alexdoesstuff

Small nitpick: the models probably make some money on actual inference. Might not be a massive amount, but hard to see them not having a positive contribution margin purely on inference.

What's losing OpenAI money is paying for the whole of R&D, including training and staff. Microsoft doesn't pay that, so they get the money making part of AI without the associated costs.

It's unclear. That was never disclosed. It's similarly unclear what it means that they will no longer pay revenue share to OpenAI. Do they get the models for free now? How does OpenAI make money from the models hosted on Azure if not via revenue share?

It's kind of shocking, given financial transparency, that Microsoft gets away with not disclosing any details of this agreement (or the one it is replacing) to its shareholders. We know there's a cap on the revenue share from OpenAI to Microsoft, but we have no idea what that cap is (not whether it's higher, lower, or unchanged from the prior agreement).

We have no idea what it means to be the "primary cloud provider" and have the products made available "first on Azure". Does MSFT have new models exclusively for days, weeks, months, or years?

Both facts and more details from the agreement are quite frankly highly relevant to judge whether this is a net positive, negative or neutral for MSFT. It's unbelievable that the SEC doesn't force MSFT to publish at least an economic summary of the deal.

We updated our analysis of Meta's AI infrastructure spending after yesterday's Q4 call. Some specifics from the earnings call on where the money is going:

Ad models: They doubled the GPU cluster training their main ads ranking model. The architecture now reportedly "scales with similar efficiency as LLMs." They also consolidated multiple models into fewer, more capable ones (12% increase in ads quality) and moved their ads retrieval engine across NVIDIA, AMD, and custom MTIA chips, nearly tripling compute efficiency.

Internal tools: Management cited a 30% increase in output per engineer since early 2025, mostly from agentic coding tools. Power users saw 80% gains.

Ad products: Video generation tools hit a $10B combined revenue run rate. Incremental attribution drove a 24% lift in conversions.

On the financial side: Q4 ad revenue was $58.1B (+24% YoY). FY25 FCF was $43.5B, a 22% margin, down from 32% in FY24. They paused buybacks and signaled they may take on net debt. Management committed to FY26 operating income above FY25 in absolute terms against $162-169B in guided expenses.

For context, Meta's 2025 capex alone roughly equals the total lifetime investment in Reality Labs over 10+ years.

The article covers the full AI model stack (GEM, Lattice, Andromeda), capital intensity breakdown, and peer comparison.

This is primarily a story of a failure to supervise the creation of the report, rather than anything related to AI.

The role of the outsourced consultancy in such a project is to make sure the findings withstand public scrutiny. They clearly failed on this. It's quite shocking that the only consequence is a partial refund rather than a review of any current and future engagements with the consultancy due to poor performance.

There shouldn't be a meaningful difference if the error in the report is minor or consequential for the finding, or if it is introduced by poorly used AI or a caffeinated up consultant in a late-night session.

To expand on the overlooked point: it gives you a DB and a programming environment (however challenged) that you can use without needing sign-off from IT. In any moderately sizeable organization, getting approval to use anything but standard software is slow and painful.

Nobody wants to explain to IT that they need to install Python on their machine, or drivers for sqlite, or - god forbid - get a proper database. Because that requires sign-off from several people, a proper justification, and so on.

The $13bn investment in 2023 was so clearly structured to skirt antitrust concerns that it's unsurprising that that avenue is discussed.

Since then, MSFT has made other regulatory-aggressive investments, and the recent Meta / Scale AI is similarly aggressively designed.

Full agree!

Being close to the edge of AI usage, it's important to realize that most AI use cases are not "fully autonomous AI software engineer" or "deep research into a niche topic" but way more innocuous: Improve my blog post, what's the capital of France, what are some nice tourist sites to see around my next vacation destination.

For those non-edge use cases, costs are an issue, but so are inertia and switching costs. A big reason OpenAI and ChatGPT are so huge is that it's still their go-to model for all of these non-edge use cases as it's well known, well adopted, and quite frankly very efficiently priced.

Author here

I mostly agree on the first point. Even prior to the price race to the bottom, no AI Lab managed to make any money above marginal cost on inference, let alone recoup investment in infrastructure or model training. Clearly, investment in infrastructure and model training have been largely subsidized by VCs. It's a bit unclear how much of a subsidy inference costs had. The fact that AWS runs hosted inference at roughly similar cost than AI Labs suggests to me that there's at least not a massive subsidy going on at the moment.

I don't subscribe to the narrative that nation states (i.e. China) massively support DeepSeek. Thus, while their core business as a hedge fund is clearly profitable, they have considerably less deep pockets and willingness to front losses than the investors in VC supported AI Labs. Consequently, I expect their inference cost to at least cover their marginal costs (i.e. energy) and maybe some infrastructure investment.

All that suggests that they've managed to lower cost (and with that presumable resource and energy requirements) of inference considerable, which to me is a clear game changer.

Fully agree on the premise: there are X different ways to do anything on the web. But - prior to this - the solution seemed to be: everyone starts from scratch with some ad-hoc Regex, and plays a game of whackamole to cover the first n of the x different ways to do things.

Best of my knowledge there isn't anything more modern than Mozilla's readability and that's essentially a tool from the early 2010s.

Feels surprising that there isn't a modern best-in-class non-LLM alternative for this task. Even in the post, they described that they used a hodgepodge of headless Chrome, readability, lots of regex to create content-only HTML.

Best I can tell, everyone is doing something similar, only differing in the amount of custom situation regex being used.