AWS raised on-demand H200 GPU instance prices by 15% in January 2026. The first GPU price increase in roughly 20 years, per Cast AI.
In the same move, it cut H100, H200, and A100 prices by up to 45% for committed customers. The pricing message is singular and clear: whoever commits pays less than ever; whoever does not pays more than ever. The gap between the two worlds has never been wider.
The price became a maturity ruler
For twenty years, GPU prices only went down: capacity arrived, cost fell, and nobody needed a purchasing strategy. January’s hike ends that era: frontier GPU became a contested commodity, and the on-demand price is now the instrument separating planners from reactors.
The design is classic mature utility: full price punishes improvisation; the discount rewards predictability. And the discount, right now, is historically large.
The instruments and their numbers
Four levers, with measured discounts:
- Reserved Instances: up to 72% off, for stable long-term load.
- Savings Plans: up to 66%, with instance flexibility.
- Spot: 60 to 91%, for what tolerates interruption.
- Dedicated inference chips: Inferentia2 delivers up to 4 times better price-performance for models under 13B parameters.
The bottleneck is not knowing the instruments: it is coverage. Most companies cover only 55% of eligible compute with commitments, and the rest runs on the rate that just rose 15%.
The design that avoids the liability
The piece that organizes everything is separating inference from training. Inference is stable: 24/7, predictable demand, commitments fit. Training is bursty: concentrated peaks, interruption-tolerant, spot fits.
Contracting both with the same instrument pays a design tax in both directions: inference on on-demand wastes the discount; training on an underused reservation becomes the zero-net-benefit commitment we have already documented. The separation, per CloudZero, cuts total AI cost by 35 to 50% without switching vendors.
And the final ruler remains utilization: with the enterprise average at 5%, effective GPU cost is 20 times nominal. With the H100 at 12.29 dollars per hour and 5% usage, each effective GPU-hour costs 245.80 dollars. AWS’s hike did not create that problem; it made every lost utilization point more expensive.
Conclusion
AWS’s price hike is the marker of the GPU market’s maturity: the era where waiting a quarter solved things is over, because the rate fell on its own.
From now on, AI GPU purchasing is commitment management: separate loads, cover the measured floor, use spot for the interruptible, reassess against usage every quarter, and compare chips. Whoever structures this negotiates with the vendor; whoever does not negotiates with the invoice. Both pay, at very different prices.