DeepSeek's API Runs an 82.9% Gross Margin • The Weighted Average

DeepSeek’s first detailed financials arrived this week, and the number that should reset your vendor spreadsheet is not the revenue. The Chinese lab generated 475 million yuan, about $70.7 million, in the first seven months of 2026 — roughly ten times its full-year 2025 total — while narrowing net losses from 935 million yuan to 715 million yuan, according to figures The Information attributed to two people with knowledge of its financial data. Underneath sits the disclosure that matters: an overall gross margin of 44.6%, and an API gross margin of 82.9%.

The cheap tokens were never a loss leader

Eighty-three points of gross margin on API sales is not the profile of a company subsidising inference to buy share. It is the profile of a company that solved inference cost and priced well above it anyway. Coverage of the same financials attributes the margin to the deployment scale of the open-weight V4 series, particularly the lightweight V4-Flash, and puts DeepSeek’s API margin against OpenAI’s 39% and an estimated 63% for Anthropic.

That framing is the operator’s cue. If a lab can serve a 670-billion-parameter class model at those margins, the price you pay is a market-position decision rather than a cost floor — and it moves when the position changes. It has already moved. DeepSeek raised V4-Pro pricing this month, with peak output going from $0.87 to $3.96 per million tokens, a 4.6x increase catalogued alongside the margin figures. The current DeepSeek pricing page confirms the new tier: V4-Pro at $3.96 per million output tokens at peak and $1.98 off-peak, with V4-Flash at $1.32 and $0.66.

Here is the arithmetic nobody has published. Apply the disclosed 82.9% API margin to the new V4-Pro peak output price and the implied cost of goods is about $0.68 per million output tokens. Against the old $0.87 price, that same cost implies a gross margin near 22% — so either DeepSeek’s inference costs fell dramatically during 2026, or the earlier price was close to breakeven and the increase is the moment the flagship tier started carrying the company. Both readings point the same direction for buyers: the discount you budgeted was a phase, not a structural advantage.

What a repricing lab does to your build-versus-buy math

Context makes the move look inevitable. DeepSeek is pursuing a second round targeting 50 billion yuan at a 500-billion-yuan valuation and has hired banks for a possible listing, per the same reporting. A company preparing for public markets needs a margin story, and margin stories are built by raising prices on the tier customers cannot easily leave.

Run the revenue against the loss and the shape of that story becomes legible. Seven months of $70.7 million annualises to roughly $121 million, against a net loss that annualises near 1.23 billion yuan — about $183 million. DeepSeek is therefore losing something on the order of $1.50 for every dollar of revenue, even while its API line clears 82.9 points of gross margin. Every one of those loss dollars sits above the API business, in training runs and infrastructure, and the only lever that closes the gap without slowing model development is price. The 4.6x increase is the first pull on it.

Even after the increase, DeepSeek remains the cheap option: $3.96 per million output tokens against Kimi K3 at $15 and Claude Opus 5 at $25 per million output tokens on Anthropic’s published rates. The gap is roughly six-fold against Opus. But six-fold is not the sixty-fold that justified building a whole routing layer around Chinese open-weight models, and the direction of travel is against you. This paper documented Anthropic’s premium tier struggling to win 8% of its own customers’ spend; DeepSeek is running the opposite experiment, testing how far a cheap tier can climb before demand notices.

The counterpoint deserves weight. These figures come from unnamed sources reporting to a subscription publication, not from an audited filing, and DeepSeek has published nothing itself. The 82.9% API margin almost certainly excludes training amortisation — the company still posted a 715-million-yuan net loss on 475 million yuan of revenue, which means everything below the gross line is deeply underwater. An API margin that looks like software while the enterprise burns cash is a familiar pattern, and it usually ends with more price increases, not fewer.

The verdict for anyone routing production traffic through DeepSeek: treat the current price as a floor that has already lifted once this quarter, and re-run your unit economics at $3.96, not $0.87. If your margin only works at the old number, you were arbitraging a fundraising strategy. The evidence that would change this call is a self-published DeepSeek cost disclosure or a competitor matching V4-Pro capability below $2 — the same test applied to the gray-market Claude tokens that cost $4.62 once verified. Today’s edition’s lead makes the complementary point from the hardware side: inference efficiency gains are real but arrive on a vendor’s schedule, not yours.