The Man Who Predicted 2008 Takes on the AI Giants

The paradox at the center of the AI boom

Michael Burry’s latest fight with Big Tech is not really about whether artificial intelligence is useful. It is about whether the machines being bought to run it are being depreciated fast enough. The investor made famous by the 2008 housing crash has argued that hyperscalers are stretching useful-life assumptions for AI servers and Nvidia-based compute gear, thereby lowering annual depreciation expense and flattering reported earnings [1]. His core point is simple: a chip can still function after several years without remaining economically competitive for the same number of years.

The uncomfortable counterpoint is the H100. Nvidia introduced the H100 in March 2022, and, in an ordinary semiconductor cycle, it would be expected to lose pricing power after newer Blackwell and Vera Rubin systems entered the market. Instead, the H100’s rental price rose by nearly 40% between November 2025 and July 2026, while market trackers saw tighter availability and firmer contract pricing early in 2026 [1]. That does not disprove Burry’s accounting critique. It does, however, make the critique harder to reduce to a simple “old chip equals obsolete chip” argument.

A chip is not what customers are really renting

The key distinction is between a component and a working cluster. A cloud customer renting an H100 is not merely renting silicon. It is buying immediate access to a stack: GPUs, CPUs, memory, NVMe storage, NVSwitch or comparable fabric, networking, orchestration software, maintenance, power, cooling and a data-center shell that is ready to run [1]. That is why the economics of rented compute can diverge from the resale value or technological age of the GPU itself.

This is also why Wall Street’s interest in compute pricing has accelerated. Yahoo Finance reported on August 29 that CME Group plans to launch futures tied to hourly rental costs for Nvidia H100 and B200 GPUs on October 5, pending regulatory approval [2]. The same report put Silicon Data’s H100 benchmark at about $2.68 per GPU-hour and the newer B200 benchmark at about $5.66, underscoring that compute is becoming a priced, benchmarked and potentially hedgeable input [2]. If the futures market develops liquidity, it could become a public scoreboard for Burry’s thesis.

A strong H100 forward curve would suggest that older deployed infrastructure can retain real economic value when power and data-center capacity are scarce. A weakening curve, especially if Nvidia and hyperscaler capital spending remained high, would point in the other direction: capacity arriving faster than paid demand. The accounting question would then sharpen quickly.

Nvidia’s moat is moving outside the GPU

The current state of the story is not just that H100s remain valuable. It is that Nvidia’s advantage increasingly sits around the GPU as much as inside it. TechCrunch reported on August 29 that as AI compute grows toward gigawatt scale, orchestration has become a more complex and strategically important layer [3]. Nvidia’s Vera Rubin architecture is being positioned not only as a GPU generation but as a broader system involving CPUs, storage and networking components designed to move data efficiently through huge clusters [3].

That matters for the depreciation fight because useful life is workload-specific. A GPU that is no longer best-in-class for frontier training may still be profitable for inference, fine-tuning, batch jobs, internal AI features or lower-cost customer tiers. The H100 pricing paradox therefore reflects a broader market truth: the age of the chip is only one variable. The value of the operating environment around it can dominate the calculation.

The same logic explains why cloud providers can reach different conclusions about useful life. According to the 8news background analysis, Alphabet generally depreciates servers and networking gear over six years, Meta moved most servers and network assets to 5.5 years, and Amazon both extended and later partly reversed server-life assumptions for some equipment [1]. Burry’s broader estimate is that hyperscalers could understate depreciation by about $176 billion from 2026 to 2028 [1]. But the market evidence does not yet deliver a clean verdict. It says the answer depends on whether old clusters remain revenue-generating assets or become stranded, power-hungry inventory.

The financing layer is becoming more fragile

The boom is also increasingly debt-financed. TechCrunch reported on August 28 that Lambda, a neocloud that buys computing chips and rents them to businesses, raised $1 billion in private short-dated debt to buy Nvidia AI chips that it will lease to Microsoft [4]. The same report noted that Lambda had already closed a $1 billion secured credit facility in May and announced a separate $926 million loan for Nvidia GB300 GPUs tied to a deployment under contract with Nvidia [4]. Bloomberg data cited by TechCrunch put global AI-related debt raised by banks and tech companies above $400 billion in 2026 so far [4].

That financing model amplifies both sides of the debate. If demand stays ahead of usable capacity, short-dated borrowing against contracted GPU revenue can look rational. If rental rates fall or customers defer workloads, the same structures can become a stress point. Burry’s depreciation argument is therefore not an isolated accounting complaint. It is tied to the capital structure of the AI buildout.

New chips are arriving, but not as bare chips

Fresh supply is coming, yet it often arrives as complete infrastructure rather than loose processors. SuperX said on August 28 that it received an initial purchase order for 128 Nvidia B300 AI server clusters from Australian compute provider Ezisight, with delivery scheduled for the fourth quarter of 2026 and deployment in local Australian data centers for large-model training and inference [5]. SuperX also described its strategy as spanning hardware, liquid-cooling, power-supply solutions and end-to-end data-center delivery [5].

That reinforces the central point. The market is not merely short of GPUs; it is short of deployable systems. Newer B200, B300 and Vera Rubin capacity can eventually pressure H100 pricing, but only when sites, power, cooling, networking and customers are in place. Until then, older H100 clusters that are already powered and connected can command a scarcity premium.

What would prove Burry right or wrong?

Burry is most likely to look right if two things happen together: reported profits continue to rely on long depreciation assumptions while rental benchmarks for older AI compute roll over. That would imply the accounting lives outlast the economic lives. He is more likely to look early or too blunt if H100 and similar deployed assets continue to earn meaningful cash flows well into the Blackwell and Vera Rubin cycles.

The current evidence points to a narrower conclusion. The H100’s surprising strength does not prove that five- or six-year depreciation is always conservative. It proves that AI infrastructure ages in layers. Silicon ages quickly; powered, cooled, networked and contracted compute can remain scarce for longer. The next phase of this story will be measured less by slogans about bubbles than by rental curves, utilization, power availability and whether the debt supporting the AI buildout can be paid from real compute revenue.

Sources from the last 72 hours

1. [1]The Man Who Predicted 2008 Takes on the AI GiantsAug 30, 2026, 7:00 AM UTC
2. [2]Wall Street is turning Nvidia's AI chips into a new futures market: Chart of the DayAug 29, 2026, 11:51 AM UTC
3. [3]Nvidia’s AI advantage is moving beyond the GPUAug 29, 2026, 1:00 PM UTC
4. [4]Neocloud Lambda secures $1B in debt to buy more chipsAug 28, 2026, 8:24 PM UTC
5. [5]SuperX Secures First Batch of 128 Units of NVIDIA B300 AI Server Purchase Order from Ezisight, Marking Its Official Entry into the Australian MarketAug 28, 2026, 1:44 PM UTC

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.