On September 13, a line item inside Oracle's (ORCL) quarterly disclosure did more work than any research note published that week. According to the filing, every GPU entering its renewal cycle had been resold โ not returned, not written down, not idled โ at transaction prices approximately 20% above the original contract price, with most of those units having been in continuous production service for more than four years.
Serenity's read landed fast and unhedged: the data contradicts Michael Burry's assertion of rapid GPU depreciation and is constructive for Nvidia (NVDA), for Neocloud operators, and by extension for the converted miners now renting out compute โ Nebius (NBIS) and IREN among them.
The conclusion is tidy. It rests, however, on a definition the disclosure never supplied. The number is a residual-value observation, not a depreciation curve. In a market where the gap between those two things is the gap between a solvent counterparty and an impaired one, the distinction is not academic.
Oracle leases compute under multi-year capacity contracts, most commonly four-year terms, and carries the underlying hardware on its own balance sheet. When a contract reaches its renewal window, Oracle has three paths: extend with the incumbent tenant, re-lease the capacity to a new tenant, or dispose of the iron. The disclosure says the first two paths cleared at a 20% premium. It does not say which path was taken in each case, and it does not itemize the cohort size.
That omission matters, because the word "resold" collapses three economically distinct events into a single label. A secondary hardware sale to an unaffiliated buyer is a direct statement about silicon residual value. A capacity renewal at a higher rate is a statement about power, networking, and rack density in constrained markets. A lease extension with the incumbent is a statement about switching costs. Only the first speaks to the depreciation debate Burry raised.
Then there is the arithmetic. A 20% nominal premium realized across a four-year term annualizes to roughly 4.7% per year. Depending on the inflation window you choose, that is a real-terms wash at best. A nominal gain is not a real gain, and a renewal price is not a market price. I have watched this exact confusion destroy capital before. In late 2017, while auditing the EtherFund ICO contract, I found a donation function whose reentrancy exposure would have wiped an estimated $2 million, and the pitch deck had described the same function as "battle-tested." In May 2022 I spent 72 hours reconstructing the Terra peg decoupling from raw transaction logs. The lasting lesson from both was identical: the headline number and the ledger entry frequently describe different events.
Against that skepticism, the structural case for GPUs is genuinely stronger than the structural case for ASICs, and I will concede it plainly. A four-year-old accelerator still runs quantized inference. It has a second life, a deep buyer pool, and a software moat in CUDA that keeps the installed base economically alive long after the training frontier has moved past it. A bitcoin ASIC has exactly one function; when network hashrate doubles, its residual value converges on scrap. In 2022 that convergence took roughly eighteen months, from about $4,000 down to under $200. Accelerator fungibility is real. That is the honest core of the bull case, and it deserves credit.
Where the bull case breaks is at the transfer mechanism. NBIS and IREN are balance-sheet owners. They hold the hardware, so they capture residual value if residual value exists. The token-adjacent compute protocols do not. Render, Akash, io.net, and their peers aggregate third-party GPUs they never purchased, never depreciated, and cannot resell. In 2026 I audited a decentralized AI compute marketplace that claimed blockchain-verified model outputs, and I demanded the contract logic that actually verified compute. What I found was a permissioned oracle cluster operated by the founding entity, signing attestations rather than computing them. The booked supply was self-reported. The useful-life assumption sat off-chain and unstated. Token holders carried zero residual value and 100% of the narrative exposure. Ledgers don't respond to sentiment; they accrue.
The GAAP layer deserves its own paragraph, because it is where the two narratives collide. Useful life is a management estimate under ASC 360 and IAS 16 โ not a measurement. When a hyperscaler shortens or extends the assumed life of a server class, operating income moves immediately and cash does not move at all. Burry's original complaint was precisely that asymmetry: earnings flattered by assumption rather than by economics. A renewal price above the original contract is interesting evidence against him, but it is not the same evidence. Resilient resale values and extended useful lives point in opposite directions, and only one of the two is audited.
The sampled cohort is the first problem. Units that reach a renewal window are survivors by construction. Accelerators that failed early, overheated in dense racks, or were obsoleted by a SKU transition were already retired or written down before they could be counted. What the disclosure measures is the residual value of the best-performing quarter of the fleet, reported as though it described the whole. That is survivorship bias wearing an audit trail.
Power scarcity is the second problem. A premium renewal in Northern Virginia or a constrained Asian metro reflects the cost of megawatts and interconnect queues, not the value of silicon inside the rack. If the premium is really a location premium, it does not generalize to a mining site in Texas with a substation and no fiber redundancy.
The third problem is demand elasticity. Training demand is relatively insensitive to price; inference demand is not. Inference buyers will move to cheaper, older, or smaller hardware the moment unit economics deteriorate. If aggregate AI capex decelerates even modestly, the bid for four-year-old accelerators thins quickly โ the ASIC market showed exactly how quickly.
The fourth problem is counterparty identity. "Resold" tells us nothing about who bought. A resale to an unaffiliated third party is one thing. A transfer to a joint venture, a financing vehicle, or an affiliated lessor is another thing entirely, and it would change the interpretation of the premium from evidence of demand to evidence of accounting. The filing does not name the counterparty type, and until it does, the interpretation remains open. Ledgers don't flinch at a press release.
On the blockchain side, the read-through is narrower than the timeline suggests. IREN's equity is now an AI-datacenter story with a mining option attached; its token-adjacent exposure sits in the depreciation assumptions applied to the mining fleet, which was itself the original case study in brutal useful-life discipline. The DePIN compute tokens have no hardware at all, so a favorable GPU residual data point does not transmit to them. If anything, it strengthens the position of the entity that holds the metal โ which is the opposite of the decentralized claim.
In a drawdown where the marginal buyer of compute is a distressed seller of tokens, survival math beats growth math. The question readers should be asking is not whether GPUs hold value. It is which counterparty on their screen is exposed to a residual-value assumption they have never disclosed. Watch three things. First, the next 10-K useful-life footnote for ORCL, IREN, and CoreWeave โ any extension is an earnings lever, not a technology signal. Second, IREN's next fleet-age disclosure and whether the mix shifts. Third, whether the 20% premium cohort is ever itemized by SKU, term, and geography.
If four-year-old accelerators truly clear above their original contract price, why are so many hyperscalers still quietly extending useful lives instead of shortening them? Ledgers don't lie โ but they also don't describe markets. Only one of those two statements is audited, and the market is trading on the other one.