On October 14, 2026, TrendForce announced a forecast that capital expenditure by nine major global cloud service providers (CSPs) will exceed $1.5 trillion in 2027. That is up from roughly $1.3 trillion in its August forecast, and the projected year-over-year growth rate has been raised from about 50% to 60%. Demand for AI servers continues to expand, but the higher figure also reflects rising prices for servers and components, as well as prepayments to secure supply. To read growth in spending as growth in computing capacity, price and payment timing need to be considered separately.
What changed from the August $1.3 trillion forecast
TrendForce's 2027 capex forecast for the same nine companies was revised upward from about $1.3 trillion on August 3 to more than $1.5 trillion on October 14. The North American group consists of five companies: Google, Amazon and Meta, plus Microsoft and Oracle. The Chinese group includes ByteDance and Tencent, as well as Alibaba and Baidu.
Placing the August 3 announcement and the October 14 announcement side by side, using the same target years and metrics, shows what changed.
| Metric | August 3 forecast | October 14 forecast |
|---|---|---|
| Nine companies' 2026 capex | More than $886.7 billion | $957.0 billion |
| Year-over-year growth in 2026 | About 90% | 106% |
| Nine companies' 2027 capex | About $1.3 trillion | More than $1.5 trillion |
| Year-over-year growth in 2027 | About 50% | 60% |
| Growth in global 2026 AI server shipments | About 31% | 32.2% |
Both announcements are forecasts, and the amounts have been converted to U.S. dollars. Only the shipment figure covers the global market rather than just the nine companies. Because the figures include "about" and "more than," we have not calculated an exact revision rate for capex. The announcements also do not make clear how company-specific accounting treatments were adjusted in the aggregate.
The revision also reaches the 2026 baseline. In 2027, spending is projected to grow a further 60% from that higher base. In August, TrendForce explained that the 2027 growth rate was slowing because the prior year's investment scale had become so large. In October, it explicitly named price and payment timing as reasons for the higher spending, which stands out.
Between 36% unit growth and 90%-plus revenue growth
TrendForce forecasts that AI server shipments will grow about 36% year over year in 2027, while revenue will grow more than 90% to approximately $920 billion. Revenue growth is far larger than unit growth. The configuration and pricing of the systems buyers procure offer clues to the gap. TrendForce names rack-scale systems as the center of 2027 orders, including NVIDIA's GB300 and VR200 and AMD's Helios. It particularly expects supply shortages and price increases for key components, memory above all, to push up the average selling price of next-generation AI server racks. The view is that shipments will rise while the systems being sold also become more expensive.
As a result, the dollar composition of the server market also shifts. In 2026, AI servers are forecast to account for about $479 billion of the $639.2 billion in combined general-purpose and AI server revenue, roughly 75%. In 2027, the overall server market is expected to exceed $1 trillion for the first time, and AI servers' share of revenue is expected to rise to about 81%.
The 81% figure is a share of revenue, not of unit shipments. The forecast is that expensive AI systems gain presence in dollar terms; it would be a leap to read it as general-purpose servers falling to the same share by units. Nor can the gap between revenue and shipment growth be used to calculate a price increase for the same product, because the average price changes when the product mix changes.
Prepayments spend money before servers start running
TrendForce cited three factors behind the capex forecast increase: higher prices for finished server systems, higher component prices due to supply shortages, and prepayments to secure supply. Prepayments in particular need to be considered separately from equipment price increases, because they bring cash outlays forward to lock in components to be received in the future.
Even if a buyer pays first, the servers are not necessarily delivered and put into operation at that point. In other words, reading an increase in capex requires accounting for a time lag between when payment is made and when the equipment becomes usable, in addition to purchase volumes and prices. If spending to secure supply comes first, the period from payment until the equipment generates revenue also needs to be watched.
Still, it would be a stretch to explain the increase by prepayments alone. TrendForce says companies outside the large CSPs, such as Tesla and CoreWeave, continue to buy AI servers. In 2027, emerging cloud providers and national AI infrastructure buildouts are also expected to support demand, and new models and services from large language model developers should lead to procurement. The forecast rise in shipments shows demand expanding beyond what price and payment timing can explain.
The published forecast depicts rising procurement costs layered on top of this growing demand. However, it does not state what share of the increase comes from prices and prepayments. From the total of more than $1.5 trillion, it is difficult to derive a single figure for actual added computing capacity or the speed of investment recovery.
Microsoft's lease change shows a caution in comparing amounts
Reported capex figures can also move because of accounting treatment. In its July 29, 2026 earnings presentation, Microsoft explained that it will extend the estimated useful lives of data centers and office buildings from 15 years to 25 years at the start of fiscal year 2027. As a result, it said many future data center leases will shift from finance leases to operating leases.
Microsoft's capex metric includes finance leases but not operating leases. Reflecting this reclassification, its calendar 2026 capex outlook becomes about $175 billion. Microsoft also explained that, excluding the effect of the useful-life change, its investment outlook is unchanged. Even if contracts for using facilities continue in the same way, the scope of what is counted in the capex metric can change.
This is Microsoft's own explanation as of July, and it is unclear how TrendForce reflected it in its nine-company aggregate. There is also no basis for directly linking the two sets of figures and attributing the latest upward revision to an accounting change. Still, when comparing companies' capex outlooks, adding them up without aligning the periods covered and lease treatment risks misreading actual increases or decreases in investment.
To judge the expansion of AI infrastructure in 2027, it will be necessary to track not only updates to total investment but also when secured components are delivered and when servers go into operation. Whether the growth in spending leads to real capacity gains can be confirmed by how the equipment ramps up after the money is spent.