Frontier AI training cost trajectory 2026: the run rate, the deal stack, and the power-bound horizon

The compute scaling trajectory #

Epoch AI tracks training compute for notable models back to 1950 and finds that frontier training FLOPs grew at roughly 4x per year through the deep learning era, with the leading lab releases of 2020 to 2024 sitting closer to 10x per year. GPT-3 in 2020 trained at roughly 3.1e23 FLOPs. PaLM in 2022 trained at roughly 2.5e24. GPT-4 in 2023 trained at roughly 2e25 FLOPs on a Microsoft Azure cluster of about 25,000 A100s. Claude 3.5 Sonnet, released by Anthropic in mid 2024, sits near 3e25 FLOPs based on Epoch back-of-envelope estimates that triangulate hardware-time, parameter count, and Chinchilla-optimal token budgets. Gemini 1.5 Ultra is closer to 1e26 FLOPs given the longer training duration and the larger TPU v4 and v5p pod allocations that Google has disclosed in its technical report.

The frontier 2026 training runs land in the 1e26 to 1e27 FLOPs window. GPT-5 class models, the next Claude generation, and Gemini 2 Ultra all sit inside that band. Above 1e27 the compute budget is no longer constrained by purchase orders. It is constrained by how much continuous gigawatt-scale power can be parked in one place, by how much synchronous all-to-all interconnect bandwidth a single training job can absorb without throughput collapse, and by how much pretraining-quality token data exists at all. Epoch projects that, absent an architectural break, frontier training compute growth decelerates from 10x per year to roughly 3x to 4x per year through 2028, with the deceleration starting in 2026.

From FLOPs to dollars: the training cost stack #

Translating FLOPs to dollars requires three multipliers. First, the realized FLOPs per dollar of GPU hour, which depends on the chip generation, the precision (BF16, FP8, or FP4), and the model FLOPs utilization or MFU achieved by the training stack. Second, the multi-month rental or amortization rate for the cluster, which differs sharply between owned hyperscaler capacity, NeoCloud rental, and long-dated reserved deals. Third, the overhead stack: networking, storage, data engineering, salaries for the research and infra teams, electricity, and the substantial cost of failed runs that never ship.

Epoch and SemiAnalysis converge on a simple decomposition for frontier runs. Pretraining hardware-time is roughly half of the total training program cost. Research salaries, including the compute consumed by the experiments that informed the final architecture and data mix, are another quarter. Failed runs and ablations, which typically consume 10 to 30 percent of the FLOPs of a successful frontier run before it is launched, are the remainder. By that decomposition the headline 80 million dollar GPT-4 number understates the program cost by roughly 2x to 3x once pre-training research is included.

Cluster economics: 100,000 GPU class and beyond #

The 100,000 H100 cluster is the working unit of frontier compute in 2026. xAI Colossus in Memphis went from 100,000 H100 SXM at 700 watts each to 200,000 H100 equivalents through 2025, with Nvidia GB200 NVL72 racks added in late 2025. Meta Reservation Hopper, the company's H100-class internal capacity, exceeded 350,000 H100 equivalents by end of 2025 according to capex disclosures and reporting. Microsoft and OpenAI broke ground on Stargate Phase 1 in Abilene, Texas at a provisioned 0.5 gigawatts and have publicly described a 5 gigawatt Stargate vision spanning multiple sites. Google does not publish chip counts but its TPU v5p pods and the newer Trillium and Ironwood generations operate at roughly equivalent aggregate scale.

The all-in capital cost of a 100,000 H100 cluster runs between 3 and 5 billion dollars depending on networking choice, building shell, power, cooling, and storage. The H100 SXM5 itself draws 700 watts and a fully loaded GB200 NVL72 rack draws roughly 120 kilowatts. A 100,000 GPU cluster therefore needs at minimum 70 to 100 megawatts of IT load, which translates to 130 to 180 megawatts of grid draw after accounting for cooling and power conversion. The 5 gigawatt Stargate target implies roughly 3 to 4 million GPU-equivalents at full Blackwell density, which is why the binding constraint has shifted from chip availability to substation, transformer, and turbine lead times measured in years rather than quarters.

Pretraining versus post-training cost ratios #

Pretraining still dominates frontier training cost in 2026, but the ratio is moving. Through 2023 a frontier run was roughly 95 percent pretraining and 5 percent post-training (instruction tuning, RLHF, RLAIF, red teaming, evals). By 2025 the post-training share for the Claude, GPT, and Gemini families has grown to roughly 15 to 25 percent of the total compute budget, driven by long-horizon reasoning training, tool use training, agentic trajectories, and the synthetic data pipelines that produce training signal in domains where human labels are scarce or expensive.

Synthetic data generation now consumes a meaningful fraction of frontier compute on its own. The pattern is to use the previous frontier model as a teacher to generate filtered, verified, and rated trajectories for the next model. DeepSeek and Anthropic have both described variants of this pattern in technical reports. The compute cost of generating high-quality synthetic data for a frontier run can reach 10 to 20 percent of the pretraining budget, but it substitutes for human labeling that would otherwise dominate the post-training line item.

The deal stack: who is buying what from whom #

The frontier compute supply chain in 2026 is structured by a handful of multi-year, multi-billion dollar bilateral deals that have replaced the older spot market for hyperscaler capacity. Anthropic and AWS announced Project Rainier, a cluster of more than one million Trainium 2 chips, with Anthropic committing to AWS as its primary training partner and AWS taking a substantial equity stake. OpenAI and Oracle anchored the Stargate Abilene buildout, with Oracle sourcing Nvidia capacity and providing the operating layer; Microsoft remains OpenAI's exclusive cloud for inference but is no longer the exclusive training cloud. Anthropic also announced an expanded Google TPU commitment, giving it a multi-vendor training portfolio across Trainium 2, TPU v5p and Trillium, and Nvidia H100 and B200.

Meta procures Nvidia directly at scale and is also building Catalina and MTIA accelerators for inference, with training still dominantly Hopper and Blackwell. xAI is exclusively Nvidia. Google trains Gemini almost entirely on TPU. The pattern that emerges is a pivot away from a single Nvidia on a single cloud toward a portfolio of accelerators sourced through bespoke deals, with custom silicon (Trainium 2, TPU, MTIA) capturing a growing share of training, while Nvidia retains the leading position in the merchant tier and in models that need maximum framework flexibility.

NeoClouds, the merchant training tier, and Chinese frontier players #

Below the hyperscalers and the integrated frontier labs, the NeoCloud tier (CoreWeave, Lambda, Crusoe, Nebius, Applied Digital, and a long tail of regional operators) provides Nvidia GPU capacity on shorter contracts to second-tier labs and to enterprises running large fine-tunes. CoreWeave operates more than 250,000 GPUs across more than 30 sites as of early 2026 and has anchored several multi-year contracts with Microsoft and OpenAI, effectively acting as an external balance sheet for hyperscaler training capacity. Crusoe is building the Abilene shell that hosts Stargate Phase 1 in partnership with Oracle. The NeoCloud margin profile is thinner than hyperscaler cloud, but the depreciation cycle is faster (typically 4 to 6 years on the GPU itself) and the time to revenue from new capacity is measured in quarters, not years.

The Chinese frontier picture is shaped by export controls and by domestic substitution. DeepSeek V3, released in late 2024, claimed a final pretraining run cost of about 5.6 million dollars on roughly 2,000 H800 chips, a figure that became politically salient because it was orders of magnitude below the implied cost of US frontier runs. The 5.6 million number is the marginal hardware-time cost of the final successful run, not the total program cost; SemiAnalysis and others have argued the full DeepSeek program (including failed runs, research salaries, and infrastructure) is closer to a few hundred million dollars. The point that survives the critique is that algorithmic and systems efficiency gains can compress the marginal cost of a frontier-class run by roughly an order of magnitude relative to the leading US labs, and DeepSeek, Qwen, and Kimi have demonstrated that the gap between Chinese and US frontier capability is now measured in months, not years.

The training versus inference compute split, and the power horizon #

Microsoft's AI revenue mix is the cleanest public window into how training and inference compute are splitting at the hyperscaler level. Through 2024 the majority of AI infrastructure spend at Microsoft was directed toward training capacity for OpenAI and toward Copilot inference scaffolding. By late 2025 the disclosed mix had tilted toward inference, with management commentary describing inference as the larger and faster-growing line. The same shift is visible in AWS, Google Cloud, and Oracle commentary. Frontier training is a concentrated, lumpy capital line item; inference is a distributed, continuously growing operating line.

The implication for the training cost trajectory is that the compute fleets being built in 2026 and 2027 are dual-use by design. A Blackwell GB200 NVL72 cluster trains a frontier model for six to nine months, then serves inference for the remaining three to four years of its useful life. FP4 quantization, supported natively on Blackwell, turns the same hardware into a far more efficient inference engine after training, which is why model distillation from a large frontier teacher into a smaller FP4-served student has become standard practice. The economics of the training budget therefore depend on the inference revenue that the same fleet generates after the training run completes.

The binding constraint on the next order of magnitude in frontier training is power, not silicon. Wells Fargo and Morgan Stanley equity research both estimate that US data center power demand will roughly double by 2030, with AI training and inference responsible for the majority of the increment. The 5 gigawatt Stargate vision, if executed, would consume roughly 0.5 percent of total US electricity generation on its own. ERCOT, MISO, and PJM interconnection queues for new gigawatt-scale loads are now measured in five to seven year wait times for transmission upgrades, which is why hyperscalers and frontier labs are pursuing behind-the-meter gas turbines, nuclear PPAs, and direct utility partnerships. Chinchilla optimality told labs how to allocate FLOPs between parameters and tokens; the 2026 to 2028 question is how to allocate them between sites, given that no single substation can host the full next-generation training run.

Sources #

• Epoch AI: notable AI models database and compute trends
• SemiAnalysis: AI cluster economics and frontier training cost models
• Anthropic: Claude 3 and 3.5 model card and infrastructure notes
• OpenAI and Microsoft: Stargate announcement
• Google DeepMind: Gemini 1.5 technical report
• Microsoft 10-K filings (capex and AI revenue commentary)
• Meta capex and AI infrastructure disclosures (10-K, 10-Q)
• MLPerf training benchmark results
• DeepSeek V3 technical report
• Wells Fargo and Morgan Stanley equity research on AI power demand