Nvidia GPUs: Hyperscalers Hoard Billions in Unused Chips 2026

Nvidia GPUs have become the hottest commodity in tech, but a growing pile of evidence suggests that hyperscalers like Microsoft, Amazon, Google, and Meta are sitting on billions of dollars' worth of chips they can't actually plug in. According to analyst Ed Zitron, Nvidia and Broadcom have sold over $560 billion in AI chips and equipment since early 2023, yet roughly half may be uninstalled or lacking power.

This isn't just a niche concern. It raises questions about the real state of AI infrastructure buildout, the sustainability of Nvidia's record sales, and whether we're looking at a bubble or a temporary bottleneck.

What Exactly Is Happening with Nvidia GPUs?

Hyperscalers—the massive cloud providers—have been spending hand over fist on AI infrastructure. Nvidia's data center revenue has skyrocketed, and the company has become a trillion-dollar giant on the back of this demand. But the physical deployment of these GPUs is lagging behind the purchases.

Ed Zitron's analysis, based on public financials, estimates that a significant portion of the chips sold since 2023 are still in their boxes. Microsoft alone may have $50-100 billion in GPUs sitting idle, according to Zitron. Bloomberg has reported that Microsoft has about 12 GW of total capacity, but only around 2 GW is dedicated to AI chips. And in November 2025, Microsoft CEO Satya Nadella admitted the company had chips it couldn't plug in.

Why Can't Hyperscalers Plug In Their Nvidia GPUs?

The bottleneck isn't the chips themselves—it's the surrounding infrastructure. Data centers require massive amounts of electricity, cooling, and physical space. Building new facilities takes years, and power grids in many regions are already strained.

Even if a company has the land and the permits, getting enough electricity to run thousands of high-end GPUs is a monumental challenge. Each Nvidia H100 GPU consumes up to 700 watts, and a rack of them can draw as much as a small neighborhood. Multiply that by tens of thousands, and you're talking about gigawatts of additional load.

The Power Crunch

Utilities and grid operators are struggling to keep up with the surge in demand from data centers. In some areas, new data center projects are being delayed or rejected outright due to power constraints. This creates a logjam: hyperscalers have the chips, but they can't turn them on.

Data Center Construction Lags

Building a hyperscale data center takes 18-36 months, and the permitting process can add years. Even with accelerated timelines, the current pace of construction can't match the rate at which chips are being purchased.

The Financial Implications for Nvidia and Hyperscalers

If a large portion of Nvidia GPUs are sitting idle, that's a red flag for the entire AI supply chain. For hyperscalers, it means capital is tied up in depreciating assets that aren't generating revenue. For Nvidia, it could signal that future orders may slow once the backlog is cleared.

Some analysts see this as evidence of an AI bubble, where spending has outpaced practical utility. Others argue it's a temporary mismatch—the infrastructure will catch up, and the chips will eventually be deployed. The truth likely lies somewhere in between, but the risk of a correction is real.

• Capital inefficiency: Billions in GPUs sitting idle represent sunk costs and missed revenue opportunities.
• Depreciation risk: GPUs lose value quickly; idle time accelerates write-downs.
• Supply chain ripple effects: A slowdown in hyperscaler orders could hit Nvidia and its partners.
• Energy constraints: Power availability is becoming the limiting factor for AI growth.
• Bubble concerns: Investors are watching closely for signs of overspending.

What the Experts Are Saying

Ed Zitron's estimates are controversial but have gained traction. Bloomberg's reporting on Microsoft's capacity adds credibility to the idea that chips are underutilized. Meanwhile, Satya Nadella's admission in late 2025 confirmed that even the most sophisticated buyers are facing deployment challenges.

Not everyone agrees with the pessimistic take. Some industry observers point out that hyperscalers are building for the long term, and temporary idleness is a normal part of rapid expansion. They argue that once power and data center capacity come online, the chips will be put to work.

How This Affects the Broader Tech Industry

The implications extend beyond the hyperscalers. If demand for Nvidia GPUs slows, it could impact the entire semiconductor ecosystem, from TSMC to memory makers. It could also cool investment in AI startups that rely on cloud credits and GPU access.

On the other hand, if the infrastructure bottleneck is resolved, the AI boom could continue unabated. Much depends on how quickly power generation and data center construction can scale.

Note: Figures are estimates based on public reports and analyst projections. Not all companies disclose exact numbers. The background to this story is set out in Nvidia Invests $3.5 Billion in MediaTek to Build GPU Alternatives. A related thread runs through Mistral Raises €3 Billion to Build AI Infrastructure and Challenge US Giants. Earlier coverage in Apple AI Server Market: M8 Ultra Chip with Nvidia Tech shows how the argument developed.

What to Watch Next

Investors should monitor hyperscaler capital expenditure guidance and any announcements about data center delays. Nvidia's earnings calls will be scrutinized for signs of slowing demand. Additionally, power grid updates and regulatory decisions on data center approvals will be key indicators. A closer look at the same subject appears in Nvidia RTX 5090 Discontinued: What the Leak Means for Gamers. Anyone following the wider picture can start with AI Startups Turn Idle GPUs Into Chatbots as Data Centers Lag. The same theme surfaces in China's AI Infrastructure Strategy Expands to Brazil with Alibaba Cloud, which takes a different angle.

If the bottleneck persists, we may see a wave of consolidation or a pullback in AI spending. If it eases, the AI revolution could enter its next phase. Context on how this played out before sits in Pensions Triple Lock: Lifeline or Unaffordable Billions. A companion piece, Gigabyte RTX 50 Infinity GPUs: Premium Design, High Price, approaches the same subject from another angle.