DeepSeek releases open-source inference engine

On Tuesday, October 20, DeepSeek released DeepInfer as open source, the inference engine that runs its DeepSeek-V4.1-Flash model in production. The model’s weights were published three days earlier, on October 17.

Revenue estimates for inference cards

A sentence from DeepSeek’s post on Zhihu, which circulated on Hacker News and Reddit, stated that a single cluster processes 155,000 billion tokens per day, generating "38 dollars per card and per hour" (approximately 34 euros), calculated at off-peak pricing. These figures are presented as approximate and have not been independently verified.

Calculations based on production data

According to production statistics published by DeepSeek and cited by 36Kr, a card tasked with generating responses produces 14,000 tokens per second. At the current exchange rate, this results in approximately 30 dollars per hour. A card handling reading requests generates about 48 dollars per hour, leading to an average of around 39 dollars per hour if both types of cards are used equally.

Comparison with market prices

An H100, the current market reference, is rented for around 3 to 4 dollars per hour in 2026 (approximately 2.70 to 3.50 euros), with prices reaching up to a dozen dollars at major cloud providers like AWS, Azure, or Google. A next-generation B200 card costs between 6 and 11 dollars per hour. Purchasing a card for around 60,000 dollars (approximately 53,000 euros), including the server, amortized over five years, costs approximately 1.4 dollars per hour.

Economic implications for inference business

An article on the Chinese portal Sina calculated that renting an H200 at 5.40 dollars per hour (approximately 4.80 euros) would leave 85.8% of the revenue for the provider. The article concluded that the business model for inference could be "much better than what the market imagines."

Costs and profitability for home users

An RTX 5090, launched at 2,349 euros in January 2025, now costs between 6,500 and 8,500 euros among French resellers due to memory shortages. When in use, the PC hosting it consumes about 600 watts, costing approximately 12 cents per hour in electricity. However, the RTX 5090 is not supported by DeepInfer, which targets only data center chips and Huawei’s Ascend 950.

Alternative options for using AI at home

For home users, alternatives to DeepInfer include platforms like Vast.ai, RunPod, or Salad, which allow renting graphics cards. On Salad, an RTX 5090 can be rented for 0.69 dollars per hour, but the platform retains a portion of the revenue. After paying for electricity, about 13 cents per hour remain, or roughly 1,000 euros per year for a machine running day and night.

Economic comparison between DeepSeek and home users

A DeepSeek card earns about 38 dollars per hour, while an RTX 5090 rented on Salad earns about 0.25 euros per hour. Electricity costs about 0.12 euros per hour for the RTX 5090. Running a medium-sized model like Qwen3 32B on an RTX 5090 generates about 60 tokens per second, costing around 14 euros per million tokens, compared to 0.53 euros for DeepSeek’s Flash model during off-peak hours.

Profitability and business models

DeepSeek’s profitability comes from processing thousands of requests simultaneously and reusing cached data instead of recalculating everything. In contrast, an RTX 5090 would take five to six days to generate the same amount of tokens as a DeepSeek card does in one hour. OpenAI, despite being profitable on inference, is expected to lose billions due to high costs for training and research.

Challenges and limitations

The revenue figures refer to a single cluster at full capacity, excluding reserve cards. The calculation uses off-peak pricing, which is significantly lower than peak hours, which DeepSeek limits to weekdays. Additionally, the Flash model is much cheaper and more optimized than top-tier models like GPT or Claude. DeepInfer supports specific hardware, and performance figures are based on Nvidia cards, not Huawei’s Ascend 950.