Tech • AI • Robotics
The Man Who Predicted 2008 Takes on the AI Giants
7/10TL;DR
The rising rental price of Nvidia H100 chips in 2026 suggests that AI compute scarcity is being driven less by the age of the chip than by shortages in power, cooling, networking and ready-to-use data center capacity.
KEY POINTS
An older chip is getting more expensive
The H100, introduced in March 2022, should normally be losing value after the arrival of Blackwell and then Vera Rubin systems. Instead, its rental price rose by nearly 40% between November 2025 and July 2026. Market trackers including Silicon Data and SemiAnalysis also observed tighter H100 availability and firmer contract pricing as 2026 began.
The dispute goes beyond silicon performance
The debate intensified after Michael Burry argued in late 2025 that a chip can keep working for six years without remaining economically competitive for six years. His point targeted depreciation policies at cloud giants, where accounting life is not about whether hardware still powers on, but how long it stays useful enough to justify spreading its cost over several years.
Longer depreciation meaningfully lifts profits
Major platforms have extended server lives. Alphabet generally depreciates servers and networking gear over six years. Meta moved most servers and network assets to 5.5 years, a change that reduced annual depreciation expense by $2.9 billion and lifted net income by $2.6 billion. Burry’s broader claim is that hyperscalers could be understating depreciation by about $176 billion from 2026 to 2028.
Amazon and Meta reached opposite conclusions
Amazon extended server life from five to six years effective January 1, 2024, citing better hardware, software and data center design, with a $3.2 billion reduction in annual depreciation expense. But roughly a year later it reversed part of that decision for some servers and network equipment, returning to five years and taking about $1.4 billion in extra depreciation expense, cutting net income by roughly $1 billion, mainly at AWS. Just nine days before Amazon formalized that move in February 2025, Meta went the other way.
A rented H100 is really rented infrastructure
A cloud customer is not buying a bare GPU. A full DGX H100 system includes 8 H100s, CPUs, system memory, NVMe storage, NVSwitch fabric and network cards, then depends on cluster networking, orchestration software, maintenance, power, cooling and the building itself. AWS P5 instances package H100s with up to 3.2 terabits per second of networking and connect them into ultra clusters of up to 20,000 H100 or H200 GPUs.
That distinction explains the pricing paradox
The secondary-market value of a chip reflects the residual value of the component itself. Rental pricing reflects the value of compute capacity that is already installed, connected and available immediately. A chip can depreciate technologically while the hourly price of usable compute rises because customers are paying for instant access to a scarce operating environment, not just for older silicon.
Different operators can extract different economic lives
AWS offers multiple generations side by side, including P5 for H100, P5e for H200, P6 for B200 Blackwell, and even older P4d instances based on A100. That lets customers choose by workload and budget. Meta, by contrast, mainly deploys much of this infrastructure inside its own data centers for internal AI and core services, giving it more discretion over how to reassign older assets across uses.
The real bottleneck is power and physical infrastructure
Data centers cannot swap generations freely because each site has fixed envelopes for electricity, cooling and rack density. A DGX H100 SuperPod setup may run around 40 kilowatts per rack with air cooling. A GB200 NVL72 system can approach 120 kilowatts, roughly triple the power density, often with liquid cooling. Upgrading may require new electrical distribution, cooling loops and heavier structural support.
The constraint extends far beyond data centers
Even when companies are ready to spend, key electrical equipment is scarce. GE Vernova reported 116 gigawatts of equipment backlog and reservation agreements in the second quarter, up from 10 gigawatts a quarter earlier, with about 20% tied to data center customers. Wood Mackenzie estimates average U.S. transformer lead times around three years, with some large units beyond 160 weeks. In Northern Virginia, securing grid capacity can take three to five years, and in parts of the U.S. Northwest eight to ten years.
The H100 boom may be real but not permanent
The current market weakens the idea that GPUs automatically become economically obsolete after two or three years. Yet it does not prove they deserve five- or six-year lives on the books. If more power, cooling, data center space and more efficient new systems arrive fast enough, the scarcity premium supporting H100 rental rates could fade even while the chip remains functional.
CONCLUSION
The H100’s surprising strength in 2026 reflects a shortage of deliverable AI compute, not a suspension of technological progress. The central question for cloud groups is whether demand will keep outrunning new infrastructure long enough to justify today’s long depreciation assumptions.
Explain this