Inside an AWS rack in 2028, Amazon's own accelerator will read from memory NVIDIA designed. NVIDIA and Amazon's Annapurna Labs are extending the NVLink Fusion interconnect to a new NVIDIA custom memory technology, which would give Trainium access to faster, more power-efficient memory. Trainium is the chip Amazon built so it would need fewer GPUs, and NVIDIA is now selling into it.
The quarter behind that agreement was the largest NVIDIA has reported, with Data Center revenue of $89.0 billion, up 117% from a year ago. The guide for the current quarter is $108.0 billion, and it assumes no Data Center compute revenue from China. The number everyone read in the agreement itself was the GPU count, 2 million additional NVIDIA GPUs across AWS's global infrastructure in 2027-2028.
For a buyer the effect shows up in the spread. When the accelerator, the interconnect, the memory and the CPU all come from one vendor, a rack quote stops being a set of competing bids. The rest of that agreement is a parts list: AWS is also taking NVIDIA Vera CPU-based infrastructure, alongside the silicon Amazon designed itself.
The support-assistant floor went to $416 a month from $137, after the cheapest listed route for DeepSeek V4 Pro repriced to $1.60 and $3.20 a million tokens from $0.53 and $1.05. All four neocloud GPU ranges held for a third week.
