Cheap intelligence, expensive trust

Last week I split my time between Cloudflare Connect at Moscone West, AI/Tech Investing Forum, hosted by Opal Group at SVB, and a few other events, dinners and networking sessions. The different settings could hardly have been more different. At the forum, family offices and fund managers talked about where the money is going. At Cloudflare, 5,500 people from 84 countries talked about how to get agents into production without breaking anything.

I brought one question to both: what happens when AI meets a regulated workflow?

By the end of the week I had an answer in five parts. Money has gone to the top of the stack. The moat has moved from code to data and trust. Trust has become the bottleneck for agents. Many of the hardest problems are physical. And buyers are moving, but they start small. All five point to the same place: the work at the bottom of the stack, where models meet operations.

1. The money went to the top

The forum opened with how concentrated the money has become. The host said a few hundred companies took about 80% of venture dollars in the first half of this year, and roughly 40% went to three AI companies. George Bandarian of Untapped Ventures put AI at about $500 billion of the roughly $700 billion invested so far this year. "AI isn't the sector anymore," he said. "It is the venture asset class."

The host described a dinner with managers of a $2 billion fund and a $10 billion fund. A few years ago they hunted for outliers. Now, he said, there's a "sheep mentality": be in the fastest-moving fields, because those are the companies that reach tens of billions.

One panelist went further and argued that venture capital matters less than it used to. By his estimate, the whole industry has created about $1.2 trillion of value in 25 years, and a handful of companies, Elon Musk's and the biggest AI labs among them, now dwarf it. Another panelist pushed back. For the investors in that room, he said, the lesson is to shift left and get into companies while they're still private. Bandarian made the same point with an example. In one AI exit, the early investors made thousands of times their money, and some of the most sophisticated strategic investors made about 2x. The only difference was when they got in.

I think both views hold. The frontier is a scale game, and most investors can't play it. But every frontier model still has to be carried into a batch record, a nurse's shift schedule or a quality control lab before it does any work. That last mile is slow to sell and specific to each industry, and people who have done the work have an edge there.

2. The moat moved from code to data and trust

A panel at the forum was blunt about code. "Codex or Claude can literally reproduce any tech within a few hours," said the moderator, who runs a 200,000-member AI founder community, and she was shipping features to her own platform while we talked. If code is cheap to copy, the panel agreed, the moat is distribution and proprietary data.

The data argument came up all day. One investor said the big models have scraped only a small fraction of the world's data, the open web. The rest is private: a company's email, a pharma company's compound library, a manufacturer's process history. Another panelist added a caution I agree with. Owning data isn't enough, because someone may find another route to the same data within a venture timeframe. What holds up is the expertise to use the data and the trust that it's right.

At Cloudflare, Lélio Renard Lavaud, VP of engineering at Mistral, took this into the model itself. When an enterprise trains a model on its own data, he said, it should own the weights, because everything that makes the company unique gets baked into them. Mistral's customers include companies like ASML and BMW, and he described the work as turning hard industrial problems into AI, not summarizing email. To him, sovereignty is less about geopolitics than about a company keeping control of what it builds.

In regulated operations this is familiar ground. The data comes out of validated processes, the operator owns it, and the right to touch it is earned through compliance work: validation, audits and change control, which a competitor can't copy in a few hours.

3. Trust is the bottleneck for agents

Bandarian cited an IDC forecast of a billion AI agents by 2028. A founder in the audience asked the obvious question: what do you do when agents don't do what you told them? The panel's answer was that a whole category is forming around agent identity and intent, the question of whether an agent is who it says it is and still doing what its owner meant.

California State Senator Josh Becker was more direct. He compared today's AI to a car that mostly stays in its lane: we wouldn't let that car on the road. He wants systems that can be audited, and he wants liability spread sensibly. Developers should take reasonable care, employers should understand their systems' limits, and users should use them responsibly. He also previewed legislation, the Frontier AI Safety Act, that would require major frontier labs to publish the share of their R&D computing that goes to safety and let the state set minimum standards.

The version closest to my own work came from Peter Steinberger, who created OpenClaw, the open-source personal agent, and now works on agents at OpenAI. His project was flooded with AI-written pull requests and security reports. Now he trusts contributions more when he can see the contributor's agent sessions, the reasoning behind the code. If he can see how you got from A to B, he said, he trusts the work far more than finished code thrown over the wall. He also has agents check each other's work. On the same stage, Mistral described watching millions of agent sandboxes at once and using classic machine learning to flag threats in real time.

"Show me how you got from A to B" is what FDA and EU inspectors have asked of every batch record for decades. A panelist at the forum made a related point: banking and healthcare have made the most progress on agent safety, because they already work under strict rules. Europe is now rewriting its good manufacturing practice (GMP) rules for computerized systems (Annex 11) and adding new ones for AI (Annex 22). Regulated industries have a head start with agents, because they've been doing the hard part for years.

4. AI is getting physical

The best question at the forum took three seconds. Someone asked Bandarian: "How are you going to power all this?"

The next session answered it. Keith Kaplan and Ali Bushehri of Maykr build the robots that move wafers around chip fabs. Power, one of them said, is the coin of the realm. Energy, metals, supply chains and delivery times used to move independently, and now they move together. In their experience, lead times for circuit boards have gone from about two months last year to between 26 and 52 weeks.

They also pointed to a gap few investors watch. Most new money goes to 300mm fabs for AI compute. SEMI, the chip-equipment industry association, expects spending on 300mm fab equipment to reach $149 billion this year and $227 billion by 2029, with AI chips among the main drivers. The older 200mm fabs make the chips inside sensors, actuators, power systems, cars, ICU equipment and medical devices, and every robot depends on them. By SEMI's figures, the Americas were forecast to hold about 14% of global 200mm capacity in 2023 and China about 22%. The two wafer sizes need different equipment, so moving a fab from 200mm to 300mm means ripping everything out, at a cost of billions.

Figure 1. SEMI forecast China at 22% of 200mm capacity in 2023 and the Americas at 14%. Forecast share of global 200mm fab capacity, by region, 2023.

The car chip shortage of 2021 was largely a shortage of older chips, many of them made on 200mm wafers. About a year ago, when China blocked exports from the chipmaker Nexperia, Honda cut production in North America, and Volkswagen and Volvo warned of possible stoppages. Kaplan and Bushehri compared it to rare earths: the US let processing move to China in the 1980s and 1990s, and the know-how never fully came back. They don't want 200mm to go the same way. For medtech, that makes 200mm a supply chain issue. An ICU monitor or a medical device still needs a 200mm chip before it ships.

The demos showed where the physical work is going. Figure's newest humanoid robot has started sorting and carting parts at BMW's plant in Spartanburg, South Carolina. AI cameras on production lines are giving managers live data on productivity, safety and the root cause of errors. One investor described a portfolio company whose cameras, by his account, multiplied a medical-device maker's output of tiny surgical parts more than tenfold and pushed escaped defects toward zero. Zipline came up as deep tech that is scaling. It started in 2016 by flying blood to clinics in Rwanda, carried medical supplies for Novant Health in North Carolina during COVID, and now delivers food, retail and health products to homes in the US.

I spent three years on the software that ran a DNA factory across two sites. The model was never the bottleneck. Hand-offs were, along with equipment uptime, inventory and the people who keep a line running. AI gets its value from the physical layer, and that's also where it most often stalls.

5. Buyers are moving, one small win at a time

On the buyer side, Cloudflare's co-founder and president, Michelle Zatlyn, said the past three quarters have changed how her company works. Engineering is no longer the only team building. Finance, people and sales teams build too, because the people closest to a problem can now solve it themselves. Cloudflare's security chief has run more than 100 workshops for other companies' security teams, and Zatlyn said more than 17,000 security people outside Cloudflare now build with AI.

Her advice for organizations stuck at the start was practical. You don't have to solve everything at once. Pick a place where the stakes are lower and get one win. One of her customers, based in Asia, tried a new AI service in Latin America, where its brand was small and a failure would have been quiet. It worked, and senior management came around. She added that in many traditional enterprises, AI transformation is owned by the business owner or even HR, because they own the workflows. She also told the room to use this moment to prune legacy vendors and put their energy into the partners who matter next.

Budgets are shifting as well. The forum's host said CFOs now sign off on AI spending and follow token prices, because AI shifts cost from headcount to compute. Jobs are less clear than the headlines. Becker said LinkedIn's data shows entry-level hiring dipped in 2022 and 2023 and has been flat since. He worries more about people losing skills they stop practicing. When the moderator asked each panelist to pick one sector, one chose healthcare. A CEO had told her that more than half of medical labs still send lab reports by fax. Another pointed to bioscience and to vertical AI in healthcare, finance and insurance, industries that haven't modernized in decades.

In healthcare and life sciences, the constraint is people. Hospitals and labs aren't trying to replace staff. They're trying to cover for staff they can't hire. A buyer pruning old vendors and looking for a low-risk first win is the opening a young company needs. And the first pilot will usually come from whoever owns the workflow, often not IT.

Get on the plane

The session I keep thinking about wasn't about AI. An anonymous analyst who goes by Analyst #3 told Cloudflare's audience how an argument about the price of oil ended with him flying to Oman to watch the Strait of Hormuz himself, after the internet had decided an oil shock was coming. He came back with a different answer than the consensus. "Sometimes you have to get on the plane," he said. It's getting harder, he added, to be right just by reading the internet.

Over the week, intelligence kept looking cheaper and more abundant. Trust and power did not, and neither did people who know how the work gets done. Finding them still means getting on the plane and walking the factory floor.