AI inference is obviously profitable

2026-06-02 · Source: All posts - seangoedecke.com RSS feed · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Corporate Strategy & Leadership, Entrepreneurship & Start-ups · Depth: Intermediate, short

Summary

Contrary to claims that AI inference is unprofitable and relies on investor subsidies, analysis suggests it is demonstrably profitable. Frontier AI providers report 70-80% gross margins on inference. Detailed estimates for running a dense 70B model on four Nvidia A100 GPUs show power and cooling costs around 13 cents per hour, plus GPU amortization of \$1.80 per hour over five years, leading to an approximate total inference cost of one dollar per million tokens. This contrasts with OpenAI's GPT-5.4-mini charging \$4.50 per million tokens, making high profit margins plausible. Further evidence comes from open-weights Chinese LLMs like DeepSeek-R1, which claim over 80% profit margins despite API pricing less than half of major providers, with market costs for DeepSeek-V4-Pro around 87 cents per million output tokens. AI labs like OpenAI and Anthropic maintain high inference prices to subsidize their substantial model training costs, not due to inherent unprofitability of inference itself.

Key takeaway

For AI/ML Directors evaluating LLM deployment strategies, recognize that AI inference is fundamentally profitable, often with significant margins. You should scrutinize API pricing from frontier model providers, as their high costs frequently subsidize their training efforts, not reflect actual inference expenses. Consider leveraging open-weight models or building your own inference infrastructure to achieve substantial cost savings, potentially reducing per-token costs to under one dollar. This approach can free up budget for other strategic investments.

Key insights

AI inference is inherently profitable, with high margins often used by labs to fund model training.

Principles

• Inference costs are significantly lower than current API pricing.
• Open-weight models drive down inference market prices.
• AI labs subsidize training with inference profits.

In practice

• Evaluate open-weight LLMs for cost-effective inference.
• Compare API token pricing against subscription models.
• Consider GPU amortization over five years for cost planning.

Topics

• AI Inference Costs
• LLM Profit Margins
• GPU Amortization
• Open-weight LLMs
• DeepSeek AI
• AI Lab Business Models

Code references

• deepseek-ai/open-infra-index

Best for: Entrepreneur, CTO, VP of Engineering/Data, Director of AI/ML, Investor, Consultant

Related on AIssential

• Winners and losers in the coming AI margin collapse (part 2) — Aggressively priced "good enough" AI models are collapsing inference margins, shifting value to hardware and users.
• Tokenomics Is Dead. The Lie Of Affordable Inference — Current "cheap inference" is a VC and hyperscaler-subsidized illusion, masking significant AI industry losses.
• The State of The Inference Economy — AI's economic value is shifting from training to inference, which now drives revenue and dominates compute costs.
• Cheaper AI Does the Opposite of What You'd Expect — Cheaper AI inference drives increased consumption and integration, leading to better user experiences and higher demand.
• Inside the Forward Pass: Pre-Fill, Decode, and the GPU Economics of Serving Large Language Models — LLM economics are shifting from one-time training costs to perpetual inference expenses due to massive token consumption.
• The Inference Reckoning: How to Stop Burning Millions on Cloud LLM Tokens — High-volume LLM inference demands dedicated infrastructure to avoid escalating cloud token costs and gain control.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by All posts - seangoedecke.com RSS feed.