Anthropic is the darling of the tech world right now. We are also massive fans. Claude has been our house model for years, and we use Claude Code, Claude.ai, and Anthropic’s other tools across almost every part of our business.

That has led some people to believe Anthropic and the other frontier labs will end up replacing every other tech company. It is obvious to us that they will not, and here is why.

The case for the labs eating software is not silly. Old software was code that encoded the workflows of a business. New AI does the same job through skills, plugins, and agentic scaffolding. Either way, the work is encoding a workflow to get a job done. If a single platform ships the model, the skill format, the marketplace, and the orchestration, the application layer becomes a thin wrapper the lab eventually absorbs.

It is a serious argument, and people we respect make it. We think it is wrong, because it underestimates what it takes to make a model useful for a specific job.

The mustang and the harness

Tomasz Tunguz, the founder of Theory Ventures, recently shared an insightful analysis of why the apps-eat-the-labs case misses the point. In an essay on harnessing AI, he put it like this: the model is a powerful, unpredictable mustang. Harnessing the power means domestication. The harness is what turns wild capability into a specific repeatable outcome.

Tunguz counts seven distinct components, and the model is only one of them. The application has to feed the model the right context (the business’s own data, exceptions, history), wire in tools with permissions on anything sensitive, run an orchestration loop, persist state so a long task survives a crash, sandbox the model from what it should not touch, give a human the visibility to step in, and manage the cost of every call. Seven engineering problems, every one tailored to the specific job.

I run a huge part of SuperSeed in Claude Code. Anthropic built it as a brilliant coding harness for the work they understand best, and it is. I am also using it to manage a company, which is not what it was built for. It works, but only because I have spent months building a harness that business management needs: structures, rules, policies, files of prior work. The model is the same. The harness is built for coding. The seams are visible when I drag it elsewhere. The harness explains the friction.

That is the small version of the bigger picture: every industry that deploys AI will need a harness built for it, by someone who understands the business.

The number with no ceiling

The harness tells you why building the application is real work. Siddharth Ramakrishnan from ScaleVP recently shared a mental model for why that work keeps growing as the model improves, rather than shrinking toward nothing. A good application company does not wrap a model, it wraps a metric: it owns a number the customer cannot stop caring about, and turns raw model capability into a measurable improvement on it, faster resolution, higher conversion, lower cost per claim.

That decides who gets eaten. If a job is bounded enough to describe in a sentence and rebuild in a weekend, the model absorbs it; summarise this meeting, turn this thread into a ticket, and those become features, not companies. A metric with no ceiling behaves differently. Before VisiCalc, an analyst built one financial model by hand in pencil, and everyone assumed the spreadsheet would mean fewer analysts. Instead one model became twenty, and forty years later analysts are still chasing the same number. Cheaper work never satisfied the appetite for a better metric; it raised the ceiling.

This is what the doom case misses. The numbers software settles for today, 30% support resolution, a three-day procurement cycle, are not what customers want; they are what the cost of doing better has made acceptable. Drop that cost and they stop being good enough. The same force that kills the thin wrapper strengthens the company built around a metric, and every better model lets it push the number further. The model improves, and the application improves with it.

Anthropic loses most verticals

Dario Amodei has sequenced Anthropic’s product strategy with real discipline. In eighteen months they have shipped Claude Code, Claude Code for Work, Claude Design under the new Anthropic Labs banner, a verified legal plugin, the Skills marketplace, the MCP protocol, a multi-chip compute strategy, and, on 4 May 2026, a new enterprise services company. That is a lot of surface area, and unlike OpenAI, which is running in every direction at once, every piece of it fits a plan. If any lab were going to absorb the application layer, it would be this one.

It still will not be enough. The reason is structural, and it is visible in every trillion-dollar tech company in history.

Every substrate has tried this

In 1972, five engineers left IBM in Germany to build packaged enterprise software IBM did not want to pursue, and founded SAP. IBM was the dominant computer company on the planet, with the hardware, the operating systems, the customer relationships, and the brand; SAP started with five people. Half a century later SAP is one of the largest software companies in the world. IBM let the application layer go because it did not think it would amount to much, then spent the late 1980s and early 1990s trying to take it back. It built an applications business at real scale, but it was now competing with the independent software vendors whose products were the reason customers bought IBM in the first place. The ecosystem turned against it. When Lou Gerstner took over in 1993, one of his first moves was to shut the applications business down and restore the partnerships. The lesson, as he later wrote, was that a substrate that competes with its own ecosystem destroys the thing its value depends on.

Microsoft is the most instructive case, because Microsoft actually won one of these. Office took the productivity market from Lotus 1–2–3 and WordPerfect, the incumbents of the day, because Windows was already on every desk and Office rode on top of it. That is the exception that defines the rule. A substrate captures the layer above only where it confers a structural, compounding advantage in that specific market. Where it does not, it loses, and Microsoft has lost plenty: Dynamics, with the operating system, the cloud, the email client and the developer ecosystem all behind it, still sits at a tenth of Salesforce’s revenue, and Bing remains a rounding error next to Google after twenty-five years. Google and Amazon tell the same story: a few captured adjacencies, a graveyard of everything else.

The two forces

There are two reasons this repeats. The first is internal. Running a substrate business and a deep vertical business inside one company creates a conflict for management attention, capital, and prioritisation. The substrate side wins, because it is what defines the company; the verticals get less than they would as standalone businesses and lose to focused competitors. The market calls this the conglomerate discount.

The second is external. A substrate’s value depends on the ecosystem of applications built on top of it, and a substrate that competes with its own ecosystem destroys the thing that makes it valuable. Gerstner learned this the hard way. Anthropic appears to have learned it without having to. The Skills marketplace, the MCP protocol, the plugin architecture, and the decision to put the new services company outside the core business are all the moves of a company that wants the application layer above it to thrive. The services company is the tell: Anthropic did not build a mid-market services arm in-house, it spun up a separate entity backed by Blackstone, Goldman, and a roster of other private-equity and banking giants. That is the cap table of a services business, not a frontier lab.

For the labs, that compounding advantage lives in exactly one place, and Siddharth names it precisely: the verticals where customer use loops back into the model’s own improvement. Every hour Anthropic spends making Claude Code better also makes Anthropic better at building Anthropic; the work compounds inside the lab. Resolving a bank’s billing tickets does not. So the labs go absurdly deep in coding, evals, agent infrastructure, and research workflows, and stay shallow everywhere else, not because they cannot go deep, but because the same engineers are worth more pointed at work that compounds. It is comparative advantage, the oldest argument in economics: the lab can be better at almost everything and should still do only the few things worth more to it than to anyone else.

The early evidence fits. Coding is the vertical Anthropic has gone deepest in, and the one that produced the fastest-scaling software company anyone has measured. Cursor went from a $29bn valuation in November 2025 to a $60bn deal five months later. The labs building aggressively in coding did not crush it; it grew faster than any of them, and when a substrate giant finally moved to capture that value, SpaceX agreed to buy it for $60bn rather than out-build it. The substrate could not displace the application, so it paid for it. Anthropic wins where the work loops back into the model, coding, consumer chat, agent infrastructure, perhaps one or two more; everywhere else it is the supplier, and the application companies own the customer. That held at IBM, Microsoft, Google, and Amazon, and a smarter model does not change it.

Physical AI will not be different

The same dynamic is already playing out one layer up. The next substrate after the language model is the world model, which learns from sensor data and gives a robot or a vehicle a working internal picture of its environment. The leading world-model labs are not the language labs: Yann LeCun left Meta in 2025 to start AMI Labs on a single conviction, “Real intelligence does not start in language. It starts in the world”; Fei-Fei Li founded World Labs to build spatial intelligence; Physical Intelligence is building foundation models for general-purpose robots. Three of the most credentialed teams in AI, all betting that the next substrate is a model of the physical world rather than of text. They will be superb substrates. They will not be all of Physical AI.

A world model is still just a model. It does not design the gripper that lifts a panel off the line, certify a surgical robot, write the safety case for an autonomous tractor, or take the call when a warehouse fleet stops at 3am. Every Physical AI vertical needs its own harness, built by people who understand the job, which is why the field fractures into specialists: Wayve builds its own world model for driving, and that depth is exactly why it will never also build humanoid manipulation or surgical robots. The world-model labs will supply some of these companies, the language labs others, and the application companies will own what the labs cannot, the customer, the regulatory standing, the accountability when something breaks. The same split, one layer up, against the same two forces.

Where we put our money

The labs have never been more powerful, and the application layer has never been bigger. Both things are true at once. Customers buy outcomes, not models, and outcomes are built by companies that know the customer better than the lab ever can. That does not change when the model gets smarter, and it does not change when the substrate gets bigger. The application layer is where the work is, and it is where we invest.

So be a fan of Anthropic. We are. Just don’t mistake the engine for the car.