The Price of Know-how

In July 2025 Google paid about $2.4 billion to Windsurf, a startup that made an AI coding tool. It didn't buy the company. It got a non-exclusive license to the technology, which Windsurf could keep selling to others, and it hired the chief executive and part of the research team. The customers, the product and roughly $82 million of annual revenue stayed behind, and within days another startup, Cognition, bought all of that.

So Google paid about 29 times Windsurf's revenue and didn't take the revenue. It paid for people and what they knew.

Windsurf wasn't unusual. Microsoft paid Inflection about $650 million in 2024 and hired most of its staff. Google paid Character.AI a reported $2.7 billion and hired back Noam Shazeer, a co-author of the paper that introduced the transformer, the design behind today's language models. Meta paid $14.3 billion for 49 percent of Scale AI, in shares with no votes, and hired its 28-year-old chief executive. Nvidia paid a reported $20 billion in a similar deal with the chip maker Groq.

The usual reading is that Big Tech panicked and overpaid. Sometimes it did. But every one of these buyers had the money and the engineers to try building the same thing itself, and chose to pay instead. That choice has a logic, and it applies far below Big Tech, to any company deciding whether to hire a team, buy a tool or build.

Here it is in one sentence: the price of know-how is set by how long it would take to rediscover it, and by what that time is worth in a market that changes every month. And because that rediscovery time shrinks every month, know-how is perishable.

What these deals actually buy

In a normal acquisition the buyer pays the shareholders and the whole company comes across: people, code, customers, contracts, liabilities. These deals work differently, and the difference tells you what the buyers wanted.

The big company pays the startup for a license to use its technology. Separately, it offers jobs to the founders and the people it wants most. The startup uses the fee to pay back its investors, and what's left (most of the staff, the product, the customers, the brand) carries on under a new chief executive. This is called a license-and-hire deal, or a reverse acqui-hire.

Scroll sideways to see the whole figure

Look at what the buyer leaves behind. Customers, revenue and the brand are normally the reason one company buys another. Here they stay put. The price is mostly a price for people and what's in their heads.

There's a second motive the companies don't advertise. In the US, buying a company above a certain size requires a filing with regulators before the deal closes, and a license plus job offers has generally been treated as outside that rule. That escape is narrowing: the UK regulator reviewed Microsoft's Inflection deal as a merger, and in early 2026 the head of the Federal Trade Commission said it was looking very closely at these deals. The structure still saves time, but it now carries a regulatory risk that belongs in the price.

Why rediscovery takes so long

What the buyer really pays for is the time its own people would need to get to the same place.

Imagine a large company trying to match Windsurf in-house in mid-2025. It would hire a few dozen engineers who had built coding agents, from a small pool of people who all had other offers, and turn them into a team. Then it would have to learn what the first team learned the hard way: which approaches fail in real codebases, which editor integrations developers put up with, which internal measurement actually predicts whether a developer is still using the tool a week later. Money shortens the first steps. It barely shortens the last one, because that step consists of running into problems and noticing them.

Licensing a startup's code without its people is like buying a restaurant's recipes without its cook. The recipe lists the ingredients. The cook knows the moment the sauce is about to split.

The philosopher Michael Polanyi put it in one sentence in 1966: "we can know more than we can tell." He called it tacit knowledge. In an AI team that has shipped a product, it includes the approaches that failed and why (failures rarely make it into a document, because they don't ship), which measurements actually track what users do, the quirks of the data pipeline, and a feel for how the model goes wrong, built from thousands of examples. That's why none of these deals was a license alone. A license gives you the artifact. The people carry the judgment about how to use it, change it, and tell when it's wrong.

Teams matter as units, too. Star stock analysts who changed firms, one study found, performed worse for years afterward, unless they moved with their teams. Analysts aren't AI engineers, but part of what makes any team good only moves if the team does.

What a month is worth

So whether a premium makes sense comes down to two numbers nobody knows precisely: how many months the deal saves, and what a month is worth.

Here's a worked example with illustrative numbers for a large software company that wants a coding agent for its developer tools.

The premium is $850 million: the $1 billion deal minus the $150 million it would cost to build. Fourteen months of delay at $35 million a month cost $490 million. On that arithmetic alone, build.

But a late arrival loses more than the months. Some customers who adopt a competitor in the meantime never come back. If arriving fourteen months late costs the company a fifth of its share for the following five years, the delay is worth about $808 million in today's money. Close to the premium, still slightly under.

Now be honest about in-house projects, because they slip. Give it an even chance that building takes thirty months instead of eighteen. The deal now saves twenty months on average, and the cost of delay rises to about $1 billion, comfortably above the premium. Buy.

Look at where the answer comes from. The verdict turns on two inputs that managers estimate badly: how long the in-house project would really take, and how much of the market is lost for good by arriving late. Change either one moderately and the decision flips.

A premium is a bet on the value of a month.

You can also run the arithmetic backwards to read what a price says about the buyer's beliefs. Google's $2.4 billion bought none of Windsurf's revenue. If the deal saved it about a year in coding agents, Google valued a year of earlier position in that market at roughly $2 billion or more. That's a statement about how much coding agents mattered to its models and its cloud business, and unlike the panic story, it can be checked against what happened next.

Know-how goes off

Rediscovery time isn't fixed. Every month some of what a leading team knows leaks into the field, through papers, open models, engineers changing jobs, open-source tools, and the next general model handling a problem that once needed a specialist. Know-how that took a team two years to build may be rediscoverable in six months a year later.

That explains the timing of these deals. The price is highest in a narrow window: after the problem is widely recognized, so the buyer knows it needs the capability, and before the solution has spread, so the buyer can't get it cheaply. In spring 2025 Windsurf was reportedly worth $3 billion to OpenAI, while the big labs were already shipping coding agents of their own. A year later the same know-how would likely have been much easier to rebuild. A buyer who waits for certainty pays less for the know-how and more for the delay.

How it goes wrong

Two years in, the record is mixed, and it says more about which assets survive a deal than about whether the buyers were wise.

In several cases the hired people became the core of the buyer's own program: Microsoft AI under Inflection's co-founder Mustafa Suleyman, Gemini with Shazeer among its leads, and Google's Antigravity coding environment, which shipped about four months after the Windsurf team arrived. But in none of them did the deal replace building. Microsoft's first in-house models still took about seventeen months. Bought know-how accelerates an internal program. It doesn't stand in for one. Meanwhile the companies left behind mostly shrank or changed course: Inflection and Character.AI stopped building frontier models. Windsurf's remaining staff did well only because another buyer showed up within days.

The same mechanisms that justify a premium explain how it fails.

The team never integrates. A team bought for speed lands inside a large organization with its own priorities. In a study of about 4,000 US startup acquisitions, acquired employees left far more often than ordinary hires, about a third in the first year against roughly an eighth, and less so when the founders stayed and the unit was kept separate.

The team leaves after vesting. Retention packages vest over a few years, so a premium buys a window. Within about two years Shazeer had left Google for OpenAI, and the Adept co-founder who led Amazon's agent lab had left Amazon. If what the team knew never becomes part of the organization, it walks out with them.

The know-how was scaffolding. Part of what any AI team knows is how to work around today's models' weaknesses, and the next model often makes that unnecessary (Same Model, Different Results). Adept spent two years building agents that operate software through its screens. By early 2025, Anthropic and OpenAI had shipped general models that could do that, imperfectly, without a specialist product.

The buyer paid for a reputation. When the UK regulator reviewed the Inflection deal, it found Inflection hadn't been able to grow its chatbot's users. That says nothing against the people Microsoft wanted, only that the product was a weak basis for the price.

The deal destroyed the asset. Part of Scale AI's value to the AI labs was neutrality: a lab sending work to a data vendor can reveal unreleased model outputs and hints of its roadmap. Once a direct competitor owned nearly half the vendor, Google, its largest customer, was reported to be cutting ties within days, and others pulled back.

Some assets disappear the moment someone buys them.

What lasts

Put the cases together and a license-and-hire deal turns out to buy a bundle of assets with very different lifetimes.

Scroll sideways to see the whole figure

People and their judgment last as long as they stay and keep learning. Tacit know-how moves only inside those people, and partly expires with each model generation. Code and model weights survive the deal but rarely the next model; Character.AI's switch from its own models to other companies' is the clearest case. Customers stay behind with the startup. Trust can't move at all.

The durable exception is the evaluation set: a collection of real cases with known good outcomes that records what "good" means for real users in a specific job. It keeps its value when the model underneath changes, because it's the instrument you use to judge the next model. It's also the closest thing to tacit judgment that can be written down. If a buyer can insist on only one artifact beyond the people, this is the one.

The view from both sides of the table

If you run a small company, the same logic suggests a strategy. A large AI company's tightest constraint is usually time, more than money or talent. Its teams focus on models, scale and safety. But an agent that works in production needs dozens of smaller components that are too critical to ignore and too specific for a big company to prioritize early: tracing why an agent made a decision across many steps, recovering cleanly when a tool fails mid-task, adapting outputs to European regulation, running code in a tightly sandboxed environment. A small company that solves one of these properly, runs it in production, documents where it stops working and has evaluation results that hold up, holds exactly what the arithmetic prices: months the buyer would rather not spend. The senior engineers who decide these deals are persuaded more by an honest account of limits than by a demo.

But building only to be acquired is a poor plan. The window closes as rediscovery time falls, a lab may absorb your component into its next model, and these deals pay founders and investors first and can leave everyone else with little. My advice: build a critical component so well that if a large company needs it tomorrow, you're the obvious choice, and make sure it's worth having if nobody calls. Scale's experience adds a reason: being a trusted supplier to several large companies can be worth more than being owned by one.

If you run an ordinary company, you face the same choice at smaller scale: hire a team or an experienced lead, buy a tool, bring in a vendor or consultancy, or build. Four steps.

Estimate rediscovery time honestly. How long would your own people take to learn what the outside team knows, including its failures? Ask the vendor for its list of known failure cases and its evaluation set. If it has neither, there may be less know-how to buy than the price suggests.

Estimate what a month is worth to you, without the headline numbers. Google faced a market where arriving late could cost share for years. Most companies don't. For a mid-sized insurer automating claims triage, a month of earlier deployment might be worth tens of thousands of euros in handlers' time, and no customer leaves because triage arrived in spring instead of winter. Without that lasting loss, a premium has to be justified by direct value alone, so most companies should pay less for speed than the headlines suggest.

Ask whether the know-how is scaffolding. If the next model will make it unnecessary, rent it and keep the contract short. If it encodes your own advantage (your data, your processes, your customers' particular needs), build it or hire for it (Who Should Build It).

Decide in advance what must survive the engagement. Evaluation sets, documented decisions, trained people. This matters most in departments you've already outsourced, where the know-how often lives entirely with the supplier and you pay the rediscovery premium again every time you switch.

Code and weights age fastest. What still takes long to rediscover after the next model is judgment about what good work looks like for particular users, the evaluation sets that record it, and the trust that lets someone hand you their data. Those are worth paying for, and worth building. The question to ask before paying anyone for know-how is what they'll still know, a year after the next model ships, that your own people won't.

Check yourself

Why did Google pay about 29 times Windsurf's revenue and leave the revenue behind?

It was buying people and what they knew, not customers or the product. The price stood for the time Google would have needed to rediscover that know-how.

In the worked example, what makes the answer flip from build to buy?

The two inputs managers estimate badly: how long the in-house project would really take, and how much of the market is lost for good by arriving late.

Which part of a bought team's know-how is most likely to lose value after the next model?

Code, weights and scaffolding, which the next model often makes unnecessary. Judgment and an evaluation set last longer.