The Torch Has Passed

Anthropic filed for an IPO at a $965 billion valuation. OpenAI is preparing its own. Venture money that used to flow toward incremental improvement is now being placed on foundational bets. If you’ve been watching the AI industry long enough to remember when these were scrappy research labs with no clear revenue path, the scale of what’s happening now is genuinely difficult to process.

Most of the coverage treats this as a market story. It isn’t, or at least not primarily.

What the valuation numbers signal is something more fundamental: a generational shift in which companies get to define the technological infrastructure everyone else builds on. That kind of shift has happened before — most recently when Google, Amazon, Facebook, and Apple consolidated the internet layer — and the companies that emerged from it didn’t just win markets. They set the terms for every business that came after them. We’re at the same kind of inflection with AI, and it’s worth being clear about what that means and what it doesn’t.

A New Class of Foundational Companies

The torch has passed. Anthropic and OpenAI represent the first genuinely new class of foundational technology companies in roughly two decades. The models they’re building sit beneath everything else in ways that parallel what the database, the operating system, and the browser did in their eras. That’s real.

But the old guard — Microsoft, Google, Apple, Amazon — still controls something the new companies are only beginning to build: existing software relationships with virtually every enterprise on the planet. Google has bundled Gemini into every Workspace plan at no extra charge. Microsoft is folding Copilot permanently into Microsoft 365 bundles. Incumbent AI doesn’t have to win a procurement battle — it arrives as a line item in a contract enterprises already signed.

What makes this complicated is that the incumbents know they can’t build frontier capability on their own, and their responses reveal how differently each one has positioned itself.

• Microsoft bet early and heavily on OpenAI — Copilot runs on OpenAI’s models, which means one of the most widely deployed enterprise AI products in the world depends on a company Microsoft no longer has exclusive rights to.

• Google built Gemini in-house and made the most serious attempt among the incumbents — but real-world performance on knowledge work hasn’t matched the frontier labs.

• Apple is the starkest case: a company controlling a billion devices spent years promising Apple Intelligence features that didn’t work, then signed a $1 billion-a-year deal with Google to fill the gap.

Controlling the device layer didn’t solve the capability problem. Distribution gets you into every enterprise’s workflow. It doesn’t close the gap. The innovator’s dilemma isn’t theoretical anymore — it’s playing out in product announcements and earnings calls in real time.

The Moat Isn’t the Model

For senior leaders, the practical implication of all this is not “which foundation model should we use.” That question is less important than it sounds, and I want to be direct about why.

Today’s frontier models can handle tasks that would have required substantial teams of specialists just three years ago. But for the high-volume, well-specified work that makes up most enterprise knowledge work — research synthesis, first-draft document production, structured analysis — the frontier premium yields diminishing returns past a certain threshold. The race to build the most capable model is a different business from the race to build enterprise value.

The real moat is orchestration, institutional knowledge, integrations, and the surrounding system — not the frontier model itself. I run my Compound Agent OS on a combination of models, and the routing decisions — when to reach for frontier capability versus when a smaller, faster, cheaper model is sufficient — matter far more than the top-line quality of any single model. The AI arriving in your existing software contracts connects to your tools and your data. It does not connect to your judgment or your specific way of working. That gap is where the compounding advantage lives.

The first handoff is what the IPO filings signal. The torch passed from the incumbents to the foundation labs, and most senior leaders understood that one. But there’s a second handoff happening inside AI itself — from whoever holds the best model to whoever builds the best system around it. Most organizations are watching the first handoff. The second one is already underway.