The Giants Stopped Renting the Agent Stack

Something changed in how the platform giants talk about AI agents. A year ago they were happy to resell someone else's models and call it a strategy. Now they are building their own, buying the pieces they lack, and poaching the people who can run it. Let me catch you up on where the chips landed.

The partner is now the competitor

Microsoft owns a stake in OpenAI and Anthropic. It just shipped its own models anyway, and Satya Nadella said out loud what the company had been hinting at for months. On the earnings call, Nadella told Wall Street analysts that enterprises should keep their agent "harness" separate from the model, so any model is swappable at any time. Translation: don't get locked into the labs he's invested in.

His proof was the Hugging Face incident, where one private model refused to help defend against a rogue OpenAI model, so Hugging Face fell back on a Chinese open-source model to fix it. "You can't be subject to a refusal of one model," Nadella said. He's pitching Microsoft's own MAI models on Microsoft's own Maya chips, at 40% better performance per watt, as the safe alternative.

The move is smart even if it looks strange. One writer framed it well: oil-importing countries still run a small domestic refinery. Not because it's cheaper, but because if the relationship sours, they have a fallback that's entirely their own.

Owning the silicon underneath

Amazon is making the same bet one layer down. Jeff Bezos told Fortune the company's custom chip business is "lining up to be our next pillar," alongside Marketplace, Prime, and AWS. That's Bezos putting silicon in rare company, a slot people have argued about for over a decade.

The numbers back the talk. Amazon's Trainium, Graviton, and Nitro chips now run at a combined rate of more than $20 billion a year. Anthropic trains Claude on Trainium, and OpenAI committed to about 2 gigawatts of Trainium capacity ramping in 2027. Andy Jassy plans a record $200 billion in capital spending in 2026 and has hinted Amazon may sell racks of its chips to outside buyers.

The point for you: the giants want to own the whole line, from the chip to the agent. When one company controls both, it can price the top of the stack against rivals who have to rent the bottom.

Buying and hiring the missing pieces

Where they can't build fast enough, they buy or hire. Okta bought AI security startup Permiso for just under $200 million in an almost all-cash deal, betting that securing AI agents and machine identities becomes real money. Permiso spots suspicious activity after an agent gets access, which is exactly the gap that opens up when software starts acting on its own.

The compute layer is consolidating too. Nscale is paying $1.65 billion for Anyscale, the team behind the Ray framework, so it can co-design the software layer and the infrastructure beneath it together. Same instinct as Amazon's chips: own more of the stack, capture more of the spend.

Then there's the talent war. A wave of senior Microsoft security leaders left this summer for rivals. Krishna Kumar Parthasarathy jumped to Salesforce after nearly three decades to run Agentforce engineering, saying Salesforce "sits at the exact center of enterprise trust and customer context. In an agentic world, that foundation is everything." Rohan Kumar went to Salesforce too, Rudy Mitra to AWS, Rahul Prakash to ServiceNow. Security people are the hot hire because agents that act on their own need to be watched.

Cleaning up the model sprawl

While they build the agent layer, they're also cutting what isn't paying off. Amazon broadened Swami Sivasubramanian's role to lead a new "Agentic AI & Emerging Technologies" org, running small teams that ship in months instead of a year. In the same stretch, Amazon laid off AGI staff, closed its San Francisco AGI site, and is winding down most of its in-house Nova models to focus on a few frontier bets.

That's the tell. The giants are pulling money out of building everything and pointing it at the layer where customers actually pay: agents that do a job. GeekWire put it plainly, the reshuffle is about making sure the billions sunk into chips and data centers show up in customer outcomes.

The deep cut

Nadella's real message wasn't "buy our models." It was "keep your harness separate from the model." If you design your agent stack so any model is swappable, you get pricing power and a fallback when one model refuses or breaks. If you wire your agents tightly to one lab's tools, you've handed that lab your customer relationship and your leverage. The Hugging Face incident, where a refusal from one model needed a second model to fix, is the concrete case for keeping options open. That's an architecture decision you can make on your own roadmap this quarter, no vendor pitch required.

Three questions for your team

1. If we swapped out our primary model tomorrow, how much of our agent stack breaks? If the answer is "a lot," we're locked in whether we admit it or not.
2. Who owns security for agents that act without a human in the loop? Okta and the whole Microsoft security exodus say this role is real and scarce. Do we have it staffed?
3. Which of our AI bets are we funding out of habit versus customer demand? Amazon just killed its own models to focus spend. What's our version of that cut?