The space-heater ledger

Five predictions on AI inference before 2030

1. 1The six-year GPU book will not survive the decadeopen

   Grades true if: before 2030, a hyperscaler or a neocloud with more than $1 billion of annual revenue discloses a useful life of four years or fewer for a class that names accelerators, GPUs or AI servers. Amazon's 2025 cut from six back to five does not count: the number has to start with a four.

2. 2AI will get its hashpriceopen

   Grades true if: before 2030, a filed credit agreement, indenture or securitisation prospectus references a third-party published rate for a named accelerator generation inside a covenant, borrowing base or collateral-valuation clause. A rate quoted in a press release, an investor deck or a research note does not count. It has to be load-bearing in a document somebody can be sued over.

3. 3Custom silicon takes the payroll and the GPU keeps the frontieropen

   Grades true if: before 2030, at least two of Alphabet, Meta, Amazon, Microsoft and OpenAI state in a filing, an earnings call or an official engineering post that most production inference for a named flagship model family runs on silicon they designed. Their own words, about a named model, not an analyst's estimate of chip mix.

4. 4The secondary market prints the number everyone is avoidingopen

   Grades true if: before 2030, a public company books an impairment or accelerated depreciation charge of $1 billion or more that names accelerators, GPUs or AI servers as the asset, in a quarter where the same company reports record revenue or record capital expenditure. Both halves in the same filing.

5. 5Most neoclouds will not die as cloudsopen

   Grades true if: before 2030, at least three of CoreWeave, Nebius, Crusoe, Lambda, Together and Applied Digital are acquired, restructured under creditors, or re-segment their reporting so that power, land or data-center capacity is the primary reported business rather than GPU rental.

Nothing new has printed yet. NVIDIA reports in late August. The hyperscalers and CoreWeave file for the September quarter between October and November. Useful-life changes tend to show up in annual filings early in the new year.

Recorded before publication

The filings the essay already cites, so the ledger does not start empty.

• 2025Amazon shortened a server and networking subset from six years back to five, citing faster AI/ML change; the reversal added $1.4 billion to 2025 depreciation. Amazon 10-K
• Feb 2026CoreWeave disclosed a six-year technology-equipment life beside an about-five-year weighted-average committed customer term; the earning life of the same fleet appears nowhere. CoreWeave 10-K
• Aug 2026Nebius said pricing for older GPUs improved more than 30% from the prior quarter, with expected payback of about twenty-two months. Nebius Q2 shareholder letteragainst
• Aug 2026CoreWeave's credit stack priced the tail: $8.5 billion at investment grade against contract-wrapped GPUs in March, then $2.6 billion below investment grade on a roughly five-year loan against contracts averaging about three years, with no residual-value guarantee in the filed agreement. CoreWeave 8-K
• Mar 2026SemiAnalysis's H100 rental index, distributed on Bloomberg: one-year contract rates rose nearly 40% from their October 2025 low, and on-demand capacity has been sold out since February. SemiAnalysis H100 indexagainst
• Aug 2026NVIDIA's financing post made residual value the official pitch: a support mechanism for up to 25% of an opportunity, the factory described as flexible and fungible, the A100's economic life stretched toward a decade, and the one-year H100 rental series quoted by the vendor itself. NVIDIA financing post

Disclosed useful lives

Every change I have found disclosed, oldest first in each group. These are management estimates for asset classes that are not alike, and none of them isolates a GPU.