AI Capital Prediction Note — 2026-07-31

Correction probabilities

• 6 months: 40% ↑
• 12 months: 71% ↑
• 24 months: 85% ↑

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:

The strongest new evidence FOR the forecast

1. The bond market has broken from equity on AI infrastructure survival. CoreWeave's 5-year CDS hit 855 basis points on July 29 — a standard pricing model yields ~50% cumulative default probability. Meanwhile, 21 equity analysts maintain Buy/Strong Buy ratings with an average price target of $145.76 against a closing price of $61.53. This is not a disagreement about magnitude. It is two markets pricing incompatible realities. When credit and equity diverge this sharply, credit has historically been correct first (2007 MBS vs. equity analyst ratings, 2008 CDS vs. bank stocks). The bond market is saying: the largest GPU neocloud — the entity that buys Nvidia chips with Nvidia-financed debt to serve Nvidia-backstopped capacity — has a coin-flip chance of going bankrupt within five years.

2. OpenAI cut Luna API prices by 80% and Terra by 20% on July 30. This is simultaneously the most encouraging and most alarming signal of the week. Encouraging because the efficiency gains are genuine — 20% reduction in end-to-end serving costs, 15%+ improvement in token generation efficiency, achieved by GPT-5.6 Sol autonomously rewriting production kernels. Alarming because you do not cut prices 80% when demand is absorbing supply. This is the telecom playbook: build capacity, cut prices to fill it, discover that revenue per unit falls faster than unit volume rises. With OpenAI spending ~$50B on compute in 2026 against ~$13B in revenue, an 80% price cut on its highest-volume tier compresses the revenue trajectory precisely when the financing gap is widest.

3. Meta's free cash flow collapsed 91% year-over-year. Meta reported Q2 2026 revenue of $60.8B (beat) but EPS of $6.18 (missed $7.10 by 13%), with FCF falling 91% due to AI capex. 2026 capex guidance was raised to $130-145B. This is the clearest data point yet that AI capex is eating hyperscaler cash flow — not in theory, not in projections, but in a reported quarter. When a company generating $60.8B in quarterly revenue sees its free cash flow nearly vanish, the question of "who pays for the buildout" moves from abstract to existential. Meta's response — launching Meta Compute to sell excess GPU capacity — is itself a threat to CoreWeave, the neocloud whose CDS is already at 855 bps.

4. Moody's formally warned that AI spending threatens the credit quality of six companies carrying $460B in direct debt and $1.2T in lease commitments. This is not a blog post or a tweet. It is an institutional credit rating agency putting Microsoft, Amazon, Alphabet, Meta, Oracle, and CoreWeave on formal notice. AI-linked CDS trading totaled ~$650M in Q2 2026, up 600% year-over-year (DTCC data). The credit market is building a real distress-pricing infrastructure around AI capital exposure.

5. Michael Burry publicly flagged Nvidia's CDS as "going parabolic." When the investor who called the 2008 housing crash posts about AI circular financing on the same day that Nvidia's CDS records its largest single-day move since the contracts launched, the signal is no longer confined to credit desks. Retail and institutional attention is broadening.

The strongest new evidence AGAINST the forecast

1. Hyperscaler AI revenue is genuinely accelerating. Microsoft: $90B revenue, Azure +43% YoY, $678B commercial backlog. Amazon: $200.6B revenue, AWS +37% ($42.2B, fastest growth in 18 quarters). Alphabet: Cloud $24.8B (beat by $4.8B). These are not speculative reservations or circular financing — they are reported quarterly revenue from the largest cloud platforms on Earth. AI contributed 22-26 percentage points of Azure's growth. The demand is real at the platform layer.

2. Anthropic achieved its first operating profit. $559M operating profit on $10.9B Q2 revenue (more than double Q1's $4.8B) — a 5% operating margin, arriving two years ahead of the timeline Anthropic had told investors the previous summer. Revenue run rate reached ~$30B by April/May. This is the first evidence that frontier model unit economics can work. Anthropic projects $17B in positive cash flow on $70B revenue by 2028. If the #2 lab can reach profitability, the "AI is a money pit" thesis needs qualification.

3. OpenAI's price cut unlocks genuine enterprise adoption. At $0.20/$1.20 per million tokens, Luna makes high-volume agentic workflows economically viable for the first time. Customer testimonials from Replit ("intelligence too cheap to meter"), Notion, Ramp, Cognition, and Dust describe real production use cases — tool-calling agent loops, coding automation, document analysis — that were not economical at previous price points. The deflationary spiral that hurts model provider revenue may simultaneously validate the demand layer by making AI cheap enough to deploy at scale.

4. The grid, not demand, remains the binding constraint. Gartner estimates global data center electricity demand will exceed 1,000 TWh in 2026 — more than Japan's entire consumption. ~50% of planned US AI data centers cannot connect to the grid. When the bottleneck is physical infrastructure rather than customer orders, the risk is delay and cost overruns, not a demand mirage. This caps the speed of oversupply buildout and provides a natural throttle.

Which layer became more fragile or more defensible

• More fragile — dramatically: Venture / Private Credit / Circular Financing. This layer has crossed from "structural concern" to "active credit market stress." CoreWeave at 50% default odds, Oracle one notch from junk, Nvidia's own CDS going parabolic, Moody's institutionalizing the warning, Burry publicly shorting — the credit market is building a distress-pricing apparatus around AI infrastructure debt. The self-reinforcing loop (wider CDS → higher borrowing costs → harder refinancing → wider CDS) is now live. Apollo's Torsten Slok warned that rising yields could force the AI capex cycle to "self-throttle" — and with the 10yr Treasury at 4.65% (98th percentile of its 12-month range), the macro backdrop is hostile.

• More fragile: Frontier Model Economics. OpenAI's 80% price cut is a pricing-power erosion signal. The $1.4T / 8-year data center commitment against $13B revenue means the gap between spend and income is widening, not closing. Anthropic's profitability is a counterweight, but Anthropic is spending $15B/year on Musk's Colossus data centers — its path to profit depends on continued revenue acceleration that may not survive a broader AI spending pullback.

• More defensible — significantly: Enterprise Adoption / Revenue. This is the strongest signal in the system. Azure +43%, AWS +37%, Google Cloud $24.8B, Anthropic's first profit. The price cut makes high-volume use cases viable. This layer has moved from "speculative" to "demonstrated" in the span of one earnings week. The caveat: revenue is growing at 37-43% while capex is growing at 77%. The gap is widening even as both rise.

• More fragile: Chips / Hardware (near-term). KOSPI halted, Nasdaq near correction, China producing advanced chipmaking equipment, $2.7T erased from semiconductor market caps since June 22. The long-term demand story (Samsung/SK $950B deals) remains intact, but the near-term is deteriorating rapidly.

Assumptions that changed

• From: Circular financing is a structural concern that could trigger stress if end-user demand disappoints.
  To: Circular financing is an active credit market stress event. CoreWeave's 855-bps CDS is not a warning — it is a price. The bond market is now explicitly pricing default risk across the AI infrastructure stack (neoclouds → Oracle → Nvidia). The question is no longer whether credit markets will price AI risk but how fast it spreads from neoclouds to hyperscalers.

• From: OpenAI's pricing power supports the frontier model revenue trajectory.
  To: OpenAI is in a deflationary pricing spiral. Cutting Luna 80% signals that demand is not absorbing capacity at previous price points. This compresses the revenue trajectory precisely when the financing gap is widest ($50B+ compute spend vs. $13B revenue). The telecom parallel is exact: build the network, cut prices to fill it, discover that revenue per unit falls faster than unit volume rises.

• From: Hyperscaler cash flow can absorb AI capex without strain.
  To: Meta's 91% FCF collapse proves that even the most profitable hyperscaler is being consumed by AI capex. The combined hyperscaler capex of ~$725B in 2026 (+77% YoY) is growing nearly twice as fast as cloud revenue (+37-43%). This is not sustainable without either a revenue acceleration miracle or a capex reset.

Underappreciated second-order consequence

The neocloud death spiral is the 2008 MBS cascade in disguise — and Meta just pulled the trigger.

CoreWeave's business model depends on its $21B Meta contract running through 2032. On July 17, Meta launched Meta Compute — a commercial GPU cloud selling excess capacity to third parties. Meta's $130-145B in 2026 AI capex means it is building far more GPU capacity than it needs internally. Selling the excess is rational. But it makes Meta both CoreWeave's largest customer and its most direct competitor.

If Meta pulls workloads in-house (economically rational given its own capacity), CoreWeave's $99.4B revenue backlog becomes partially phantom. As CoreWeave's CDS widens (already 855 bps), its refinancing costs rise, which worsens GPU economics, which widens CDS further. CoreWeave needs $31-35B in 2026 capex against $6.2B in trailing revenue — it cannot survive a refinancing shock.

If CoreWeave defaults, Nvidia inherits the GPU collateral — the chips it financed CoreWeave to buy. Nvidia then has to mark down that collateral on its own balance sheet, directly hitting its $5T market cap. This is structurally identical to 2008: the entity that created the financing structure (Nvidia = investment banks creating MBS) ends up holding the bag when the SPV (CoreWeave = SIVs) blows up.

The critical difference: in 2008, the "houses" behind the mortgages were real, physical assets in established markets. In 2026, the "assets" behind the GPU debt are data centers that take 3+ years to build, require grid connections that don't exist yet (~50% can't connect), and serve an end-user market whose demand trajectory is uncertain enough that the frontier model provider just cut prices 80%. The collateral is less liquid, less proven, and more temporally stranded than housing ever was.

What to watch next

• CoreWeave August earnings call: Will the company address the CDS crisis directly? Any guidance cut or customer concentration disclosure could be the trigger.
• Oracle credit trajectory: BBB- is one notch from junk. A downgrade to junk would force index fund selling and dramatically widen the refinancing window.
• Nvidia Q2 FY2027 earnings (expected August): The circular financing structure means Nvidia's revenue partly depends on customers it finances. If those customers slow purchases, Nvidia's revenue decelerates — which is simultaneously the cause and the effect of the credit spiral.
• Anthropic IPO pricing: A ~$965B valuation at IPO tests whether public markets will fund frontier labs at dot-com-level multiples. A weak IPO or a down-round would close a critical capital tap.
• OpenAI IPO timeline: If OpenAI's IPO is delayed or repriced downward, the entire financing stack loses its exit liquidity assumption.