What happened
Amazon and Nvidia announced an expanded partnership this week. According to TechCrunch, Amazon will add another 2 million Nvidia GPUs to its AWS data centres in 2027 and 2028. The deal includes Blackwell Ultra, Rubin, and Rubin Ultra chips.
This new agreement comes just five months after an initial order for over 1 million GPUs. Nvidia stated that "demand has exceeded those expectations." The companies cited "surging demand" from startups, enterprises, and AI labs as the reason for the expanded deal.
How the room's reading it
The consensus among infra watchers is that this deal confirms Nvidia's continued dominance. Even as Amazon develops its own Trainium and Graviton chips, it's clear they can't escape Nvidia's gravity for top-tier AI workloads. This massive, front-loaded order signals to the market that hyperscalers are cornering supply for years to come — a move that will likely keep GPU costs high and availability low for everyone else.
Some analysts point to the deal's breadth, which includes CPUs and robotics platforms, as proof of Nvidia's successful ecosystem lock-in. The prevailing view on X is that for builders, this means the compute crunch isn't ending soon. It's a clear indicator that securing GPU access will remain a primary strategic challenge.
Sailfish's take
We're reading this as a clear signal to stop waiting for compute costs to drop. The hyperscalers are buying up the next two years of supply, and the trickle-down economics won't favour startups. For builders, this means the game has changed from chasing scale to chasing efficiency. We've shipped enough products to know that banking on cheaper, faster GPUs arriving just in time is a losing strategy.
The sharpest teams we see are already moving. They're prioritising model optimisation, aggressive caching, and architectures that get more done with less inference. If your entire business model depends on abundant, cheap access to top-tier GPUs, this is the week to rethink it. The real competition isn't about having the biggest model — it's about having the leanest one.