AWS On-Demand GPU Prices Spike 20%, Spot Market Next

On July 1, 2026, Amazon Web Services raised the per-accelerator hourly price of its GPU instances by roughly 20 percent. The P6-B300, its top-tier Nvidia-powered instance family, now bills at $14.04 per accelerator-hour. The P6-B200 moved to $12.355. The P5, the prior generation, hit $5.191 in US regions. These are on-demand, no-commitment list prices, and they represent the second hike AWS has pushed through this year, following a 15 percent increase earlier in the spring, as The Next Web reported alongside the updated rate card.

The proximate cause is the memory supply chain. Tom's Hardware reported last October that AI data centers are consuming memory and storage at a pace that is outstripping legacy DRAM production capacity, and the situation has not eased. HBM3e stacks, the high-bandwidth memory that every H100, H200, and B300 depends on, remain tightly allocated. Add a helium shortage that Forbes detailed in April, which constrains the semiconductor fabrication equipment that depends on helium for cooling during lithography, and you have a supply-side vise that gives hyperscalers both the cover and the motive to raise prices. AWS did not blame the memory crunch in its customer notice. It did not need to. The invoice does the talking.

The price increase widens a gap that has been growing for eighteen months: the spread between what a hyperscaler charges for on-demand GPU compute and what a neocloud charges for a one-year reserved commitment. Chris Campbell, senior director of AI solutions and GPU-as-a-service at World Wide Technology, put the neocloud math on the record with CRN in June. A one-year deal for 1,000 Nvidia B300s at $5 per chip-hour comes to roughly $45 million. At the new AWS on-demand P6-B300 rate of $14.04 per accelerator-hour, the same thousand chips would cost $123 million over a year, before any reserved-instance discount. Even after the standard hyperscaler one-year commitment discount, the neocloud price per GPU-hour lands somewhere between 50 and 60 percent below the hyperscaler on-demand equivalent.

That spread is not a market inefficiency. It is the structure of the GPU allocation market in mid-2026. Hyperscalers price on-demand compute like a convenience good, and enough buyers, especially inference-heavy startups that cannot forecast token volume six months out, pay the on-demand rate. Neoclouds price like a commodity lease. They want the committed contract, the predictable utilization, and the customer who will sign for a year and bring their own orchestration layer.

The neocloud market, led by CoreWeave, Nebius, Crusoe, and Lambda, reached $9 billion in revenue in the fourth quarter of 2025, a 223 percent year-over-year increase, according to data from Synergy Research Group cited by CRN. Synergy now forecasts the category will hit $400 billion by 2031, a 58 percent compound annual growth rate sustained over five years. Jeremy Duke, Synergy's founder and chief analyst, described the trend as a structural realignment of cloud architecture itself. "Neoclouds are, in effect, an architectural response to those constraints," Duke said, referring to the rigidity of AI workloads compared to the generalized elasticity that traditional hyperscale clouds were designed around.

The structural point matters because it explains why spot pricing, the traditional cloud escape valve for price-sensitive workloads, has not developed into a liquid GPU market the way it did for general-purpose EC2 instances. AWS, Google Cloud, and Azure all offer some form of interruptible GPU capacity. But the spot pools are thin. A training run requiring 512 H100s with tightly coupled InfiniBand fabric cannot be gracefully preempted with two minutes of notice the way a stateless web server fleet can. The workloads that dominate GPU demand are stateful, synchronized, and long-running. The spot discount, when available, is often not worth the coordination cost.

Instead, a different kind of allocation mechanism is emerging. TechCrunch reported in May that the CME Group and Intercontinental Exchange are designing derivative products around AI tokens, treating computational output as a raw material input analogous to electricity or bandwidth futures. If an exchange-listed token futures contract materializes, it would create a forward price curve for inference compute, allowing a model provider to lock in per-token costs six or twelve months out and hedge against exactly the kind of on-demand rate hike AWS just imposed.

This re-framing of compute as a commodity is also visible in the physical supply chain. Despite the memory crunch, discrete GPU shipments from NVIDIA, AMD, and Intel remained relatively flat quarter-over-quarter in Q1 2026, according to a Jon Peddie Research report published in June. NVIDIA maintained a 90 percent share of the add-in board market. The flat shipment number, combined with rising prices at every tier, tells a clear story: unit volume is not the constraint. Memory allocation is. And whoever controls the memory supply chain, from HBM3e procurement through to the finished server node, controls the price.

Imagine a one-year deal for 1,000 Nvidia B300s chips at $5 an hour. That would be $45 million for one year. So if you just take 5 percent of that, it's a couple million dollars for WWT., Chris Campbell, senior director of AI solutions and GPU-as-a-service, World Wide Technology, to CRN, June 2026

Channel partners like WWT are building margins into the spread between the neocloud wholesale rate and the enterprise customer's willingness to pay for a managed GPU deployment. Campbell told CRN that customers are not asking for dozens or even a hundred GPUs. They need 300, 500, or 1,000. Nebius, for its part, told CRN that for every GPU cluster it brings online, it has four to five customers already lined up to take it. Laurelle Roseman, Nebius's vice president of global partnerships, said the company is finalizing a formal partner program with tiered incentives and co-selling opportunities, scheduled to launch in summer 2026.

The channel is gravitating toward neoclouds for a reason Roseman articulated bluntly: hyperscalers offer partners low margin and limited access. Nebius is promising double-digit margin potential when partners wrap their own services around the compute. The TD Synnex distribution deal Nebius signed earlier this year, which CRN described as covering over 1,000 Nvidia GPUs made available through the channel, is a template for how reserved GPU capacity is being sliced into smaller, partner-mediated commitments rather than sold as a single hyperscaler contract.

Then, on July 1, the same day AWS's price hike took effect, Bloomberg and Reuters reported that Meta Platforms is developing plans to sell its surplus AI computing capacity to outside customers. Meta is weighing two models: selling developers access to AI models hosted on its own infrastructure, or renting raw GPU capacity directly. The news sent neocloud stocks tumbling. CoreWeave and Nebius shares fell sharply, as TechTimes reported, on the straightforward logic that a hyperscaler with Meta's capital budget, which the company has guided to as much as $65 billion in 2026 capex largely directed at AI infrastructure, entering the capacity resale market would be a supply shock that neither neocloud margins nor hyperscaler on-demand rates could withstand.

The Meta entry, if it materializes, would recast the GPU allocation market in three ways. First, it would add a large new supplier of what is effectively distressed inventory, GPUs procured for internal training runs that run intermittently. Second, Meta's cost of capital and its existing infrastructure amortization mean its floor price could undercut neoclouds whose business model depends on a spread between their own GPU acquisition cost and the rental rate. Third, it would accelerate the commoditization of GPU compute at exactly the moment when exchanges are designing financial products to trade it.

Who Captures the Margin

The GPU pricing stack in mid-2026 can be read as a margin map. At the bottom, the chip. NVIDIA's data center GPU gross margins have been estimated by Wall Street analysts in the high 70 percent range. The memory suppliers, SK hynix, Samsung, and Micron, are capturing scarcity premiums on HBM3e that flow through to the BOM cost of every accelerator. At the next layer, the server OEM and the datacenter operator. Then the cloud provider, whether hyperscaler or neocloud, which layers on power, cooling, networking, and a margin for operating the fleet. At the top, the model provider and the application layer.

The question is where the AWS price increase actually sticks. Enterprise buyers with predictable inference workloads, think a customer-service chatbot serving a known daily query volume, can avoid the on-demand rate entirely by signing a one-year or three-year reserved instance commitment with AWS, or by taking their workload to a neocloud at $5 per B300-hour. The buyers who absorb the $14.04 rate are the ones who cannot forecast: the research lab running a month-long ablation study that might need 2,000 H100s for two weeks and then nothing for six; the startup chasing a benchmark score ahead of a fundraising round; the inference provider whose token volume spikes unpredictably. These buyers are paying a liquidity premium, and AWS is pricing for it.

AMD's position in this landscape is worth noting. Seeking Alpha reported in late June that AMD's AI GPU business is being strengthened by hyperscaler diversification needs, citing multi-year, multi-gigawatt deals from Meta and OpenAI for AMD's MI300-series and upcoming MI400-series accelerators. If AMD can offer a credible second source at a lower per-accelerator price, even at lower absolute performance per chip, the math for a hyperscaler running a mixed fleet changes. The per-token cost of an AMD MI300X may be higher than an H100 on a raw FLOPs basis, but if the chip is available and the Nvidia equivalent has a six-month lead time, the availability premium flips the equation.

What to Watch

The checkpoint to monitor is not AWS's next price move. It is whether Meta proceeds with its cloud business and at what price floor. If Meta offers B300-equivalent capacity at $3.50 to $4.00 per GPU-hour on a one-year commit, the neocloud price of $5 becomes vulnerable, and the hyperscaler on-demand rate of $14 begins to look like a legacy artifact. The second checkpoint is the CME token futures contract. A listed forward price for inference compute would, for the first time, give every buyer in the stack a public benchmark against which to measure whether their GPU contract is fairly priced. At that point, the opacity that currently lets on-demand rates drift upward while committed rates hold steady becomes harder to sustain.

The final checkpoint is the HBM supply chain. Jon Peddie Research's flat Q1 GPU shipment number, reported by MSN, suggests the industry is shipping every accelerator it can build, constrained not by wafer starts but by memory attach rates. If HBM3e capacity expands meaningfully in the second half of 2026, as SK hynix has guided, the entire pricing stack from chip to cloud could loosen. If it does not, the divides between on-demand, reserved, neocloud, and hyperscaler-excess pricing will sharpen further, and the buyers who can forecast and commit will capture an even larger spread over those who cannot.