The Economics of LLMs: Training, Inference, and Who Pays

Current stance (June 2026): The AI industry’s economics are defined by a structural tension: training costs are rising exponentially while inference prices are collapsing. This creates a narrow window where only well-capitalized labs can afford to build frontier models, but the resulting models become cheap enough for anyone to use. The strategic implication is clear — if you’re building, you need a moat beyond model quality. If you’re buying, the market is your friend.

The training cost problem

Training a frontier AI model is one of the most expensive single endeavors in the history of technology. The numbers are staggering and growing:

The trend is clear: training costs have grown 2–3x per year for the past eight years, according to Epoch AI. Each new generation of frontier models requires more compute, more data, and more time on expensive GPU clusters. Companies spent 28x more training their most recent flagship model compared to the predecessor, based on data from OpenAI, Meta, and Google.

The DeepSeek V3 number deserves special attention. At $5.6M in GPU rental costs, it claimed to produce a model competitive with GPT-4 — roughly 1/30th the cost. But that figure covers only the final training run. It excludes R&D, prior experiments, failed runs, salaries, data acquisition, and infrastructure that existed before the training cluster was assembled. The true cost is higher. The question is how much higher, and whether the approach is replicable.

For most organizations, the answer is: you cannot train a frontier model. The capital requirements alone — hundreds of millions in compute, plus the talent to use it — exclude all but a dozen or so labs worldwide. This is not an accident. It is the structure of the industry.

The inference economics

If training is a fixed cost, inference is a variable cost — and it is falling fast.

The pricing landscape as of June 2026 spans an enormous range:

Three things are visible in this table:

1. The spread is enormous. GPT-5.5 Pro costs 75x more per output token than DeepSeek V3.2. Both claim frontier-class performance. The market cannot sustain this spread — it will compress.

2. Output tokens dominate the bill. Output pricing is 3–5x input pricing across every provider. For conversational and agentic workloads where output outnumbers input 2:1 to 4:1, output pricing drives 75–90% of total spend.

3. Open-source models are pulling the floor down. DeepSeek V3.2 at $0.28/$0.42 per million tokens is roughly 24x cheaper on output than GPT-5.4. When the cheapest viable model is also the most efficient, the entire pricing structure is under pressure.

The ROI question

The fundamental question for every AI lab is: how long does it take for inference revenue to recoup the training investment?

Consider a simplified model. If a lab spends $100M training a model, and that model generates $10M per month in inference revenue, the payback period is 10 months — before operating costs. If the model generates $50M per month, payback is two months. If it generates $1M per month, the model is a loss leader.

The data suggests that revenue is concentrating at the top. Menlo Ventures reported that enterprise model API spending jumped from $3.5B to $8.4B in the first half of 2025 alone — more than doubling in six months. But this spending is not evenly distributed. Anthropic has overtaken OpenAI in enterprise usage (32% vs. 25%), while Google holds 20%. The rest of the market fights for scraps.

The economics favor the winner-take-most dynamic. Training costs are fixed. Inference costs scale with usage. The lab that captures the most usage spreads its training cost across the most revenue. This creates a natural incentive toward consolidation — not because of monopoly power, but because of math.

Who wins, who loses

The current landscape breaks into three tiers:

Tier 1: The training clubs. OpenAI, Google, Anthropic, Meta, xAI, and a handful of Chinese labs (DeepSeek, Alibaba/Qwen, Moonshot/Kimi, MiniMax, Zhipu/GLM). These organizations can afford to train frontier models. Some do it as a business (OpenAI, Anthropic). Some do it as a subsidy (Meta, Google). Some do it as a geopolitical imperative (Chinese labs). The common thread: they have the capital and talent to play at the frontier.

Tier 2: The inference optimizers. Companies like Groq, Cerebras, DeepInfra, Together.ai, Fireworks, and a growing list of serverless inference providers. They do not train models. They serve them more efficiently — faster, cheaper, or both. Their moat is infrastructure, not intelligence. They benefit from the training clubs’ investment without bearing the cost.

Tier 3: The users. Everyone else. Application developers, enterprises, researchers, individuals. For this tier, the economics are pure upside: model quality improves while prices fall. The question is not whether to use AI, but which model to use for which task.

The losers in this structure are the middle-tier labs — organizations that tried to train their own models but cannot sustain the capital requirements. They either need to find a niche (a specific domain, a specific modality, a specific geography) or accept that they will be users, not builders.

The compute allocation tradeoff

Not all training compute is equal. The industry is increasingly distinguishing between three phases:

1. Pre-training: Teaching the model the structure of language and the world. This is the most expensive phase, consuming the majority of compute.

2. Mid-training: Fine-tuning, RLHF, constitutional AI, and other alignment techniques. Less expensive than pre-training but critical for model quality and safety.

3. Post-training: Inference-time scaling, chain-of-thought, tool use, and agentic capabilities. This is where the model learns to do useful work rather than just predict tokens.

The balance is shifting. Pre-training costs still dominate, but the marginal returns are diminishing. The biggest quality gains in the last year have come from post-training — reinforcement learning with verifiers, reasoning chains, and tool use. This is good news for the economics: post-training is cheaper than pre-training, and its effects are more targeted.

The implication: labs that optimize their compute allocation across all three phases will outperform labs that pour everything into pre-training. This is a solvable problem — it is engineering, not science — and it creates a competitive axis beyond raw capital.

What this means

For anyone making strategic bets in AI, the economics point to three conclusions:

1. Model quality is necessary but not sufficient. If your only advantage is that your model is slightly better, you are one training run away from irrelevance. The winners will combine model quality with distribution, reliability, ecosystem, and trust.

2. Inference is a commodity; training is a bet. The inference market will continue to compress toward cost. The value is in the training — in creating the model that everyone else wants to serve. But the training bet requires capital on a scale that most organizations cannot match.

3. The real money is in the application layer. The companies that build on top of these models — the ones that solve specific problems for specific users — will capture more value than the model providers themselves. The model is infrastructure. The application is the product.

This is not a new pattern. It is the story of every technology stack. The companies that build the foundation (semiconductors, operating systems, databases) capture some value. The companies that build on top (applications, services, experiences) capture more. AI is following the same arc, just faster.

Last reviewed: 2026-06-04. This is a living document. As the economics shift, this perspective will be updated.