The AI Graveyard Is Demand Proof — but the Write-Offs Land on Credit, Not on Silicon

TechCrunch's running ledger of AI projects that failed, first published Sept. 15, 2026 and updated Oct. 24, 2026, is the first public record that maps where AI spending stopped rather than where it was promised. It is more useful than a shutdown count because the entries separate products that never touched hardware from the handful that did, and only the second group carries a vendor invoice. It shows AI demand failing at the application layer first.

Counting the ledger's entries by who was left holding the invoice changes the conclusion. Banana.dev wound down as GPU costs rose and Forefront Chat closed when API costs outran free GPT-4 and Claude access, but neither owed anyone for silicon. Builder.ai, Stability AI, Character.AI and Atos each died with a vendor on the other side of an unpaid bill, and each of those vendors had already agreed to a haircut. Dead demand does not delete capacity, it reprices the contract attached to it.

The pattern is consistent: projects that died on GPU economics were the ones that had rented capacity, and projects that died on product economics left no bill behind. Every case with a named vendor on the other side of the invoice is a case where the loss was agreed, renegotiated or capped — never a surprise. Vendor losses get negotiated, which is why they surface slowly.

NVIDIA's second quarter of fiscal 2027 is the cleanest test of the thesis, and the result is not a write-down. Revenue of $96.2 billion was up 106% year over year, gross margin expanded to 75.0% from 72.4%, and management guided to $108.0 billion for the fiscal quarter ending in late October while assuming no data center compute revenue from China. Provisions for inventory and excess purchase obligations fell to $2.1 billion in the first half from $6.3 billion a year earlier.

The same filing discloses the exposure that has not yet matured. Supply and capacity commitments jumped from $119 billion to $279 billion in one quarter, and $175 billion of that comes due in fiscal 2028 and fiscal 2029 alone. Concentration cuts both ways: one direct customer was 16% of quarterly revenue, and the $0.4 billion H200 excess-inventory and purchase-obligation charge shows what happens when a product cycle turns before the paper catches up. The write-off is scheduled, not absent.

Because capacity is reserved and prepaid rather than ordered on demand, the memory and packaging chain is paid for the build before the graveyard's demand loss shows up anywhere. NVIDIA's supply-chain announcement on Oct. 21, 2026 describes reservation advances and plant expansions, not forecasts, and the memory vendors are converting them into revenue and margin right now. Upstream vendors collect the cash before the cancellations arrive.

The adjustment therefore lands first in price, then in volume, and only later in write-offs. The first crack is already visible in how the market treats a record quarter: SK hynix's shares fell about 10% on the day of its best-ever profit report, which is what peak-cycle pricing looks like from inside. Revenue keeps compounding while the multiple discounts the ending

Oracle's $638 billion of remaining performance obligations is the single largest claim on future AI revenue, and about $75 billion of the most recent $85 billion increase came from customers who prepaid for GPUs or supplied their own hardware. That structure protects the vendor and endangers the equity: a customer that fails takes its prepayment with it while the data center it financed stays on the balance sheet. Cloud infrastructure revenue still grew 77% to $18.1 billion in fiscal 2026, so this is dead demand being absorbed, not demand disappearing. Prepayment moves risk off the vendor and onto the balance sheet.

The same pattern runs through Amazon and Microsoft, which are named creditors on Builder.ai's filings and are funding the build with debt and thinner free cash flow — Amazon's first-half 2026 free cash flow was negative $27.0 billion, and its interest expense reached $1.3 billion in the quarter from $516 million a year earlier. Every layer of the chain now carries somebody else's default. Vendor credit is the new transmission channel for AI demand failure

The credit market reached that conclusion before the vendors did. CoreWeave's five-year credit default swaps closed near 1,038 basis points on Oct. 15, 2026 after touching 1,140, and its 9.75% bond due 2031 yielded 15.78% — spreads that price a real probability of default rather than a repricing of risk. CoreWeave had more than $51 billion of debt and obligations behind $63.3 billion of property and equipment, and its shares set a 52-week low of $53.50 on Oct. 22, down 54% over twelve months. Credit is pricing failed AI projects, and vendor income statements are not.

Gartner's arithmetic is the framework the market is trading against: roughly $1 trillion of AI infrastructure spending in 2026 and about $6 trillion more through 2030 requires $8.5 trillion of cumulative revenue to cover capital spending, financing costs and a 12% return hurdle, against at most $3.6 trillion of AI revenue attributable to AI data centers. Neocloud shares were the loudest casualty of that arithmetic, cooling sharply through October as the sector's newest listings stalled. Quantifies how far AI revenue must travel to justify what has already been spent

Over the next two quarters the ledger only matters if it produces accounting. Watch for the first neocloud or landlord disclosure of a customer default, a lease termination or an idle-capacity charge, and watch NVIDIA's $108.0 billion outlook for the quarter ending in late October. If the guide holds and no write-down appears, the ledger is describing churn in applications rather than a collapse in infrastructure demand. The next two quarters decide which story the data supports.

One to three years out, the fiscal 2028 and fiscal 2029 maturity schedule is the pressure point: $87 billion and $88 billion of NVIDIA supply and capacity commitments come due against a build whose incremental buyers are increasingly balance sheets rather than startups. The base case is consolidation, in which failed projects hand their contracts to surviving labs. Consolidation keeps volume intact while compressing price