AWS GPU Prices Jump 20% as Spot and Reserved Markets Split

On July 1, 2026, Amazon Web Services will raise the hourly price of renting its most powerful GPU instances by 20 percent. The new rates, reported by Yahoo Finance on June 26, peg the P6-B300 instance at $14.04 per accelerator-hour, the P6-B200 at $12.355, and the P5 in US regions at $5.191. It is the second increase AWS has applied to GPU capacity this year, following a 15 percent hike earlier in 2026, according to The Next Web. The cumulative effect is a roughly 38 percent price escalation inside six months on the world's largest public cloud. For a startup running a 32-GPU training job on P6-B200 instances for two weeks, the bill just went from approximately $99,500 to $119,400, before any reserved-instance discount.

The hike is not a surprise to anyone tracking the physical inputs. Memory prices have been climbing since late 2025 as AI data centers absorbed an outsize share of global DRAM and NAND flash production. Tom's Hardware reported in October 2025 that legacy DRAM supply was lagging behind demand, sending prices upward across all memory categories. Separately, Forbes reported in April 2026 that helium gas, essential to advanced semiconductor fabrication, had entered a supply crisis, tightening the entire chip supply chain. These are not abstractions. They are line items on a foundry invoice that eventually land on a cloud bill.

What makes the July 1 increase instructive is not the number itself. It is what the number reveals about the structure of the GPU compute market in mid-2026. There are now effectively two markets. The first is the reserved-instance and long-term contract market, where hyperscalers, frontier labs, and sovereign AI projects commit to multi-year, multi-gigawatt deals at negotiated rates. The second is the on-demand and spot market, where smaller labs, inference startups, and academic researchers buy compute by the hour at whatever price clears. The 20 percent hike applies to the second group. The first group, by and large, does not see these increases on their invoices.

The spot market for GPU compute has always been volatile. What is new is the institutionalization of that volatility. In May 2026, CME Group announced a partnership with data provider Silicon Data to launch cash-settled futures contracts tied to GPU computing costs, CNBC reported. Days later, Intercontinental Exchange, the owner of the New York Stock Exchange, teamed with index provider Ornn to announce a competing suite of compute futures, as covered by The Next Web. By the end of 2026, a quant fund will be able to hedge exposure to P6 instance pricing the way an airline hedges jet fuel.

This financialization is a rational response to an irrational procurement landscape. A mid-sized AI company that needs 64 H200 equivalents for a three-month fine-tuning run has no way, today, to lock in that price before the credit card is charged. The cloud providers offer reserved-instance discounts of 30 to 50 percent for one-year or three-year commitments, but those commitments are denominated in instance-hours, not in tokens or flops. A company that signs a three-year reserved-instance deal for P5 capacity is making a bet that P5 will still be the right SKU for its workloads in 2028. That is an increasingly difficult bet to make.

The supply side is not helping. NVIDIA's Blackwell generation GPUs are effectively sold out through mid-2027, eWeek reported, with hyperscaler pre-orders consuming the entire allocation pipeline. At the consumer end, the situation is equally tight. TechTimes reported on June 21 that NVIDIA had cut GPU supply to manufacturing partners by as much as 20 percent for the RTX 50 series, with two mid-range cards effectively delisted from the channel. Consumer graphics and data center AI accelerators share the same TSMC CoWoS advanced packaging capacity. When one market is starved, it is usually because the other is being fed.

AMD has positioned itself as the beneficiary of this concentration risk. A Seeking Alpha analysis published June 26 noted that AMD's AI GPU position is strengthened by hyperscaler diversification needs, citing multi-year, multi-gigawatt deals from Meta and OpenAI. The logic is straightforward: no hyperscaler wants its entire inference fleet dependent on a single silicon vendor whose supply lead times stretch past 18 months. AMD's MI300X and MI400 series accelerators are not performance-matched to Blackwell on a per-chip basis, but on a per-available-watt basis at the datacenter level, the calculus shifts.

Who captures the margin

The 20 percent price increase on AWS GPU instances raises a question about the margin stack. If the cost of the underlying silicon has not changed, and if the cost of power and cooling is largely fixed by long-term utility contracts, then the incremental revenue from the price hike flows almost entirely to AWS's operating margin. Business Insider noted on June 27 that the price increases come as memory chip costs soar, which suggests AWS is passing through some component cost inflation. But the magnitude of the pass-through, 20 percent on the instance price against what is likely a single-digit percentage increase in AWS's per-instance cost basis, implies margin expansion.

The neoclouds, CoreWeave, Lambda, Voltage Park, and the dozens of smaller GPU-as-a-service operators, sit in a different position. They do not own the silicon. They lease it from NVIDIA or from hyperscaler surplus, and they resell it by the hour or by the reserved block. When AWS raises on-demand prices, the neoclouds gain pricing headroom. A startup quoted $14.04 per P6-B300 accelerator-hour on AWS might find CoreWeave offering an H200 equivalent at $9.80, with availability guaranteed inside a week. The neoclouds are not charities. That $9.80 still carries a margin. But the spread between hyperscaler on-demand and neocloud reserved represents the price of not having a multi-year contract.

CoreWeave, which went public in early 2025 and reported a $99.4 billion contract backlog as of mid-2026, according to a Seeking Alpha analysis published June 20, occupies the middle ground. Its business model is built on acquiring GPU capacity in bulk at preferential rates and reselling it with a margin. But its contracts with NVIDIA are themselves subject to the same allocation dynamics that push customers toward the neoclouds in the first place. If Blackwell supply remains constrained through 2027, CoreWeave's ability to expand its fleet at competitive rates becomes the binding constraint on its growth.

The compute futures contracts being developed by CME and ICE are designed, in part, to give this middle layer of the market a hedging instrument. A neocloud that commits to a fixed price for a 12-month reserved GPU block can use futures to offset the risk that spot prices fall below its contracted rate. A frontier lab that expects to spend $800 million on inference compute in 2027 can buy futures to cap its exposure. In theory, this deepens the market and lowers the cost of capital for everyone. In practice, the contracts will be cash-settled against benchmark indices whose methodology has yet to be published. The composition of those indices, which GPU SKUs are included, at what batch size and sequence length assumptions, will determine who benefits.

What the per-token price implies

The AWS price hike filters down to the per-token economics of inference in ways that are not always visible to end users. At $14.04 per P6-B300 accelerator-hour, assuming a model serving at batch size 1 with a time-to-first-token of 200 milliseconds and 50 output tokens per second, the raw compute cost per million output tokens is approximately $0.78. Add the cost of the model provider's margin, the application layer's margin, and any middleware tax, and the retail price per million tokens that a customer sees on an invoice might be $2.50 to $5.00. A 20 percent increase at the infrastructure layer, if not absorbed by someone in the stack, becomes a 20 percent increase at the application layer.

The assumption set matters. At batch size 32, the per-token cost collapses. The same P6-B300 instance, running inference at batch size 32 instead of batch size 1, might deliver per-token costs closer to $0.08. The difference between $0.78 and $0.08 is the difference between a consumer chatbot that loses money on every query and an enterprise API that runs at 65 percent gross margin. Every benchmark published without disclosing batch size, sequence length, and hardware SKU is, intentionally or not, cherry-picking a regime.

The GPU price increases also interact with model architecture decisions. If per-token inference costs rise 20 percent, the economic incentive to distill a large model into a smaller one, to quantize from FP16 to INT8, or to move from dense to mixture-of-experts architectures increases proportionally. A model provider that can serve GPT-5-class performance at half the per-token compute cost captures not just margin but market share. The cloud providers, by raising prices, are effectively taxing model inefficiency. That is not a policy objective. It is an emergent property of a supply-constrained market.

For the rest of 2026, the checkpoint to watch is the publication of the CME and ICE contract specifications. The benchmark indices will determine which GPU SKUs are tracked, at what granularity, and with what lag. If the indices are daily snapshots of on-demand list prices, they will reflect the sticker shock but not the discounting that happens in private contract negotiations. If they incorporate neocloud and broker pricing, they will be more representative but harder to construct. Either way, by early 2027, the price of an H200-hour in Northern Virginia will be a number that trades on a screen, with a bid and an ask, settled in cash, monitored by analysts. That is new. It is also, given the scale of capital flowing into AI infrastructure, overdue.