Most investors are still asking the wrong question about defensibility. They ask whether a competitor could rebuild the product. In 2026, the more important question is whether the customer could.

The test changed while we were not looking.

Retool surveyed 817 builders this year and found that 35% of enterprises have already replaced at least one SaaS product with something they built internally. Seventy-eight percent plan to build more. They are replacing workflow automation, admin tools, BI, CRM, project management and support software. These are not bad categories. These are categories venture capital has funded for fifteen years. McKinsey independently puts the build-instead-of-buy number at 32%.

The buyer did not suddenly become a software company. The buyer got a coding agent.

So when a founder tells me the moat is feature depth, integrations or UX, I now apply a simple customer test. If your largest customer decided to clone the product next quarter, would their version be materially worse? If the honest answer is no, you do not have a moat. You have a head start.

What can a customer not build for itself?

I keep coming back to four things.

The first is accountability. When an AI agent acts autonomously inside a customer's systems, someone has to be responsible when it gets something wrong. An internal tool gives you nobody outside the company to hold accountable. In high-consequence workflows, the startup is not just selling software. It is selling the transfer of risk.

The second is data the customer cannot create alone. Not its own data sitting inside your product. I mean data that only exists because you sit across many customers, markets or counterparties. Benchmarks. Fraud signals. Pricing intelligence. A customer can reproduce what it teaches you. It cannot reproduce what everyone else teaches you.

The third is owning the record, not just the interface. AI is making interfaces incredibly cheap to build. But systems holding state and history remain painful to remove because removing them means reconciling records and figuring out what breaks downstream.

The fourth is multi-party position. If you sit between parties that do not want to integrate directly, a customer cannot rebuild your product because it cannot rebuild the other side of your network.

Put these together and the objective has changed.

Stop trying to be hard to copy. Start trying to be hard to remove.

Removal is an organizational problem, not a technical one. I care less about whether rebuilding your software takes six months or six days. I care about how many people must approve removing it, what breaks when you leave and who gets blamed when the migration goes wrong.

This changes how I do diligence.

I want to see the MSA. Who is responsible? Is there an indemnity? What is the liability cap? A contract with real teeth tells me the customer is buying more than functionality. It is transferring responsibility.

I ask whether customers have discussed rebuilding the product internally. I ask why churned customers left and how many rebuilt themselves. I ask how many departments use the product. And I ask the simplest question: what happens when your software is wrong?

There is an uncomfortable implication here.

If defensibility increasingly comes from accountability, trust and relationships, then part of the AI moat is human. That sits awkwardly with our current obsession with tiny AI-native companies.

A company may be able to build its product with almost no employees. It cannot necessarily hold the customer relationship, absorb liability and earn the right to act autonomously on a customer's behalf with almost no employees.

There is an accountability floor below which companies cannot shrink.

We are about to find out where it is.

So stop asking whether a competitor can copy the startup.

Ask whether the customer can.

Then ask who is on the hook when the software gets something important wrong.

Increasingly, I think that person is the moat.

Dario Amodei may be right about AI and still be wrong about medicine. His argument is that intelligence is about to become abundant. Millions of AI scientists could read the literature, generate hypotheses, design molecules, plan experiments and work continuously at a speed no human research organisation can match. I think we should take that possibility seriously.

So assume he is right. Assume AI scientists become better than most human scientists. Assume automated laboratories improve. Assume virtual cells become useful. Assume AI compresses years of early drug discovery into months. Even then, curing most human disease in five to ten years does not necessarily follow, because intelligence and evidence do not scale in the same way.

AI can manufacture hypotheses. It cannot manufacture ground truth.

For most of scientific history, good ideas have been scarce. There are only so many scientists, only so many hours in a day and far more biological possibilities than humans can investigate. AI could reverse that scarcity. We may eventually have far more credible hypotheses than the scientific system can physically test.

DeepMind has already described the beginnings of this as a validation bottleneck. As AI becomes better at generating scientific conjectures, our ability to verify them does not automatically improve at the same rate. Medicine makes this problem especially difficult because validation is not purely computational. It consumes laboratories, biological samples, manufacturing capacity, clinicians, capital and eventually patients.

Imagine an AI system generates one million genuinely plausible cancer treatments. Not hallucinations. One million interventions supported well enough by existing biology that each deserves investigation. We will not test one million of them. We may not even test one thousand.

This changes the economics of science. The question is no longer only, what can we discover? It becomes, what deserves one of our scarce opportunities to find out whether it is true?

The obvious answer is that AI will rank the candidates too. It will use genomics, mechanistic models, animal data, biomarkers, simulations and previous clinical outcomes. That will make the selection process much better. But all of those remain representations of reality. The reason experiments exist is because our representations are incomplete.

At some point prediction ends and measurement begins.

This matters because biomedical development already loses most candidates before approval. Even if AI dramatically improves those odds, it does not make biological uncertainty disappear. A system that moves success from one in twenty serious candidates to one in five would be extraordinary. It would still mean that four of every five serious bets consume time, capital and experimental capacity without becoming medicines.

Now scale candidate generation by 100 or 1,000 times.

The constraint moves.

An AI can generate another molecule in seconds. It cannot generate another patient with a rare disease. It cannot manufacture another person with precisely the mutation required for a clinical study. It cannot turn five years of survival data into five minutes. In rare diseases and narrow biomarker populations, eligible patients are already a limiting resource.

This suggests the next important scaling variable in science is not intelligence.

It is reality throughput.

Reality throughput is the rate at which scientific ideas can encounter enough high-quality evidence to determine whether they are actually true. If AI drives the marginal cost of generating hypotheses toward zero while experiments remain expensive, physical and slow, the value of each opportunity to test reality rises.

That is the real consequence of scientific abundance.

We may soon have too many good ideas, not too few.

This also means we may be benchmarking AI scientists incorrectly. Instead of asking a system to generate 1,000 novel hypotheses, give it 100,000 credible hypotheses and a budget for only 100 experiments. Ask it what to test first. Reveal the result. Then ask again.

The metric should not be how many ideas the AI produces. It should be how much truth it extracts from a fixed amount of reality.

That is a very different kind of intelligence.

The best AI scientist may not be the system that generates the most discoveries. It may be the one that wastes the fewest experiments.

Now return to Amodei’s idea of a country of geniuses in a datacenter. Imagine one million superhuman cancer researchers working continuously. Then add another million. Then ten million more.

Eventually the datacenter contains more intelligence than the physical world can answer.

The scientists form a queue at reality.

This is not an argument against Amodei’s optimism. AI may become the most important technology medicine has ever acquired. But every technological revolution removes one scarcity and exposes another. The internet made information abundant and attention scarce. Generative AI is making cognition abundant and judgment more valuable.

If AI makes scientific hypotheses abundant, ground truth becomes scarce.

And if that happens, the next frontier in AI for science will not be building systems that can think faster.

It will be building systems that know which pieces of reality are worth spending next.

By the time a frontier AI researcher’s departure reaches TechCrunch, the best-connected investors may already be months into the relationship.

That is the problem with treating talent movement as news.

News tells you what happened. Sourcing infrastructure should tell you what is beginning to happen.

Frontier AI labs have become some of the most productive founder factories in venture history. Anthropic came out of OpenAI. Safe Superintelligence was founded by former OpenAI chief scientist Ilya Sutskever and reached a valuation above $30 billion before launching a product. Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, raised $2 billion at a reported $12 billion valuation in July 2025.

The capital was not underwriting revenue. It was underwriting the people who left.

THE DEPARTURE IS NOT THE SIGNAL

A single researcher leaving a lab is noise.

Three researchers leaving the same twelve-person team within six weeks, after years of working and publishing together, is different.

It might be a reorganisation. It might be coordinated recruiting by a rival. Or it might be a founding team assembling before incorporation.

The data already exists to distinguish between these possibilities.

Researchers are nodes. Co-authorship, shared projects and team membership are edges. Departure timing, seniority, technical ownership, new domains, corporate registrations and open-source activity are features.

The objective is not to label someone a “future founder.” It is to estimate how the probability of company formation changes as new evidence appears.

One unexplained departure creates a small increase in probability. A second departure from the same team increases it further. A new domain registration, company filing or renewed collaboration among former colleagues raises it again.

This is not a static founder score. It is a time-to-event problem.

THE RAW LIST WILL NOT BE THE EDGE

The data layer is already being productised.

Evertrace monitors company registries, GitHub, patents, grants, domains and social activity. Specter tracks new founders, stealth hires and job changes. Dealroom has built a talent graph covering millions of companies and profiles.

That means the departure list will become available to everyone.

Buying access to the same list as every other fund is not alpha. It is software procurement.

The real advantage moves to three things: scoring, timing and access.

First, how likely is this person or cluster to form a company?

Second, how likely is that company to attract institutional capital?

Third, at the valuation the market is likely to assign it, will the investment generate venture returns?

These are different questions.

Prestigious lab experience may predict fundraising success while also producing an entry price that absorbs most of the upside. A model that only predicts who will raise money may consistently find the most expensive deals, not the best ones.

THE MODEL DOES NOT GET THE MEETING

Even a good model only creates a timing advantage.

Knowing that three researchers quietly left the same team is not particularly useful if your first interaction is a cold LinkedIn message after the system flags them.

The funds that win these deals already have relationships with the researchers, their former colleagues, professors, founders and angel investors.

The system tells them where activity is forming.

The network gets the meeting.

Reputation converts the meeting into an allocation.

THE BOTTOM LINE

Build the defection graph as infrastructure, not as a Slack alert.

Continuously ingest career transitions, collaboration histories, company registrations, domains, patents and technical activity. Score clusters with a time-dependent model. Test separately for company formation, fundraising probability and investment returns.

Then build relationships with the people the model suggests are most likely to move, before they have decided to raise.

The list will become available to everyone.

The enduring edge belongs to the firms that interpret it correctly, act before the signal becomes obvious and already have the phone number.