The AI industry keeps pretending this is a story about genius models fighting it out in a fair arena. Chat, image, video, agents, whatever. Very cinematic. Very investor-deck. But if you look at what’s actually happening, the real fight is lower in the stack and way less glamorous.
It’s about who owns the chips, who rents the cloud, who controls the app surface, and who gets to quietly steer users toward their own products once everyone is trapped in the ecosystem.
That’s why the interesting story this week is not some breathless new demo reel. It’s that EU antitrust chief Teresa Ribera is looking beyond chatbots and into the entire AI stack — models, training data, and the cloud pipes underneath them. Reuters reports that she met with the CEOs of Google, Meta, OpenAI, and Amazon while warning that dominant firms may be using their own platforms to favor their AI services and squeeze rivals out before those rivals get oxygen. Good. Because chat is this real: if the same handful of companies own the hardware, the hosting, the distribution, and the defaults, you do not have an open market. You have a toll road.
And once you see it that way, a bunch of other headlines suddenly line up.
Nvidia’s Jensen Huang, fresh off another stadium-scale keynote, is selling the vision that every company now needs an agent strategy. He’s probably right about that part. The next interface really may be software that acts instead of merely answering. But the other half of the pitch matters too: Nvidia wants to remain the layer everybody depends on while making it sound like dependency is freedom. That’s not evil. It’s just business. Still, let’s not do the corporate cope routine. If one company becomes the indispensable substrate for every ambitious AI deployment, the future gets expensive and fragile real fast.
Then you’ve got Alibaba launching Accio Work, pitching AI taskforces for small and medium-sized businesses. That story is actually more useful than most flashy AI coverage because it reveals where this market is going: away from parlor tricks and toward permissions, workflows, and operational reliability. Reuters noted that Alibaba emphasized guardrails around payments and private files. Exactly. That’s the real product. Not vibes. Not “wow it wrote a sonnet.” The product is whether an agent can do meaningful work without accidentally emptying your checking account or emailing your confidential docs to half the internet.
Which gets to the part the market still doesn’t want to admit: the best AI companies of the next few years may not be the ones with the most theatrical demos. They may be the ones that build boring, sturdy permission systems and refuse to let autonomous tools turn into chaos goblins.
That’s also why the antitrust question matters so much. If the biggest firms can bundle the model, the cloud, the compliance story, and the built-in distribution, then startups aren’t really competing on product quality. They’re competing against gravity. Against procurement defaults. Against enterprise lock-in. Against “nobody ever got fired for buying the incumbent.” Legacy tech has been running that scam for decades. AI just gives it better branding.
And yes, there’s a broader political layer here too. Export controls, licensing, safety rules, and data governance are all becoming hidden product features. When governments shape who gets compute and under what terms, regulation stops being some external force and starts becoming part of the architecture. That can be necessary. It can also calcify markets if we let incumbent-friendly compliance become another moat.
So what would actually work?
Simple. Separate layers of power as much as possible. Encourage open standards. Make portability normal instead of painful. Treat cloud switching, model interoperability, and permission transparency like first-class competitive issues. If a customer wants to move an agent workflow from one provider to another, that should be annoying at worst, not a six-month hostage negotiation.
The useful future is still on the table. Agents really could help small teams do serious work. Better tooling really could lower the cost of building companies, products, and weird ambitious things. But only if we resist the lazy version of the future where five firms own the rails and everyone else rents ambition from them.
AI doesn’t need more magic. It needs less gatekeeping.
Sources: Reuters on EU AI antitrust scrutiny (