Dan Niles — "You need to have a viable open source community"

Niles tests cloud businesses on two things at once, and Google Cloud is the only one that cleared both by a wide margin. The first test is whether growth is speeding up rather than merely fast — its revenue growth accelerated by 19 percentage points, and it is now growing 82%. The second is whether the profit margin is rising while that happens — margins expanded 3 points. "You're seeing both growth and profitability at that layer. And I really like that."

Why both matter together: a cloud provider can always buy faster growth by cutting prices or absorbing the cost of the chips, and the margin is where that shows up. Growth that accelerates while the margin widens means customers are paying up for capacity that is genuinely scarce — the supplier has pricing power, not just volume.

It also sits on the right side of his bigger argument. Google runs one of the platforms that will automatically send each customer request to whichever model is cheapest and good enough. The customer never picks a model vendor; the platform does. That makes the platform the toll booth and the model a substitutable input — and it is a business Google owns end to end, chips included.

One caveat he raises is about timing, not the business: Google reported roughly a week before the market overhang he was waiting on cleared, so the print landed into a worse tape than it deserved.

AWS is Amazon's cloud-computing arm — it rents out servers, storage and, increasingly, AI capacity. It is one of the "infrastructure players" Niles says the money ends up with, on the logic that if AI models become interchangeable, the profit moves to whoever owns the computers the models run on.

The June quarter passed his two-part test. Growth accelerated by 9 percentage points versus the March quarter, to 37% — the pace picked up rather than just staying high. And, in his words the more important half, "the operating margins also expanded by a percent." Selling more while keeping a bigger slice of each dollar is the combination that says demand is outrunning supply.

The plain reading: AWS is being paid more for capacity it is also selling more of, which is exactly what you would expect if the binding constraint in AI has moved from clever models to raw compute.

Azure is Microsoft's cloud platform, and it shows the same shape as AWS: growth accelerated about 4 percentage points to 43%, with operating margins up a point. Same conclusion — faster and more profitable at the same time.

Azure has a second role in his argument, and it is the more interesting one. He names it as one of the services that will "route whatever you're trying to do to the best model available for the task" — meaning the platform quietly decides, request by request, whether your question goes to an expensive frontier model or a cheap open-source one. The customer just gets an answer.

That routing layer is what turns the individual AI labs into replaceable suppliers. Whoever controls the routing keeps the customer relationship and the pricing power; the model behind it becomes a component you swap out the moment a cheaper one is good enough. Microsoft owns one of the few such control points.

Anthropic is the private AI lab behind the Claude models, sold mainly to businesses. Asked whether its private valuation — and OpenAI's — is too high, Niles refuses to argue about the price and argues about the position instead.

His case: open-source models cut the cost of producing AI output by roughly 90%, and most of what companies actually ask for is easy. "You don't need a Ferrari to go to the corner store to get milk." Once cloud platforms automatically send the easy work to the cheap model and only the hard work to the frontier, the frontier labs lose the routine volume that would have made them enormously profitable — "Anthropic and OpenAI become more commoditized over time, and the value accrues to more of the infrastructure players."

Read this as a bet against the layer, not against the company. Four days earlier he named Anthropic the corporate-side winner of AI on hard financial evidence — profitability and an unprecedented revenue ramp — and nothing here contradicts that. What he is saying is that being the best model vendor is worth less than owning the platform that decides which model gets used.

Anthropic also appears in his governance warning: it sat on Figma's board, resigned, and shipped a competing product about a week later. His conclusion for corporate customers — keep your proprietary data proprietary — is a caution about handing your data to any model vendor, this one included.

OpenAI, the maker of ChatGPT, gets the identical verdict by the identical mechanism. In a world where "90% of the companies aren't using the leading edge models and they're switching to using open source models," the volume that would justify a frontier lab's valuation goes to somebody cheaper for everything that is not genuinely hard.

The frontier keeps the hard end — "if you want to design a new database, then yes, you need their highest end models for that." But the trivial end, which is most of the requests, drains away to free or near-free alternatives: "you don't need them for what's three plus three."

What makes the drain automatic rather than a customer decision is the routing done by Azure and Google Cloud. Nobody has to actively switch away from OpenAI; the platform simply stops sending it the easy work. That is what commoditization looks like in practice — you keep the prestige jobs and lose the volume, and the margin lives in the volume.

Figma makes collaborative design software. It appears here only as a cautionary tale, with no view expressed on the stock — but the lesson is the reason Niles gives for his whole data-governance position.

The episode: Anthropic held a seat on Figma's board, stepped off the board, and roughly a week later launched a product competing with Figma. Whether or not anything improper occurred, the optics are the point — a board seat carries deep visibility into a company's roadmap, customers and internals.

The generalisation he draws for every corporate buyer of AI: your model vendor may become your competitor, so keep "all their proprietary data proprietary and not share it, because that's where problems get created." This is also, quietly, part of the bull case for open-source models — a model you run yourself never sends your data anywhere.