Premium: What If…We’re In An AI Bubble? (Part 3) – A Critical Look at the Numbers

The Core Argument: Compute Demand Must Match NVIDIA’s Sales

Zitron starts from a simple arithmetic exercise. NVIDIA’s roadmap for the Blackwell and Vera Rubin GPUs targets over $1 trillion in sales by the end of 2027. Assuming a power‑usage‑effectiveness (PUE) of 1.35, that translates to roughly 40 GW of data‑center capacity—about 30 GW of GPUs and the supporting infrastructure. At an industry‑wide estimate of $12 million per megawatt, the market would need $435 billion in annual compute spend just to keep those chips busy.

Outside the two biggest AI labs—OpenAI and Anthropic—Zitron sees only a few billion‑dollar pockets of demand. The numbers suggest a massive gap between the compute NVIDIA is selling and the actual revenue that can be generated to pay for it.

Why NVIDIA’s Revenue Model Is Fragile

• Long build cycles – Even a modest 40 MW data‑center can take 24 months to become operational. Companies that buy GPUs now may not see returns for years, turning NVIDIA’s sales into a bet on future construction timelines.
• Multi‑year cloud contracts – NVIDIA’s latest 10‑Q shows it has committed $30 billion in multi‑year cloud compute agreements with the very partners it sells chips to. Those contracts are a safety net, but they also expose NVIDIA to the risk that those partners cannot meet their own build schedules.
• Warehousing risk – Zitron argues that NVIDIA is likely stockpiling at least a million Blackwell GPUs. If the data‑center pipeline stalls, those chips become idle assets, inflating inventory and eroding margins.

The Two‑Company Dependency

Zitron points out that OpenAI and Anthropic together account for roughly 89 % of AI‑startup revenue. Their projected compute spend is staggering:

Company	2026 Compute Spend (est.)	2029 Revenue Target
OpenAI	$50 B (per Greg Brockman)	$284 B
Anthropic	$15 B on Musk’s Colossus + other clouds	$174 B

Even if both double their spend next year, the market would still need two more OpenAI‑sized customers to justify the GPU pipeline. Zitron finds no evidence of any other organization spending even $100 M annually on AI compute.

Cash‑Flow Mismatch: A Billion‑Dollar Burn

The numbers become stark when you look at the total compute commitments:

• OpenAI: $852 B in compute spend projected through 2030.
• Anthropic: $330 B in cloud contracts plus additional hardware purchases, totaling $423 B.

Both firms claim they will generate $100 B‑plus in annual revenue by 2028, but the cash required to meet compute contracts far exceeds those revenues. Zitron cites internal statements from Anthropic’s CEO that a mis‑timed growth forecast could push the company into bankruptcy.

Token‑Based Billing – A Double‑Edged Sword

Anthropic switched to token‑based pricing in Q1 2026. Early adopters—large enterprises and AI‑first startups—have burned through their token budgets in months. Uber’s COO, Andrew Macdonald, recently admitted that measuring ROI on AI spend is “very hard”, and that the company exhausted its annual token allocation in just four months.

The problem is twofold:

1. No clear ROI – Companies cannot tie token spend to concrete business outcomes, making budgeting a guesswork exercise.
2. Ceiling on growth – If enterprises cannot justify the expense, Anthropic and OpenAI lose the “sticky” revenue stream they need to fund their compute commitments.

What If the Bubble Bursts?

Zitron sketches several downstream effects:

• Data‑center debt crisis – If construction slows, lenders tied to data‑center loans (private credit funds, mortgage‑backed securities) could face massive write‑offs.
• Hardware supply chain shock – ODMs in Taiwan that assemble servers may see a sharp drop in orders, echoing past GPU‑related downturns.
• Legal exposure for NVIDIA – Over‑shipping or misrepresenting GPU shipments could trigger shareholder lawsuits.
• Collapse of AI‑centric startups – With the two dominant players struggling, the broader AI startup ecosystem could see a funding freeze, leading to a wave of bankruptcies.

Counter‑Perspectives

While Zitron’s calculations are meticulous, several analysts offer a more optimistic view:

• Emerging compute‑efficient models – New architectures (e.g., transformer‑lite, sparsity‑focused designs) could reduce the total megawatt‑hour demand, stretching the existing GPU pool further.
• Alternative hardware – Companies like Cerebras, Graphcore, and custom ASICs may capture a slice of the compute market, easing pressure on NVIDIA’s sales forecasts.
• Revenue diversification – Both OpenAI and Anthropic are expanding beyond pure API access into enterprise tools, consulting, and licensing, which could generate higher‑margin income than token spend alone.
• Policy and regulation – Governments may step in with subsidies or tax incentives for AI‑related infrastructure, effectively lowering the cost of data‑center builds.

Bottom Line

Zitron’s piece is a sobering reminder that the AI boom is being financed by a set of assumptions that may not hold: massive compute demand, relentless revenue growth from a handful of firms, and a data‑center construction pipeline that can keep pace. If any of these pillars wobble, the financial fallout could reverberate across hardware manufacturers, credit markets, and the broader tech ecosystem.

The next installment promises to explore the debt‑market fallout and potential legal battles, which will be essential reading for anyone with exposure to AI‑related investments.

---

For a deeper dive into the numbers and source material, see the original newsletters and the recent filings from NVIDIA, OpenAI, and Anthropic.