AWS sells priority capacity as GPU economics split on the card you rent

AWS sells priority access as 2027 capacity sells out

AWS capacity for 2027 is largely sold out, Jassy said on 23 October, and part of 2028 has gone too. The Bedrock price list, updated on 3 November, sells that scarcity.

Priority access costs a 75% premium to Standard pricing. Flex comes at a 50% discount, and Batch is half price on selected models. The tiers let AWS sell the same model to a buyer who needs guaranteed speed and to one who can wait. Demand is the reason. AWS grew 39.3% year over year in the third quarter, according to Amazon's results. That was its sixth straight quarter of acceleration and its fastest growth in nearly five years. Amazon planned roughly $220 billion of capital spending for 2026. On 30 October it said it was installing an AI accelerator about every 14 seconds.

AWS keeps a reserve for on-demand use, EC2 and Bedrock inference. It has not said how large that reserve is. The price list is now the clearest public signal of how tight capacity is.

The GPU card decides whether hosting pays

Analysis published on 3 November found that an H200 card loses money even when fully busy. A B300 makes money once it is 35% busy. On an 8xB300 node running at 50% to 70% utilization, gross margin lands between 31% and 51%. Which card you rent matters more than which inference engine you run, the analysis argues. OpenRouter lists 29 providers selling DeepSeek V4.1 Flash. The cheapest input is $0.056 per million tokens, against about $0.30 for most. That is below cost by any math. That model alone took roughly 30% of all tokens on OpenRouter in the week of 19 to 25 October.

Token prices keep falling for buyers

For buyers, the direction runs the other way. Epoch AI puts the fall at about 47% a quarter since 2023 for a fixed level of AI performance, roughly 13 times a year.

Stanford's 2025 AI Index puts the trend in longer perspective. A query at GPT-3.5-level performance cost $20 per million tokens in November 2022. By October 2024 Gemini 1.5 Flash-8B cost $0.07 for the same work, a decline of more than 280 times. GPT-6.1 Terra, released on 26 October, lists at $1 per million input tokens and $5 per million output. That is 50% below GPT-5.6 Terra on fresh input, 58% below on output and 75% below on cache reads. In a worked agent session of 10 million input tokens, 97% of them cached, plus 200,000 output tokens, Terra costs $1.79 against $4.94 for GPT-5.6 Terra.

The margin squeeze behind the price cuts

Cheap tokens do not make an easy business. Citadel Securities' Tokenomics 2.0 notes that AI token prices collapsed by roughly 50% in six months while GPU rental costs moved the other way.

Modal CEO Erik Bernhardsson says 90% gross margins are gone, and that Modal targets 40% to 60% over the long term. His argument is that selling tokens on their own carries no switching cost. So the winners will bundle sandboxes, post-training and batch inference around the meter. Investors are still paying for growth. Modal raised $750 million after growing 50 times over the prior twelve months, according to Bernhardsson.

What happens next: optical trials and sold-out capacity

Some of the answer is meant to come from hardware, not from pricing. ARIA, the UK public R&D funding agency, will deploy Lumai Iris optical inference servers in its Scaling Inference Lab. The plan comes with claims of 90% lower data center power consumption and a total cost of ownership around one-tenth of conventional GPU architectures. Those are agency and vendor claims. The test is a live commercial setting rather than a lab bench.

On the demand side, Jassy said AWS could grow above 40% in both 2027 and 2028. Until new capacity lands, the tier prices are where that expectation shows up first.