Technology

nvidia, microsoft, google, oracle, etc are all profitable

• NVIDIA is a hugely profitable hardware company. Not an AI model provider. They're selling the shovels for the gold rush. And their profits will tank as soon as one of the model providers fold, though pivoting back to consumer hardware is definitely an option for them.
• Microsoft is a hugely profitable business and consumer software company. Again, not an AI model provider. They have trained a couple models, but have made no profit on serving them. Their add-on "Copilot" services aren't profitable (hence the recent enormous price hikes for GitHub Copilot, which has resulted in a bunch of companies scaling back their AI usage). All of Microsoft's profit comes from software sales.
• Google is a hugely profitable software company as well - but again, not with AI. Their model, Gemini, isn't remotely profitable, and has similar costs to maintain as Anthropic and OpenAI see for their models. Heck, they just did a fundraising round for the first time in years to support it, which is NOT a sign of a healthy AI business.
• Oracle is also profitable in software, but they've staked their company on hardware roll-outs supporting data centers for OpenAI. They've booked future profit from OpenAI owing them almost a trillion dollars. If OpenAI can't pay that bill when the time comes, Oracle is completely fucked.

inference is profitable, training is not

When I don't include the costs of doing business, my business is profitable! That's silly. Inference might be very slightly in the green now when viewed by itself (although that's deeply questionable; no actual GAAP accounting has shown it to be so). But since training is an ongoing expense that frontier model providers have to constantly engage in, their companies are - and will remain - very deeply in the red.

And without seeing GAAP accounting showing where all the money goes in support of inference, I am highly doubtful that it's profitable.

training is what the majority of data centre spend is going towards

if they want to be profitable pull back the training but right now they are competing for market share

They can't. Ever. Pulling back on training means allowing model drift. You need to understand that models are obsolete the moment they're released. Their training data is set in stone. New version of Typescript ships? Some celebrity dies? Big election happens? The model not only doesn't know about any of it, it can't be updated. The best you can manage is throwing MCP and RAG at it in the hopes that the model will pay attention to it, but the point of diminishing returns on that arrives almost instantaneously. You have to train. Constantly.

feel free to look back at all the times lemmy predicted the end of spotify because it wasn’t profitable, now they turn around and cry it’s making money

Bad comparison. Spotify has already been a profitable, publicly-traded company for years.

And - this part's important - I'm not Lemmy. The platform we're having this conversation on has nothing to do with whether or not the AI model providers are profitable.

At work nobody is talking like this, everyone is talking about claude and it makes sense, it’s the best thing since vscode

Anecdotes aren't data. But as long as we're swapping anecdotes, here's mine: I work with actual machine-learning engineers. They're the ones who bag on Anthropic and OpenAI the most. And they use Qwen, Gemma, and a few other small, open-source, open-weight models. Have you looked at Hugging Face? Its community is huge, and growing daily. No one wants to be locked in to Claude Code or any other proprietary development tool when the service has been unstable and the pricing has becoming ridiculous in their desperate attempts to become profitable.

The cost for using Qwen tokens is $0, no matter how many tokens they use.

You say no one talks like this... Are you sure you're listening?