DeepSeek says its V3 and R1 models' cost of inferencing relative to sales during a 24-hour-period on February 28 put “theoretical” profit margins at 545%

Chinese artificial intelligence phenomenon DeepSeek revealed some financial numbers on Saturday, saying its “theoretical” …

Bloomberg 2025-03-02 Saritha Rai

Context & Ripple Effects

Earlier coverage positioned DeepSeek around an efficiency claim for its open-source V3 model, emphasizing training with fewer chips. This disclosure extends that narrative from model creation to the cost of serving users.

The stated figure is explicitly theoretical and based on a single 24-hour period, so it is a signal about claimed inference economics rather than a full view of sustainable profitability.

First-order effects

DeepSeek gains a public unit-economics claim for V3 and R1 that can support its positioning with users and prospective commercial partners.
The disclosure puts attention on the gap between revenue and inference expense for these models, while leaving other operating costs and the durability of the result unaddressed.

Second-order effects

Rival model providers and customers have a sharper incentive to compare serving costs, not just benchmark performance, when assessing model offerings.
Low claimed serving costs could create room for lower tool prices or greater investment in distribution; DeepSeek later cut pricing alongside a new attention technique, illustrating how efficiency and pricing can reinforce one another.

Third-order effects

If such economics prove repeatable at scale, AI competition could increasingly turn on inference cost of goods sold and the ability to translate model efficiency into price or margin advantage.
The more durable test will be whether disclosed costs remain competitive as models, hardware compatibility, and usage volumes change; DeepSeek later described a model customized for next-generation Chinese-made chips, linking model economics to deployment hardware.

The trend: This is one data point in the shift from headline training costs toward inference economics as the commercial measure of AI models.

Discussion

@timkellogg.me Tim Kellogg on bluesky
DeepSeek did the “one more thing” 🙄 — but guys, check this out, they go into detail on how they run inference on V3/R1, how they partition the experts across lots of nodes and pipeline attention and.. — just read this 🤯 — github.com/deepseek-ai/...
@trashfirefurry @trashfirefurry on bluesky
Sure Chinese companies have a rational plan to become profitable before it causes a global economic crash but the nerd guy in a sweater is telling me that if we just duct tape fossil fuel power plants attached to data centers like we are playing Factorio they will maybe make a pr…
@kwerb.com Kevin Werbach on bluesky
I was going to say that DeepSeek making $200 million/year on >500% profit margins is almost as big a deal as its model performance to training cost improvement. — Then I saw that's “theoretical.” — www.bloomberg.com/news/article...
@sbyrnes Sean Byrnes on bluesky
If “theoretical” profits are now a legitimate measure, the value of AI companies are now measured in the Quintillions of dollars. — Also, I have a bridge to sell you. [embedded post]
@carnage4life Dare Obasanjo on threads
DeepSeek published an an article about how their inference architecture works and ended it by saying if not for off peak discounts and not charging for their app, their profit margins would be 545%. This is a rather bogus argument given that usage wouldn't exist if people had to…
@deepseek_ai @deepseek_ai on x
🚀 Day 6 of #OpenSourceWeek: One More Thing - DeepSeek-V3/R1 Inference System Overview Optimized throughput and latency via: 🔧 Cross-node EP-powered batch scaling 🔄 Computation-communication overlap ⚖️ Load balancing Statistics of DeepSeek's Online Service: ⚡ 73.7k/14.8k
@doodlestein Jeffrey Emanuel on x
OK, so now we know just how much more efficient DeepSeek is for inference in terms of total tokens per second processed. They are doing roughly 7-8x more tokens per second on an H-800 (a crippled, export-control version on the H-100) than the open-source state of the art on H-100
@deedydas Deedy on x
BREAKING DeepSeek just let the world know they make $200M/yr at 500%+ profit margin. Revenue (/day): $562k Cost (/day): $87k Revenue (/yr): ~$205M This is all while charging $2.19/M tokens on R1, ~25x less than OpenAI o1. If this was in the US, this would be a >$10B company. [ima…
@therealadamg @therealadamg on x
Let's not repeat misunderstandings of math again. It's an 84% profit margin. Impressive, but not 545%.
@ns123abc Nik on x
Sam Altman: Please DeepSeek no more releases I need people to pay $150 for 1m tokens DeepSeek: Here's our 545% profit margin inference code [image]
@zephyr_z9 @zephyr_z9 on x
Wenfeng is devious China can meet all its AI demands in less than 250K GPUs And yes Jevon's paradox got fucked in the ass [image]
@zephyr_z9 @zephyr_z9 on x
The reason why I'm saying all this is becuz the whale is processing 600B tokens & outputting 150B tokens per day on 300 H800 nodes (2400 H800s) 100x (240K) will get u 60T tokens & output of 15T tokens per day The world does not have this high AI demand
@modestproposal1 @modestproposal1 on x
“They have 545% margins”
@dorialexander Alexander Doria on x
Some EU influential person just posted on LinkedIn: “Although Deepseek claims to have trained R1 for a bare $5m (...) some analysts claim that this might only be the cost of computational capacity used for the training excluding research and personnel” => they literally said so.
@bindureddy Bindu Reddy on x
In 6 months, we will go from $150 to $1.50 per 1M tokens - pure inference efficiency will increase by 700% - model sizes will decrease by 1000% - test time compute will become 1000% more efficient for 95% of queries Frontier tech moves at lightning speed
@theahmadosman Ahmad on x
I love DeepSeek. The secret sauce was delivered to our doors this week. They wiped billions of dollars of future revenue for a lot of (predatory) companies.
@ai_for_success AshutoshShrivastava on x
What the heck, DeepSeek's cost-profit margin is 545%. They don't charge $200, and everything is open source. Just imagine what DeepSeek would do if they got $500B. Respect the 🐋.
@hajekd David Hajek on x
Combined peak node occupancy for V3 and R1 inference services reached 278, with an average occupancy of 226.75 nodes (each node contains 8 H800 GPUs). Thats not “thousands” GPUs at all.
@tphuang @tphuang on x
Impressive. DeepSeek achieves maximum utilization of its H800 nodes during 24 hr period. Maximizing cache hit to deliver more for less -> can achieve theoretical margin of 545%! Even accounting for free web/app access + nighttime discount, DS still probably profitable. [image]
@vllm_project @vllm_project on x
Amazing system! It is now the north star for LLM inference 🌟. We will get there, quickly.
@dl_insider Jose Lopez on x
Amazing sharing by DeepSeek. They are truly sharing everything. This hit below the waterline of every company whose moat was to serve an LLM efficiently. [image]
@niklas_sikorra Niklas Sikorra on x
That means on a world scale: At 7.5 Trillion tokens / day or 86.8 mn tokens per second 1,178 GPUs would be needed or a total investment of $100mn to serve the world marked? Is this correct?
@madiator Mahesh Sathiamoorthy on x
Everyone in the US is like.. “we don't know how to serve DeepSeek-R1” and “it's too hard and its so expensive”.. and these guys have a profit margin of 545%. WOW. WOW. WOW
@bytebot Colin Charles on x
How is DeepSeek making so much theoretical profits? 545% is amazing, in a time when AI companies are burning so much cash. Out on Bloomberg now too [image]
@jiayi_pirate Jiayi Pan on x
It's truly inspiring to see how a small, sincere, and talented team can shake up the entire world, even in an industry as fiercely competitive as AI 成事在人
@hanchunglee Han on x
deepseek dropped over $100b valuations of work over its open source week. insane.
@valmianski Ilya Valmianski on x
This is insane. Deepseek v3 is cheaper than gpt4o and still has like 80% margin, on an inferior H800 node!
@oyattia Omar Attia on x
They must be lying bro. They're a CCP psyop bro. Believe me bro I invested in 20 SaaS AI companies and I need this to be true or my LPs will be furious at me bro.
@junxian_he Junxian He on x
They even release the details of profit and cost of their deployed system. New level of openness
@jobergum Jo Kristian Bergum on x
What a week of open source release from @deepseek_ai . Lots of novel AI infra. From a distributed file system to how they scale inference with a «few» hundred H800. [image]
@tariqrauf Tariq Rauf on x
you can now have a colossus clusters inference output at 20% of the cost v/s last night human ingenuity knows no bounds, especially when resource constrained
@adamlogs @adamlogs on x
It's like a new big guy coming to the street : YO, LISTEN UP, YOU LOT! DEEPSEEK JUST DROPPED A TASTE OF THEIR NEXT BIG HIT, AND I'M CALLIN' IT—THEY'RE PACKIN' HEAT! THIS AIN'T NO WEAK STUFF; IT'S A FISTFUL OF AI POWER! STEP UP, ANY OF YA—BRING YOUR BEST, 'CAUSE I'M READY TO
@glennluk Glenn on x
What's more likely? The company that pushes out its methods onto open source is lying about its compute needs Or research analyst guessed wrong on the number of chips they inputed into an excel table
@gavinsbaker Gavin Baker on x
The endless quote tweets commenting on the “545% profit margin” are quite funny. Reminds me of 2022 when there was a viral interview with seed stage VCs who said “we are trying to get really smart, really fast on what exactly gross margins are” or something to this effect.
@jacquesthibs Jacques on x
Unreal...they just...they just released it “We hope this week's insights offer value to the community and contribute to our shared AGI goals.” “⚡ 73.7k/14.8k input/output tokens per second per H800 node 🚀 Cost profit margin 545%”
@nembal Balázs Némethi on x
Deepseek is both Silicon Valley's dream child and its worst nightmare. It operates at a staggering 515% margin, with revenue that would justify a $10 billion valuation—while simultaneously releasing open-source tooling and foundational models at a level that would typically
@robinzhong42 Robin Zhong on x
Amazing! DeepSeek's efficiency inference service is top-notch, delivering great performance and profitability: ⚡ 73.7k/14.8k input/output tokens per second per H800 node 🚀 Cost profit margin 545% They're incredible! It reminds me of the Google talents at the beginning, where [ima…
@longtonylian Long Lian on x
Quote: “If all tokens were billed at DeepSeek-R1's pricing (*), the total daily revenue would be $562,027, with a cost profit margin of 545%.” [image]
@stuartreid1929 Stuart Reid on x
Holy shit. Ngl the short thesis on Nvidia is starting to make sense. Why invest 10x on CAPEX when there are 10x software gains seemingly around every corner!
@mvvvqv @mvvvqv on x
545% margin if all requests are for R1 [image]
@yuchenj_uw Yuchen Jin on x
holy shit, DeepSeek is able to get 73.7k tokens/s input and 14.8k tokens/s output throughput per H800 node! Their profit margin is 545%, while OpenAI is bleeding money despite charging so much?? that's how awesome their inference stack is. [image]
@kimmonismus @kimmonismus on x
DeepSeek has a cost profit margin of over 500%! What the! Holy moly [image]
@oyattia Omar Attia on x
What deepseek is releasing for free is enough to build a $500M startup, maybe more. Just out there for free. That's why the VC bros were freaking out. This is not normal.
@bindureddy Bindu Reddy on x
DEEPSEEK CATEGORICALLY PROVES WE DON'T NEED ALL THOSE GPUs OpenAI - GPT-4.5 - $150 per 1M tokens - profit margin - 0% Deepseek - R1 - $2 per input token - profit margin - 545% DeepSeek is almost 300x more efficient!! 🤯
@dorialexander Alexander Doria on x
In one year we went from will China catch up? to well the US catch up?

Chronicles