The rumor hit the developer circles like a flash loan on a volatile pool. DeepSeek, the Chinese AI lab known for its aggressive pricing and open-weight models, allegedly pulled in $70 million in revenue during July. The figure claims a tenfold year-over-year increase. No official statement. No audited financials. Just a number floating in the ether, waiting to be verified or debunked.
Let me be clear about what I do when I see a number like this. I do not take it at face value. I tear it apart like a smart contract audit, looking for edge cases, overflow errors, and unintended consequences.
Context: The Machine Behind the Myth
DeepSeek, or Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd., emerged from the High-Flyer quantitative hedge fund ecosystem. Founded in 2023, it quickly became a thorn in the side of established AI giants. Their models, particularly the V2 and V3 series, leverage a Mixture-of-Experts (MoE) architecture that dramatically cuts inference costs. This is not a marketing gimmick. MoE allows the model to activate only a fraction of its parameters per token, reducing compute overhead significantly.
Their pricing strategy has been nothing short of disruptive. By undercutting established players by an order of magnitude, they forced the entire Chinese API market into a price war. Competitors like Alibaba's Qwen and Baidu's Ernie had to slash prices to stay relevant. DeepSeek also embraced open-source distribution, releasing model weights that developers could download, fine-tune, and deploy on their own infrastructure. This created a viral loop: developers used DeepSeek for personal projects, then pushed for its adoption in enterprise environments.
The core of this commercial engine is the API. It is the primary revenue stream. But the question remains: is $70 million per month a realistic number, or is this just another token-inflated narrative? Let's run the math.
Core: Dissecting the Revenue Claim
Based on my experience auditing protocol economics, the first thing I look at is the gross-to-net ratio. If $70 million is gross revenue, the net figure after channel costs, compute, and operational overhead could be significantly lower. If it is net revenue, that implies a much higher gross volume. The rumor mill does not specify. This ambiguity is a red flag.
Second, consider the customer structure. In the AI API space, a common pattern is the Pareto distribution: 80% of revenue comes from 20% of customers. If DeepSeek's revenue is concentrated in a few large enterprise contracts, the monthly figure could be volatile. One lost contract could mean a 20% hit. If it is spread across a long tail of small and medium developers, the growth is more sustainable but harder to achieve quickly.
Now, let's compare against industry benchmarks. SenseTime, a major Chinese AI company, reported its entire generative AI business generated approximately 1.2 billion RMB (about $170 million) for the full year of 2023. If DeepSeek is generating $70 million per month, that annualizes to roughly $840 million. That would put them at nearly five times SenseTime's entire AI business, just two years after founding. The claim is not impossible, but it is aggressive.
The technological angle supports some of this. DeepSeek's MoE architecture gives them a significant cost advantage. Their training and inference optimization is reportedly excellent. They have been known to use high-end chips like the H800, and their engineering team has deep expertise in model parallelism and efficient deployment. This means they can offer lower prices while maintaining healthier margins than competitors who rely on less efficient architectures.
But there is a limit. Even with perfect engineering, serving millions of API calls requires massive compute infrastructure. The revenue growth implies a corresponding growth in GPU demand. This aligns with the broader trend of AI companies hoarding compute resources. However, it also means that DeepSeek is exposed to supply chain risks. US export controls on advanced chips could throttle their capacity expansion, creating a bottleneck between revenue growth and service delivery.
Contrarian: The Security Blind Spot
The contrarian angle here is not about DeepSeek's technology. It is about the narrative itself. In my years auditing protocols, I have learned that when a single metric is highlighted to the exclusion of all others, there is usually a reason. This rumor surfaced on "Dongcha Beating AI," a source with no established track record. No mainstream outlet like Reuters or Caixin has confirmed it. This looks like a signal flare, possibly launched by investors looking to boost sentiment before a funding round, or by competitors trying to pressure DeepSeek into revealing more than they want.
The blind spot is the assumption that high revenue equals healthy fundamentals. A company can buy growth by subsidizing usage. If DeepSeek is running a promotional campaign where API credits are sold below cost, the revenue figure would be inflated. The real question is about profitability and net revenue retention. We have no visibility into that. We also do not know if this revenue includes large-scale private deployments, which are often one-time deals with high margins but low recurrence.
There is also the issue of accounting standards. In China, some AI companies recognize revenue on a gross basis, including pass-through costs. If DeepSeek is reselling cloud compute or third-party services as part of their offerings, the reported revenue could be padded with zero-margin pass-through amounts. The code is law, but the revenue is the human exception. The ledger remembers what the wallet forgets.
Takeaway: The Vulnerability Forecast
The $70 million figure is a hypothesis, not a fact. It needs rigorous validation. My recommendation is to treat this as a market signal with a confidence rating of C-minus. The information is plausible enough to affect sentiment, but not reliable enough to base investment decisions on. If this rumor is true, DeepSeek has proven that an open-source, high-efficiency strategy can generate massive commercial traction. If it is false, it is a lesson in how quickly narratives can outpace reality.
I am watching three signals. First, whether DeepSeek's official channels or a mainstream financial outlet validate the number within the next month. Second, whether their API pricing remains stable or adjusts upward, which would indicate they have pricing power. Third, whether their competitors, particularly the cloud giants, respond with further price cuts or capacity expansion. Code is law, but bugs are the human exception. The bug here might be the revenue report itself. Until it is patched, I remain skeptical and alert.
The ledger remembers what the wallet forgets. I will not forget this figure until I see the transaction history.