Analysis
Amazon Web Services and Nvidia are expanding their infrastructure partnership again, with AWS planning to deploy an additional 2 million Nvidia GPUs across its global data-center footprint in 2027 and 2028, according to reporting on the companies' latest agreement, FX Leaders reported. The announcement landed alongside Amazon's decision to raise its 2026 capital-expenditure forecast to roughly $220 billion, up from an earlier estimate near $200 billion.
The Scale of the Commitment
Two million additional GPUs is a number worth sitting with: it's on top of whatever Nvidia hardware AWS has already deployed across its existing fleet, and it commits Amazon to years of chip supply, power procurement and data-center construction before most of that capacity generates a dollar of cloud revenue. Pulse has tracked Amazon's AI infrastructure buildout as capex guidance has climbed through 2026, and this raise -- the second upward revision this year -- fits a pattern common across all four hyperscalers: guide up, spend more, and ask investors to trust that demand catches up.
Amazon isn't alone in that posture. Yahoo Finance has tracked the rest of the hyperscaler capex race for 2026:
- Alphabet -- guiding as high as $205 billion
- Microsoft -- tracking toward roughly $190 billion
- Meta -- guiding to $115-135 billion
Combined hyperscaler AI spending is now running well past $700 billion for the year. What separates Amazon's raise is the direct pairing with a specific, named Nvidia hardware commitment rather than a vaguer capex number, giving investors an unusually concrete unit -- GPUs, not just dollars -- to underwrite.
Investors Aren't Fully Buying It
The market's reaction has been lukewarm at best: Amazon's stock has retreated from around $287 toward the mid-$250s even as the company keeps raising its own guidance, a sign that “we're spending more” is no longer, on its own, the reassurance it was a year ago. That's the same skepticism showing up in Friday's chip-sector selloff following Marvell's cautious guidance -- investors want spending translated into deployed, revenue-generating capacity, not just bigger forward commitments.
Competitively, the GPU race among hyperscalers is now as much about securing allocation as building differentiated infrastructure: Microsoft, Google and Meta are each pursuing custom silicon (Maia, TPU, MTIA respectively) specifically to reduce dependence on Nvidia supply, even as they keep buying more Nvidia hardware in parallel -- Amazon's own Trainium chips follow the same hedge-while-still-buying logic. A 2-million-GPU commitment this large suggests AWS still sees Nvidia as the primary lever for the next two years of capacity growth, custom silicon investments notwithstanding.
The Policy Wildcard
The deal also lands amid fresh uncertainty: Commerce Secretary Howard Lutnick has signaled the administration is weighing tariffs on semiconductors and the finished electronics built with them, a policy shift that would raise costs across exactly the kind of large-scale hardware deployment AWS just committed to. Nothing in the reported agreement addresses how a new tariff regime would affect its economics -- a multiyear infrastructure commitment signed before a major trade-policy decision carries real execution risk regardless of how sound the underlying demand thesis is.
For startups building on AWS, the practical read is capacity: two million more GPUs landing in 2027-2028 should ease allocation pressure for compute-heavy workloads over that window, assuming Amazon's demand assumptions hold and the tariff question resolves without disrupting supply chains already stretched thin across the industry.
