Training a frontier model in 2020 cost somewhere in the low tens of millions of dollars. By 2023, GPT-4-scale runs were estimated in the $50M to $100M range. Today, credible estimates for the most capable frontier runs sit north of $100M, with some whispered figures approaching $500M for the largest clusters. The numbers are climbing in absolute terms. What is less obvious is that the cost-per-unit-of-capability is compressing sharply, and that compression is the structural dynamic worth watching.

Where the Cost Actually Lives

Training run economics break into three buckets: compute (GPU or TPU hours), data pipeline and curation, and engineering labor. Compute has historically consumed 60 to 80 percent of total spend on large runs. That figure is shifting as data quality becomes the binding constraint at scale and human-generated curation labor scales less cleanly than hardware procurement.

The H100 cluster economics that dominated 2023 and 2024 are giving way to GB200 NVL72 rack-scale configurations, where memory bandwidth and interconnect architecture matter more than raw FLOP counts. A training run that required 10,000 H100s for a given model class now completes on roughly 4,000 to 5,000 GB200s with comparable wall-clock time. Fewer chips, denser interconnect, lower total energy draw per token processed.

• Compute cost per FLOP at the chip level has declined roughly 2.5x to 3x over the H100-to-B200 transition.
• Inference-optimized architectures (mixture-of-experts, speculative decoding) are reducing the amortized cost of post-training serving, which changes the ROI math on the initial training investment.
• Synthetic data pipelines are compressing data acquisition costs, though they introduce new quality-control failure modes that labs are still working through.

The Frontier Consolidation Dynamic

When training a competitive frontier model costs $200M or more in compute alone, the viable entrant pool is not startups. It is sovereign wealth vehicles, hyperscalers, and a small number of well-capitalized independent labs with committed capital from strategic partners. This is not a temporary condition. The scaling thesis, even under efficiency improvements, points toward runs that will cost multiples of today’s figures within 24 to 36 months if capability curves hold.

The market structure this produces is familiar from semiconductor fabs and pharmaceutical discovery: high fixed costs, winner-concentration, and a long tail of application-layer businesses built on top of the infrastructure layer’s outputs. The interesting operator question is not who trains the next frontier model. It is who extracts durable margin from the application surface those models expose.

What Efficiency Gains Do to the Market

Efficiency improvements do not flatten competitive moats at the frontier. They tend to compress them at the tier below. When training a mid-tier capable model drops from $10M to $3M, the population of entities that can produce a domain-specific fine-tuned model expands. Enterprise verticals with proprietary data and a clear inference use case become structurally interesting. The model is no longer the moat. The data and the deployment context are.

Distillation from frontier models, which OpenAI, Anthropic, and Google have moved to restrict in their terms of service with varying degrees of enforceability, was compressing mid-tier training costs faster than organic efficiency gains alone. That dynamic is not fully resolved.

The Operator Read

The training run cost trajectory rewards a specific kind of patience. Frontier capability is concentrating around entities with sovereign-scale capital access. Below that layer, efficiency curves are opening a window for well-resourced domain specialists. The structural observation is that the value migration is moving from model training toward data infrastructure, deployment infrastructure, and the enterprise integration layer, where margins are less visible but arguably more defensible.