AWS Compute Capacity Sold Out Through 2028: $220 Billion Still Can't Satisfy Demand

On October 23, 2026, Amazon CEO Andy Jassy said in a video podcast hosted by a16z co-founder Ben Horowitz and general partner Erik Torenberg that AWS's cloud compute capacity for 2026 is fully booked, most of 2027 is sold out, and a substantial portion of 2028 has already been locked up in advance.

Amazon's capital expenditure plan for this year is about $220 billion, a record for the company. Jassy said: "If we could build more, we would; this would not be an inefficient use of capital." The construction cycles for land, power, and cooling infrastructure lag behind demand growth.

The Structural Logic Behind the Supply-Demand Gap

The cloud computing industry has previously experienced periodic capacity tightness every few years, usually alleviated through expansion within 12 to 18 months. What is different this time is that capacity for 2027 and 2028 has been booked in bulk through multi-year contracts, with customers locking in long-term deals to hedge against future compute risk.

Jassy described current demand as "barbell-shaped": at one end are AI labs using large-scale compute for model training and inference; at the other are enterprise customers applying AI to tasks such as customer service automation and fraud detection. The middle layer—the broad embedding of AI into everyday business software—remains blank, but he believes this layer could eventually become the largest source of demand.

Not Just an Amazon Problem

Microsoft Azure has a backlog of about $80 billion in orders it cannot deliver in time because power and cooling infrastructure deployment is lagging. Microsoft's capital expenditure for fiscal 2026 is annualized at about $150 billion, while Google's 2026 capex guidance exceeds $75 billion. The simultaneous structural capacity tightness at the three major cloud providers indicates that demand has established a new baseline.

A Paradox: Sold Out, but Still Too Expensive

Jassy said AI inference costs still need to fall to one-tenth of current levels for widespread adoption. Existing compute has been locked up through multi-year contracts by a small number of high-consumption AI labs and large enterprises, leaving smaller developers with limited room for cost declines and lower priority over the next two years.

In-House Chips: A Hedging Path or a Strategic Bet

Jassy positioned Amazon's in-house Trainium chips as a core means of reducing AI compute costs, saying their price-performance is 30% to 40% better than existing solutions. This logic is consistent with the earlier development of Graviton chips. However, in-house chips typically take three to five years from development to mass production, and the window in which Trainium can fill the gap may overlap with a period of demand saturation.

The larger significance of in-house chips lies in building differentiation: against the backdrop of rising Nvidia GPU prices and lengthening lead times, they offer customers an alternative path that does not depend on a single supplier.

What Multi-Year Pre-Sales Really Mean

Cloud compute is usually sold on a monthly or annual basis, so bulk prepurchases spanning two to three years are unusual. Such contracts show that buyers are willing to accept time risk in exchange for capacity guarantees at locked-in prices. If AI demand proves to be a short-term bubble, these contracts will become a heavy burden. Therefore, multi-year bookings themselves constitute a statement of market confidence: the signatories believe they will still need this compute over the next two to three years.

The Risk That Truly Deserves Attention

Jassy said AWS will reserve a portion of capacity for on-demand access and Bedrock inference services. That reserved proportion determines how much compute developers without long-term contracts can obtain over the next two years and at what price. Once a structural shortage is established, pricing power shifts to the supply side, and upward pressure on the cost curve will persist until a new wave of data centers is built and brought online.

Independent Judgment

$220 billion in capital expenditure still cannot meet demand, and multi-year contracts have locked up capacity for the next two years. These facts indicate that investment decisions upstream in the industrial chain have undergone an irreversible shift. Infrastructure construction cycles are three to five years, and supply will be released on a large scale around 2028 to 2030. Over the next two years, access to compute will depend on customer identity and the degree of advance locking.