A GPU can appear in a supplier’s revenue long before it processes a paying customer’s request. Between those events lie construction, electricity connections, cooling systems, networking, testing and software deployment.

That gap is central to writer Ed Zitron’s critique of the AI infrastructure boom. In his interview, he estimates that $200 billion to $300 billion worth of Nvidia GPUs could be sitting in warehouses or unpowered data centers. That is his estimate, not an established industry-wide inventory count. The financial disclosures he discusses cannot, by themselves, prove how many GPUs are idle.

But the underlying question deserves attention regardless of whether that estimate holds up: how much purchased AI hardware has become usable, revenue-producing capacity?

GPU sales answer only the first question

Discussions of AI investment often compress several separate milestones into one story. A company buys chips, so demand must be strong, capacity must be expanding and future revenue must follow.

In practice, there are at least five distinct stages:

1. Procurement: Hardware is ordered or purchased.
2. Delivery: Equipment reaches a warehouse or project site.
3. Installation: Servers, networking and cooling are assembled.
4. Commissioning: The facility is powered, tested and ready to operate.
5. Commercial use: Customers use that capacity at prices that can support its costs.

Strong chip sales demonstrate purchasing demand. They do not independently establish the timing of deployment, customer utilization or profitability.

Equally, hardware awaiting installation is not automatically evidence of a bubble. Large infrastructure projects require advance procurement, and buying scarce equipment early can be sensible. The important distinction is between a planned inventory buffer and an expanding backlog whose deployment dates keep slipping.

What construction-in-progress figures can—and cannot—reveal

Zitron builds part of his argument around construction in progress, an accounting category generally covering assets that are not yet ready for their intended use. Large balances can help identify how much capital a company has committed before the associated assets enter service.

However, construction in progress is not a GPU inventory report. Depending on the company’s accounting policies and disclosures, it may include buildings, electrical infrastructure, equipment and other project costs. Data center hardware also encompasses much more than Nvidia accelerators.

A reliable estimate of uninstalled GPUs would require company-specific detail: the hardware share of those balances, the accelerator mix, delivery status and expected commissioning dates. Total capital expenditure is no substitute because it can fund many other assets.

Large unfinished-asset balances justify further investigation. They do not establish concealment, wrongdoing or a precise quantity of unused chips.

Electricity is a separate construction project

Buying servers does not secure the electricity needed to run them. An AI facility may depend on transmission upgrades, substations, transformers, generation capacity or fuel infrastructure. Each dependency has its own approvals, equipment orders and construction schedule.

This creates a practical problem: a building can look nearly finished while remaining commercially unusable. A delay in one essential system can hold back equipment that is otherwise ready.

The interview’s discussion of Oracle’s Project Jupiter in New Mexico illustrates this concern. Zitron describes a notice concerning possible force majeure relief tied to power availability. The transcript does not include the notice or the governing agreements, so its precise legal and financial consequences cannot be established from the interview alone.

Force majeure is a contractual mechanism whose scope depends on the agreement and applicable law—not a universal right to cancel a project or stop paying. A notice may preserve contractual rights without establishing that a project will fail.

The broader issue is risk allocation: if electricity arrives late, who continues paying—the tenant, developer, equipment owner, lender or utility customer?

Why delays become a financing problem

For a debt-funded project, a missed opening date can mean more than postponed revenue. Financing costs may continue while equipment remains unavailable for commercial use.

Consider a simplified example: $10 billion of fully drawn debt at 7% annual interest carries approximately $700 million in yearly interest before principal repayments. Actual projects use different draw schedules, financing structures and interest arrangements, but the basic point remains: borrowed money has a cost even when construction is unfinished.

Refinancing can add another pressure. If lenders become less confident about completion dates, customer payments or collateral values, they may demand higher rates, additional security or tighter conditions.

A credit downgrade can also narrow the pool of eligible investors. It does not mean every bondholder must sell, nor does it automatically change the coupon on existing fixed-rate debt. Its most direct consequences often emerge when a company needs new financing.

Accounting depreciation does not stop hardware from aging

Equipment generally begins depreciating when it is ready and available for its intended use. That can create a lag between capital spending and depreciation expense.

But accounting treatment does not freeze economic value. Hardware awaiting deployment can face newer alternatives with better performance or energy efficiency. Buyers may seek lower prices for older capacity even if the equipment has barely operated.

Older GPUs can still perform valuable work; a new generation does not make them worthless. Nevertheless, a delayed project may begin earning revenue in a more competitive market than its financing model anticipated. Impairment considerations may also arise separately from routine depreciation.

The disclosures that would make the picture clearer

Investors need a bridge between purchasing announcements and operating results. Useful disclosures would include:

• Deployment status: Equipment ordered, delivered, installed and placed in service.
• Power status: Proposed capacity versus electricity actually available to commissioned facilities.
• Capacity definitions: Whether a figure describes an entire campus, IT equipment load or AI-specific infrastructure.
• Commercial performance: Utilization, realized pricing and customer concentration.
• Delay exposure: Payment obligations, financing costs and remedies if commissioning slips.

These distinctions also explain how chip stockpiles and shortages of usable compute could coexist. Equipment can be plentiful in the wrong place—or without the infrastructure needed to operate it.

The real test is conversion, not purchasing

The interview’s strongest argument does not require accepting its largest estimate or its predictions of an industry collapse. It requires recognizing that hardware purchases, functioning infrastructure and sustainable customer demand are different measures.

AI investment succeeds when companies convert capital into reliable capacity, and that capacity into sufficient cash flow. Until disclosures make those transitions clearer, GPU sales should be treated as evidence of spending—not proof that the entire economic model works.

---

This article was inspired by Hyperscalers are hiding a massive secret about uninstalled GPUs | Ed Zitron from The Tech Report. Please visit the original video for the creator’s full presentation and context.