The AI spending boom: more computing, or more expensive computing?

Spending more money on AI infrastructure and buying more computing are not the same thing. A company can double its capital budget and receive less than double the processing power if prices rise, if it buys more expensive configurations, or if it pays suppliers earlier to secure parts.

TrendForce's latest AI server forecast, published on 14 October 2026, makes that distinction unusually visible. Its projections suggest that much of next year's growth in AI server revenue comes from higher value per system rather than from more systems. That does not mean the spending is unproductive. It means the headline numbers measure dollars, and dollars and computing capacity have started to move at different speeds.

What TrendForce is forecasting

TrendForce tracks nine large cloud service providers (CSPs): Google, Amazon, Meta, Microsoft, Oracle, ByteDance, Tencent, Alibaba and Baidu. It projects their combined capital expenditure to rise 106% to US$957 billion in 2026 and to surpass US$1.5 trillion in 2027, a further 60% increase. Both figures were revised up from earlier estimates.

Capex is broader than servers. It covers data-centre buildings, power equipment, networking and other long-lived assets. So the capex forecast and the server forecast should be read as two separate views of the same boom, not added together.

On servers, TrendForce expects AI server revenue of about US$479 billion in 2026, nearly 75% of a US$639.2 billion total server market. For 2027 it projects AI server revenue up more than 90% to nearly US$920 billion, about 81% of a server market that would pass US$1 trillion for the first time.

Volume, price and mix

Three forces sit inside a revenue forecast. Volume is the number of units shipped. Average selling price (ASP) is revenue per unit. Mix describes what kind of units are shipped. In AI servers, mix is moving fast: TrendForce expects 2027 orders to centre on rack-scale systems such as NVIDIA's GB300 and VR200 and AMD's Helios series. A complete rack of tightly linked accelerators is a very different product from an older stand-alone server, so a higher price per unit partly reflects more computing in each unit.

Putting TrendForce's numbers side by side shows the gap. AI server shipments are forecast to grow nearly 36% in 2027, while AI server revenue grows more than 90%. Dividing one growth factor by the other gives an approximate 41% rise in revenue per unit shipped. That figure is a rough derivation from rounded forecasts. It mixes price, configuration and component cost, and it cannot say how much of the increase is pricing power.

Units versus dollars

Indexing both series to 2026 makes the divergence plain. On TrendForce's forecasts, shipments reach about 136 in 2027 while revenue reaches about 192. The difference between the bars is the space occupied by higher value per system.

Shortages and prepayments

TrendForce attributes its upward revision of 2027 capex growth, from 50% to 60%, to three factors: rising prices for complete server systems, higher component quotations amid supply shortages, and advance payments some component suppliers require to secure supply. It singles out memory as a key component whose shortage and price rises are lifting rack ASPs.

The third factor is about timing rather than quantity. If cloud providers pay suppliers earlier, cash leaves their accounts sooner, raising reported capex in a given year without changing how many systems eventually arrive. Part of the 2027 surge may therefore be spending pulled forward rather than extra capacity.

Who could benefit along the chain

Higher system value flows to the suppliers inside each rack. TrendForce names accelerator designers through the rack systems it expects, and memory through its price commentary. Networking, advanced packaging, power delivery and cooling all scale with rack-level systems, but the press release does not quantify any supplier's share, so none is assigned here.

The counter-argument

A more expensive system can still be cheaper per unit of work. If a new rack delivers far more processing per dollar or per watt than the servers it replaces, rising revenue per unit is consistent with falling cost per AI task. Price increases and productive investment are not mutually exclusive.

It is also worth separating three measures that are often blurred: physical data-centre capacity (power and floor space), servers shipped, and usable computing capacity. A forecast in dollars tells us about the first only indirectly and about the third not at all.

What to watch

The useful indicators are concrete. Watch actual AI server shipments against the 36% forecast; the mix between rack-scale and conventional systems; capacity utilisation at cloud providers; memory contract prices; cloud providers' free cash flow as capex rises; and, above all, revenue from AI services. If service revenue grows alongside the hardware bill, the boom buys more computing. If it lags while prices climb, it is mostly buying more expensive computing.