The great compute grab

Why the newest gigawatt is worth more

Electricity is only the container. The amount of useful work inside a power budget keeps improving. Newer accelerators can deliver several times the throughput per megawatt of older systems. NVIDIA says GB300 can produce about five times more throughput per megawatt than Hopper on one DeepSeek-R1 benchmark; SemiAnalysis has measured a similarly large jump for Vera Rubin on another inferenceⁱ comparison. S19 S03

This creates a compounding advantage. A laboratory that secures a large share of the newest power capacity can gain still more ground if every new watt also performs more work. Its share of usable computing power can therefore rise much faster than its share of electricity.

Value capture has already moved around the stack. In 2023 and 2024, chip and infrastructure suppliers captured much of the visible gross profit while several model providers were still selling tokens at poor or negative gross margins. By 2026, SemiAnalysis estimated that Anthropic’s inference economics and revenue per megawatt had improved sharply. S35 S02

Yet the biggest beneficiary may still be the customer. A trading firm, advertiser, manufacturer or pharmaceutical company can pay a model provider only a fraction of the value created by a better decision. Jane Street is an unusually vivid example because small improvements can be worth enormous sums in financial markets; OpenAI lists the firm among early users of GPT-5.6 Ultrafast. S29

Lithography is only one choke point. Leading accelerators also depend on advanced logic wafers, high-bandwidth memory, packaging, substrates, networking, cooling, transformers, turbines and grid connections. SemiAnalysis estimated that AI-related products would consume just under 60% of TSMC’s N3 output in 2026. S04

The commercial effect resembles a bullwhip travelling backwards through the supply chain. Model economics improve first; owners of scarce compute can then charge more; chip, memory, foundry and equipment suppliers respond in turn. Prices can adjust in weeks. New factories, turbines and lithography capacity take years.

Consider the public-company thought experiment used in the source arithmetic. A 20GW fleet reallocating 10% of capacity from inference to trainingⁱ would move 2GW. At $100bn of annual revenue per inference gigawatt, that appears to sacrifice $200bn of present revenue. The bet is that the next model can be worth still more.

Raw gigawatts can be deceptive. A gigawatt filled with older accelerators can deliver far less useful AI work than a gigawatt of the newest American hardware. National comparisons therefore need a quality adjustment as well as a power figure.

The policy constraint is concrete. American export rules introduced in 2022 and tightened thereafter restrict advanced accelerators, semiconductor-manufacturing equipment and high-bandwidth memory. S21 Estimates of smuggling, diverted foundry output and domestic Chinese production are much less transparent and should be treated as market intelligence rather than audited statistics.

The most aggressive arithmetic assumes 100GW of annual compute additions and puts critical IT spending alone near $5tn. Pre-funding the data centres and power plants required for later years pushes the scenario toward $7–10tn a year. It is intentionally extreme, but it illustrates how quickly exponential infrastructure demand collides with the size of the real economy.

More conservative external baselines are still enormous. Goldman Sachs models about $765bn of AI capital expenditure in 2026, rising to $1.6tn in 2031. Morgan Stanley has estimated nearly $3tn of global AI-related infrastructure investment through 2028, while PwC puts annual data-centre capital expenditure around $800bn in 2026. S14 S15 S31

The historical analogy is the Volcker-era debt shock. The source conversation cites roughly 40 countries defaulting in the 1980s. World Bank research finds 40 countries fell into arrears and 27 restructured, while Federal Reserve research documents the wave of emerging-market crises around the early-1980s tightening. S27 S28

The forward-looking step is far more speculative. One scenario moves hyperscalerⁱ borrowing costs from roughly 5–6% toward 8%, an increase of about 250 basis pointsⁱ, and assumes part of that rise spreads through credit markets. Reuters has already reported concern that large American hyperscaler bond issuance could crowd out other borrowers in Europe. S32

A laboratory that is slightly ahead may enjoy several reinforcing advantages at once. It can monetize scarce compute better, buy more of it, learn from wider deployment and use stronger internal models to help create the next generation. Each advantage makes the next one easier to obtain.

There is no neat governance answer in the arithmetic. Private concentration, public control and regulation each carry their own risks. The narrower conclusion is more robust: decentralisation should not be assumed. It has to overcome strong economies of scale.