September 25, 2026. DeepSeek, the Chinese lab whose January 2025 model release was the moment cheap frontier-class AI became a buying assumption, has more than doubled its annualized revenue run rate to $1 billion, The Information reported on September 23, citing people who heard CEO Liang Wenfeng share the figure with investors. The growth followed, not preceded, a price increase. For the agencies and SMBs that built personalisation, enrichment and classification pipelines on the cheapest capable model, that ordering is the whole story: the low-cost provider raised prices by multiples, kept its customers, and grew.
The reported numbers
- Run rate: $1 billion annualized, more than double the pace reported a few months earlier, and roughly ten times 2025's revenue according to The Information's August 26 report, which put the first seven months of 2026 at about 475 million yuan ($70.7 million).
- Price rise: API fees increased in August by 2.3 to 4.5 times depending on the model, with peak-hour pricing on the flagship V4 models. Liang told investors the increase did not reduce the customer base, and the company's prices remain among the lowest for major models.
- Fundraise: a second round targeting 50 billion yuan, about $7.5 billion, at a 500 billion yuan valuation, roughly $75 billion, with a close aimed at the end of October per RuntimeWire's summary of the same report, ahead of a Shanghai Stock Exchange listing.
- Who pays: revenue is primarily developers paying for API access, not the consumer chatbot.
- What is not disclosed: how much of the growth came from higher prices, more usage or new customers, and what a full year of booked revenue looks like. A run rate is a pace, not a result.
Why this is the opposite of the OpenAI and Anthropic story
On September 22 OpenAI halved GPT-6 Sol and Luna and Anthropic cut Opus 5.5 by 40% on typical workloads; we covered the whole price war this week. DeepSeek moved the other way in August and reports that demand held. Both facts fit one market: the frontier labs are pricing to pull volume down from above, and the low-cost labs are pricing to fund the compute that keeps them competitive. A GPT-6 Luna call at $0.10 per million input tokens and $0.50 output is now close enough to a DeepSeek call that the price gap no longer pays for the integration risk. The cheap-model assumption, that there will always be a provider at a fraction of frontier price with no strings, is the assumption that broke this month.
What it means for operators
If a client pipeline depends on a single low-cost model, its economics now sit with a vendor that has demonstrated it can multiply prices and keep growing. Three moves follow. Abstract the model behind a routing layer so a price change is a config change, not a rebuild. Benchmark the cheapest frontier tier, Luna or Haiku, against the low-cost incumbent on your actual tasks, because the September cuts closed most of the gap. And keep the parts of the pipeline that do not need a model at all off the model: verification, deduplication and sending live in infrastructure, not in prompts. That last point is why we treat cold email infrastructure as the fixed layer and the personalisation model as a swappable one; a campaign that sends 50,000 emails a month should not have its deliverability, its domains or its warmup depend on whose API is cheapest this quarter. For the industry, a $75 billion valuation on a $1 billion run rate says investors believe the cheap tier will keep its customers at higher prices. Operators should plan as if they are right.
Frequently Asked Questions
The Information reported on September 23, 2026 that DeepSeek's annualized revenue run rate has reached $1 billion, more than double the pace reported a few months earlier. Its August 26 report put the first seven months of 2026 at about 475 million yuan, roughly ten times 2025's revenue.
Yes. According to The Information, DeepSeek raised API prices in August 2026 by 2.3 to 4.5 times depending on the model, adding peak-hour pricing on its V4 models. CEO Liang Wenfeng told investors the increase did not reduce the customer base.
The reported second funding round targets 50 billion yuan, about $7.5 billion, at a 500 billion yuan valuation, roughly $75 billion, with a close aimed at the end of October ahead of a planned Shanghai Stock Exchange listing.
Put a routing layer between the pipeline and the model so a price change is a configuration change, benchmark frontier budget tiers such as GPT-6 Luna against the low-cost incumbent on real tasks, and keep deliverability, verification and sending infrastructure independent of the model vendor.