AI Capital Prediction Note — 2026-08-03 - Artifacts

Correction probabilities

6 months: 41% ↑
12 months: 73% ↑
24 months: 86% ↑

Definition: A material AI-capital correction = a broad repricing, financing stress event, or capex reset large enough to break the assumption that demand will grow into the current $2.5T+ annual spend trajectory.

Forecast ledger trajectory:

Date	6-mo	12-mo	24-mo	BPI
Jul 12	31%	43%	57%	—
Jul 13	33%	45%	60%	—
Jul 14	30%	47%	63%	51
Jul 15	35%	52%	65%	56
Jul 16	34%	53%	66%	53
Jul 17	32%	55%	70%	52
Jul 18	31%	56%	72%	53
Jul 21	33%	61%	77%	57
Jul 25	35%	64%	80%	59
Jul 27	37%	67%	83%	61
Jul 31	40%	71%	85%	64
Aug 3	41%	73%	86%	65

Claim check: the great defection — verified, with receipts

Yesterday's brief flagged credit as the active stress. Today's lead is the demand side: reports that multi-million-dollar customers of OpenAI and Anthropic are switching to Chinese open-weight alternatives, driven by token costs and by the labs' habit of competing with their own customers. We ran the claims down. Verdicts:

Claim	Verdict	Evidence
Enterprise customers are switching to Chinese open-weight models on cost	CONFIRMED	Chinese models have held >30% of US token share on OpenRouter every week since Feb 8, peaking at 46% — vs. an 11% average over the prior 12 months and 4.5% in H1 2025 (CNBC, July 7). Lindy migrated 100% of its traffic from Claude to DeepSeek — "you could see that cost curve crash to the ground… millions saved within months," CEO Flo Crivello (CNBC, June 26). Vercel: GLM 5.2 was the fastest-adopted model of 2026 — 27× daily token growth, 80× customer growth in its first full week — landing within 1pp of Opus 4.8 on a watched agentic benchmark at ~1/5 the cost. Chinese open models run 60–90% cheaper than frontier APIs (OpenRouter). DoorDash is building on Moonshot ("better quality… cheaper cost" — CTO Andy Fang, Fortune, July 17); Cursor routes workloads to Chinese models.
Switching is motivated by the labs cloning customers' business models	BEHAVIOR DOCUMENTED; MOTIVE INFERRED — one clean case	Anthropic launched Claude Design (April) — a standalone product competing with Figma (80–90% market share), Adobe, and Gamma, at free/$20 pricing that undercuts Figma teams (The Information, VentureBeat, PYMNTS). Anthropic's Claude Science plus its own drug-development program (STAT, June 30) and the ~$400M Coefficient Bio acquisition (April, ex-Genentech team) put it in competition with pharma customers. The clean case: Claude Code competes with Cursor — and Cursor now routes to Moonshot. That is vendor-as-competitor converting directly into defection.
"Several multi-million customers" are at stake	DIRECTIONALLY RIGHT, SCALE UNPROVEN	Confirmed migrations are startups (Lindy), developer platforms (Vercel ecosystem), and traffic routers — real revenue, but not yet verified big-enterprise contract cancellations. The largest single signal: Microsoft is evaluating a Microsoft-hosted DeepSeek V4 as a lower-cost engine for Copilot Cowork — currently powered by OpenAI and Anthropic models (Rest of World). When the keystone partner hedges, the question stops being whether defections matter.
Bonus: the capability gap argument	WEAKENING FAST	Chinese models trail the US frontier by ~6–9 months (Brookings; CAISI's May eval of DeepSeek V4 Pro says ~8). For most production workloads, "8 months behind at 1/5 the price" is a trade procurement will take every time. And the security establishment already routes around US labs: Hugging Face contained OpenAI's rogue-model incident using Z.ai's GLM 5.2 after Anthropic's Fable 5 failed (CNBC, July 24).

Why this matters for the capital stack: the compute commitments propping up CoreWeave's facilities, Oracle's debt, and Nvidia's guarantees are underwritten on OpenAI/Anthropic revenue trajectories. Those trajectories assume premium token pricing. The defection data says premium pricing is being arbitraged on a measurable weekly time series — not by skeptics, but by the labs' own customers and partners.

The strongest new evidence FOR the forecast

The demand base is repricing its loyalty, quantifiably. The OpenRouter time series (30%+ weekly share, 46% peak vs. 11% a year ago) is the first hard, continuous measurement of the migration this model has tracked as a structural risk since July 15. Combined with OpenAI's 80% Luna cut (July 30), the pattern is unambiguous: frontier labs are cutting price to defend share against a supply curve that costs 60–90% less. That is a pricing war, and pricing wars are how financing gaps become financing events.
Vendor-as-competitor is now a documented switching motive. Claude Design vs. Figma, Claude Science vs. pharma, Claude Code vs. Cursor — and Cursor defecting to Moonshot. The labs' vertical expansion, pitched to investors as revenue diversification, functions in the market as a customer-repellent. Every new vertical Anthropic enters manufactures another cohort of motivated open-weight adopters.
CoreWeave's contradictions compound ahead of Aug 11. A $3.1B GPU-backed facility closed at Ba2/BB+ — junk ratings marketed as "asset-class maturity" while CDS prices ~50% default odds. Q2 insider selling surfaced: co-founder Venturo ~$734M, CEO Intrator ~$447M (10b5-1), plus a ~$407M Nvidia director sale. A ~60-day construction delay at Denton, TX (WSJ). $14B of Oracle-backed data-center debt repriced to higher yields. Nebius -10%, CoreWeave -9% in recent sessions. The equity/credit divergence is now joined by an insider/outsider divergence.
The bear frame has gone institutional-mainstream. Zitron's "OpenAI is the Lehman of the AI bubble" is now Quartz and Business Insider copy, with a specific ledger: $852B declared spend by 2030, ~$750B of it tied to partner performance obligations, $50B+ 2026 compute = over half of global AI compute spend by his math. Whether or not the math is exact, the adoption of this frame by mainstream outlets changes refinancing psychology — which is itself a transmission channel.
Even Anthropic's golden number has an asterisk. The $47B "annualized revenue" driving the $1.08T secondary consensus is a run-rate, not contracted ARR (The Information). At the valuation the IPO market is being asked to absorb, the difference matters.

The strongest new evidence AGAINST the forecast

Chip demand is printing records, not reservations. Global semiconductor sales hit $120.6B in May, +104.1% YoY, the 15th consecutive monthly record. Broadcom guides AI semi revenue +200% YoY to $16B this quarter; Micron guides a $50B quarter. This is physical, shipped demand — the strongest bull evidence in the system.
Anthropic's slope is still extraordinary. $30B annualized in April → ~$47B now, with a first operating profit already banked, a $1.08T secondary consensus, and Amazon holding $100B+ in decade-long commitments including up to 5GW of capacity. If the #2 lab monetizes at this slope, frontier economics are not uniformly broken — and the Anthropic IPO could reopen the capital tap for the whole stack.
OpenAI's revenue is real and growing — on corrected numbers. ~$25B annualized by mid-2026 (vs. the $13B audited FY25 figure this ledger previously carried), enterprise at 40% of revenue targeting 50% by year-end, S-1 confidentially filed June 8. The $50B compute spend against $25B annualized revenue is a 2:1 gap — enormous, but half the ratio previously assumed here. See the ledger correction below.
Defection is migration, not destruction. Tokens routed to GLM, Kimi, DeepSeek, and Qwen still get consumed — and cheaper tokens historically mean more total consumption (Jevons). The demand layer for compute survives the pricing layer's collapse. That caps the downside for chip volume, even as it guts frontier-lab margins.
Credit is still being supplied. A $3.1B facility closed for CoreWeave, at junk ratings — but it closed. Private credit keeps underwriting the buildout. A market pricing 50% default odds that still lends is a market that has not yet decided.

Which layer became more fragile or more defensible

More fragile: Frontier Model Economics. Pricing power is being arbitraged from below (open weights at 1/5–1/10 the price) and from above (hyperscaler self-supply, Microsoft's DeepSeek hedge). The labs' answer — vertical integration — accelerates the very defections it needs to stop. Anthropic's profit slope is the only counterweight, and the entire IPO market is about to price it.
More fragile: Venture / Private Credit / Circular Financing. Insider selling, junk-rated GPU collateral expansion, Oracle debt repricing, and an Aug 11 earnings tripwire with CDS at 855 bps. The self-reinforcing loop is live.
Newly active layer: Chinese open-weight supply — now the marginal price-setter for good-enough intelligence. This is the day's structural upgrade: from parallel curve to active driver.
More defensible: Chips / Hardware (near-term demand) and Enterprise Adoption (as volume). Neither helps the financing layer, which is priced on rent, not tokens.

Assumptions that changed

From: Chinese open-weight competition is a structural parallel supply curve worth monitoring. To: It is a quantified, active migration — 30–46% of routed US tokens, named companies, and the frontier labs' keystone partner hedging. It now sets the marginal price of good-enough intelligence.
Ledger correction (accountability): Prior notes used OpenAI's ~$13B audited FY25 revenue and the $1.4T headline commitment as operative figures. Corrected: ~$25B annualized revenue by mid-2026; the investor-facing compute target has been ~$600B by 2030 since February (CNBC), with $852B in declared intentions (Zitron's count). The gap is ~2:1, not ~4:1. Still unbridgeable without external capital — but this ledger now carries the honest ratio, and the correction cuts against the bear case it supports elsewhere. That's what the ledger is for.
From: Frontier-lab vertical expansion (apps, design, drugs) diversifies revenue and strengthens the story. To: It is a defection accelerant. Vendor-as-competitor is now a documented switching motive, compounding the price motive.

Underappreciated second-order consequence

The collateral behind the AI debt stack was never the GPUs — it was premium token pricing. And that is what is evaporating. The defection wave reroutes demand rather than destroying it, which sounds benign until you trace the underwriting: Oracle's OpenAI backlog, CoreWeave's GPU-backed facilities, Nvidia's guarantees were all priced on US-frontier pricing power. If GLM-class models set the marginal price of good-enough intelligence at 1/5 to 1/10 of frontier rates, the volume survives and the rent does not. Compute demand keeps growing; the layer that services the debt loses the pricing power the debt was priced on. And Microsoft's DeepSeek evaluation is the tell that the political firewall — the assumption that security politics would keep Chinese models out of the US enterprise perimeter — is already being negotiated away by the very company that backstops OpenAI. When the firewall goes, premium pricing loses its last structural defense.

What to watch next

CoreWeave Q2 earnings, Tuesday Aug 11: management tone on the Meta contract, the Denton delay, refinancing plans — against a CDS curve pricing 50% default odds. This is the week's tripwire.
OpenRouter / Vercel weekly share data: does Chinese-model share of US tokens cross 50%? The slope of this line is now the cleanest leading indicator of frontier revenue compression.
Microsoft's Copilot Cowork decision: hosting DeepSeek V4 would normalize Chinese models inside the US enterprise perimeter — and signal the keystone partner's confidence in OpenAI's pricing defense.
Anthropic IPO pricing vs. the run-rate question: does the market absorb $1.0–1.2T on annualized run-rate revenue, or does the ARR asterisk get priced?
OpenAI S-1 disclosures (confidentially filed June 8): the first audited look at the post-reset $600B spending plan against actual 2026 revenue.