DeepSeek, the Chinese AI phenomenon that's been keeping Silicon Valley executives awake at night, just dropped another bombshell. The startup claims its AI models could rake in theoretical profits of 545% per day. The word "theoretical" is doing a lot of heavy lifting here.
The Hangzhou-based company chose Saturday to peel back the curtain on its finances, revealing details about its V3 and R1 models' performance. The numbers are striking: potential daily revenue of $562,027 against costs of $87,072. That's enough margin to make even luxury fashion brands jealous.
Doing More With Less - Much Less
This revelation comes from a company that previously shocked the industry by claiming it spent less than $6 million on chips to train its models. For context, that's pocket change compared to what U.S. companies like OpenAI are spending. The kicker? DeepSeek did it with Nvidia's H800 chips - the AI equivalent of bringing a Toyota to a Formula 1 race and somehow winning.
The timing couldn't be more pointed. AI stocks outside China took a nosedive in January after DeepSeek's chatbots started winning popularity contests worldwide. Their secret sauce? Doing more with less, apparently.
The Fine Print Behind Those Eye-Popping Numbers
But before investors start dumping their U.S. tech stocks, there's some fine print to consider. DeepSeek quickly added that its actual revenues are "substantially lower" than the theoretical figures. It turns out running a profitable AI company isn't as simple as their initial math suggests.
The company processes a whopping 608 billion input tokens daily, producing 168 billion output tokens. Each H800 node - their computational workhorse - handles about 73,700 tokens per second during initial processing. These numbers are impressive, but they come with more asterisks than a Hollywood contract.
Here's where the rubber meets the road: most of DeepSeek's services aren't even monetized. They're giving away their product for free on web and mobile apps, like a high-tech food truck handing out samples. They also offer discounts during off-peak hours when demand drops, and their premium R1 model's pricing isn't applied across all services.
Their infrastructure runs on H800 GPUs, costing about $2 per GPU per hour. During peak times, they deploy 278 nodes, each packing 8 GPUs. At night, they scale down and repurpose the idle computing power for research and training - like turning your restaurant into a cooking school after hours.
A Symphony of Semiconductors
The company's technical innovations focus on throughput and latency, using something called cross-node Expert Parallelism. Think of it as an orchestra where each section plays perfectly in time, but with semiconductors instead of strings.
This revelation comes at a crucial moment. Major players like OpenAI and Anthropic are still experimenting with their business models, trying everything from subscriptions to usage-based pricing to licensing fees. Meanwhile, investors are getting antsy, wondering when - or if - these companies will turn a profit.
In an unusual move for the secretive AI industry, DeepSeek also shared details about its technical innovations this week. While American companies guard their secrets like dragons hoarding gold, DeepSeek has embraced transparency - at least about their technical achievements, if not their actual financial performance.
At current rates, DeepSeek's theoretical revenue could hit $200 million annually. But that's like calculating your salary assuming you work every minute of every day - technically possible, practically impossible.
Why this matters:
- DeepSeek's numbers are the first public glimpse into potential AI profit margins, but they're about as realistic as a Hollywood movie budget. The gap between theoretical and actual profits highlights the challenging economics of the AI industry.
- The company's ability to achieve impressive results with less powerful hardware raises serious questions about U.S. AI companies' massive chip spending. Either DeepSeek has figured out something revolutionary, or someone's math doesn't add up.
Read on, my dear:
- DeepSeek on Github: Day 6: One More Thing, DeepSeek-V3/R1 Inference System Overview
