A commodity market hiding inside a technology product
Cloud computing is sold with the vocabulary of software, but it is structurally a capacity market with all the classic features: perishable inventory, a spot price, forward contracts at a discount, and price discrimination by commitment level.
An idle server earns nothing and the hour cannot be recovered, exactly like an empty airline seat or a ship sailing in ballast.
Three price tiers
On-demand — pay by the second, no commitment, highest price. Reserved or committed use — commit to one or three years for a discount often approaching 50–70%. Spot — bid for unused capacity at a large discount, on the condition that the provider can reclaim it with a couple of minutes' notice.
Spot pricing is a genuine auction-like market that varies by instance type, region and availability zone, and sophisticated users architect workloads specifically to be interruptible so they can buy at spot rates. That is capacity arbitrage in the ordinary commodity sense.
AWS operates a marketplace allowing standard reserved instances to be resold, which makes committed capacity a genuinely transferable asset.
The GPU squeeze
Demand for AI training and inference turned high-end accelerators into the constrained resource in the entire industry from 2023 onwards. Access to H100-class hardware became a competitive advantage in itself, with multi-year capacity contracts, allocation queues and a proliferation of specialist providers buying hardware to rent out.
This is the standard shape of a commodity shortage — scarcity, allocation by relationship, entry by new suppliers, heavy capital investment — applied to a product that did not exist in that form five years ago.
Why the shape matters
The real constraints are increasingly physical: electrical supply, grid connections, cooling water and planning permission for data centres. Cloud capacity is converging with the electricity market, and in some regions the binding constraint on new compute is a grid connection date years in the future.
How you would actually trade it
For a business buying capacity, the decisions are the same as any commodity procurement: what proportion of expected consumption to commit forward, what to buy on demand, and what to run interruptibly on spot.
Committing three years for a 70% discount is a forward purchase. It is a good trade if your usage is predictable and a bad one if your workload changes — which is why the secondary marketplace for reserved capacity exists.
For investors, exposure comes via the hyperscalers, the chip designers and manufacturers, and increasingly through data centre REITs and the power generators supplying them.
What goes wrong
- Commitment risk. Multi-year reservations become stranded costs if usage changes.
- Egress fees and lock-in. Moving data out is priced to discourage leaving, making provider switching expensive.
- Spot interruption requires genuinely fault-tolerant architecture.
- Hardware obsolescence is rapid; a three-year GPU commitment can be outdated.
- Concentration risk in a market dominated by three providers, which the CMA has examined in a market investigation.
This entry describes how a market works. It is not a recommendation, a price forecast, or advice about whether this belongs anywhere near your money. How entries are written.