Who Will Beat Anthropic

In 1995 the most famous software company on Earth was Netscape. 90% browser share. Their IPO was the moment people decided the internet was real.

Microsoft was the boring one. Everyone already owned it. Nobody was excited about it.

Five years later Netscape was worth almost nothing and Microsoft was the most valuable company in the world. The mindshare winner and the margin winner were not the same company. Most investors couldn’t tell those apart in 1995. Most investors can’t tell those apart in any cycle.

Which brings us to Anthropic.

Anthropic is the safest-feeling trade in AI right now. Family offices are pounding J.P. Morgan for pre-IPO placement. Wealth managers call it the IBM of AI. The mindshare is real. The revenue is real. The crowd is forming up around it.

The crowd is information about mindshare. Not about margin and the eventual winner

So who beats Anthropic at the layer that ends up mattering.

OpenAI is the obvious guess and probably wrong. They have the scale but they follow mindshare, they don’t set it. The number two in mindshare in cycle one is rarely the winner in cycle two. They’re like the Yahoo of this round.

Microsoft won’t beat Anthropic directly. Hyperscalers never have the model talent. They have the distribution. They’ll capture value somewhere in this stack, not by out-modeling the labs.

Meta is silent. Google is dunking infrastructure and hasn’t figured out the layer above the model. Both are too big to move at the speed this cycle rewards.

Which leaves Jensen.

Not because he has GPUs. Because he is the only operator on the field who reasons two clicks deeper than the question being asked.

He told this story himself. Nvidia built a billion-dollar mobile business during 3G. The numbers looked great. The customers were the right customers. Then Qualcomm shut them out during 3G to 4G because Qualcomm owned the modem, and the modem was the chokepoint, not the application processor. Jensen says, in his own words, he should have reasoned two more clicks before committing. The application processor was never the control point in a phone. The shutout was inevitable from day one. The billion dollars of revenue was false reinforcement. Revenue isn’t strategy. Revenue can mask a position with no terminal defensibility.

The mobile mistake is why Hopper exists. Pre-training was going to be a problem so large no existing supercomputer could touch it. Nvidia designed a multi-billion-dollar system with zero customers and built it on first principles. Two clicks deeper than the visible demand curve. Then the customers arrived. Then the customers became the entire AI industry.

He did it again for Grace Blackwell. Token generation needs aggregate memory bandwidth way beyond what one chip can provide, so they ganged up seventy-two of them and invented the rack-scale computer. He’s doing it again with Vera Rubin for agents. Agents need long memory in storage connected directly to the fabric, and extreme single-threaded CPU performance, because a multi-billion-dollar GPU system can’t sit waiting on slow CPU tool calls.

Each generation answers a question the rest of the market hasn’t asked yet. That’s the Jensen Loop. Observe, reason to first principles, ask so what, build the mental model, work backwards from the system that workload will need. The strategy books should retire OODA and put this in its place. OODA was built for dogfights. Fixed terrain, fixed game, faster reaction wins. That terrain is gone. The terrain now is one where the chokepoint moves every six months and the workload one generation ahead doesn’t exist yet. Reacting faster on the old map doesn’t help. The Jensen Loop is the loop for terrain that’s still forming. He doesn’t predict demand. He predicts the shape of the workload one generation ahead and builds the machine that workload will need.

The line he keeps repeating is the cleanest strategy frame I’ve heard from a sitting CEO. “The opportunity cost of pursuing a strategy is the real cost. So you have to ask yourself, how can you be smart enough such that the opportunity cost is reduced and your optionality is increased.”

Every yes is a no to something else. The discipline is making sure the no you’re saying is smaller than the yes.

Now look at Anthropic against that frame.

Anthropic ships Claude Code. Cursor’s revenue stops being revenue and becomes a timer. Revenue was the wrong dimension to fight on. Cursor optimized revenue and got eaten. The right dimension was position. The yes is owning the application layer. The no is being trusted by every other application that could have built on Anthropic. That is a big no. The platform predator trade may win the application layer for Anthropic. It forecloses the ecosystem trade that would have made them the cognitive operating system winner. Jensen would not make that trade. He says it directly. He wants multiple cloud providers and multiple token factories to coexist because he sells to all of them. He owns the substrate and refuses to compete with the layers it enables. The opportunity cost of eating his ecosystem is larger than the revenue of eating his ecosystem.

This is the deeper reason Dario won’t win the next cycle. Not that he banned OpenClaw. He reasoned one click and stopped. The one click is “AI is dangerous so we have to control it.” Defensible on its own. The second click was “and the way to make it safe is to make sure the layer above the model is open so the customer can audit and switch.” He didn’t go there. He went straight to control. Anthropic became the platform that wants to own everything because the original click never got tested.

There’s a second thing about Dario that goes against him. The future he paints is one people don’t want to build inside. Mass job loss. Civilizational risk. He may be right about the substance. He’s wrong about the rhetoric. People don’t rally around fear. Operators don’t move companies for a CEO who is afraid of the thing the company is building. Capital doesn’t compound for a decade behind a vision whose closing argument is “this might destroy us.”

Jensen’s future is also serious. He’s not pretending the transition is small. The future he paints is one where the world has more compute, more capability, more sovereignty, more agency. People want to build inside that. The best operators flow toward it. The best capital flows toward it. The best students choose it. The best people move toward the future they want to live in, years before the market figures out where they went.

Dario has the model. Jensen has the strategy method, the execution muscle, and the vision people want to be part of. Two of those compound. One doesn’t.

Back to the layer question.

Claude Code beats Cursor. Not because of the model. The same Sonnet runs inside both. The harness is what people pay for. The harness is what they switch back for.

SOTA models ship every couple of weeks. Inference price falls 10x a year. The model is the part that’s commoditizing. The harness is the part that isn’t. The harness is where the customer’s switching cost lives. The harness is where the margin will land.

Same pattern in every previous cycle. The database commoditized and the application captured the margin. The OS commoditized and the application captured the margin. The cloud commoditized and the SaaS application captured the margin. No obvious reason to think the model layer will be the exception. Obvious reason to think it won’t.

Enterprises will figure this out the way enterprises always figure these things out, which is too late and at great expense. A few will go bust running their entire cognitive surface through a vendor who ships the next model trained on their workflow. The survivors will want to own the house instead of renting the panopticon.

Jensen sees this and is positioning for it. Organically and inorganically, eighteen months early, while the rest of the market is still staring at the model layer. He doesn’t need to build the harness himself. He needs the harness alternatives to exist, on his stack, available to the enterprise, so the enterprise can run its software factory without renting it from Anthropic.

So who beats Anthropic.

Not directly. Nobody beats Anthropic at the model layer. They probably hold that for the cycle.

The cycle isn’t where the money is. The cycle after the cycle is. The chokepoint sits one layer above the model, in the harness, in the deployment standard, in the substrate that lets the enterprise own its own intelligence instead of leasing it. That layer is open. That layer is where Jensen will be when everyone else arrives.

Anthropic has momentum. Inertia carries it for another cycle.

The IBM of AI is a label, not a moat. IBM was the IBM of computing for thirty years, until the chokepoint moved to the operating system and Bill Gates inherited the decade. The chokepoint always moves. The person who inherits the next decade is the one who saw it move before the market did.

Bet on the operator who reasons two clicks deeper than the layer you can see.