Hook
The number hit my terminal at 6:42 AM Barcelona time. DeepSeek—the Chinese AI lab that's been quietly terrorizing the LLM pricing charts—allegedly pulled in $70 million in July. Market whispers claim a 10x revenue surge across 2025. No official filing. No audited statement. Just a "market rumor" sourced from a channel called "Dongcha Beating AI."
Speed beats analysis when the graph is vertical. But this graph isn't vertical. It's a phantom.
I've seen this playbook before. In 2022, FTX was "generating $1 billion in revenue." The number sounded real until it wasn't. The difference? FTX had order books you could audit. DeepSeek has API endpoints and a reputation for aggressive pricing.
Here's what I actually know: DeepSeek's MoE architecture—specifically the V2 and V3 series—delivered inference costs that made competitors look like they were pricing in 2021. Their API pricing triggered at least three rounds of price cuts across the Chinese market. When a lab becomes the industry's price anchor, revenue follows volume. The question is whether $70 million monthly is volume or vaporware.
Context
DeepSeek isn't a typical Chinese AI startup. Founded in 2023, the lab built its reputation on two pillars: open-source model releases and a pricing strategy that borders on predatory. The DeepSeek-V3 release was a watershed moment—benchmark scores competitive with models trained at 10x the compute budget.
The Chinese AI market operates differently from the West. Baidu and Alibaba dominate through cloud integration. Startups like Zhipu, Moonshot, and MiniMax fight for scraps. DeepSeek carved a different path: pure API economics. No consumer app. No enterprise sales team. Just a dev-friendly platform with token prices that undercut everyone.
In this context, a $70 million monthly run rate means two things. Either DeepSeek has captured an enormous share of price-sensitive API demand, or someone is telling a story they want investors to believe. The truth probably sits between those poles.
Core
Let's run the numbers with a trader's eye. I don't read whitepapers; I read order books.
The $70 million figure, if monthly, annualizes to $840 million. For context, SenseTime—a publicly traded AI company with thousands of employees—generated approximately $170 million from its entire generative AI business in 2023. DeepSeek claiming 5x that revenue with a fraction of the headcount demands extraordinary evidence.
The API usage math works against the rumor. At DeepSeek's pricing—roughly $0.14 per million input tokens for V3—you'd need astronomical token volumes to hit $70 million monthly. We're talking quadrillions of tokens. That's not impossible in aggregate, but it requires an installed base of enterprise customers that conflicts with DeepSeek's dev-first reputation.
More plausible: the figure includes enterprise deployments, private model licensing, or government contracts. Chinese AI companies often bundle API access with infrastructure deals. The revenue is real but the composition matters. Recurring API revenue trades at a premium to one-time licensing. If $70 million includes non-recurring items, the actual run rate is significantly lower.
Based on my audit experience, when a rumor surfaces without accompanying cost structure data, treat it as marketing. The missing context—customer concentration, gross margins, churn rates—tells you more than the headline number.
Contrarian
The real news isn't DeepSeek's revenue. It's that nobody can verify it—and nobody seems to care.
That's the story here. We're in a bull market for AI narratives, not just crypto. The market desperately wants a Chinese AI champion that doesn't depend on Nvidia's latest chips. DeepSeek fits that narrative perfectly: algorithmic efficiency overcoming hardware constraints.
This rumor does exactly what it was designed to do. It tests the market's appetite for a "China AI breakout" story. The response determines whether more detailed—and potentially fabricated—data points follow.
Consider the timing. AI Act enforcement in the EU is ramping up. US export controls are tightening. DeepSeek needs to signal resilience to both domestic and international audiences. A revenue story accomplishes this without revealing technical specifics.
There's also a secondary effect the market overlooks: DeepSeek's success pressures every other Chinese AI startup. Zhipu, Moonshot, MiniMax—they all need to justify their valuations against this alleged benchmark. If DeepSeek is doing $840 million annualized, why are you investing in competitors with thinner revenue trajectories?
Takeaway
Watch the API pricing over the next 90 days. If DeepSeek raises prices, the revenue story gains credibility—they have pricing power. If prices drop further, they're buying revenue at the expense of margins. The best news is the news that moves the price, but the absence of verifiable data is itself a data point.
The only thing worse than chasing a rumor is ignoring the signal it represents. China's AI sector is consolidating around cost efficiency. DeepSeek is winning that game. Whether the $70 million number is real or aspirational, the direction is clear. The market just needs to decide what that direction is worth.
Speed beats analysis when the graph is vertical. This graph is horizontal. Wait for confirmation. Then move.