Unverified50% confidenceFactExact time
In the inference system description DeepSeek published in February 2025, the cost was calculated as follows: an average of 226.75 nodes running, rented at 2 US dollars per card per hour, for a cost of 87,000 US dollars per day; billed at R1's price at the time, revenue was 562,000 US dollars per day, a cost profit margin of 545%.
1
Sources
50%
Confidence
Long-term
Relevance
10/23/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe author's monthly running cost for local LLM inference in the first three months (hardware amortization plus electricity) was about $58 to $73, versus a cloud API monthly bill of about $180 to $240 at equivalent usage, a net saving of $107 to $180 per month, breaking even around the fourth month.75% similarUnverifiedOutsiders speculate that DeepInfer's average revenue per GPU per hour may be around $46, corresponding to annual revenue of about $400,000.75% similarUnverifiedThe model's key assumptions include: reviewer calls are 2000 input + 300 output tokens; system-under-test calls are 1500 input + 500 output tokens; a mid-tier model ($2/$10 per million tokens) is used by default; offline runs use the batch API for a 50% discount; online review uses standard pricing; human review is 4 minutes per case at $45 per hour; tooling costs follow Langfuse list prices; and the loaded cost of US full-time labor is $16188 per month.72% similarUnverifiedThe monthly running cost of standard models is about $0.004 per conversation, and about $0.008 per conversation for advanced models, plus about $150/month in hosting and maintenance fees.72% similarUnverifiedOfficial public cost estimate: each H800 node processes about 73,000 input tokens or 14,800 output tokens per second; if all traffic were billed at the most expensive price, the theoretical profit margin would be about 545%.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1088578API
curl https://kongchang.com/api/v1/knowledge/claims/1088578MCP
get_claim(id=1088578)