An unaudited sentence repriced an industry last month. A market account attributed a number to Oracle: every GPU that reached the end of a four-year contract was renewed, at roughly twenty percent above its original price. No page number. No card model. No denominator. No split between training and inference demand. No disclosure of whether the transaction was a re-lease or a resale. By the time the number reached me, it had already been compressed into a thesis — depreciation is slower than the bears say, AI capex has a longer tail, buy the compute complex.

I have watched this pattern before, from the receiving end. In May 2022, when UST broke its peg and roughly forty percent of DeFi TVL evaporated, I spent eleven days auditing our student DAO's treasury against the liquidation curves of Aave and Compound. What destroyed those positions was never the code. It was the collateral schedule — a number everyone accepted because nobody had asked who counted it, or how.

Crisis is just code with a high gas fee. So is a quarterly earnings narrative.

Context: the argument everyone is having

Burry's short case against the AI complex rests on a single accounting claim: that hyperscalers and GPU renters are depreciating compute assets too slowly, extending useful lives to flatter earnings while the silicon ages out underneath them. It is a legitimate argument. Depreciation is not a measurement. It is a policy. A server has a physical half-life and a book half-life, and only one of them appears in a filing.

The counterargument is the one Oracle allegedly supplied. If a four-year-old card can be re-leased at a premium, then four years was conservative, the write-down schedule is aggressive rather than generous, and the earnings the bears call inflated are actually understated. On that reading, the sector re-rates.

The asset at the centre of this is doing two jobs at once. A GPU is a physical object with a thermal design power, a memory bandwidth ceiling and a perf-per-watt curve. It is also a financial instrument with a book value, a lease term and a residual assumption. Those two things decay at different rates, and most of the shouting conflates them.

The players matter too. Oracle sells compute through OCI, a distant fourth in cloud share, but wrapped in database and ERP switching costs that no pure-play renter can replicate. The neoclouds — Nebius, IREN and their peers — rent compute as their core product, financed with debt, priced against spot, and constrained by power. NVIDIA sits above all of it, supplying the scarce input and capturing the margin regardless of who is right about depreciation.

Which is why the honest question is narrower than the one being debated. Is a twenty percent renewal premium a fact about compute demand, or a fact about contract structure? And is it a fact at all, or a number that cannot be verified by anyone outside the room?

Core: three things the headline cannot tell you

One. A GPU does not depreciate on a clock. It depreciates on a workload migration path.

If the event occurred in 2025 and the card is four-plus years old, the plausible silicon is an A100 or a V100 deployed in 2020 or 2021. Take the A100 80GB SXM: roughly 2 TB/s of HBM2e bandwidth, a 400-watt envelope, and a tensor throughput that an H100 SXM exceeds several times over, with about 1.6 times the memory bandwidth at a 700-watt envelope. On a pure perf-per-watt basis the old card loses badly — which is precisely why the naive read is wrong.

Because the workloads that clear matter more than the peak numbers. A model doing batch inference, embedding generation, video transcoding, small-model fine-tuning or rendering is frequently memory-bandwidth-bound, not FLOP-bound. On those jobs, an A100's cost per unit of work against a quoted dollar-per-hour rate is often competitive with, and occasionally better than, an H100's — not because the A100 is fast, but because the H100 is over-provisioned for the task. That is the migration path. Frontier training leaves; inference, serving and commodity transforms stay.

A card's residual value is not a function of its age. It is a function of how many workloads still clear at its cost curve — and that curve is set by electricity, utilisation and contract length, not by a depreciation table.

Two. "All GPUs that entered the renewal stage were resold" is not a statistic. It is a selection.

Read that sentence the way an auditor reads it. Renewal stage is not a random sample of the fleet. It is the subset of contracts where the counterparty still wanted something. Cards that died, cards sitting in a region whose power price tripled, cards whose tenant shut down, cards that were impaired and scrapped — none of them reach the renewal stage, so none of them appear in the numerator. They appear as a line somewhere else, or they do not appear at all.

Run the base rates. Deploy a hundred units. Twenty come up for renewal. Twenty renew, at a premium. The headline is "one hundred percent renewed at plus twenty." The truth is "twenty percent of the fleet renewed." Both sentences describe the same event, and only one of them is a depreciation signal.

Survivorship bias is not a data error you can scrub out. It is a data shape. The pool is defined by its winners, and the denominator is the story.

There is a second denominator hiding underneath. Twenty percent above what? If the renewal price bundles power, network, storage, orchestration, support and the switching cost of moving an enterprise off a database ecosystem, then the premium may be the price of a service relationship, not the price of silicon. Oracle's enterprise clients do not rent compute. They rent continuity. That is exactly the kind of number that surfaces in a renewal negotiation and has nothing to do with H100 supply.

And a third ambiguity, the one I would flag first in any audit: re-lease and resale are different economic objects. A re-lease is a continuing cash flow with a customer attached. A resale is an exit — often to a broker, often at a discount, often the moment an operator decides the asset is no longer strategic. The same phrase, "all of them moved," can describe either.

Three. The ledger this argument needs already exists in crypto. Nobody in AI capex is using it.

Here is the part that irritates me. The AI industry is repricing hundreds of billions of dollars of capital assets using secondhand narrative, while a parallel industry has spent five years building the exact primitive that would settle it: attestable, machine-signed evidence about physical compute.

The decentralised compute networks are imperfect — token incentives distort pricing, utilisation is lumpy, and most of them have yet to prove they can serve enterprise-grade workloads at enterprise-grade reliability. But architecturally they make utilisation a first-class, observable, priced object. Render routes rendering and inference jobs. Akash runs a permissionless container market over contributed GPUs. io.net aggregates cluster capacity. Bittensor builds incentive markets around model output rather than hardware. Whatever you think of their economics, they treat compute as something that produces a verifiable receipt.

To make a twenty percent premium auditable, you would need three components. A hardware root of trust — confidential computing modes and TEE attestation, so that the card reporting utilisation is provably the card doing the work. Signed work receipts — input hash, output hash, model hash, timestamp — aggregated into a utilisation proof that a third party can verify without trusting the operator. And a price oracle covering multiple rental venues, with an explicit latency budget.

That last one is where I get nervous, because I have watched it fail before. Oracle feed latency is DeFi's Achilles' heel; stale prices triggered cascading liquidations in 2020 and again in 2022. A compute price feed would be worse. GPU rental prices move on hours and days, not on blocks, and they are quoted privately across hundreds of bilateral contracts. A naive index would be arbitraged instantly, and a leveraged market built on it would fail in a cascade.

The missing ledger is not a blockchain problem. It is an attestation problem — and blockchain happens to be unusually good at exactly that shape of problem.

Contrarian: the premium could be real and bearish

Everyone is arguing about whether the twenty percent is genuine. That is the wrong axis. Assume it is genuine. It can still mean three things the bulls do not want to hear.

A renewal premium on aging capacity is what a supply bottleneck looks like, not what durable demand looks like. When new capacity cannot be delivered on schedule, whoever already owns racks extracts rent from whoever needs them. That rent is a tax on the entire downstream stack, and it tells you the demand curve has outrun the supply curve. Bottlenecks resolve. New fabs, new capacity and inference-optimised silicon are all routing toward this market. A four-year-old card commanding a premium today can be repriced down forty percent in two quarters without a single demand forecast changing.

A premium on a fully depreciated asset is also a different economic object from a premium on a new one. On a fully written-down card, twenty percent is nearly pure margin, and it accrues to whoever owns the depreciation schedule — which is Oracle, not the renter. On a new card, the same twenty percent merely means capex is recovered faster. The bulls are treating one company's margin improvement as an industry-wide residual-value upgrade. Those are not the same claim.

And the blind spot nobody is pricing: if compute becomes verifiable, it becomes collateralisable. Verifiability plus a price index plus signed receipts equals a lending market. Which sounds like progress until you remember what happened the last time an industry pledged correlated assets against a mark-to-model price. Crypto learned in 2022 that correlated collateral does not unwind in a trickle. It unwinds in a cascade. Compute is the most correlated collateral imaginable — every borrower posts the same card, marked against the same rental index, priced by the same latency-prone feed.

Open source is a promise, not a product. An auditable compute price index is the same thing: a promise until someone publishes the methodology.

Takeaway

Three things to track. Oracle's next filing — specifically whether the useful-life assumption on server and GPU assets changed, and whether the GPU disclosure includes a denominator this time. NVIDIA's data centre revenue against the A100 and H100 spot rental curves — if rental prices hold while new supply expands, the demand thesis survives; if they diverge, the premium was a bottleneck dressed as a trend. And whether any compute network publishes a hardware-attested utilisation and pricing index that an outside party can verify without trusting the operator.

The number that repriced a sector will be forgotten long before the ledger that could prove it exists. The only question worth asking is which one arrives first.