Microsoft's Snubbed Oracle GPUs Give Burry's Warning Its First Real Data Point

On November 6, 2026, Microsoft declined to renew a computing arrangement with Oracle, and the reported reason cuts straight to the biggest accounting fight in AI: the older GPUs no longer suited its needs. Oracle pushed back. It said demand remains strong and noted that some GPU contracts have been renewed at 20% premiums.

One lapsed contract is thin evidence. But it is exactly the kind of real-world pricing data that Michael Burry's famous warning has lacked until now.

Burry, the investor known for calling the 2008 housing crash, argued that hyperscalers are stretching chip depreciation schedules past the chips' realistic economic lives. By his estimate, that practice could understate depreciation expenses by $176 billion through 2028. He named two companies specifically: Oracle (NYSE:ORCL) and Meta Platforms (NASDAQ:META).

The logic is simple. If chips lose value faster than the books assume, then reported profits are too high. And if reported profits are too high across the largest AI spenders in the market, the entire sector's earnings picture is built on softer ground than investors believe.

That is why a single non-renewal at Oracle matters. It is a market transaction, not a model or a forecast. The best-informed buyer in the compute market looked at older GPUs and walked away.

What is Microsoft actually doing with its hardware?

Microsoft is refreshing its silicon at a pace that makes Burry's concern harder to dismiss. In its fiscal first quarter, property and equipment additions hit $39.37 billion. That is more than double the $19.39 billion from a year earlier.

Chief Financial Officer Amy Hood said about three-quarters of that spending went to short-lived assets, mainly CPUs and GPUs. She guided to over $50 billion in spending next quarter and roughly $170 billion for the calendar year. Free cash flow came in at $18.6 billion, against $58 billion of operating cash flow — a gap of nearly $40 billion in a single quarter.

The hardware mix tells the same story. GB300s now make up the majority of Microsoft's GPU deployments. The company has also docked its first Vera Rubin chips, the newest generation of Nvidia's lineup. When the buyer with the deepest insight into AI workloads keeps choosing newer silicon over older capacity, it shows how quickly the frontier moves on — and how fast last year's flagship becomes this year's surplus.

Are the other hyperscalers already feeling it?

Across the group, cash is running out faster than it comes in. The numbers from the most recent quarter:

• Alphabet (NASDAQ:GOOGL) spent $51.36 billion on capex and posted free cash flow of negative $9.07 billion.
• Meta went negative $2.12 billion.
• Oracle went negative $5.40 billion.
• Amazon (NASDAQ:AMZN) reported a trailing free cash outflow of $42.4 billion.

Amazon added a detail on its call that should ring a bell. The company said it is finding ways to extend useful lives of its equipment. That is precisely the accounting lever Burry flagged. Stretching out depreciation schedules lowers annual expenses and lifts reported profits, without changing how fast the hardware actually loses its economic value.

What does the other side of the argument say?

The rebuttal has real evidence behind it. The leading AI chip supplier — Nvidia, which carries $279.00 billion in supply obligations and has the most at stake in this debate — has pushed back forcefully on depreciation fears.

In September 2026, Nvidia's then-CEO Jensen Huang cited a 22% monthly increase in H100 rental prices and A100 capacity contracted through 2029. Huang resigned today, according to the report, amid allegations of extramarital affairs, including one with a purported Chinese spy and another with an Nvidia executive. His departure does not change the data he cited, but it removes the sector's most vocal defender from the field.

Oracle, for its part, reported GPU utilization of 97.9%. It also said the majority of its renewed GPUs were 4 years or older. On paper, that is a strong answer to Burry: nearly full utilization, aging chips still under contract, and renewals coming at premiums.

But Burry saw this defense coming. In a July 9, 2026 Substack essay, he wrote: "A chip can rent and still depreciate very fast economically."

That sentence is the crux. His claim is narrower than the one critics attack. Older GPUs keep earning, yes. But that rental income may still fall short of the value the chips carry on the balance sheet. Utilization measures activity. It does not measure whether the asset earns back what the books say it is worth.

Which reports could settle the question?

Microsoft's exit from the Oracle arrangement is relevant but not conclusive. Oracle could rent the same chips to someone else, possibly at attractive rates. The deciding question is whether Oracle earns back its original investment over the useful life it assumed for the hardware.

Two events will carry most of the signal. First, Oracle's next quarterly report should show whether the capacity Microsoft released found a new tenant — and at what price. Oracle has already called fiscal 2027 and 2028 its peak capex years, which frames the stakes: maximum spending, maximum depreciation assumptions, maximum exposure if the chips age out early.

Second, watch Meta's fourth-quarter call and any 2027 capex guidance. A large spending increase paired with unchanged depreciation schedules would push Burry's question to the center of the AI investment debate.

The fork in the road is narrow and clean. If old chips re-lease at a premium, Huang was right and the depreciation alarm was overblown. If they re-lease at a discount, the profits hyperscalers have reported are overstated — and the correction Burry predicted, $176 billion of it through 2028, starts to look less like a contrarian's hobbyhorse and more like a reckoning with a date attached.