Deep Dive — DeepSeek's leaked P&L shows frontier AI can run on a shoestring

A lab that spent the year as everyone's favorite underdog just handed the industry its first real balance sheet, and it reads like an argument about what AI should cost.

DeepSeek generated roughly $70.7 million in revenue over the first seven months of 2026 while losing $106 million, according to figures The Information reported this week, citing sources. The company has not confirmed the numbers, and no filing backs them yet — treat them as sourced, not filed. But if they are even roughly right, they describe something none of DeepSeek's Western rivals have ever volunteered: a public look at what it actually costs to run a frontier AI lab, at least one built in Hangzhou on constrained hardware and stubborn engineering economics.

The numbers land with extra force because of where DeepSeek may be heading: multiple reports through July pointed to a planned Shanghai STAR Market listing, with a Chinese regulatory filing implying a valuation around $52 billion. A leak like this, three months before a possible filing window opens, is either carelessness or pre-marketing. Given how carefully the lab guards everything else, assume the latter — and read it the way a fund manager would.

What the numbers actually say

Start with the shape, not the size. Revenue of roughly $70.7 million for January through July is about ten times what the lab reportedly earned in all of 2025 — around $7 million against a $139 million annual loss, per the same reporting. Two curves move at once here: revenue is scaling fast while the burn actually shrank year over year. Most companies show you growth or discipline. Showing both simultaneously is the whole pitch.

The revenue itself comes overwhelmingly from selling API access — token consumption, batch inference endpoints, and private deployment partnerships built on the V3 and R1 families and the newer V4 line, which had its own eventful month. We covered the launch of its paid flagship in DeepSeek ships V4 Pro, ending its flagship's four-month preview, and the pricing reset that followed in China's LLMs quit the price war as DeepSeek API jumps 11x. That sequence is the financial story in miniature: raise list prices by an order of magnitude, then watch revenue grow anyway because demand is compounding faster than price. The earlier step in that ladder — when we covered how DeepSeek hikes API prices up to 4.7x with new peak-hour billing — now reads less like opportunism and more like a deliberate march back toward margin.

One more number deserves attention precisely because nobody said it out loud: usage. The reporting describes revenue but not token volume, so any claim that "demand exploded" rests on inference being the product — reasonable, but unverified. And the loss figure carries real ambiguity too. It is a net figure across seven months that almost certainly includes training runs for models whose revenue arrives later; whether $106 million represents "sustainable operations" or "operations subsidized by training spend" depends on accounting choices nobody outside the company has seen.

Why the burn is so small

The interesting question is not whether $106 million is a lot — by US frontier-lab standards it is rounding error — but how DeepSeek keeps it that small. The answer is architectural, and it predates the money.

DeepSeek's model families are built around efficiency techniques that were born of necessity. Mixture-of-experts routing means only a small fraction of total parameters activate per token, so each query costs a fraction of what a dense model of equal size would. Multi-head latent attention compresses the key-value cache — the working memory a model keeps during generation — into low-dimensional vectors, cutting the high-bandwidth memory footprint that makes long-context serving expensive. Native low-precision (FP8) training and inference stretch the same silicon further. None of these tricks is secret; DeepSeek's own papers documented them years ago, and its open-source releases turned them into a public playbook that labs from Beijing to San Francisco have spent two years copying.

What changed is that the constraint became the strategy. US export controls cap the compute DeepSeek can legally buy, so the lab optimizes relentlessly for work per chip — and that discipline happens to be exactly what the rest of the industry now needs as inference demand explodes and power budgets tighten. When Nvidia benchmarks its next rack against DeepSeek's models, or when OpenAI tests its custom silicon on DeepSeek R1, the efficiency-first architecture is doing double duty as the industry's reference workload. We saw the same dynamic from the hardware side yesterday, when Nvidia's Vera Rubin posted a 30x agentic jump in first hardware tests running agentic coding work on a 1.6-trillion-parameter DeepSeek model. Efficiency stopped being a workaround and became the benchmark.

There is also a quieter structural advantage: DeepSeek does not carry Western-scale go-to-market costs. No sales army, no consumer-app marketing blitz, no enterprise-success organization. Its distribution is largely self-service API plus its own reputation in the developer community — which is why a tenfold revenue jump can arrive without a matching expense line. Whether that stays true at IPO scale is a fair question; public markets tend to demand growth infrastructure that pure research shops never had to build.

The comparison problem

Here is where the skeptics get their turn, and they should. Comparing DeepSeek's burn to OpenAI's or Anthropic's is not apples to apples — it is apples to a different fruit grown in a different orchard under different weather.

US frontier labs report losses measured in billions per quarter, but those losses buy things DeepSeek is not buying: frontier-scale training clusters measured in gigawatts, compensation packages that anchor top researchers in the Bay Area, and global compliance, safety, and enterprise organizations. Some of that spending is choice; some of it is the cost of competing at the absolute edge of capability. If DeepSeek chose tomorrow to train a model requiring 100 times its current compute budget, its loss would look very American very quickly.

Capability is the other half of the comparison problem. Cheap tokens matter only if the tokens are good enough for the job. On raw intelligence indexes, the leading closed models still sit above DeepSeek's best; the lab's edge is the price-performance frontier — delivering maybe 95 percent of frontier usefulness at a tenth the cost, which is plenty for most commercial workloads. That gap has been narrowing, but it has not reversed, and the moment anyone stops measuring price-adjusted scores, DeepSeek's advantage looks smaller.

And there is the funding asymmetry. DeepSeek's external round this year — reported at tens of billions of yuan, with Tencent and CATL among the backers — bought it room, but its historical model was a quant fund's trading profits, patient capital by startup standards. OpenAI raises more than DeepSeek's implied valuation in single rounds. Whose model survives contact with the next capability jump is genuinely unknowable.

What an IPO would change

The listing reports are the real story underneath the leak. DeepSeek has historically been funded by founder Liang Wenfeng's quant fund High-Flyer, and only recently took meaningful outside capital. An IPO on Shanghai's STAR Market — filings could come as early as late 2026 by some accounts, with trading possibly in 2027 — would make it the first frontier-scale AI lab whose costs and revenues any investor can inspect line by line.

That transparency cuts both ways. For bulls, audited financials would prove the efficiency thesis in public: cheap tokens, thin losses, compounding demand. For bears, quarterly disclosure works in reverse — every training cycle's cost lands on the record, every margin squeeze becomes visible, and a management team used to answering to one man suddenly answers to analysts asking why the loss widened. Frontier labs everywhere watch this with interest, because DeepSeek would become the first live test of whether an AI lab can be a normal public company rather than a venture-subsidized research project.

It also matters geopolitically. A listed DeepSeek gives Chinese retail investors direct exposure to the AI buildout, deepens domestic capital markets' stake in homegrown models, and hands Beijing a showcase listing at exactly the moment export controls keep tightening. Wall Street has been waiting for the AI index trade; Shanghai may get there first.

What to watch next

Three signals worth tracking. First, confirmation: if DeepSeek files for the STAR Market, audited versions of these numbers will surface, and the gap between leak and filing will tell you how much pre-marketing was happening. Second, pricing discipline: whether the lab continues raising API rates as demand compounds — the pattern behind both the 11x jump we covered and the current trajectory — or re-enters a price war to defend share against Kimi, GLM, and Qwen. Third, the capability race: whether the next DeepSeek flagship narrows the raw-intelligence gap to the closed leaders, because the entire economic story assumes near-frontier quality at a fraction of the cost.

Would you rather own a lab that loses billions chasing the frontier, or one that loses millions selling picks and shovels to everyone else? Tell us in the comments.

Sources: The Information · LLMs Blog · Reuters · Nvidia Developer Blog