Skip to content
World clockEU--:--UK--:--USA--:--CN--:--PLDEFRIT中文EN

portal about AI and technologyevents · analysis · interviews · technical background

Search
LIVE
›

Recalculation puts DeepSeek inference revenue at $39.10 per GPU-hour against $38 disclosed

An analysis published on 9 November recalculates DeepSeek's disclosed inference revenue at about $39.10 per GPU-hour, against the $38 the company reported on 20 October. The gap is small, and the reason the price is so low is that 96% of input tokens are cache hits.

AI & modelsExplainerRachel NwosuPublished: 9 November 20263 min readSources 18
Recalculation puts DeepSeek inference revenue at $39.10 per GPU-hour against $38 disclosed

What DeepSeek published, and what the check found

DeepSeek published part of DeepInfer, its inference framework, on 20 October 2026.

Anyone comparing per-token API prices is partly comparing how well each provider reuses cached context, not just how fast its GPUs run. A workload of short, non-repeating prompts would be priced very differently from DeepSeek's average request. The company says one cluster processes 155 trillion tokens a day.

Yukio Takagi's analysis, published on 9 November, rebuilds that $38 from the disclosed throughput and the public price list. A prefill GPU and a decode GPU together bring in $78.19 an hour, or $39.10 each. It is not independent confirmation: both numbers rest on DeepSeek's own measurements.

The cache is what makes the price work

Cached input costs $0.003 per million tokens off-peak. Uncached input costs $0.15, fifty times as much, and output costs $0.60.

The average request carries 77,000 input tokens. Roughly 95% of processed tokens are cached input. They produce only about 20% of revenue, while uncached input and output bring in about 80%. A single short question reuses little of its input, so the same GPU would earn less per hour on that traffic.

Cache hits change throughput too. A prefill GPU handles 1.5 million tokens a second with hits counted and 60,000 without them, a 25-fold difference. DeepSeek keeps the cache on about 100 PiB of SSD, for three days on agent workloads and seven days on chat.

What the margin leaves out

DeepSeek does not publish what it pays for GPUs. Any margin therefore depends on an assumption. At $2 per GPU-hour, the figure DeepSeek used in a February 2025 calculation, gross margin on GPU time is about 95%. At $7, a rate cited from a post by Yuchen Jin on X, it is about 82%.

The calculation also leaves out cache storage, inter-node bandwidth, research and staff costs.

The wider price collapse behind it

Epoch AI estimates the price of a fixed level of AI performance has fallen about 47% a quarter since 2023. That is roughly 13 times a year.

DeepSeek's cache hit rate was about 56% when it published V3 and R1 figures in February 2025. The current rate may reflect a different model and measurement.

Together says its Compiler reached 1,205 tokens a second on DeepSeek v4.1-Flash, part of a 2.6x gain across frontier open models in under 24 hours. Edoardo Vacchi of Red Hat says ordinary Kubernetes load balancers break LLM efficiency because they ignore cached conversational context.

Comments 0

Sources

18
  1. 01Reading DeepSeek's Inference through Resource Allocation and Pricing Systems: A Verification of DisclosuresEN
  2. 02Together Compiler: Continual Compilation for AI InferenceEN
  3. 03Podcast: Distributed LLM Inference on Kubernetes with llm-d, Wasm, & TornadoVMEN
  4. 04AI Model Cost Statistics: Training, Inference and Token PricingEN
  5. 05Musk's Starlink Embraces 'Dynamic Pricing,' Further Jacking Up CostsEN
  6. 06GPU Infrastructure: Cloud Costs and Scaling for AIEN
  7. 07Can Optical AI Cut Data Center Inference Power Use?EN
  8. 08Green Hydrogen: India’s Big Ambitions, Costly RealitiesEN
  9. 09Anthropic's FlashProve proves which model wrote a reply, and its 72x costEN
  10. 10AWS Adds NVIDIA GPU Partitioning to SageMaker Inference EndpointsEN
  11. 11Gemini Omni 1.2 Flash: Fast and cost-effective with longer clipsEN
  12. 12E2E Networks Launches Blaze to Offer Pay-Per-Token AI Inference ServicesEN
  13. 13VMO2 to axe 30% staff costs, halt fiber build in £800M savings planEN
  14. 14Optimizing Reference Deletions in GitEN
  15. 15Launch HN: Rivet (YC W23) – Scale and Orchestrate Agents on S3EN
  16. 16Launch HN: Momentic (YC W24) – Agents that catch bugs by using your appEN
  17. 17Rivian, Ford and GM are using 3D printing to reduce costsEN
  18. 18Here's how much Trump's Iran war has added to Americans' fuel costsEN

All figures and quotations in this text come from the sources listed below.

Content prepared by the editorial team with AI assistance.

Rachel Nwosu

Rachel Nwosu

AI, models and technology

Rachel Nwosu covers AI, models and technology for FLASH24, working from public model documentation, benchmark releases and repository histories rather than press summaries, and she skips announcements that arrive without reproducible numbers. She checks training-data claims against dataset cards and reruns reported metrics where code is available. She spends much of her week interviewing researchers and engineers, tracking model launch calendars, and comparing vendor benchmarks with independent evaluations. Outside the desk she runs 3D printers, restores old computers, and tests how models learn from internet junk. She does not publish benchmark figures she cannot trace to a source.

Newsroom →

Comments

0
  1. No comments yet — be the first.

Write a comment

Comments are public. We do not publish abuse, spam or advertising.