unchanged with no CoreWeave reporting event in-window — the next scheduled print is the Q3 filing. Nebius's $5B convertible print on August 19 is the category's parallel move and confirms the tenor pattern that made backlog conversion the binding lever, without changing the CoreWeave numbervs $104.2B as of June 30, up 4.8% sequentially from $99.4B and 246% year over year, excluding more than $25B of commitments added in early Q3 — but conversion crossed the threshold: FY2026 revenue guidance of $12.4-13.2B implies 11.9-12.7% annual conversion, below the 15% floor

    Threshold: Sequential decline, or conversion below 15% annually

    What this measures

    Backlog is the neocloud category's core collateral and its conversion rate is the number its financing implicitly assumes. Held flat this week because the primary issuer had no filing event; the category's structural picture remains what W33 documented, with Nebius's convertible print reinforcing rather than changing it. Next filing-driven update is expected with the Q3 print.

  • NVIDIA quarter-over-quarter data center revenue

    $75.2B for Q1 FY27, unchanged with no earnings event in-window — the Q2 print and guide land August 26, and this week's 8-K contingent-obligation disclosure raises a separate question about how the residual-value line is footnoted in the accompanying 10-Q exhibitsvs $75.2B for Q1 FY27, unchanged with no earnings event in-window — the Q2 print and guide land August 26, and reports of reduced Rubin Ultra memory configurations make delivered content per GPU the variable to watch alongside unit volume

    Threshold: Two consecutive quarters of sequential decline

    What this measures

    The cleanest read on whether AI infrastructure demand is still compounding. August 26 is the scheduled event that resets this lever; the additional watchpoint this week is the residual-value guaranty disclosure and whether the accompanying 10-Q exhibits translate the SB Energy commitment into a per-quarter contingent-obligation accrual disclosure. The lever moves on revenue, but the auditability of the residual-value bet now moves alongside it.

  • Open-weight to closed-model capability gap on coding

    Still narrowed on paper and widened in practice. Ornith-1.5 arrived MIT-licensed at 397B/35B-active with vendor-run coding and terminal benchmarks against DeepSeek V4 Pro and GLM-5.3, but with no independent tracker ranking — a third consecutive week where a claimed open-weight advance did not clear the substitutability testvs Narrowed on paper and widened in practice. DeepSeek's 0813 build reports DeepSWE at 62.7 against the April build's 12.8, and GLM-5.3 claims the best open-source Terminal Bench 3.0 result — but neither model's weights were obtainable in the benchmarked form this week, so the gap that closed is between closed models and models you cannot yet download

    Threshold: Open weights within 2 Index points of the closed leader

    What this measures

    Measures whether a self-hosted model can substitute for a frontier API. Held flat again this week because Ornith's headline scores are vendor-run and no independent leaderboard has ranked the model, and moving the lever on a vendor harness would restart the substitutability problem the metric was built to avoid. See The Model Pulse for the full Ornith read.

  • Sovereign AI program commitments

    ~15 programs and ~$186B, unchanged — no new national program announced in-window. Pennsylvania's Executive Order 2026-05 is a state-level regulatory posture rather than a compute commitment; the White House NSSTS action on August 18 continues to circulate in secondary coverage without a primary text supporting the 'data centers removed from CET list' framingvs ~15 programs and ~$186B, unchanged — no new national program announced in-window. The week's state-level activity was regulatory rather than capital: Colorado opened ADMT and chatbot rulemaking and Taiwan's digital ministry confirmed AI agents were used in attacks on government sites

    Threshold: Above 20 programs or $250B committed

    What this measures

    Tracks state-directed AI capital as distinct from corporate capex. For a third consecutive week state capability showed up as governance and regulatory motion rather than compute commitments — a consistent enough pattern to name and continue to track without moving the number. The PA EO is a friction-adding rather than capital-committing rule and belongs in the interconnect lever, not this one.

  • PJM capacity auction clearing price

    $325.00 per MW-day for 2028/29, unchanged with no auction and no in-window filings. Pennsylvania EO 2026-05 does not change the clearing price but does add a permitting-side stage that compounds the queue length behind the same pricevs $325.00 per MW-day for 2028/29, unchanged with no auction and no in-window filings

    Threshold: A second consecutive auction clearing at the cap

    What this measures

    The clearest market price for grid scarcity in the largest US market. No new auction data this week; the next scheduled read is the following auction. Carried forward unchanged rather than estimated, with the note that state-level permitting changes affect queue length behind the same price rather than the price itself.

  • Time from interconnection request to energization

    60-84 months, held flat with a lengthening bias. PA EO 2026-05 adds GRID compliance as a binding stage for DCs above 25 MW, removes qualifying DCs from Permit Fast Track and conditions the tax exemption on GRID — regulatory friction lengthening the effective queue for the largest US hyperscale market rather than shortening itvs 60-84 months, unchanged — no new interconnection data in-window. Cisco's call commentary that campuses are fragmenting under power limits, driving scale-across optics demand, is indirect evidence the queue is shaping architecture rather than shortening

    Threshold: Below 48 months in two or more major queues

    What this measures

    The hard limit on how fast AI capacity can actually arrive. Held flat quantitatively because no queue-length figure was published in-window, with the explicit lengthening bias attached because the largest-market policy change this week added a permitting stage to sites above 25 MW. If any PJM-region operator publishes an updated queue estimate reflecting PA EO 2026-05, this lever should move down (longer) rather than flat.

  • Cost per task, frontier reasoning model

    Falling further at the top of the market on a scheduled cut: OpenAI reduced GPT-5.6 Sol pricing by more than 20% on August 21 for a dated three-month promotional window with the reversal date published at launch. The efficiency component of the lever is unchanged; the rate-card component now has three published increases within six monthsvs Falling sharply at the top of the market for the first time since April, on turn efficiency rather than rate cards: Artificial Analysis measured Grok 4.6 resolving long-horizon agentic tasks in ~53 turns and ~0.5B input tokens against ~103 turns and ~2.0B for Claude Opus 5, at $2 / $6 against Sol's $5 / $30 — but two published increases land within five months

    Threshold: A frontier-tier reasoning model below $1 per million output tokens

    What this measures

    Tracks real unit economics rather than headline token prices. Moved down again this week on a promotional cut with a published reversal, which continues to argue that the rate-card component of the metric is a schedule rather than a market. Efficiency component (turns per task) has not moved and remains the durable input; procurement should treat the two components as separate durations.

  • Custom silicon share of hyperscaler AI compute

    ~34-37%, held flat with a directional-up note. Marvell's warrant to Google indexes ~7% of the silicon vendor's equity to Custom Products revenue on a $120B implied ceiling and no purchase obligation — a structural argument for share moving upward on Google's roadmap, but the vesting mechanism is forward and not yet a booked shipmentvs ~34-37%, unchanged — reports of Microsoft seeking 300,000-plus Maia 300 units from TSMC would move this materially if confirmed, but the company disputed the reported figures and capacity talks are not booked wafers

    Threshold: Above 45% share

    What this measures

    Measures how much hyperscaler AI compute escapes merchant accelerator pricing. Held flat because the Marvell warrant is a forward-vesting instrument on a customer's optional purchasing rather than a booked share change, and it belongs in the underwriting story rather than in the share number until Google reports Custom Products revenue attributable to it. Directional bias is up on the strength of the instrument's shape and the Anthropic Salek hire in the same week; wait for FY2027 Q1 disclosure before moving.

Predictions

What we expect next.

7 predictions for the next 30-90 days, confidence 22%-72%.

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

Prediction 01

72%

confidence

Hardware

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.

Deadline: By October 31, 2026

Trigger: NVIDIA's next Form 10-Q filed on EDGAR containing exhibit or footnote text that specifies the guaranty structure disclosed in the August 17 8-K, including the aggregate cap, ready-for-service conditions, and identification of the tenant.

Prediction 02

34%

confidence

Hardware

Marvell discloses the first attributable Custom Products revenue tranche under the Google warrant agreement in a filed earnings release or subsequent 10-Q, by November 30, 2026.

Deadline: By November 30, 2026

Trigger: A Marvell Form 10-Q or earnings release identifying a $500M-plus Custom Products revenue amount and confirming a warrant tranche has vested under the Google agreement disclosed August 19, 2026.

Prediction 03

43%

confidence

Software

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.

Deadline: By September 15, 2026

Trigger: A Hugging Face repository under Z.ai containing model weights for GLM-5.3 distinct from GLM-5.2, dated on or before September 15, 2026.

Prediction 04

58%

confidence

Software

Anthropic or a named enterprise customer publishes an independently attributed Skills API production case study with a stated manual baseline and a measured post-Skills outcome (time, cost, or completion), by October 31, 2026.

Deadline: By October 31, 2026

Trigger: A published case study, press release or customer-authored blog post that identifies a named enterprise customer, cites a specific SOP or workflow, includes a numerical baseline for the pre-Skills-API state, and includes a measured post-implementation result attributable to Skills API rather than to a general workflow redesign.

Prediction 05

39%

confidence

Power

A second US state with major hyperscale exposure (Virginia, Texas, Georgia, Oregon, or Illinois) publishes a permitting or interconnection rule that adds a binding compliance stage above a stated MW threshold, comparable in kind to Pennsylvania Executive Order 2026-05, by December 31, 2026.

Deadline: By December 31, 2026

Trigger: An executive order, PUC ruling, or legislature-passed and governor-signed statute in one of the named states that binds an additional permitting, grid-compliance, or interconnection-review stage on data centers above a numeric MW threshold, published on or before December 31, 2026.

Prediction 06

22%

confidence

Hardware

Anthropic publicly confirms a tape-out or first-silicon milestone on the in-house accelerator program being led by Amir Salek, by June 30, 2027.

Deadline: By June 30, 2027

Trigger: An Anthropic release, blog post, executive statement at a scheduled event, or major-media report attributed to Anthropic that identifies a tape-out, first-silicon, or physical prototype milestone on the in-house accelerator program.

Prediction 07

47%

confidence

Networking

A second major networking or accelerator vendor (Broadcom, Marvell, Cisco, or Arista) publicly discloses production shipments of co-packaged-optics Ethernet or scale-up switches to a named hyperscaler by March 31, 2027.

Deadline: By March 31, 2027

Trigger: A press release, earnings statement, or product-page update from Broadcom, Marvell, Cisco, or Arista identifying a co-packaged-optics switch product shipping in volume to a named hyperscaler customer, distinct from NVIDIA's Spectrum-6 CPO family.

Track record

Scoring prior predictions.

11 prior predictions: 0 hit, 0 miss, 0 partial, 11 pending. Hit rate —.

11 predictions across issues so far. Hit rate: . Hits 0, misses 0, partials 0, pending 11.

Prediction 01

83%

confidence

Software

Artificial Analysis publishes Endpoint Accuracy Index results covering at least two models beyond the initial GLM-5.2, gpt-oss-120b and DeepSeek V4 Pro set by October 31, 2026.

Deadline: By October 31, 2026

Trigger: Published Artificial Analysis Endpoint Accuracy Index pages or articles showing measured endpoint results for at least two models not in the launch set.

pendingNo Endpoint Accuracy Index expansion observed in-window. Deadline is still 10 weeks out. Ornith-1.5 is a strong near-term candidate for inclusion given the served-versus-benchmarked divergence its release documents.

Prediction 02

24%

confidence

Software

OpenAI publicly assigns its Astra model a final Preparedness Framework cybersecurity rating of Critical by December 31, 2026.

Deadline: By December 31, 2026

Trigger: An OpenAI system card, Preparedness Framework update, or official post stating that Astra has been assessed at the Critical cybersecurity capability level, as distinct from the possibility not being ruled out.

pendingNo Astra rating published in-window. OpenAI's release cadence this week (Sol pricing and ChatGPT ads) did not touch Preparedness Framework disclosure. Deadline is 18 weeks out.

Prediction 03

31%

confidence

Networking

A major model-serving platform or AI gateway publishes per-endpoint accuracy, precision, or output-token-limit disclosures for the open-weight models it serves by January 31, 2027.

Deadline: By January 31, 2027

Trigger: Public documentation from Azure AI Foundry, Amazon Bedrock, Google Vertex AI, or a major independent gateway disclosing per-endpoint serving configuration or measured accuracy against reference weights.

pendingNo gateway disclosure in-window. The need got sharper this week: Ornith-1.5 is now the second open-weight release in a month whose only benchmark numbers come from a vendor harness.

Prediction 04

46%

confidence

Hardware

AMD publicly names a Taalas-derived product or roadmap item tied to a specific model or model class by June 30, 2027.

Deadline: By June 30, 2027

Trigger: An AMD announcement, roadmap disclosure, or earnings statement naming a model-specific inference product derived from Taalas technology, with an identified model or model family.

pendingNo AMD announcement in-window; the acquisition is not expected to close until Q4 2026. Deadline is over ten months out.

Prediction 05

44%

confidence

Capital

At least two frontier labs publish network isolation or containment requirements for third-party cyber evaluation partners by January 31, 2027.

Deadline: By January 31, 2027

Trigger: Published policy documents, system cards, or safety framework updates from two or more of OpenAI, Anthropic, Google DeepMind, Meta or xAI specifying containment or network isolation requirements for external evaluation vendors.

pendingNo published containment policy from any lab in-window. Anthropic's Skills API GA sandboxes skills but does not address third-party evaluator containment.

Prediction 06

27%

confidence

Power

Alphabet discloses the size of its investment in Discovery Loop in an SEC filing or official release by December 31, 2026.

Deadline: By December 31, 2026

Trigger: An Alphabet 10-Q, 10-K, or official press release stating a dollar figure for its investment in Discovery Loop.

pendingNo Alphabet disclosure in-window and no Alphabet reporting event in the window. The next scheduled opportunity is the Q3 10-Q.

Prediction 07

64%

confidence

Software

DeepSeek publishes the V4-Pro-0813 build weights to Hugging Face by September 30, 2026.

Deadline: By September 30, 2026

Trigger: A Hugging Face repository under the DeepSeek organization containing a model card or config identifying the 0813 build, distinct from the April 2026 preview weights.

pendingNo 0813 weights on Hugging Face in-window. DeepSeek's in-window release was V4-Flash-Vision-Exp, an API-only experimental multimodal model — evidence that the vendor's current release cadence favours served endpoints over downloadable weights.

Prediction 08

69%

confidence

Capital

A second publicly traded neocloud discloses, in an SEC or equivalent filing, GPU-backed debt whose maturity extends beyond the stated weighted-average or characteristic tenor of the customer contracts securing it, by December 31, 2026.

Deadline: By December 31, 2026

Trigger: A 10-Q, 10-K, 8-K, 6-K or equivalent filing from a neocloud other than CoreWeave disclosing both a facility maturity and a contract tenor where the former exceeds the latter.

pendingNebius priced $5B of convertibles due 2030 and 2034 on August 19 in a filed 6-K, which is directionally consistent with the prediction. Scoring remains pending pending confirmation that the accompanying disclosure explicitly compares facility maturity against contract tenor as required by the trigger; the base convertible print does not, on its own, meet the specific comparison language.

Prediction 09

37%

confidence

Hardware

NVIDIA publicly confirms a Rubin Ultra memory configuration at or below 512GB, or an 8-high HBM4E stack option, in official specifications or an earnings disclosure by March 31, 2027.

Deadline: By March 31, 2027

Trigger: An NVIDIA product page, technical brief, GTC announcement, or earnings-call statement specifying a Rubin Ultra SKU with 8-high HBM4E stacks or total memory at or below 512GB.

pendingNo Rubin Ultra memory specification confirmed in-window. NVIDIA's August 26 Q2 FY2027 earnings call is the next scheduled opportunity for such a confirmation; this week's 8-K was about a residual-value guaranty rather than product specifications.

Prediction 10

61%

confidence

Networking

The OCP Open Silicon Photonics for AI Systems workstream submits its first specification by December 31, 2026, meeting the Q4 2026 target stated at launch.

Deadline: By December 31, 2026

Trigger: A specification document published or formally submitted to the Open Compute Project under the Open Silicon Photonics for AI Systems workstream, dated on or before December 31, 2026.

pendingNo OCP specification submission observed in-window. Hot Interconnects 2026 confirmed NVIDIA CPO shipping in production without an equivalent multi-vendor spec landing; the OCP workstream remains on its Q4 target with no interim milestone this week.

Prediction 11

26%

confidence

Software

A frontier lab publishes turn count or task-completion cost as a headline metric alongside benchmark scores in an official model card or launch post by January 31, 2027.

Deadline: By January 31, 2027

Trigger: An official model card, launch blog post or documentation page from OpenAI, Anthropic, Google, Meta or SpaceXAI reporting average turns, average tokens per completed task, or cost per completed task as a primary reported metric rather than as third-party commentary.

pendingNo lab-authored turn-count metric published in-window. OpenAI's Sol announcement was rate-card driven; Anthropic's Skills GA release quotes a vendor-relayed customer time-savings figure but does not adopt turn count as a reported model metric.

Track record

The full ledger, misses included.

29 of 94 predictions resolved: 7 hit, 13 partial, 9 miss, 6 overdue.

Every prediction this publication has ever made, scored against its own written trigger when the deadline passes — ambiguity resolves against us. Overdue means we haven’t adjudicated yet; it stays visible until we do.

94

predictions made

47%

hit rate (partial = half)

0.159

Brier score (0 = perfect)

6

overdue, unresolved

Calibration by confidence band

  • Bold (<55%)

    No resolved predictions yet — a gap the craft rules now force us to fill.

  • Core (55-80%)

    29 resolved · hit rate 47% vs mean confidence 66%

  • High-conviction (>80%)

    No resolved predictions yet — a gap the craft rules now force us to fill.

Recently resolved

  • partial80% 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.

  • hit66% called

    Samsung's HBM4 supply to NVIDIA is publicly confirmed — via earnings call, company statement, or multi-source supply-chain reporting — by August 31, 2026.

    Hit on the multi-source-reporting trigger: Korean press (Seoul Economic Daily, Korea Herald) reported alongside Samsung's record Q2 guidance that HBM4 — in mass production since February for NVIDIA's Vera Rubin — reached $1B in sales within four months. Caveat: Samsung's Jul 30 divisional results would make it unambiguous from the company itself.

  • hit62% called

    GPT-5.6 reaches broad GA with the Terra tier priced at or below $2.50/$15 per MTok — half of GPT-5.5's rate — confirming a closed-lab repricing cycle rather than a one-off Sonnet 5 cut, by August 31, 2026.

    Hit, seven weeks early. GPT-5.6 went GA Jul 9 with Terra at exactly $2.50/$15 per MTok. Grok 4.5's $2/$6 launch the day before makes it a three-vendor repricing cycle (Sonnet 5, Terra, Grok 4.5), not a one-off.

  • hit66% called

    At least one major enterprise platform ships an admin control specifically for scheduled/background coding or app-building agents by August 31, 2026.

    Hit. GitHub shipped Copilot agent session streaming to public preview (Jul 2) — SIEM/Purview streaming of all agent sessions — on top of its agent control plane, and GitHub also added AI-credit session limits covering background agents (Jul 1, per Agent Techniques coverage).

  • partial65% called

    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.

  • partial60% called

    Expanded Beam Optical MSA publishes a v1.0 spec within 90 days of launch (May 12), with at least one in-production deployment announced by a hyperscaler member (AMD, Cisco, Meta, Oracle).

    EBO MSA membership expanded 17 to 23 vendors May 18 (HPE marquee addition, Bellwether, JPC Connectivity, Mixx, TIME, TFC). v1.0 spec not yet published. Member growth is positive signal but spec + in-production deployment still pending. On track.

Watchlist

On the radar this week.

7 catalysts to watch, starting Aug 24.

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

  • Aug 24

    ChatGPT advertising rollout across 31 EU countries begins

    The first live deployment of ChatGPT ads in the EU market — a direct test of whether consumer-tier AI monetization can be built on advertising rather than subscription in a jurisdiction with distinct data-privacy expectations. Watch complaint volume, national regulator response, and any adjacent product surfaces (Teens, Autopilot) that reprice on the same schedule.

  • Aug 26

    NVIDIA Q2 FY2027 earnings and 10-Q exhibits

    The single largest scheduled datapoint of the quarter, and this time the residual-value guaranty disclosure changes what the print measures. Watch for 10-Q exhibits translating the SB Energy PORTS-Pike guaranty into per-quarter contingent-obligation disclosure, guidance commentary on Rubin Ultra memory configurations, and any change in framing on the ~$150-200B revenue-opportunity language attached to the transaction.

  • Sep 1

    OpenAI Daybreak hardware security key requirement takes effect

    Physical-authentication gate on frontier model access lands. Whether Anthropic or Google adopt equivalent controls within the quarter determines whether physical authentication becomes a norm for dual-use agent access or stays an OpenAI-specific control. The Skills API GA raises the stakes because more capability is now behind the identity gate.

  • Sep 15

    GLM-5.3 weights promised window closes

    Z.ai's August 14 announcement promised weights within two weeks of the release. September 15 is the outside edge of that promise plus a two-week buffer. If GLM-5.3 weights are still not on Hugging Face by that date, the pattern of announced-but-unavailable open-weight releases moves from a curiosity to a category-level trend, and the substitutability question this publication has been holding open should be reopened.

  • Sep 30

    DeepSeek V4-Pro-0813 weights deadline (Prediction p83)

    The scoring window for the outstanding prediction from W33 closes. If the 0813 build weights are on Hugging Face by September 30, the served-versus-downloadable divergence remains a timing artifact; if not, DeepSeek's release cadence has structurally changed toward API-only endpoints and the substitutability lever this publication tracks needs to be revised.

  • Nov 21

    Dated Sol reversal exposes November business cases

    The three-month scheduled step-up on the more-than-20% cut lands. Any November-priced agent business case must be modelled against the reversal, and watch whether a competing frontier vendor uses the moment to publish a comparable dated cut. Full architecture read lives in the Model Pulse.

  • Q4 2026

    OCP Open Silicon Photonics first specification submission

    The stated target from the nineteen-company coalition launched at OCP APAC. Meeting it puts a multi-vendor CPO path on the table for 2027-28 refresh cycles; missing it effectively hands the next fabric refresh generation to single-vendor co-packaged optics by default. This is the fork in the road that the Big Story identifies as needing an explicit written decision by any operator refreshing fabric inside eighteen months.

Companion reads

The rest of the spine.

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

Edits this issue

  • House-original claim moves from framework restatement to filing translation this week: the duration-mismatch thesis introduced in W33 now sits inside NVIDIA's own 8-K under Item 2.03, and the issue leads with the residual-value credit-support-per-IT-GW derivation from the primary filing rather than with the circulating framings.
  • Predictions carry a mandatory courage entry per the standing rule: p89 (Marvell Google first tranche) sits at 34% confidence and p93 (Anthropic in-house silicon tape-out by June 30, 2027) sits at 22%, both outside the 55-80% comfort band on purpose.
  • The interconnect lever gains an explicit 'lengthening bias' note because PA EO 2026-05 adds a permitting stage without producing a numeric queue-length update — flat with a bias is more honest than moving the range on a rule that has not yet been measured in a queue.
  • signalVsNoise adds an explicit '1' entry (SB Energy $50B IPO rumor) rather than assuming a lowest-tier signal is present by implication — see the AUTHORING.md noise-call bar.

About this brief

Compiled from public announcements, SEC filings, earnings transcripts, and official lab and vendor publications. Every quantitative claim is graded 1–5 on source quality. Claims graded 2 or below are flagged as noise. The thesis the brief defends is published separately and updated only when a hypothesis materially changes.

Authorship

Written by Brian Letort. Independent analysis. All sources cited are public. Not investment guidance.

Operate. Publish. Teach.