Microsoft turning away AI training workloads – inferencing makes better money Microsoft has explained that its method of funding the tens of billions it's spending on new datacenters and AI infrastructure is to shun customers who want to rent GPUs to train new AI models. CEO Satya Nadella revealed that strategy on Wednesday during the software behemoth's Q1 2025 earnings call, during which execs …
"We're not actually selling raw GPUs for other people to train," Nadella said. "In fact, that's sort of a business we turn away because we have so much demand on inference"
I'm twitching here... isn't that basically saying they don't want anyone to use their hardware to train a product that might compete with one of theirs? Not sure that's actually breaching any sort of law, but it's definitely on brand for Micros~1.
I'm sure that's just a side-effect of the policy...
It makes sense though. If you're charging more for "AI" software than just the cost of the server time alone - then why take the lower priced work of training other peoples' engines - when you can be selling the use of yours on the same hardware? More profit for you. This is particularly true if there are supply bottlenecks (that they complained about) in getting more lovely GPUs to cram into your datacentres. They're getting harder to build as well, because of high electricity and water demand.
Of course, if you see "AI" as a bubble, then surely the best thing is not to be selling the "AI" itself - but to take the lesson from the goldrush and go into selling picks and shovels? So you should be selling server time to everyone. But the problem with that is that when the bubble bursts - you're left holding the datacentres. So it also makes sense for Microsoft to try and scale their server capacity to the demand they temselves can sell in the long term - although they probably buy the hype and don't see it as a bubble.
Am I misreading something? They spent $20 billion this QUARTER on equipment that is forecast (fantasised?) to deliver a $10 billion ANNUAL revenue - not profit, revenue. Revenue is what you bill the customer, profit is what is left after accounting for the costs you incurred...
Assuming 10% profit (generous!) it will take 20 years to make enough profit to cover the costs of the stuff they bought in just 3 months
First, this explains why they are trying to bill everyone for stuff they don't want
Second, why don't they just buy all the GPUs that are no longer needed by the craptocurrency fantasists now they have moved on...oh, hang on, they businesses pumping AI ARE the ones with all those GPUs that were used to mine shitcoins....
Hmmmm
Everything these days suggests that AI is becoming CI (Cash Intelligence) ...
... but the older world is much better for so many people these days, my first ThinkPad laptop was only sold for $50 but I was offered $5,000 for my IMSAI 8080 with two 8" floppy drives. The big advantage of the old IMSAI environment was that anyone trying to steal gigabytes of data would need a big truck, that data is on all the shelves. We were so much safer in the old days, AI sees the old world as irrelevant.
I can imagine that the financial strategy from Micros~1 at the moment looks a lot like crossed fingers and repeating "oh please please please please please" quietly under their breath, hoping that the bubble doesn't burst. (spoiler: it probably won't matter whether it does or not, it looks like they might be losing a lot of money on this one)
Your suggestion of 10% profit and 20 years payback is nonsense.
If you ignore the costs of buying the PCs and building the data centers (the thing they spent $20b on), then the only real ongoing cost is power and a tiny amount of staff costs. On that accounting, almost all the revenue is "profit". Far more than 10%, more like 90%.
So 2 and a bit years to break even. Maybe 3 years at a stretch. Which is fine. The servers will probably last 5+ years, and the data center itself will last decades.
I don't think so. Those GPUs are not being made available to third parties for any purpose, not just training rival products, so in a sense MS just don't have a product on the market at this point.
You can't force a company to create a product just because you'd like to buy it.
That's the same day the news broke that VSCode will incorporate OpenAI, Anthropic, and Google as alternatives to MS' own co-pilot.
Conclusion: the news that inference, e.g. copilot for coding, is profitable is not true - it is still losing money at current subscription rates. If MS isn't making money at it, it is questionable whether OpenAI, Anthropic, and Google can either at the present of software and hardware technology. They can raise prices, and will, but that will cut demand, and therefore reduce advantages of economy of scale.
Having to adjust the level of service to make a profit is not a bad thing. In the long run, new improvements will depend on reducing power requirements and capital costs - as it should be.
The only downside would be if individual subscribers are charged at a much higher rate than corporate subscribers.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
