Amazon Web Services Capacity Sold Out Through 2028; Jassy: Server Procurement Can Be Flexibly Adjusted
nashnova research
Most AWS compute capacity is booked through 2028, with 2026 fully sold out. CEO Andy Jassy says server procurement needs only months of lead time, letting Amazon scale back if demand softens.
Capacity sold through 2028 — what does that actually mean?
AWS capacity for 2026 is fully sold out. 2027 is largely sold out. A significant portion of 2028 is already locked in.
This means → AWS has unusually high revenue visibility for the next two to three years — large customers have committed upfront, sharply reducing top-line uncertainty.
In plain terms = the nightmare for any cloud builder is "we built it and nobody came." Right now the problem is the opposite: built and still not enough.
If everything is sold, can on-demand users still get in?
Jassy confirmed AWS will still reserve capacity for on-demand services and its inference offerings Bedrock and EC2.
This means → bulk capacity went to long-term contract customers, but the spot-access channel stays open — just with tighter available headroom.
This reflects a deliberate trade-off: long-term contracts lock in certainty, while keeping on-demand access alive preserves ecosystem stickiness.
What if demand slows down?
Jassy noted that AWS needs only a few months' lead time to procure servers and networking gear — no need to fill a data center all at once.
His words: "If we don't ultimately need that much capacity to fill the data center, we just won't buy the server networking equipment."
In plain terms = the "shell" of a data center — land, power, building — is built ahead. How many servers go inside is decided later. If demand doesn't show up, Amazon skips the hardware; the loss is idle-shell cost, not an entire production line.
Was the custom-chip bet the right call?
Jassy called the acquisition of chip-design firm Annapurna Labs more than a decade ago perhaps "the best acquisition Amazon has ever made."
Annapurna Labs is the technical foundation for Amazon's custom AI chip Trainium and its CPU line Graviton.
This means → Amazon is no longer just a cloud "landlord" — it is building its own "appliances." Custom silicon is gradually reducing its single-source dependence on Nvidia for compute cost.
Anthropic and Trainium — who helped whom?
Jassy revealed that Amazon's investment in Anthropic has in turn driven iteration on the Trainium chip.
Anthropic is already training on Trainium, and the two sides have "learned a great deal from each other" technically.
In plain terms = Amazon funds Anthropic; Anthropic runs models on Amazon's chip; the process reveals where the chip falls short. The investment doubled as a free stress-tester.
市场有风险，内容仅供研究参考，不构成投资建议。