Chinese language AI startup DeepSeek has reported a theoretical every day revenue margin of545% for its inference providers, regardless of limitations in monetisation and discounted pricing constructions. The corporate shared these particulars in a latest GitHub put up, outlining the operational prices and income potential of its DeepSeek-V3 and R1 fashions.
Primarily based on DeepSeek-R1’s pricing mannequin—charging $0.14 per million enter tokens for cache hits, $0.55 per million for cache misses, and $2.19 per million output tokens—the theoretical income generated every day is $562,027.
Nonetheless, the corporate acknowledged that precise earnings had been considerably decrease because of decrease pricing for DeepSeek-V3, free entry to net and app providers, and automated nighttime reductions. “Our pricing technique prioritises accessibility and long-term adoption over instant income maximisation,” DeepSeek stated.
In line with the corporate, DeepSeeks inference providers run on NVIDIA H800 GPUs, with matrix multiplications and dispatch transmissions utilizing the FP8 format, whereas core MLA computations and mix transmissions function in BF16. The corporate scales its GPU utilization based mostly on demand, deploying all nodes throughout peak hours and decreasing them at evening to allocate assets for analysis and coaching.
The GitHub put up revealed that over a 24-hour interval from February 27, 2025, to 12:00 PM on February 28, 2025, 12:00 PM, DeepSeek recorded peak node occupancy at 278, with a median of 226.75 nodes in operation. With every node containing eight H800 GPUs and an estimated leasing price of $2 per GPU per hour, the entire every day expenditure reached $87,072.
The above revelation might have an effect on the US inventory market. The launch of DeepSeek’s newest mannequin, R1, which the corporate claims was educated on a $6 million price range, triggered a pointy market response. NVIDIA’s inventory tumbled 17%, wiping out almost $600 billion in worth, pushed by issues over the mannequin’s effectivity.
Nonetheless, NVIDIA chief Jensen Huang, through the latest earnings name, stated the corporate’s inference demand is accelerating, fuelled by test-time scaling and new reasoning fashions. “Fashions like OpenAI’s, Grok 3, and DeepSeek R1 are reasoning fashions that apply inference-time scaling. Reasoning fashions can devour 100 occasions extra compute,” he stated.
“DeepSeek-R1 has ignited international enthusiasm. It’s a superb innovation. However much more importantly, it has open-sourced a world-class reasoning AI mannequin,” Huang stated.
In line with a latest report, DeepSeek plans to launch its subsequent reasoning mannequin, the DeepSeek R2, ‘as early as attainable.’ The corporate initially deliberate to launch it in early Could however is now contemplating an earlier timeline. The mannequin is claimed to supply ‘higher coding’ and cause in languages past English.
The put up DeepSeek Studies 545% Day by day Revenue Regardless of Free AI Companies appeared first on Analytics India Journal.
