Shadow debt jumps 90% in six months to $3.2 trillion as Anthropic files a loss statement ahead of IPO

Start with the most anomalous number. According to Nikkei's estimate based on the latest financial reports of five U.S. tech giants, their combined AI-related shadow debt is about $3.2 trillion, up roughly 90% in six months, with some figures estimated. This is not the leverage of any single company but an off-balance-sheet expansion by a group. Shadow debt is precisely the part of accounting most easily glossed over in PR releases: it does not enter cost of goods sold, it does not enter current profit, yet it genuinely consumes future cash flow. The market is celebrating model iterations while balance sheets quietly add leverage.

Anthropic states the bill more bluntly. Its S-1 discloses infrastructure spending commitments of up to $900 billion, with costs growing faster than revenue; investor Tomasz Tunguz, citing the same S-1/A, notes long-term obligations of $895.4 billion and a swing from profit to loss in the third quarter as short-term compute procurement costs surged. A company about to IPO admits in its prospectus that costs are outrunning revenue — something that would require explanation in any industry, yet in AI is waved away as an 'investment phase.' What officials have not disclosed: the payment schedule for the $900 billion commitment, the locked-in compute prices behind it, and the segment breakdown on the revenue side.

Bring the lens back home, and the supply-side math is just as tight. Bloomberg reported that Alibaba, Tencent, ByteDance and other Chinese firms collectively ordered about 6 million Nvidia RTX Pro 5500 chips but will receive only about 1% of deliveries over the next four months. In the same period, Nvidia required cloud service providers to cap China business at no more than 49% and ensure non-Chinese customers account for at least 51%; Oracle, under pressure from the Trump administration, canceled an AI chip leasing agreement with Tencent worth about $7 billion. Three threads point to the same thing: the 'rent' channel for compute is being narrowed by geopolitics, leaving only 'buy' and 'debt' — and both consume capital.

When computing power shifts from 'rent' to 'debt,' the payer of this feast has already changed from VC to the bond market.

A clear gap opened this week between management rhetoric and the language of filings. The public narrative is one of surging demand: ChatGPT search scale is about 20% of Google's, 250 million people use ChatGPT, and the company plans to merge Chat and Work; Google's Gemini 4 Carbon, according to Tom's Hardware, topped the intelligence leaderboard and edged ahead of Claude Opus 5.5; Tencent's Hunyuan said Hy4 entered the strongest open-source tier for the first time but did not disclose parameter scale or benchmark results. These are revenue-side stories, and they hold up.

But the cost-side landmines are buried in the S-1 footnotes. FT's reading of Anthropic's filing: costs are growing faster than revenue, and investors are using that to scrutinize the economics and business-model sustainability of the AI boom. The weight of that sentence lies in the fact that it is not an outside critique but a direction acknowledged by the prospectus itself. On that basis, Tomasz Tunguz proposed a four-layer AI economy framework, arguing that rising financing costs could shift the center of value toward the software layer — translated into buyer's language: capital returns at the hardware layer are being eaten by debt costs, and profit must be found at the application layer.

Boundaries need to be drawn: a single-quarter swing from profit to loss does not equal a structural loss, and Anthropic has not fully disclosed its fiscal year, its capitalization policy for compute procurement, or whether the $900 billion commitment includes cancelable clauses. Extrapolating one third-quarter figure into 'the collapse of the AI business model' is overreading. But treating it as a one-off disturbance is equally dangerous — because the 90% increase in shadow debt is measured over six months, not a single quarter.

AI's bill is not only on the debt side but also on the materials side. DRAM bulk transaction prices rose for nine consecutive months, reaching about twice their level at the end of 2025 by September 2026; Apple's manufacturing costs rose by $33 billion due to memory price increases, with pressure likely to persist into 2027. Samsung and SK Hynix together accounted for about 83% of global HBM revenue in the second quarter, and South Korea's October exports rose 93.7% year on year, the largest increase since January 1974. The center of profit is shifting upstream to memory — the combined operating profit of the world's five largest memory makers exceeded that of the five largest U.S. tech giants for the first time in the third quarter.

This means the cost curve of downstream hardware companies is locked in by upstream suppliers. Apple's $33 billion is the quantified part, while the Raspberry Pi 5 16GB version rising by $80 to $385, more than triple its launch price, is a sample of costs passing through to end devices. As memory capacity is pulled from consumer-grade production into HBM, one end of the price-increase rope is tied to AI data centers and the other is choking consumer electronics — this is not a demand problem but a capacity-allocation problem.

The second rope is credit. FT reported that the AI data center credit market is cooling, with investors demanding higher returns; shares of new cloud players such as CoreWeave and Nebius fell. Read the two threads together: the money providing leverage for AI infrastructure is starting to reprice risk. And of more than 10,000 public comments on data centers from over 410 U.S. local governments, more than 90% were negative, only 21% mentioned artificial intelligence, water resources were the primary concern, and 73% of related meetings had no supporters present — community resistance is adding a non-financial cost to site selection and approvals.

One end of the price-increase rope is tied to AI data centers, the other is choking consumer electronics.

The valuation question comes back to cash flow. When the AI shadow debt of five giants rises 90% in six months and Anthropic's long-term obligations approach $900 billion, the multiples the market assigns are essentially pricing 'future compute rent.' If financing costs rise and the credit market demands higher returns, then the same future cash flow should be discounted at a higher rate, and valuations should be revised down. This is not sentiment; it is arithmetic.

But the possibility of an unjustified selloff also exists. The pricing logic of the software layer differs from the hardware layer: ChatGPT search scale is about 20% of Google's with 250 million users, and if merging Chat and Work can raise per-user monetization, the marginal cost at the application layer is far lower than at the compute layer. Tomasz Tunguz's four-layer framework holds that the center of value may shift toward software — if true, hardware valuations currently suppressed by debt would instead give the software layer room for revaluation. Our estimate: if the software layer's revenue share rises 5 percentage points over the next two quarters, rough calculations based on public pricing suggest it could partially offset the impact of rising discount rates at the hardware layer; this is a hypothetical projection and has not been confirmed officially.

There are two signposts to watch. First, whether issuance spreads on AI data center credit continue to widen over the next two quarters — widening would confirm rising debt costs and pressure hardware-layer valuations; second, whether Anthropic's IPO pricing and updated S-1 break out the payment schedule for the $900 billion commitment — the more detailed the disclosure, the better the market can price it. What officials are not saying now is precisely what should be watched most closely.