TrendForce Lifts 2027 AI Server Revenue Past $920 Billion as Memory Shortage Inflates Rack Prices

TrendForce raised its forecasts for AI infrastructure again on October 14. Combined capital expenditure among the world's nine leading cloud service providers, Google, Amazon, Meta, Microsoft, Oracle, ByteDance, Tencent, Alibaba and Baidu, is now seen rising 106% year on year to more than $957bn in 2026 and a further 60% to over $1.5 trillion in 2027. The 2027 growth rate was revised up from 50%, which TrendForce attributes to three things: higher prices for complete server systems, higher component quotations driven by shortages, and prepayment demands from component suppliers that force providers to bring cash expenditure forward.

Volumes are growing, but value is growing faster. TrendForce expects AI server shipments to rise 32.2% in 2026 and by nearly 36% in 2027 as hyperscalers are joined by Tier-2 data centre operators such as Tesla and CoreWeave, plus neocloud providers, sovereign AI projects and large language model developers. AI server revenue is put at roughly $479bn in 2026, about 75% of the projected $639.2bn total server market. In 2027 total server market revenue is expected to pass $1 trillion for the first time, with AI server revenue surging more than 90% to nearly $920bn and taking its share to about 81%.

The difference between the shipment number and the revenue number is rack average selling price. TrendForce expects 2027 end-customer orders to centre on rack-scale platforms such as NVIDIA's GB300 and VR200 and AMD's Helios, and states plainly that shortages and rising prices of key components, particularly memory, are driving up the ASPs of next-generation AI server racks. Buying a rack rather than a server moves bargaining power to system integrators with full-rack integration capability, gives suppliers of memory, power management and thermal modules longer demand visibility, and shifts the capital burden onto them through customer prepayment requirements.

For anyone modelling AI infrastructure cost per unit of compute, the implication is that the familiar deflationary curve has inverted. Memory has moved from a fitted component to a strategic input, just-in-time inventory is being replaced by stockpiling to secure allocation, and the cost of expanding a training cluster is increasingly set by HBM and enterprise SSD availability rather than by GPU list prices. TrendForce's numbers imply that general-purpose servers are being squeezed into a shrinking share of a market that is doubling, which is as much a warning for incumbent CPU-centric vendors as it is a forecast for AI hardware suppliers.