--- title: AI inference is obviously profitable | daily.dev description: A back-of-the-envelope cost analysis argues that AI inference is genuinely profitable, contrary to popular claims that it requires VC subsidies to survive.... canonical: https://daily.dev/posts/ai-inference-is-obviously-profitable-4xqzohtxz twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: AI inference is obviously profitable | daily.dev og:description: A back-of-the-envelope cost analysis argues that AI inference is genuinely profitable, contrary to popular claims that it requires VC subsidies to survive.... og:url: https://daily.dev/posts/ai-inference-is-obviously-profitable-4xqzohtxz og:image: https://api.daily.dev/og/posts/4XqZoHtxz.png og:image:alt: AI inference is obviously profitable og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # AI inference is obviously profitable **[sean goedecke](https://daily.dev/sources/seangoedecke)** · 6 min read · 16 upvotes · 4 comments ## Summary A back-of-the-envelope cost analysis argues that AI inference is genuinely profitable, contrary to popular claims that it requires VC subsidies to survive. Using A100 GPU power consumption, amortized hardware costs, and industrial electricity prices, the author estimates inference costs at roughly $1 per million output tokens — well below the $4.50+ that providers like OpenAI charge, implying 70–80% gross margins. DeepSeek's open-weights models and competitive API pricing further corroborate these margins. The key distinction: inference itself is profitable, but AI labs like OpenAI and Anthropic use those margins to subsidize expensive model training. Pure inference providers without training costs could remain profitable even if the current AI investment bubble deflates. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### Is AI inference actually profitable or is it subsidized by investor money? Inference is profitable on its own. Rough estimates put the cost of serving a 70B-class dense model at about one dollar per million output tokens, factoring in A100 power draw (400W, ~13 cents per million tokens) and amortized GPU cost over a five-year lifespan (~$1.80/hour). OpenAI charges $4.50 per million tokens for GPT-5.4-mini, implying a large profit margin on inference itself, even though the companies overall may be unprofitable due to training costs. _Track the real economics behind AI inference pricing and margins on daily.dev._ ### Why do OpenAI and Anthropic charge so much more than the estimated cost of inference? Because inference margins have to subsidize training costs. AI labs use profits from serving existing models to fund the enormous capital investment required to train new frontier models, which is why inference margins run 70-80% even though actual serving costs are much lower. A standalone inference provider without training costs would not need such high margins. _Compare AI provider pricing strategies and margins with daily.dev when evaluating API costs._ ### How does DeepSeek's inference pricing compare to OpenAI or Anthropic and what does that reveal about actual serving costs? DeepSeek reports over 80% profit margin on R1 inference while charging less than half of what OpenAI or Anthropic charge, and DeepSeek-V4-Pro sells on the open market for around 87 cents per million output tokens. Because DeepSeek's weights are open and competitors can undercut inflated pricing, that near-cost price suggests real inference costs are close to that figure, supporting lower cost estimates than raw GPU math alone implies. _Follow open-weight model pricing trends like DeepSeek's on daily.dev to gauge true inference costs._ ## Community discussion Top comments from developers on daily.dev. **@jensroland** · 25 upvotes > That article is just one false assumption after the next: > > - Nothing in a datacenter draws power apart from the GPUs > - Staff costs don't matter, neither operational nor development. We all know top lab AI researchers are cheap hires, right? It's not like top talent could [cost 2.7 BILLION dollars](https://bloomcapitalreview.com/2026/06/25/the-limits-of-the-2-7-billion-acqui-hire-noam-shazeer-leaves-google-for-openai/) to acqui-hire. > - The listed GPU is the $20k A100, and the author claims an A100 bought today will still be in service in five years. Noone is buying A100s in 2026 expecting to... --- Tags: [#llm](https://daily.dev/tags/llm), [#openai](https://daily.dev/tags/openai), [#gpu](https://daily.dev/tags/gpu), [#ai-inference](https://daily.dev/tags/ai-inference), [#deepseek](https://daily.dev/tags/deepseek) [View this post on daily.dev](https://daily.dev/posts/ai-inference-is-obviously-profitable-4xqzohtxz) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"AI inference is obviously profitable","url":"https://daily.dev/posts/ai-inference-is-obviously-profitable-4xqzohtxz","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/ai-inference-is-obviously-profitable-4xqzohtxz"},"datePublished":"2026-06-26T12:07:48.255Z","dateModified":"2026-09-13T18:48:45.014Z","description":"A back-of-the-envelope cost analysis argues that AI inference is genuinely profitable, contrary to popular claims that it requires VC subsidies to survive....","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/479e95ee2af112c350fa09a341f0c630?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/479e95ee2af112c350fa09a341f0c630?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"sean goedecke","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"sean goedecke","logo":"https://media.daily.dev/image/upload/s--WNgiZyDx--/f_auto,q_auto/v1765713419/logos/seangoedecke","url":"https://daily.dev/sources/seangoedecke"},"commentCount":4,"discussionUrl":"https://daily.dev/posts/ai-inference-is-obviously-profitable-4xqzohtxz","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":16},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":4}],"keywords":"llm,openai,gpu,ai-inference,deepseek","timeRequired":"PT6M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"sean goedecke","item":"https://daily.dev/sources/seangoedecke"},{"@type":"ListItem","position":3,"name":"AI inference is obviously profitable"}]} {"@context":"https://schema.org","@type":"WebPage","@id":"https://daily.dev/posts/ai-inference-is-obviously-profitable-4xqzohtxz","comment":[{"@type":"Comment","text":"That article is just one false assumption after the next:\n\nNothing in a datacenter draws power apart from the GPUs\nStaff costs don’t matter, neither operational nor development. We all know top lab AI researchers are cheap hires, right? It’s not like top talent could cost 2.7 BILLION dollars to acqui-hire.\nThe listed GPU is the $20k A100, and the author claims an A100 bought today will still be in service in five years. Noone is buying A100s in 2026 expecting to run them for 5 years. They are buying B200s, H200s, maybe AMD MI300Xs. Those are not $20k. He is citing last generation numbers expecting current-generation longevity, derailing the entire downstream calculation.\nInference is sold at list prices - the cited $4.50 per million tokens and up. As of 2026, every dev I know is on a fixed subscription, and a lot of us are happily spending 10x the tokens we could buy if we had to use the pay-as-you-go model. The author does mention this, but he still bases his profitability numbers on API pricing.\nCheaper chinese models are used as a proof point, under an assumtion that these are hosted under similar competitive conditions as U.S. ones. Really? You think those government-aligned labs are paying anything resembling market rates for water and electricity? You think they need to be profitable or do you think Beijing is willing to subsidize the losses for the appearance of AI supremacy and the chance to embarrass the West? China is not treating AI as some Walmart competitor. They are treating it as the Apollo program of the next millennium.\n\nI don’t have insider information about whether or not inference is profitable at scale, only the public numbers which look very running-at-a-loss-shaped. But the author claiming that “inference is obviously profitable” is somewhere between simply-not-supported-by-the-facts and completely delusional.","datePublished":"2026-06-26T14:43:37.483Z","url":"https://daily.dev/posts/4XqZoHtxz#c-IYENH494B","author":{"@type":"Person","name":"Jens Roland","url":"https://daily.dev/jensroland","image":"https://avatars.githubusercontent.com/u/210009?v=4"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":25}}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/ai-inference-is-obviously-profitable-4xqzohtxz#faq","mainEntity":[{"@type":"Question","name":"Is AI inference actually profitable or is it subsidized by investor money?","acceptedAnswer":{"@type":"Answer","text":"Inference is profitable on its own. Rough estimates put the cost of serving a 70B-class dense model at about one dollar per million output tokens, factoring in A100 power draw (400W, ~13 cents per million tokens) and amortized GPU cost over a five-year lifespan (~$1.80/hour). OpenAI charges $4.50 per million tokens for GPT-5.4-mini, implying a large profit margin on inference itself, even though the companies overall may be unprofitable due to training costs. Track the real economics behind AI inference pricing and margins on daily.dev."}},{"@type":"Question","name":"Why do OpenAI and Anthropic charge so much more than the estimated cost of inference?","acceptedAnswer":{"@type":"Answer","text":"Because inference margins have to subsidize training costs. AI labs use profits from serving existing models to fund the enormous capital investment required to train new frontier models, which is why inference margins run 70-80% even though actual serving costs are much lower. A standalone inference provider without training costs would not need such high margins. Compare AI provider pricing strategies and margins with daily.dev when evaluating API costs."}},{"@type":"Question","name":"How does DeepSeek's inference pricing compare to OpenAI or Anthropic and what does that reveal about actual serving costs?","acceptedAnswer":{"@type":"Answer","text":"DeepSeek reports over 80% profit margin on R1 inference while charging less than half of what OpenAI or Anthropic charge, and DeepSeek-V4-Pro sells on the open market for around 87 cents per million output tokens. Because DeepSeek's weights are open and competitors can undercut inflated pricing, that near-cost price suggests real inference costs are close to that figure, supporting lower cost estimates than raw GPU math alone implies. Follow open-weight model pricing trends like DeepSeek's on daily.dev to gauge true inference costs."}}]} ```