DeepSeek puts 70% of compute on training, not your calls

Two numbers from a private investor meeting tell you more about your token bill than any pricing page. DeepSeek reportedly raised prices by up to 4.5x without losing customers, and puts over 70% of its computing capacity into training rather than serving requests. If cheap inference is load-bearing in your stack, both numbers are about you.

What actually happened

The Information reported on September 24, citing two people with direct knowledge, that CEO Liang Wenfeng told investors DeepSeek's annualized revenue run rate reached $1 billion — more than double a figure that sat under $500 million months ago. The disclosures came as the company works on a second funding round targeting 50 billion yuan at a 500 billion yuan valuation by the end of October.

Three details matter more than the headline figure:

• Prices went up 2.3x to 4.5x across models, and Liang said the increases did not shrink the customer base. Most of the revenue jump is repricing, not new volume.
• Over 70% of compute goes to training, under 30% to inference — serving the models already released.
• The company still positions itself as one of the cheaper providers at the frontier, which is how a 4.5x increase survives contact with customers.

These figures come from a closed investor meeting relayed by one outlet. Treat them as reported, not audited — but the direction is consistent with DeepSeek's public price moves earlier this year.

Why inference capacity matters for your business

A 4.5x increase that kept every customer is a proven elasticity. That is the finding, and it generalizes past DeepSeek. Every provider watching this now knows that customers who wired a cheap model into production do not leave over a 4x change, because migrating is harder than paying. Price your automations against a token cost 3-5x higher than today's and see which ones still clear. The ones that do not are not businesses; they are subsidies with an expiry date.

Under 30% of capacity serves you. Inference is the minority tenant at your token supplier. That is a rational allocation for a lab racing to the next model, and it is a capacity risk for anyone whose customer-facing feature depends on a response arriving. It shows up as throttling during a launch week, not as an outage page. If a workflow has to complete on a clock, it needs a fallback provider configured and tested — not documented.

Cheap tokens were a land-grab phase, not a floor. The pattern across the last year is consistent: aggressive entry pricing, adoption, then repricing once switching costs exist. Budget for the repricing. The companies that get hurt are the ones whose unit economics only worked at introductory rates.

Decouple the model from the workflow, and keep evals you own. One internal interface per capability, provider as a config value, and a small eval set in your repo that proves a swap does not degrade output. We build this into every AI system for exactly this reason: the leverage in a repricing conversation comes from being able to leave, and you only have that if the migration is a config change you have already tested.

Sources: The Information via Investing.com, PYMNTS.