DeepSeek's Revenue Jumped Tenfold to $70.7M — and Its API Business Is More Profitable Than OpenAI's

DeepSeek, the Chinese AI lab that upended the industry in January 2025 when it matched frontier performance at a fraction of the usual training cost, has just disclosed the financial details behind its quiet ascent. According to a report from The Information published August 25, the company generated RMB 475 million (approximately $70.7 million) in revenue across the first seven months of 2026 — roughly ten times its total for all of 2025.

The number itself is modest by the standards of Silicon Valley’s AI giants. OpenAI’s annualized revenue run-rate crossed $20 billion earlier this year; Anthropic is not far behind. But the shape of DeepSeek’s business is what should catch the attention of investors and competitors alike: the company is achieving frontier-scale growth on a cost structure that its Western rivals simply do not have.

The Numbers Behind the Surge

The Information’s report, drawing on internal financials, sketches a company hitting its stride:

• Revenue: RMB 475 million ($70.7M) from January through July 2026 — about 10x its full-year 2025 figure, and already ahead of the ~$500M annualized run-rate analysts were estimating in mid-July.
• API gross margin: 82.9%. This is the headline number. DeepSeek’s API business — selling access to its models over HTTP — posts a gross margin higher than most software companies, and notably higher than the gross margins OpenAI and Anthropic report on their API products.
• Consolidated gross margin: 44.6%. Below the API figure because it blends in the cost of serving free chat users, research infrastructure, and training runs.
• Net loss narrowed to RMB 715 million. The company still loses money overall — heavily subsidizing free consumer access — but the loss is shrinking as API revenue compounds.

An 82.9% API gross margin is not just good; it is structurally significant. OpenAI’s and Anthropic’s API businesses are widely believed to run gross margins in the 40–60% range, dragged down by compute costs at Western cloud providers. DeepSeek achieves its margin the same way it achieved its famous sub-$6 million training runs: radical engineering efficiency. Sparse mixture-of-experts routing, multi-head latent attention, aggressive quantization, and cheap domestic inference hardware keep the marginal cost of a token extraordinarily low. When you sell tokens at competitive prices but they cost you a fraction of what rivals pay, margins like 82.9% become possible.

The Road to a $74 Billion Valuation

The financial disclosure did not arrive in a vacuum. It lands weeks into DeepSeek’s second major fundraising effort of the year:

1. May 2026 — DeepSeek wrapped its first external round, raising roughly $7 billion at a $52 billion valuation, with investors including CATL (Contemporary Amperex Technology), Guozhi Investment, and IDG.
2. July 14, 2026 — Bloomberg and the Financial Times reported the company had begun preparations for an onshore IPO, with filing possible as soon as this year, and was seeking up to $3 billion from US-based investors.
3. July 15–18, 2026 — Reuters confirmed the new round targets a valuation of about RMB 500 billion (~$74 billion), a 42% markup in two months.
4. August 6, 2026 — PYMNTS reported the round had resumed with an $8 billion target, and that annualized revenue was approaching $500 million as July closed.

Read in sequence, this week’s revenue report is the due-diligence backbone for that $74 billion ask. A tenfold revenue jump and an 82.9% API margin give underwriters a clean narrative: this is not a lab burning cash on prestige research — it is a business with software economics attached to the most efficient frontier model stack in the industry.

Why the Margin Gap Matters

The economics of serving large language models is becoming the central question of the AI industry. Every major lab faces the same tension: frontier training runs cost billions, inference at scale costs billions more, and pricing pressure from commoditized open-weight models compresses what anyone can charge per token.

DeepSeek’s answer has been to attack the cost side rather than the price side. Its open-weight releases — V3, R1, and their successors — forced every Western lab to cut prices in early 2025. But because DeepSeek’s per-token cost is so low, it can remain profitable at price points that would be margin-destructive for rivals running on US cloud infrastructure. The 82.9% API margin is the empirical proof of that strategy working at commercial scale.

There are caveats. DeepSeek’s consolidated margin of 44.6% reflects the drag of its free consumer chat product, which serves a user base estimated at over 100 million monthly actives. Its net loss of RMB 715 million means the company is still not profitable on a bottom-line basis. And a $74 billion valuation on a $70.7M seven-month revenue base implies a revenue multiple that would make even the most optimistic AI investor pause — the bet is entirely on the growth curve continuing.

What Comes Next

The IPO timeline is the thing to watch. An onshore listing — likely on the Shanghai STAR Market or Shenzhen’s ChiNext — would make DeepSeek the first major Chinese AI lab to go public, and would provide the clearest public window yet into the actual economics of frontier AI development. Filing “as soon as this year” means prospectus-level financials could surface within months.

For the broader market, DeepSeek’s numbers complicate the narrative that frontier AI is a money pit. The company has shown that a different cost structure — built on efficiency research, domestic hardware, and open-weight distribution as a customer-acquisition strategy — can produce software-grade margins. Whether Western labs can respond, or whether the industry bifurcates into a high-margin East and a high-spend West, may turn out to be the defining economics question of this AI cycle.

Either way, the lab that started 2025 by embarrassing Silicon Valley with a $5.5 million training bill has now shown it can embarrass them on gross margin too.