For a few years the AI business had a tidy division of labour. Nvidia made the chips, a handful of labs made the models, the hyperscalers rented the cloud, and a swarm of startups built the apps on top. Monday was the day that neat picture fell apart in public. In the space of a few hours, a chipmaker became a model company's rival, a software giant announced it was building its own silicon, and a buzzy startup put itself up for sale. The investor Chamath Palihapitiya summed up the mood in a single line: "In AI, it is increasingly true that everyone competes with everyone."
The clearest signal came from Microsoft. Chief executive Satya Nadella called the company's expansion of its homegrown Azure Maia chips "our biggest bet yet on our own silicon," unveiling a new Maia 300 processor that Microsoft says is more energy-efficient than leading GPUs. From 2027, those chips are meant to become the main hardware running AI across Microsoft's products. For a company that has spent tens of billions buying Nvidia accelerators, building its own is both a hedge against supply and a quiet declaration of independence.
On the same day, Meta moved in the opposite direction along the stack. Mark Zuckerberg announced Muse Wonder and Muse Wonder Full, which he called the company's most advanced models yet, built to handle multi-step tasks like navigating insurance claims, reviewing leases and assembling working apps. Fourteen months after a sweeping reorganisation of its AI operations, Meta is pushing to the frontier with agentic systems aimed squarely at the kind of work other labs are chasing. The model race, briefly sleepy, is crowded again.
Palihapitiya's prediction follows naturally from all this: chipmakers will drift into software, and the companies that make AI applications will increasingly build their own chips and infrastructure. When everyone is moving into everyone else's territory, the interesting question becomes what is actually defensible. "Is it distribution? Is it a data moat? Is it a physical constraint like power?" he asked. Nobody has a settled answer, which is precisely why the land grab has turned frantic.
That frenzy is showing up in dealmaking. Decagon, an AI agent startup last valued at $4.5 billion, has hired a bank and held early acquisition talks with the cloud software giant ServiceNow, among others. It would be the latest in a run of deals: Nvidia recently scooped up Hugging Face for a reported $12.9 billion, and Stripe bought the model-routing service OpenRouter for around $8 billion. After years of raising ever-larger venture rounds, the hot AI companies are increasingly choosing to be bought, often at a steep premium, by incumbents desperate to own a piece of the market before their own software businesses erode.
The throughline is vertical integration and consolidation happening at the same time. Every major player wants to control more of the stack, from the chip to the model to the point of distribution, and the fastest way to get there is to build what you can and buy what you cannot. For customers, more competition across every layer should eventually mean lower prices and more choice. For the companies, it means the comfortable boundaries that defined their business are gone. In this market, your supplier, your platform and your upstart challenger may increasingly be the same company.