GPU Price Trends for AI Projects

The most important date in AI compute right now is October 5, 2026, when CME Group will list the first exchange-traded futures based on Nvidia H100 and B200 rental prices, using indexes published by Silicon Data, a GPU-market intelligence firm supported by trading house DRW. The announcement, made with Silicon Data on August 11, 2026, provides AI compute with a public, tradable reference price for the first time. This matters more than any single per-GPU-hour quote because it turns what was previously opaque, relationship-driven negotiation into a benchmark that enterprises, hyperscalers, and investors can hedge against.

Three months ago, when this site published its first GPU spot price and capacity outlook, the headline was a $1.35-per-hour H100 SXM5 spot rate and a market that had eased from an all-out shortage into a more complex situation. The picture has changed since then. H100 rental prices have increased nearly 40% in six months, the bottleneck has shifted from memory to power, and the capital structure of the entire buildout is now uncertain after reports that Nvidia might backstop $250 billion of OpenAI’s data-center financing.

Key Takeaways

• Nvidia’s H100 one-year rental contracts rose to $2.35 per GPU-hour in March 2026, up from a $1.70 low in October 2025, according to SemiAnalysis.
• CME Group and Silicon Data will launch the first H100 and B200 rental-index futures on October 5, 2026, providing AI compute with a public reference price.
• The main constraint has shifted from chips to the memory-and-packaging stack, then to power: Gartner expects power shortages to limit 40% of AI data centers by 2027.
• CoreWeave’s contracted backlog reached $129 billion in Q2 2026, growing $29.6 billion in six weeks, while pricing increased 25%.
• Self-hosting breaks even only at high, steady use; the price gap between neo-cloud and hyperscaler on the same card now ranges from 40% to 85%.

Spot Pricing: The H100 Reversal and Two-Tier Market

The clearest sign that the shortage phase is ongoing comes from SemiAnalysis, which tracked Nvidia’s H100 one-year rental contract prices rising almost 40% to $2.35 per GPU-hour in March 2026 from a low of $1.70 per GPU-hour in October 2025, as reported by Seeking Alpha. This reverses the normalization described here in May, when GridStackHub.ai cited H100 SXM5 spot at $1.35 per hour and A100 80GB at $0.35 per hour.

Carmen Li, a former Bloomberg executive who left in 2024 to found Silicon Data specifically to track GPU prices, described the trend plainly in an April 2026 interview with Business Insider: “The price is going pretty nuts,” she said about H100 rental prices. Her explanation contrasts with the usual pattern where prices spike at a chip’s launch and then decline as supply increases. Instead, B200 pricing has remained elevated and even increased, indicating that capacity constraints persist across chips, memory, power, and data-center space.

The market has split into two tiers with very different economics. ClusterBid’s mid-2026 analysis of 84 enterprise deployments found H100 spot pricing had dropped to $1.20 to $1.80 per GPU-hour as data-center buildouts entered the market, while B200 on-demand pricing remained at $7 to $9 per GPU-hour due to ongoing scarcity, with reserved B200 contracts at $4.50 to $6.00 per hour. The lead-time spread confirms this division: H100 lead times have shortened to 4 to 8 weeks, H200 is available in 8 to 12 weeks, but B200 SXM parts still require 36 to 52 weeks from purchase order to rack-level acceptance.

The wide pricing range is the largest factor affecting unit cost. According to the eCorpIT capacity playbook, on-demand rates for the newest cards roughly doubled over the past year, while mainstream H100, H200, and A100 pricing remained within a narrower range. The price difference between neo-cloud and hyperscaler on the same card now ranges from 40% to 85%.

Li’s most notable point concerns depreciation, a concern that has affected every AI cloud valuation. In the second year, a refurbished H100 can still sell for 85 cents on the dollar, and in year three for 84 cents, she told Business Insider. “My car depreciates a lot faster than that,” she said. That residual-value floor explains why Wall Street has started treating Nvidia GPUs as near-perpetual cash-flow assets, a view that 24/7 Wall St. described as fragile in August.

The CME Moment: Compute Becomes a Tradable Commodity

The October 5 launch is a structural event that changes how the entire market prices risk. CME Group and Silicon Data will list two contracts, the Silicon Data H100 Rental Index Futures and the Silicon Data B200 Rental Index Futures, each representing a month’s worth of rent for the respective chip. The contracts will be listed under NYMEX rules, pending regulatory approval.

The reasoning, explained in the joint press release, compares compute to oil: just as crude evolved from spot trading into a global derivatives market, compute futures turn GPU rental capacity into a standardized, tradable commodity that lets AI developers and hyperscalers lock in costs.

Silicon Data CEO Carmen Li described the transparency gap this product closes: “For years, two companies buying the exact same GPU capacity could pay wildly different prices with no way to know who got the better deal. They will now have a benchmark to check that against.” Her firm collects hundreds of thousands of pricing points globally and normalizes them into indexes covering both spot rentals and longer-term contracts.

GPU pricing shifts from relationship-driven negotiation to a market with a forward curve. For an engineering manager planning a training run six months ahead, that means the ability to hedge compute costs like an airline hedges jet fuel. For hyperscalers, it means a public price that will influence both procurement and the pricing passed on to customers.

Capacity Buildout: CoreWeave, Crusoe, and the Financing Question

The supply side is expanding at a pace that would have seemed extreme a year ago, but the capital structure behind it is now under close examination. CoreWeave’s Q2 2026 earnings, reported by CNBC on August 11, showed revenue of $2.575 billion, up 112% year over year, with contracted customer demand exceeding $129 billion. The company raised its 2026 revenue guidance to $12.4 billion to $13.2 billion and increased its year-end active power target to more than 1.85 gigawatts. Pricing rose 25% year over year, and the backlog grew $29.6 billion in just six weeks.

That backlog growth clearly indicates that demand is not slowing, but it also carries a warning. The same report noted roughly 50% default risk priced into some of that contracted demand, reflecting doubt about whether every customer who signed a multi-year commitment will actually take delivery. This tension is central to the AI compute market in August 2026: order books are huge, but the probability-weighted reality is less certain.

The financing issue came into public view in late July when the Wall Street Journal reported that Nvidia was negotiating to guarantee up to $250 billion in financing for OpenAI’s data-center buildout. The report, summarized by 24/7 Wall St., caused Nvidia’s stock to drop 5% and AMD’s by 8% in one session, reigniting concerns about “circular financing”: a vendor funding a customer who then buys that vendor’s GPUs. Nvidia’s Q1 FY2027 data-center revenue reached $75.25 billion, up 92% year over year, and CEO Jensen Huang called the buildout “the largest infrastructure expansion in human history.” The market’s concern is how much of that revenue reflects independent demand versus seller-funded demand.

Crusoe represents the energy-strategy side of the buildout. The company’s Abilene, Texas campus is a multi-gigawatt AI data-center hub, and Crusoe has focused on modular, prepackaged data centers with $200 million of funding allocated for smaller deployments, according to Forbes. The company is also testing a sodium-cooled nuclear microreactor to power AI data centers entirely on-site, betting that power, not silicon, is becoming the main constraint.

The AMD angle has become significant as well. In July 2026, Anthropic agreed to purchase up to 2 gigawatts of GPU capacity from AMD, adopting MI450 series in Helios rack configurations, with AMD investing up to $5 billion in Anthropic in return, according to SiliconANGLE. Microsoft is also deploying AMD Helios racks in Azure, and Meta announced plans in February to purchase up to 6 gigawatts of AMD graphics cards. This multi-supplier approach responds directly to Nvidia supply-chain bottlenecks.

The Moving Bottleneck: HBM3e, CoWoS, and Power

The most important supply story of 2026 is that the bottleneck keeps shifting, and buyers who focus on last quarter’s constraint will misjudge the next one. GlobalData’s August 2026 analysis, reported by InvestorIdeas, explains the shift clearly: record earnings across chipmakers mask a supply constraint where memory, advanced-node capacity, and power, rather than compute demand, are determining the pace of AI deployment.

The memory figures are striking. SK Hynix’s net income rose 397.5% year over year, Samsung’s 486.7%, and Micron’s 1,398.3% on 345.7% revenue growth, all driven by an HBM shortage that has sold out DRAM through 2026. SK Hynix holds about 58% of the high-bandwidth memory market in Q1 2026, giving it significant influence over how quickly premium GPUs can ship.

CoWoS packaging is the second bottleneck. TSMC’s net income grew 77.4% alongside a 33.8% capex increase as it races to add advanced-node capacity, and its CoWoS capacity is expanding nearly 50%, according to Seeking Alpha supply-chain analysis. However, the throughput of packaging, not raw silicon, determines how fast wafers become deployable accelerators.

Power has now overtaken both as the main constraint. Gartner predicted in November 2024, now considered a baseline, that power shortages will limit 40% of AI data centers by 2027, with electricity for incremental AI servers reaching 500 terawatt-hours per year, 2.6 times the 2023 level. Constellation Energy’s 1,247.5% net income growth in the GlobalData analysis reflects the market’s conclusion: power, not chips, is where scarcity is now clearing.

Announced capacity and available capacity can differ for months. A buyer can have GPUs allocated but still be unable to rack them because megawatts, cooling, and grid interconnect are not available. Hyperscalers respond by signing multi-year power deals ahead of others, pushing mid-size buyers to the back of the queue.

Self-Hosting Versus API Pricing: Where Break-Even Actually Sits

Every serious AI buyer eventually considers whether to keep paying per token or run the model themselves. The basic calculation has not changed since May, but the numbers have shifted. A 1,024-GPU cluster of H100 SXM parts at peak pricing costs roughly $8.4 million annually in compute alone, and over $15 million once networking, storage, facilities, and personnel are included, according to ClusterBid’s budgeting analysis.

The break-even point depends on usage, not just price. ClusterBid’s survey of 84 deployments found that GPU compute accounts for 48% to 62% of total AI infrastructure cost, supporting infrastructure adds 22% to 30%, and operations contribute 12% to 18%. A B200 that trains a model 2.3 times faster than an H100 can reduce the total cost per training run despite a higher hourly rate, once compute time, storage, and operations are included.

The risk of interruptions adds a hidden cost to spot-only strategies. Teams using spot exclusively for training face 15% to 25% job interruption rates on H100 and 30% to 40% on B200 due to scarcity, according to ClusterBid. The recommended approach is to reserve 70% to 80% of baseline capacity on 12-month contracts and use spot or on-demand for the remaining 20% to 30%, capturing roughly 75% of the fully-reserved discount without locking in maximum capacity that may go unused.

There is also a human-capital factor that generic cost comparisons overlook. Self-hosting requires staff who can tune models, schedule workloads, monitor usage, and maintain clusters. APIs outsource most of that. A startup with a small platform team may rationally pay more per token because it cannot handle infrastructure complexity, while a larger company with a strong infrastructure team may choose the opposite.

The token-API side of the comparison has its own dynamics. Anthropic’s second-quarter revenue reportedly rose to more than $11.5 billion, up 14-fold year over year, according to CNBC, showing that the per-token model remains the main route for most AI consumption even as the compute market develops.

What to Watch Through the Rest of 2026

The next few quarters are unlikely to bring a sharp drop in GPU pricing. A more likely scenario is uneven softening, where mature hardware continues to get cheaper in spot markets while the newest high-performance tiers remain expensive because memory, packaging, and power constraints affect those segments first. This increases the gap between practical inference costs and frontier training costs.

Three indicators matter most. First, the CME compute futures launch on October 5 will reveal the forward curve for GPU rental prices for the first time, and the shape of that curve will indicate whether the market expects scarcity to continue or ease. Second, watch whether the $129 billion CoreWeave backlog converts to actual revenue or if default risks emerge, since that conversion rate provides the clearest indication of whether AI demand is genuine or financed. Third, monitor HBM and CoWoS production against power delivery, because the bottleneck’s location determines which suppliers control pricing power in any given quarter.

The geopolitical factor has become a permanent cost variable. In January 2026, the Trump administration invoked Section 232 to impose a 25% tariff on select advanced AI chips, and the Bureau of Industry and Security changed H200 and MI325X export reviews for China from presumptive denial to case-by-case approval capped near 50% of prior US sales. TSMC, Samsung, and SK Hynix lost their blanket Validated End-User exemptions and now require annual US licenses to ship equipment into China fabs. These are structural changes that will last beyond any single quarter’s pricing data.

For builders, the practical lesson from May still applies, but with more urgency. The shortage phase has eased in the mature tier, but the market still prices real constraints, and those constraints have shifted up the stack toward memory, packaging, and power. The capacity you need must be available under terms that fit your workload and timeline. The CME futures market, when it opens in October, will finally provide a number instead of a negotiation to answer that question.

For more on how compute access affects bargaining power across the AI sector, see this site’s earlier GPU spot price and capacity outlook, and for serving-side optimization that changes the cost-per-token calculation, the local AI inference engine guide. The capital story behind this market, including the $725 billion hyperscaler capex cycle, is covered in the AI data center buildout series.

Related Reading

More detailed coverage from this blog on closely related topics:

• Understanding System Prompts in AI
• Making AI Private with Encryption

Sources and References

Sources cited while researching and writing this article:

• Nvidia’s H100 GPU rental prices surge nearly 40% in 6 months: SemiAnalysis
• This CEO left Bloomberg to track GPUs. She explains why prices are ‘going nuts.’
• The 2026 AI compute crunch: a GPU capacity playbook
• CME Group and Silicon Data to Launch Compute Futures on October 5 to Unlock New Way to Hedge AI Risks
• AI Stocks Crash After NVIDIA Plans to Finance $250 Billion OpenAI Buildout Are Reported
• Anthropic to buy up to 2 gigawatts of GPU capacity from AMD
• AI Infrastructure Boom Shifts From Chip Race to Supply-Chain Race
• CNBC