Mature AI infrastructure has months left as a moat

The companies with the most advanced AI infrastructure on the planet just stopped competing on it.

OpenAI and Anthropic, the two labs at the front of the model race, spent two years building the deepest technical stacks anyone has assembled. This spring both pointed their next big bets somewhere else. In May, OpenAI stood up a separate deployment company worth more than $4 billion, built around buying the consultancy Tomoro and its roughly 150 forward-deployed engineers, the people who embed inside a customer’s organization to get the software actually used, according to The Next Web’s report on the deal. Anthropic launched a $1.5 billion joint venture with Blackstone, Hellman & Friedman, and Goldman on the same model. Neither move buys better infrastructure. Both buy the muscle to get someone to use what the infrastructure already does.

If you tuned out the AI plumbing news this quarter, here’s the whole of it. “AI infrastructure” is what it takes to train and run a model: rented GPU capacity (the specialized chips that do the heavy math), the pipelines that push a trained model into production, and the serving software that answers fast enough to be usable. Assembling that used to be hard and expensive, so having it looked like an edge. That belief is still load-bearing in a lot of strategy decks: we built the mature stack, that’s our moat.

That moat has a short expiry date now. Mature AI infrastructure has commoditized into table stakes, and the durable advantage moved one layer up, to how a company is organized and how it gets customers to adopt, rather than the stack it runs.

Look at how cheap the hard parts got. Running a model used to mean paying for DevOps staff to keep GPUs warm. Now Modal Labs offers serverless GPUs, rent-by-the-second compute with no servers to manage, that start up in under a second, and RunPod does it in under 200 milliseconds. The serving software that wrings speed out of a model is off the shelf: vLLM improves throughput 2 to 24 times over a naive setup, and TensorRT-LLM pushes past 10,000 tokens a second on a single high-end chip. Getting a new model into production, once a months-long job, now takes weeks on standardized pipelines. And the raw capacity is a commodity you sign for: OpenAI holds $22.4 billion in contracts with the cloud provider CoreWeave, Meta committed $35.2 billion, and Anthropic signed its own multi-year deal this April. Everyone buys from the same shelf.

The account pulling these threads together comes from the Engineering & Leadership newsletter’s July analysis of AI infrastructure patterns, and its conclusion is blunt: the stack has standardized to the point where every serious lab runs roughly the same one. When the plumbing is identical, it can’t be the thing that separates you.

Now the part that should change how you read a competitor. Six weeks ago it was reasonable to think mature infrastructure was the emerging edge, the ground the frontier was moving toward. If you worried at all, you worried the bar was too high: maturity would stay scarce, and whoever cleared it first would be hard to catch. The evidence went the other way, and the direction matters. The bar didn’t rise out of reach. It fell through the floor. Instead of staying scarce, maturity became nearly free.

That is worse for the “we built the stack” thesis than the scary version would have been. A high bar still rewards whoever clears it. A free bar rewards no one. It just moves the contest somewhere else. And where it moved is telling: to the org chart. The labs that could out-build anyone on infrastructure are spending billions to stand up deployment organizations instead, because the scarce thing is no longer running the model. It’s getting a customer to change how they work so the model earns its keep. That capability lives in people embedded in relationships, and you can’t buy it off CoreWeave’s shelf.

The strongest objection is that this only holds below the frontier, and at the very top it does. OpenAI’s $500 billion Stargate data-center buildout and its move into custom Broadcom chips are real, verified bets that proprietary infrastructure still differentiates when you’re training the largest models on Earth. If you’re one of three companies operating at that scale, the stack is still a weapon. But almost nobody is. For every company at production scale rather than frontier scale, which is nearly all of them, including nearly all of your competitors, the stack is a catalog order. And the frontier labs themselves just told you where they think the next advantage sits, by where they put their money.

So the practical read for a product person is a downgrade in what “they have great AI infrastructure” earns. It’s a table-stakes claim with a short shelf life. Respecting it as a durable moat is how you overrate a rival who’s standing on ground that’s turning to commodity under them. What survives that downgrade is the org-shaped stuff that doesn’t transfer: the platform team that lets small product teams ship without rebuilding the plumbing each time, and the deployment muscle that gets customers to actually adopt.

Which leaves the sharper question, and I don’t think it’s settled. If infrastructure maturity commoditizes this fast, is any capability-based advantage in the AI stack durable? Or is the only defensible layer the org-shaped one, precisely because you can’t replicate it by buying the same tools? I lean toward the second, but that’s a bet, not a finding. Every “we have the better stack” moat I can see right now has the same short half-life. If that’s the rule and not the exception, then the real question about any AI edge, yours or a competitor’s, is how long before everyone can rent it.