我们几个月内造出了自己的 Salesforce。为什么要取消这份60万美元合同 | Curative CEO

[Speaker?]

Lockdowns had just started. Everybody was starting to freak out. There was basically nowhere to get a test. So our chief scientific officer had, in his spare time, developed a COVID test. I think our peak day was 206,000 people tested in a single day.

1. Meet Fred Turner: thrill of winning vs. fear of losing

Fred, I'm so excited for this. You have the most wild story, and I heard it from Justin first and then from Anil. Thank you so much for joining me, man.

Yeah, thanks for having me.

I always find it very telling that entrepreneurs are often compelled either by the fear of losing or by the thrill of winning. If I were to ask you which one drives you more, what would you say?

The thrill of winning. I feel like during COVID, with some of what we built at Curative, I got a taste for the speed at which you can move when everything is behind you and all the momentum is behind you, and I've been chasing that ever since.

That is the biggest tailwind that one could have ever expected. We're going to get to that. You actually grew up in the UK and then moved to Silicon Valley very young—17?

19.

19. Okay. Could you have built the business that you did in the UK?

2. Why he couldn't have built this business in the UK

No, definitely not.

Why is that?

I just think the UK doesn't have the kind of infrastructure for startups and investing. There aren't that many people who have done a startup before and are then willing to invest in the next generation, particularly investing in younger people.

When I tried to raise a venture round in the UK, I couldn't even get meetings. This was when I was 18. I was in my first year of college, doing this startup on the side, and I couldn't even get meetings with investors. I think I got 1 fund to take an associate meeting with me.

Naturally, it went very far.

Yes. It just seemed like people were investing purely on credentials. This was more than 10 years ago, but it seemed like people were investing based on, “You came out of this university, so if you're an undergrad, how could we possibly look at this? This doesn't make any sense.”

Whereas you go to Silicon Valley and it's, “What's the possibility here? What could you envision in 10 years if everything succeeds? How big a company could this be?” It was a very different mindset. They were optimizing for how to get the best outcome, whereas I always felt like in the UK, it was optimizing for mitigating the worst downstream outcome.

Yes. How do I not get fired?

Right?

Yeah, I totally get that. That's pretty funny. And so we decided to move to the Valley. Great. Can you just walk me through how we go from cows to sepsis to COVID, just so I understand this?

3. From cattle genetics to sepsis to Covid (the origin story)

Yeah. So my first company—

I've never said that statement before in a 20VC episode, by the way.

There we go. It's a new phrase for you.

I went from initially cattle testing through sepsis to COVID. My first company in the UK was called TL Biolabs at the time. We were basically sequencing dairy and beef cows to predict various traits about the animal from an early age.

It started off with beef. You can predict that certain cows are going to have more musculature from an early age, and some cows can have too much musculature, so they have trouble giving birth. There's an optimum that you're shooting for.

I found this completely by chance. I won the UK National Science and Engineering Competition and was on TV a little bit. A farmer reached out to me because he wanted help testing his cows. He was sending his samples to the Netherlands, and it was taking weeks. It was terrible.

I initially told him, “I'm not interested in cows. I was interested in human genetics at the time. No, thank you.” He kept pressing and sent me samples with a check attached to the front, and I thought, “Okay, this is interesting.”

I did the first batch of samples for him, and then all of his friends started sending me samples. It grew from there. This was all still in the north of England, and I was in my first year of college at the time.

We started branching out into dairy and predicting how much milk animals would make. I tried to raise the first venture round for the company in the UK, but I didn't get very far. So I ended up going to the US-UK Tech Investment Conference in San Francisco. It was my first time in the States; I'd never been before.

It was my last-ditch attempt to try to raise some money. I met a bunch of VCs and didn't raise any money, but I did meet a guy who had just finished Y Combinator. He said, “You need to apply to YC. That's what you need to do. You need to move the company to the US and apply to YC. That's the only thing you can do here.”

I was familiar with YC but had never applied.

What year was this?

This was at the end of 2015. I went back to the hotel room, and it turned out the application deadline was 6 days away. I thought, “All right, it's meant to be.”

I did the application, got the interview, came back for the interview, and then moved to Silicon Valley for the Summer 2016 batch.

So how was the interview? Who was it with?

Sam, Geoff, and somebody else. It was all a bit of a blur. It was very fast.

And then you found out you got in.

Yes.

You moved to the Valley.

I moved to the Valley, and then it was the same: you pitched the same company. We were doing mostly dairy testing at that point—testing dairy cows to try to predict their milk yield, which for farmers is actually very valuable because they don't make milk until they're 18 months old.

From day 1, all your animals have to have a calf every year to keep making milk, so your herd doubles every year.

Is this still the same Curative company?

No, this is a completely different company.

Okay. I was about to say, this is where investing is so difficult, because if you hear a founder pitching milk-yield optimization—and I'm sure, logistically, it's a big TAM, I'm sure—

Well, not big enough. That was the problem.

Okay.

So we went to YC, and we raised a seed round from Andreessen Horowitz.

A seed round from Andreessen.

Yeah. Their bio fund did our seed round right out of YC, and I don't think they did the TAM calculation.

Credit to you.

They backed founders.

Chump change.

It was for the coffee.

It was—for Andreessen, it was a smaller round.

Sure.

We kept developing the technology and had customers. Then we went to raise a Series A, and people did do the TAM calculation. There are about 100 million cows in the US. If you're doing well, you could charge $15 to $20 per test. So even if you assume you could test every cow every year, you'd be at $1.5 billion in total market, which is not enough to do a Series A off of.

4. Pivoting into human diagnostics and at-home STD testing

What we ended up doing was taking some of the core DNA-testing technology that we had developed and pivoting to use it for human diagnostics. That was my first foray into health care.

We actually first launched a high-throughput STD-testing lab. We launched an at-home STD test. That was tons of fun.

I used to run a lot when I was young, and my knees were wonderful. I used to love How I Built This because they would ask questions like this: How do you go from cows and musculature in cows and milk-yield optimization to at-home STD testing? It doesn't feel like that natural a jump.

On the back end, it's more natural. All of these things have DNA in them. If you're looking to do better DNA testing, you're just looking for markets where people care more about that.

Anything human, people obviously care a lot more about, are more willing to pay for, and represents a much larger market. So we took a market-first approach: Where could there be interesting opportunities?

We narrowed in on antibiotic resistance in STDs as being a particularly interesting area, where infections are getting harder and harder to treat because of increasing antibiotic resistance. If you're doing the DNA testing, you can predict what the best drug is going to be early, treat with that drug, and then you're not using the most aggressive antibiotics.

Do we have more STDs than ever?

Yeah. Yeah.

This conversation’s pivoting somewhere I didn’t expect, but I thought we were having less sex than ever.

Yeah, but more STDs.

Wow.

Yeah, that’s worrying. And it is. It’s a while since I looked at the statistics because I’ve not been doing this for a while now. But when I was last in this, the statistics were just a steady increase, and then an increase in resistance.

And so it’s getting to the point where certain STDs are harder and harder to treat, and some of them might eventually become untreatable. You might have to be hospitalized to get a really powerful antibiotic to treat one, which is crazy. Antibiotic stewardship was a whole thing, and so we did that with STDs.

I love this conversation.

And then we found a fascinating market in sepsis. Sepsis is a disease that kills hundreds of thousands of people a year. It’s basically where you get bacteria in your bloodstream. What kills you is not actually the bacteria; it’s your own immune system.

You’re not supposed to have bacteria in your blood, right? Your blood is supposed to be sterile. When bacteria get in there, your immune system kind of freaks out and triggers this whole downstream cascade where your blood vessels start to leak and all of your organs start failing. It’s basically really bad, and that’s what kills you: your own immune reaction to the bacteria rather than the bacteria.

This is one of the leading causes of death in the US. Often, if you’re dying from something else—if you have serious cancer, for example—it’ll be sepsis that ultimately ends up being what kills you, because you get more susceptible to it with other diseases. It’s a leading cause of death, with increasing mortality. It’s incredibly expensive, and outcomes are terrible.

We were working on basically a better testing technology where, from the earliest stage, you could detect these bacteria and what antibiotic they were going to be susceptible to, and treat people faster. With sepsis, basically every hour that you don’t treat somebody is about a 12% increase in mortality, so you want to get the treatment as soon as possible.

Every hour you don’t treat someone is a 12% increase in mortality.

Yeah.

Okay. And so we start the sepsis testing.

So, we start the sepsis testing.

Does it instantly go well?

No. This company died at the end of 2019.

Oh, I’m sorry.

Yes. The testing was working great. Prototypes were pursuing the FDA approval process. We went to do a Series B round, which ended up being a strategic.

Sorry, just so I understand: we raised the Series A from a16z, and it’s still the same company as this Milky[?]?

Same company. Yeah. It changed its name from TL Biolabs to Shield.

Oh, I love it. Good. It’s a good single name. Okay, so we go to raise the Series B: bigger TAM, sepsis, death, more TAM gravitas.

Yeah. Many, many billions of dollars of TAM for testing for this. We ended up getting a term sheet from a strategic, large public diagnostic company. We signed the term sheet and did 3 weeks of work on the documents. We were in the second round of documents, and then their CEO killed it because it was too competitive with their core products.

Meanwhile, we told all the investors, “Oh, yeah, we’ve got the lead. We’re good to go. Oh, here’s the paperwork.” That was the death of the company. We had about 3 weeks’ worth of cash.

Would you be where you are today, though, if that round had come together?

No. I don’t think we would have pivoted as hard into COVID when COVID hit. It would probably have been easier because, at that point, we actually had a lab license. In the US, you need this thing called a CLIA license to run these kinds of tests, and we had gotten one of these licenses over a painstaking 2-year process.

Then, in December 2019, as part of the wind-down, I sold that license to a company in San Diego for $150,000 to pay some of the creditors. Then, 5 months later, I acquired a company in Southern California to get the same license for $27 million. So, timing is everything.

Whoa, whoa, whoa. Wait, wait. So, we’re winding down the company and we sell this license for $150,000—

Which is roughly its market value if there isn’t a pandemic.

Totally get that. Cool. Okay. We’re winding down the company. I want to go chronologically because that’s a wild number. It flies in the face of shutting down the company in 2019—now 2020, I guess.

It was kind of right at the end of 2019.

Okay, end of 2019. Shutting down the company, strategic, all of that. And then what happens?

So then I was looking at what to do next.

Were you personally devastated? This is 5 years of your life. Any lessons for founders? Reflections on that?

I think it’s a lot easier to build a company the second time around. There are so many mistakes the first time where you just don’t know how to do something: the first time you fire somebody, how to build a good interview process, how to build a pipeline. There are so many things that it’s really easy to screw up the first time around.

5. Founding Curative on a sepsis thesis, before Covid changed everything

When you’ve seen them go wrong, it’s so much easier to build it the second time around. It’s the worst thing in the world to go through, having something you poured all that time and energy into—and the 7-day workweeks and the late nights—basically go to zero. But you learn through that, and as long as you take those lessons and go solve an even bigger problem, I think you got something out of it.

Okay. And so this company is winding down. You need to find something else. What happens now?

Originally, the pitch behind Curative was that we were going to also solve sepsis, but in a completely different way.

You really focused on it, really.

Yeah. So, when we were going through all of this work with the sepsis diagnostics, one of the things that kept jumping out in the data was when you look at other companies that had tried to do sepsis diagnostics, because we were not the first.

A bunch of big pharma companies, likely including Roche, spent a couple hundred million. Likely Siemens spent $100 million. A bunch of companies spent a lot of money trying to make better sepsis diagnostics. It’s kind of this graveyard of dead sepsis companies.

When you dig into the data, you find this really interesting thing: in academic medical centers, when you try out these new sepsis tests, they work great. You see much better outcomes, you see people live longer, and it’s saving lives. Then you try to replicate that in bigger studies, and they fail.

When you dig in and look at why, it’s when you expand the aperture of who’s in the trial out of the academic medical center and into community hospitals. What’s happening in a community hospital is that they’re so understaffed and so overwhelmed with the volume, particularly in the emergency room, that they don’t suspect sepsis fast enough. As I said earlier, every hour is a 12% increase in mortality.

The intervention that they have to do is actually pretty severe. They basically put a big IV line, usually in your femoral vein. They’re pumping you full of fluids, and they’re pumping you full of nasty antibiotics that have bad side effects, so it’s a pretty aggressive treatment. But if they don’t suspect sepsis early enough and jump to that treatment, by the time they get there, it’s already too late.

If you’re in a community hospital and it’s 2 a.m. on a Saturday, is there someone on staff who actually suspects sepsis early enough, or does it wait until Monday morning? It doesn’t matter if you have a better test if no one ever runs it.

The original pitch behind Curative was: let’s take the learnings from an academic medical center and go out to community hospitals and basically build a mini-hospital in a hospital that just manages their sepsis patients.

Whenever they get somebody, we will diagnose them as having sepsis in the emergency room. We will then take on that patient. They would pay a fixed fee, so no matter what happens, we’re on the hook. If we can drive a better outcome by applying—mostly just getting doctors to follow the instructions, but at scale—then you could drive better outcomes by getting those academic-medical-center-type clinical results, but helping a community hospital actually do that.

So what happens then? We start that business.

Start that business. We raised $1 million of seed money. Justin was the first investor—you mentioned him at the beginning, Justin Mateen. He came in right as I was shutting down Shield. He was an investor in Shield, and he wanted to put more money into Shield.

I said, “No, I don’t think you should do that. I think that company is not going to make it, unfortunately, but I’m thinking of starting this new thing.” And he was like, “Yes, I’m in.” He didn’t even know what it was.

How much did he put in?

He put in, I think, $125,000 at a $3 million valuation.

Wow. So, he was the first—

First money in.

First money in. Love it.

We had a pilot set up with a first hospital in Wisconsin. This was a clinician we’d worked with before, and he was really enthusiastic. Then we got a call from his assistant saying, “This is all on hold, and I can’t speak to you for at least 3 months.”

Huh.

It was really out of character that he wouldn’t at least call us or text us, or have his assistant tell us. When we dug in, they were getting ready for this thing called COVID-19, because they were expecting to see the first patient in their hospital. That was the first inkling for me of, “Oh, crap, this is going to be a big thing. This is going to be bigger than people think it is.”

6. Covid hits: pivoting overnight into mass testing

They were shutting down the entire hospital. It started off for us as, “Okay, well, we can’t run our clinical studies. We can’t actually launch this product because all the hospitals are on lockdown. Maybe we can go help out with this testing thing for a couple of weeks until all of this blows over, and then we’ll go back to sepsis.”

And so at that point, you’re like, “We’ve moved into COVID-19 testing?”

Yes, yes. It all happened quite quickly, from me saying, “No, no, no, this is not going to be a thing. Don’t worry about it.”

What was the moment when you realized—where were you—like, “This is substantially going to be a real thing”?

I was in my apartment in San Francisco, looking at some data that I think was on Twitter. I was like, “Oh, crap. If that continues at this rate, this is going to be way more substantial than people realize.” This was probably mid-February.

I started to reach out about setting up testing capacity. First of all, we had the problem of finding a lab license, because I had just sold the lab license.

This was the $150,000 you just sold?

I had just sold the lab license, so we didn’t have a lab anymore that was capable of running these kinds of tests. We had a test.

Our chief scientific officer at Curative had, in his spare time, developed a COVID test. One of the things they’d done at a previous company was develop one of these flu tests and offer it to employees to make them feel better. He said, “Hey, can I develop a COVID test? I don’t think it’ll be very useful, but it might make our employees feel good, and it’s a good training exercise for the team.”

They had worked through January and the early part of February on a COVID test, basically in their spare time, in the evenings and on weekends. When everything started to really take off, we already had the test. What we didn’t have was a lab to deploy it in.

And so at that point, you then go back to the old one and buy it for $27 million?

No. I bought a different lab license.

You bought a different lab.

We reached out—I reached out—to a bunch of people I knew in the Bay Area who had facilities with this kind of license. Nobody wanted anything COVID-related on site. Nobody wanted anything to do with it.

I put out an email to everybody I knew. There was actually a guy who was in the same Y Combinator batch as me who had become a VC, and he connected me to a group in LA. They had this license, and they were using it for sports-doping testing. They were in what I thought was LA.

I remember telling Justin, “Oh, Justin, I’m going to LA. I’ll be in San Dimas.” He was like, “Where the hell is San Dimas?” It’s basically very far east of actual LA. It’s still in LA County. It’s a little city best known for Bill and Ted. It’s a little town of about 30,000 people.

We flew out there from San Francisco to look at that lab and at one other lab license that was, I think, affiliated with one of the universities. They had a good space, they had this license, and they were doing pretty minimal testing, so it was just kind of a blank slate.

It started off as a 50/50 joint venture between Curative and the company that had the lab license. We would bring the test and the expertise; they would bring the license. It became pretty clear quite quickly that they didn’t have the expertise to scale it up. They were actively getting in the way of scaling it up.

What did you do?

We bought them out.

And that was the $27 million. Where did you get $27 million from?

Forward revenue from customers. We were getting paid. We had our first testing contract, doing testing for the police and the fire department.

And so, do you have the chief scientific officer who’s created this brilliant test kit?

Yep.

And you go to, like, San Francisco State or the government?

So this was mostly—yeah. Actually, our very first customer was the Sheriff’s Department in San Dimas. We did some private testing for individuals who were paying for the tests, but our first government customer was the Sheriff’s Department in San Dimas.

That came about because they got wind that we were setting up a COVID lab. People were freaking out about it in the town, and one of their sheriffs reached out to me on LinkedIn and was like, “Hey, what are you guys doing?”

I connected with him and explained what we were doing, how it was very safe, and how we had a way of deactivating the COVID as soon as it went into the sample. There was no live virus on site, so we weren’t presenting a risk to the community. This was actually going to be a good thing, and we were going to be hiring a lot of people.

I got him on board that we knew what we were doing and were doing this in a safe way. Then he was like, “Well, we really need testing.” The fire department wanted testing, too.

Our first really big contract was with the city of LA, and that came about from a tweet. We had Laura Deming—

Yeah, I remember. She’s Y Combinator longevity.

Yeah, exactly. She was a friend and was trying to basically help with the pandemic. She drove me down to LA with a car full of PCR machines so I could work on a laptop, and she helped with a lot of the early development work.

She tweeted, “Hey, we’ve got COVID testing capacity. Does anybody want some?” The deputy mayor of LA slid into her DMs and was like, “Yes, please. We would like to talk about that.”

Wow. And so you speak to [laughter] the deputy mayor of LA.

Yeah. They were doing a pilot. They said, “Look, we’ve got a couple of labs. You’re going to have to demonstrate this,” because we were a complete unknown.

And COVID wasn’t at peak ramp-up yet, was it? This was—

This was early March. Lockdowns had just started, and everybody was starting to freak out. There was basically nowhere to get a test. Unless you were ultra-high-risk and in a hospital, there was pretty much no chance you were getting a test.

Everybody was freaking out. This was when everybody was still cleaning their supermarket bags with wipes, nobody knew what was going on, and everything was shutting down. It wasn’t so bad on the West Coast, but New York was really bad already by this point.

And so they’re paying ahead of time?

The best thing we could get with the city of LA—because, obviously, they’re the city and there are certain restrictions—was that they would pay after delivery, but they would pay net 1 on the invoice.

We would deliver the tests for a day, and then we’d send somebody to City Hall the next morning to pick up a check for those tests.

Wow.

The tests had been done and they were paying after we delivered them, which was not your standard net 30 or net 60 for a government contract. We were invoicing them every day for the number of tests they did, and they were having somebody in their finance department get us the check because we needed that to pay for supplies and basically grow out the testing capacity to where they wanted it to be.

What’s the single largest contract you signed?

Probably one of the Florida contracts was maybe the largest. We did a contract with the state of Florida for all of their nursing-home testing. I forget what the dollar figure was, but it was in the hundreds of millions of dollars. They put it out to bid, and we won it.

Hundreds of millions?

Yeah. They tested every employee at every nursing home across the state once a week for a 3-month period. They did a great job of basically keeping things open, keeping these nursing homes open and keeping visitation going, while making sure that the employees of those nursing homes were not spreading COVID to the people in the nursing homes.

They wanted to test every single employee who was working at those nursing homes and exclude the people who had COVID so they weren’t exposing the residents there. We ran this big program.

They put it out to bid, and everybody said, “No, that’s too crazy. That’s impossible. We cannot possibly test that many facilities with that tight a turnaround time. This is impossible.”

7. Building an "orthogonal supply chain" to scale testing 10x

We bid and were like, “Yeah, we can do that. We’ll make that work.” We delivered it.

What did you see that others didn’t?

You have to scale something like that up from scratch. The existing labs—the lab industry in general—is a very low-margin industry, and it’s built on efficiency.

You look at the big labs, Quest and Labcorp, and they are ultra-efficient machines. Some of what they do with automation is incredible. But if you’re asking them to 10x capacity, that’s literally the opposite of what they’re built for.

They are built for getting 1% extra margin by optimizing one part of the process so that it’s perfectly efficient. They’re really good at that, but if you ask them to 10x it, it really doesn’t work. The mindset isn’t there, and the people don’t know how to scale those kinds of things up.

All of the supply chain broke down, so we basically said, “Okay, start from scratch. Throw all of that away. Imagine that you’re going to have to scale this up to hundreds of thousands of tests a day. Where do you start?” We built what we called an orthogonal supply chain, which is basically a fancy way of saying we don’t use the things other people use.

You sound like a McKinsey consultant specializing in innovation: “an orthogonal supply chain.” Great.

I found that it was a good fancy word that was helpful from a sales standpoint. What it basically means is everybody was chasing the same consumables and the same supplies. Everybody was trying to use the same stuff, and if you can make 1x of that, maybe you can increase it to 1.2x. If everybody’s trying to buy that, us also trying to buy that doesn’t help. It doesn’t net increase the number of tests being done, right? It just makes us all squabble over it.

So that’s pointless. You’ve got to find other ways of doing the testing, using supplies that maybe wouldn’t traditionally be used for this kind of testing.

We were sourcing swabs from other types of vendors that were being used for electronic testing and then sterilizing them. There’s this kind of extraction material that you usually use, and magnetic beads are the default standard. There’s another way of doing it with filter plates, which is more scalable because it’s basically just glass and plastic, and you can scale that up faster than you can scale up magnetic beads, where they all come from basically 2 factories in China.

We were like, “Okay, we should never use magnetic beads because that’s not going to scale as a technology. We need to go find vendors who can scale up the plastic and glass manufacturing and partner with them to basically 10x it.” You approach every single bit of the supply chain that way, and you end up with this massive scale.

Outside of a pandemic, that doesn’t work because people don’t want 10x more testing than they wanted yesterday. But within a pandemic, you’ve got to approach it differently. We peaked—I think our peak day was 206,000 people tested in a single day.

206,000 people tested in a single day. That was December 2020, so that was within 8 months, from 0 to 206,000. The company went from about 7 to 7,000 employees in those first 9 months. 7,000 employees in 9 months.

Yeah. It was a little crazy.

Do you sleep at all?

I don’t sleep very much now.

But in that time, what was the craziest thing that you did?

Some of the hiring was pretty crazy. You have to get licensed people for certain roles, but for other, more administrative roles, you don’t need licensed people. A lot of people wanted to work on the pandemic, which was very helpful.

We’d have people line up in the parking lot, socially distanced, down the street, and then give them 5-minute interview slots. We’d just have somebody sit there with a clipboard: 5 minutes, next. It was just to get the volume of people in the door.

8. 5B in revenue, brutal margins, and losing money on vaccines

How much money did you make from COVID testing?

I think the total revenue ended up being about $5 billion over a 3-year period.

$5 billion. Were you the largest private provider?

We were, yeah. We were the largest likely non-LabCorp testing company.

That is extraordinary. What is the margin profile on a COVID test?

Really good during surges and then really bad when there weren’t surges.

When there was a peak—we’d get a new variant, or usually winter was the biggest peak, but then we started having these summer peaks, which was kind of weird—everybody would run to get tested. These were all public testing sites, so these were in parking lots. These were the drive-through tests. That was what we were doing.

If you went to a drive-through testing site, the biggest one was the Dodger Stadium site in Los Angeles. It was 7 lanes of traffic, from 7:00 a.m. to 7:00 p.m., 7 days a week. At the peak, they were testing about 10,000 people a day coming through in their cars, getting tested and coming back to the lab.

When you’re at peak capacity and you’re filling all of the lab’s volume, it’s very profitable. Then those surges subside, and you end up back at testing using 20% or 30% of your capacity. All your fixed costs are the same. You’re still paying 7,000 people. Now you don’t have to buy as many consumables, but all of that infrastructure has to be maintained for the surge.

This is again where it’s the opposite of the traditional lab industry. They have a very flat volume: Every year, people do roughly the same amount of blood work as they did last year, or maybe slightly more, but it’s within a couple of percentage points. Here, you’re building for that peak capacity, and during the lulls, maintaining that capacity is incredibly expensive.

It was necessary, and this was part of the way it was set up. They increased the reimbursement price they were paying for these tests because they needed to incentivize the capacity to be built. If you don’t build that peak capacity, when you have a surge, it all goes horribly wrong and no one can get a test.

That means you basically have to pay to overbuild, because during the dips you have to have that capacity. You can’t just shut it down, right?

You can’t build up 7,000 in 24 hours.

And so you need to maintain that. We would lose a lot of money in every one of the dips, basically.

You would actually lose money?

Yeah. We would lose money on every test during the dips.

Oh, wow.

Yeah.

Of the $5 billion, how much is profit?

After all was said and done, the money that we basically put forward into the insurance business—the health insurance company—was about $500 million that we invested into the health insurance business.

It’s absolutely astonishing.

Yeah.

Can I ask you, when we saw the vaccines roll out, did you know they were ineffective in the way that they’ve kind of turned out to be?

It wasn’t clear at the beginning, and I think it’s also changed. When nobody has had any exposure to COVID, being vaccinated probably provides a lot more benefit. Once everybody’s had COVID a few times, the benefit of the vaccines is much less because you’ve already had it. Also, the variants got weaker and weaker.

When we were first rolling them out, in December 2020, I think there was a benefit for a lot of people getting the vaccine.

Did you get vaccinated?

Yes, we did. We did 2.5 million vaccinations. That was another service we did. We also lost a ton of money on that. That was a terrible business.

Why?

Because the government wasn’t paying enough. We lost money on every single dose. It cost more to administer them than we were getting paid.

Why did you do it?

Giving back. A lot of our partners wanted it. A lot of the partners on the government side that we were working with for testing also wanted us to administer vaccinations.

This sounds awful. Was it hard, with your business with COVID obviously being eased, completely changing and forcing you to pivot again?

Yeah. That started very early for us because it wasn’t going to last very long.

You were always aware it wouldn’t last?

Yes. When we started hiring people at the beginning, we told them, “This is 3 months. You have a job for 3 months. Don’t bank on anything beyond 3 months. This is a 3-month gig, and we’re going to shut it all down in 3 months.”

The CFO and now the president of Curative joined at the beginning, and for her it was going to be a 6-month gig. She came out of retirement to help with the pandemic for 6 months. Now, 6 years later, she’s still here.

9. The pivot to health insurance after Covid winds down

It was supposed to be temporary, and every time we got through a surge, I was like, “All right, that’s it. It’ll be over now.” Then they just kept happening. We started looking at what comes next in the middle of 2020, really early.

How did that search for what comes next change? You just started looking in the middle of 2020. It’s not until the end of 2022, or the start of 2023, when that actual search is activated into a real-time plan, correct?

Yeah. I think we started probably in late 2021. That’s when we got really serious about health insurance. It just took a while to actually get the license.

Why health insurance?

It wasn’t the first idea. We looked at a bunch of other stuff. We looked at other things in the lab testing industry. Unfortunately, it’s just not that big an industry.

We had this interesting technology that could theoretically let you do a lot of lab tests that are individual tests today as just one single test, which would be scientifically quite cool. But even if you say, “Okay, I’m going to displace all of Labcorp and Quest,” that’s about $30 billion of market cap.

The largest company you could possibly build is about $30 billion, which is a big company. But coming out of what we did with COVID, I wanted to build a much bigger company than that, and there’s just not a big enough market in lab testing. The lab testing was out.

Then we briefly looked at trying to buy a hospital, or multiple hospitals. We looked at one in Florida and one in Texas, and the idea was, well, if the hospital is kind of like the health system becoming the center of where care is delivered, they’ve bought up a lot of the primary care offices.

If you can transform that with technology, can you drive much better outcomes? What we ultimately decided is that it doesn't work that well because the payer mix is too broken up. As a hospital, your customer is like 50% the government and then a whole bunch of split-up smaller insurance plans.

They all want different things, and they change their mind every 5 minutes about what they actually want, and you're trying to keep them all happy. So your ability to really change things from the hospital side is quite limited, is what we ended up deciding. When you come back to it, we looked at a bunch of preventative care things and a primary care chain, and everything sort of ends up coming back to the payer.

The payer is the one that drives behavior in the US healthcare system. If you are providing the dollars, people will go where the dollars are. If you say, “I'm going to pay for this service,” people will go do that service. If you say, “I'm not going to pay for this,” people will stop doing that. So the payer is the one that's driving things.

If you could do one thing to change the structure of the US healthcare system today—magic wand—what would you do?

I think you have to break up the negotiating into smaller units. It's gone to this point where I think it's quite an efficient system as a market when the counterparties are small. When everything gets very consolidated, it becomes incredibly inefficient.

When we look at, for example, health systems—we pay for care at health systems—some of that care you can get in other places. If we look at how much we'd pay a primary care doctor who's independent compared to a primary care doctor affiliated with a system, they get paid an average of double. Same service, same credentials.

10. Is SaaS dead? Cutting Salesforce and 80% of SaaS spend

It's just that this one is part of a hospital system, and that hospital system will use the fact that they have a ton of beds, that they have this ultra-special surgery center that you need. We need to have that capacity in our network because some people need to be hospitalized, and some people need those services. If you want to get access to that, you've got to pay me double for my primary care doctors.

When all of the payers are small, when you have smaller payers and smaller hospitals, you end up getting to reasonable negotiations. What's happened is you have these massive payers. The market is ultra-consolidated. You basically have 4 large players that control the entire market on the payer side.

Then you get these ultra-consolidated hospital systems because that's the only way for them to survive if they want to fight with Blue Cross. The only way to survive is to get really big so they have the negotiating power, and then they just reach these loggerheads where nothing gets done. Everybody's overpaying for everything, and everything's inefficient.

When you have more competition in the market, more smaller payers entering, more smaller health systems, you start to get an actual efficient market. When you're just negotiating for, “Hey, I have a third of healthcare in the state, and I have a third of all of the employees in the state,” it's not an efficient market anymore because there's no alternative. You must reach a deal.

If I am sick, is the best place to be treated in the US?

Yes, definitely.

Seriously?

Yeah. The US has access to, by far, the most cutting-edge techniques, facilities, and drugs compared with the rest of the world, and they're willing to spend a lot more.

What do you know now, sorry, that you wish you'd known when you made the pivot into insurance?

I think I wish that I knew AI was coming, because the way we designed the business in 2022, when we first started, we had no idea that this wave of AI and LLMs was coming. We were building a health insurance business because we thought it was a good business to build and we thought it needed to be built. We needed better alternatives in the market for health insurance.

Then, in the last 18 months, how we do pretty much everything is now a completely different workflow. There's so much—I mean, all health insurance does is move bits around, right? We don't have a physical product. We give you a little plastic card, but apart from that, our product is that we move bits around in a database that means care is paid for.

That's it, right?

11. How AI agents replaced entire back-office departments

And we do a lot of managing a marketplace. We work with the providers to negotiate prices. We work with employers to negotiate how much they pay, and then we try to work with employees to keep them healthy. If we can get people to stay healthy, we can avoid the long-term downstream cost of care. Essentially, it's a marketplace business, and that has been fundamentally shifted by AI. But when we first started building, we didn't know that was coming.

By AI.

So much of that back-office work, right, has been completely changed by AI. We now have entire departments that used to be people rubber-stamping things. The first one that went to 0 people was our credentialing department.

This is a process that's incredibly labor-intensive, where you have to check that all doctors who join our network have a valid medical license and aren't being sued for malpractice. This is a person going to the medical board website, checking that the license record is there, checking transcripts from their school, checking a database of who's been sued by whom, and then rubber-stamping.

That used to take us 2 to 3 months on average and cost about $50. We've now built in-house an agent that runs on Claude that does this end to end. It goes to the website, verifies the license, reads the transcript, puts it all together, and stamps it for approval.

We're now averaging about 12 hours' turnaround time for credentialing somebody, and it costs us about 20 cents. This is a mind-numbing process that payers have to do, which is important. We want to know the doctors in our network are validly licensed to practice medicine.

Historically, it's always been kind of terrible, and payers have been bad at it, right? If you're a doctor and you join a network and it takes 3 months before you can see any patients, it's just bureaucracy, right? Doctors hate that, and it's not actually adding the value that it should be adding. It's just creating paperwork.

How many people did you have in credentialing?

That one wasn't that large. I think there were 5 or 6 people. We had a few other departments that have shrunk more than that.

What other departments?

We've seen a lot on the claims side. Claims processing used to be a very manual process where claims came in and people were manually tweaking and editing.

Also on the underwriting side, the process used to be: a broker comes to us with a group—an employer that they're looking to insure—and they ask for competitive bids from multiple different insurance companies. What that means is basically sending us an email with a bunch of PDFs and spreadsheets attached: who are the employees, what current claims do they have, and what does the current insurance look like?

You'd think that over time they would develop a standardized format for how that should run, but no, every single one is a different spreadsheet format and a different PDF. We tried to solve that problem with software and build universal importers and universal intake. It kind of worked, but what we found works amazingly is literally to give the files to an agent and tell it to write Python to get these files into a standardized format, because they're not very good at parsing files, but they're incredibly good at code generation.

You can tell it to write a Python script to convert any random file into this known format, and then test it and loop and iterate on your script until it's working. Then you throw away that script. It's single-use code that never gets used again, and you just generate that code 1 time and then throw it away.

That works so well. Now, brokers, providers, and employers, when people are sending us files, we always used to insist, “Oh, you have to use our standard format for this.” They'd hate it and they'd get mad because somebody's sitting there in a provider office manually reformatting these files into our spreadsheet.

Now, send us whatever you've got, whatever format. It can be scribbles on a napkin; it doesn't matter. The model will figure it out. The model will convert it into our standard format, and it will do it in about 15 minutes.

You build these data-ingestion pipelines that used to be hundreds of people sitting and moving spreadsheets around, and it's now a model writing Python code to do that same thing. Every single time, you throw that Python away and start from scratch.

Dude, I have so many questions to ask on the back of this. The first one is, you mentioned the internal agent buildout that you've done for the company and for your specific processes.

Yeah.

Do you buy the “SaaS is dead” theory that we will—

Why?

Because I see the number of contracts we're canceling.

We just recently canceled our Salesforce contract because we have an internal CRM that was built, that was vibe-coded, that is working better, that is managing our process better, and is more integrated into what we're doing. We run our agents inside of it, and no one was using Salesforce anymore. $600,000 a year. Gone to 0.

Wow. How long did it take?

2 months.

Is it worth it? Because the argument back—I always like to do both sides.

Is it worth the engineering hours to vibe-code that and then maintain it?

The maintenance is definitely one of the most challenging pieces. I agree with that. I think for most businesses of any reasonable scale, yes, it is worth it. Now, whether they will have the tech resources to do that soon, I think that's the bigger question: when will this happen? But when you build those things custom to your workflow, they work better.

Most of these big systems, you're paying an administrator. We had a full-time Salesforce administrator, right? You're paying people whose sole job is to manage this archaic software platform. Not that Salesforce is archaic, but we have a few other internal apps that we were paying for, like industry software, that were taking multiple FTEs to maintain. You can transition that into 1 great engineer, and then whenever you want a custom feature, you just go build it.

Absolutely wild. [laughter] $600,000 a year on Salesforce. Wow. And so we're seeing pockets of software that I think persist because they are more infrastructure-based.

Which persist?

We're seeing a lot of backend stuff, like Sentry, stuff like that, right, where it's become part of your infrastructure. Slack has been notoriously hard internally for us—too many people have built integrations and workflows that are now working in Slack. I think that while they keep putting the prices up, if they put the prices up too much, then eventually it'll make sense to replace that. But what else is on the chopping block?

We're cutting about 80% of our SaaS spend this year. So, in one of our internal meetings, we have a slide of when SaaS contracts are due and whose job it is to tell them that we're not renewing this year.

Well, you can do it in one fell swoop.

Yeah. [laughter]

Well, they have renewals. We have to pay them through the renewal. Is it all legacy software like Salesforce, though?

Some of it's like that. Some of it's very insurance-specific software. So, our claims system, for example, is this massive off-the-shelf platform that we just migrated to a few years ago. This is again why, if I'd known AI was coming, we would have probably approached things differently.

And it's just very hard to use. Their API barely works. It's hard to get the data out of their database. They won't let us manage it. But that's how insurance companies are running things, and so we've built our own claims system completely from scratch in-house. We've now migrated most of the workflows off; we'll be fully off in July.

I'm a health insurer, as are other health insurers.

Yes.

I do not have the in-house capability, potentially technically, to build the agentic workforce that you're building. Am I screwed? [laughter]

I think some of the biggest insurers will struggle because they will not be able to keep up from a margin standpoint with where we can get to with agents. I think some of them do have technical expertise. It's more operational and people ops. If you've built a company of 100,000 people and, in order to get this margin improvement, 50,000 of them have to be laid off, somebody's fiefdom just got a lot smaller.

And so they will do it slowly over 10 years. It will happen. But will it happen quickly? No. And will we be able to compete more effectively in the meantime? Yes.

How do margins change?

Insurance is a very low-margin business. So, 85% of your premium that we collect must go out the door to pay for care. If we get in $1, we have to spend 85 cents. We have to, by law.

If less than that goes out the door, we have to give it back to the employer. That's another thing that's broken about U.S. healthcare, because that drives completely the wrong incentive where, from an insurance company standpoint, if your profit's capped at 15%, the only way to increase profits is to increase total spending, which is not what you want your insurance company incentivized to do.

Why would I encourage people to go to the gym, eat healthily, if actually I'm not going to get that back anyway? This was part of Obamacare. There are some good things in Obamacare, but there were a lot of things where I think the second-order consequence was not considered. It sounds like a great PR thing to say, “We've capped insurance company profits,” right? That sounds good, but it's BS.

It's like capping a CEO's fixed base pay. Sounds great. Yeah, so let's just pay them $27 million in equity compensation. That's the reason why we have such egregious comp packages for execs, because they cap the salary pay. Ridiculous.

With that, how has your Anthropic cost gone? [laughter]

This is one of, I think, the leading indicators for us: our Anthropic cost over the last 6 or 7 months has 6x'ed every month, from a base of a couple of tens of thousands of dollars now up to millions of dollars a month. Eventually, we're going to have to stop that spending increase because it'll get unreasonable, but we just keep finding new things to do with it.

And then the other thing we've found that's been fascinating is that we're seeing a lot of areas where it's not that we are necessarily replacing the team; it's that we're repurposing the team, and they are now so much more productive.

One area that has always been a particularly challenging thing that makes it hard to build a new insurance company is that we have to build this network. The network is all the doctors and all the hospitals and all the people that we have to contract with. There are about 1.2 million of those in the U.S. that you want to have contracted. That ends up being like 60,000 or 70,000 contracts that you have to do. That's just a lot of work to go out, get their attention, get them to do a negotiation, get them to sign an agreement, load all of their data, and have them in your network.

This has been one of the biggest pieces of staying power of the big health plan businesses: they built that over 100 years for Blue Cross and over 50 years for likely UnitedHealth. They did it slowly over a long period of time. If you're trying to, from scratch, come in and start a new health plan, you've got to reach out to all of those doctors and negotiate.

We have a team of about 45 people who do those network contracts, and they reach out and they negotiate. What we launched earlier this year is an agent called Gwen. Gwen does the same workflow. You give her basically a lead: “Hey, there's a primary care office over here. Here's the address.” She will Google it, research them, learn a little bit about their practice, and figure out what other payers are paying them, because there's a lot of this data out there in these transparency files now about how much they're getting paid.

She finds their email address from ZoomInfo, reaches out to them, and then basically pings them repeatedly until they answer her with custom emails like, “Hey, I know about your practice. I know what you're doing,” with customized content for them. When she gets their attention, she negotiates the rates back and forth, usually over multiple rounds of negotiation, and negotiates and redlines the language.

That's another place where we found Python is great. These models are terrible at editing Word documents, but if you tell them to write Python to edit a Word document, they're great at it. Great hack.

And then she signs the agreement. She now signs the agreements with my signature. She'll open up the DocuSign link and then click the button, and it's my signature on that agreement.

This has taken us from doing about 100 contracts a week to about 100 contracts a day. Last year, as an entire team, we did 2,300 contracts. So far, in about the last 8 weeks, the agent alone has done 3,500.

What this is letting us do is that the team doesn't go to zero. We've refocused that team to work on these bigger contracts, right? Some of these deals we can do entirely over email. This agent is email-only, and some of these providers will work completely over email to enter into an agreement. Actually, how many of them will do the whole thing over email surprised me.

There's a lot of millennials, I guess, on the other end who don't want to get on the phone and would rather do the whole negotiation completely electronically, which is fantastic because the model is great at that. [snorts]

But some of them—the bigger hospital systems, the bigger doctor groups—they want to have a phone call. They want to meet in person. They want to learn who we are. The team now gets to spend their time having those in-person meetings, developing those relationships, and working with those bigger groups.

And then, even when it gets to the paperwork, handing the paperwork off to the model, all of the smaller providers—the individual PCP over here, the small behavioral health provider here, the therapist over here—the agent just gets it done and can sign a contract end-to-end in a few hours, where you wouldn't be able to do that volume with people.

Given the transformational nature of what you're describing, if Anthropic doubled their price, would it impact your usage? When we look at a lot of the financials of these core businesses today—

Yeah.

—they are challenged businesses in their current infrastructure and pricing. If they doubled their pricing, would it stay the same?

If I say yes, I don't want our Anthropic rep to double our pricing.

[laughter] But it would work. It would be fine. Yeah. With people, it costs us about $1,500 to $2,000 on average to do a contract. The average with Gwen has been about $70, so it would still work fine. That's partly why the token use has exploded for us.

Am I being a complete idiot, then? If they 5x their pricing, if you went on the labor-displacement theory—

It would still work. It would still work. I think what they're betting, and what we've also seen, is you don't just displace the labor. Contracting is a perfect example. We've not said, "Okay, we're doing 100 a week, so we'll get the agent to do 100 a week."

What we've done is said, "Well, now that we have the agent, we can do 10 times as many contracts this year as we could do last year. So we're going to do 10 times, and then we're going to try and do 20 times." We would just do a lot more volume than you could possibly have done with a human team.

Everyone's like, "Oh, I lose my job. Lose my jobs." Do you think that's warranted?

I think for a lot of these back-office jobs, yes, because—

How do we determine between, "I'm just going to do more"?

Yeah. A lot of people say with developers, we're not going to get rid of developers. There's an insatiable appetite for more software, better software.

That side I do agree with. How do we determine between functions where we'll do more versus we'll be replaced?

What we've tried to differentiate at Curative is that there are 2 areas where we're really investing in people: technical skills and relationships. Those are 2 aspects that I don't see going away anytime soon. We still have a team that's actually deploying all of this AI. They use a ton of AI in all of their day-to-day work, right? They're not writing any code anymore. They're not even reading the code anymore. They're deploying all of this with Claude Code or Codex, and seeing incredible results.

One senior engineer now is so much more productive than they were a year ago that we're investing in having those people. At the same time, there's a side, particularly in health insurance, that is relationship-driven, and that I don't see going away anytime soon. Ultimately, we insure a member, and that member wants to be able to call and talk to a person.

We have a lot of AI they can talk to. The AI is great. They love talking to the AI, but there has to be a person somewhere in the loop. We also work with these provider groups. We have a relationship with that provider group where we're providing a chunk of your revenue. We work with you, you work with us. There's a relationship aspect there that has to be maintained, particularly for the larger groups, by a person.

Then, on the sales side, we sell through a broker, and that broker wants to have a finalist presentation. They want to go to dinner. They want to go and play golf. What we've seen is, on the sales side, that relationship is, if anything, more powerful.

Do you think they still will in 5 years? A lot of people talk about agent-to-agent transactions and how that changes the process. Do you think we will still have that heavy relationship, interpersonal sale in 5 or 10 years?

I think, in some aspects, yes, because in some aspects that becomes the foundation of trust. It's almost the scarce resource, right? If you want to do a deal that's important, then you're going to use your scarce resource of people to manage that, almost like—

It's also that the bigger the contract, the more important it is to have the whites of the eyes, and the trust in the relationship.

And most of these contracts—most employers, even our smallest employers—it's a million-dollar contract at least.

I always think, when you look at accountants and lawyers and a lot of the people who, bluntly, could be replacing some of the simpler work, you would never not have a law firm do it because if it goes wrong, they're getting fired.

Yeah. But I think you'll see it work differently, though. What we're seeing with Gwen is that we had a standard template contract that was drafted by a law firm, and then we have guardrails for what Gwen can agree to. But she just redlines it and then signs it. It doesn't go to a law firm for review.

We're signing hundreds of these contracts a day. It would be too encumbering. It would be too slow, and they would just be reviewing it with AI anyway. So we trust the agent to do that legal review within certain parameters.

In 3 years' time, knowing what you do now about the capabilities that you use it for, how big do you think Anthropic will be?

A lot bigger than they are today.

Do you think it could be $5 trillion?

I think it could be $10 trillion.

Bugger. [laughter] That's just extraordinary, isn't it?

Yeah, because I think you just find all these new things that you can do that you just couldn't do before, that it wasn't possible to do. Gwen is sending, on average, 15,000 emails a day—customized emails to providers that know about their practice, that know about their work.

One of the things we found is that the relentlessness of the follow-up is what works. A lot of providers will get them on the 9th email. There's no way that a human is going to email them 9 times, because you have to have no shame to reach out that many times.

Do you want to hear something funny? You mentioned Salesforce. I got Marc Benioff on the show because I emailed him 53 times, once every week for a year and a week. [laughter]

There we go.

I'm basically an AI model. I lost my personality.

Very effective AI model.

That is extraordinary.

But that works so well in sales, and the best salespeople will do that. It's really hard to scale that, and you end up getting people that reach out 3 times, then give up.

Yeah.

When you're trying to scale something up, if you can scale up that relentlessness, that is really valuable.

So you fundamentally buy that companies will be inherently smaller in the future, and that's why we're seeing layoffs?

Yes.

Are layoffs today just an excuse for overhiring in 2021 and 2022?

I think it's a mix. Yeah. I think there is definitely some of that, and it's also that companies are seeing valuation boosts by doing it. So that's incentivizing maybe bad behavior, but some of it, for sure, is that these workflows are changing.

How big are you today?

We're about 650 people now.

How big will we be in 5 years' time?

Well, in 5 years we'll probably be bigger. In the short term, I think we're going to be quite a bit smaller.

Smaller?

Yeah. We're not done yet with all of these backend workflows.

How does that go to 400?

Somewhere around there.

Wow.

There are some aspects of the business that are clinical workflows. All of our members get a care navigator who stays with them for their entire journey, and that is just going to grow linearly with our membership. We want you to have that human point of contact that's available.

But the care navigators are now getting significantly more useful because they can actually use the agents to do a lot of the follow-up on their behalf. They're not having to remember to reach out to this diabetic member every week about X. They can manage it at a population scale.

So there we're keeping the same headcount relative to our membership growth, but just letting them do so much more than they could do before.

That's amazing. I was speaking to a major airline where they were saying exactly that: premium-care customer service, where they're able to give so much more—recommendations for you and your wife's trip to New York—and everything's so perfected and tailored because all the work that they used to do is gone. For you, as the end consumer, it's amazing. The response time is so much better.

You get a response back in a few minutes. That's the usual place where we see people ask Gwen if she's an AI: when she responds to your email within 5 minutes. In healthcare, if you get a response the same week from an insurance company, you're doing so well.

People think that I'm an AI because I respond very quickly on email and to the point. I'm like, "No, I just have no life." [laughter]

You said about the different data inputs, "Oh, you can just send us anything now." I always thought data cleansing and data structures would be the biggest inhibitor to enterprise adoption of AI. Is that totally wrong, Fred?

I think if you approach it in the right way, then the cleanliness doesn't really matter that much, because the models are so good at cleaning up the data if you give them the right context. That's one of the things we found, actually, with migrating away from some of these SaaS vendors.

We moved away from Looker, Google's Looker product for visualizations. It's super expensive, and we moved to doing it in Snowflake, and it's been a lot cheaper. It's worked really well. Part of that migration was moving all of our dashboards.

All of the things that fed from Looker would have taken probably a year and a whole bunch of engineers and data scientists. We did most of it with an agentic workflow that would spin up, find the next dashboard, figure out how to convert it into what we needed, and then close it down on the Looker side and boot it up on the other side.

It ended up being a project for 1 or 2 people. It still took a couple of months, but it was a lot more doable because we didn't have to have somebody ingest or figure out that data. You can just feed that data into a model and let it figure out how to structure it going forward.

It's just really interesting because I often think about what role doesn't exist today that will be massive in 5 years' time. I thought data cleansing would be one of those roles. If I asked you what role doesn't exist today that you think will be very big in 5 years' time, what would you say?

Agent supervisor.

What does that mean?

One of the things we've found that's been a bottleneck is that when you launch these agent workflows, there are always things that you don't want to let them do everything on, right? With our contracting or sales workflow, there's a certain margin threshold where the sales agent can't promise a client that we'll do it at that margin, but we don't necessarily want it to say no. We want to make a business decision about whether this is the right thing to do for that client.

You end up generating this massive list of approval requests that is now much longer than it would have been because you're doing 10 times as much work. Even if you're only getting an approval request 1% of the time, you're still getting 10%—10 times as many as you were last year. One of the things we found is, how do you manage all of those exceptions that now become a really high volume? We tried agents supervising agents, which I think works to a degree.

Maybe as the models get better, you can also have a more expensive model. If we ever get one that costs $100 per million tokens, you probably wouldn't use it for the core workflow, but you could maybe use it as a supervisor. How you actually manage those agents at scale, with the volume of exceptions that they generate, is a challenge because you don't want them just rubber-stamping yes or no either way. You need a more nuanced decision there.

If you were advising your younger brother or sister on how to prepare for that role, what would you advise them to do to be adequately skilled to do that?

I think just play with the models. A lot of people severely underestimate what they're capable of because maybe they tried ChatGPT 2 years ago [snorts], and they're moving so fast and are so much better than they were even 6 months ago. If you're not relentlessly trying them, then you're going to significantly underestimate them. Where they are today is clearly not where they're going to be in a few years, so you have to skate to where the puck is going to be.

Where will they be in a few years?

Ahead of humans on most capabilities.

Are you excited? [laughter]

Yes, because I think that opens up so many possibilities—unlimited intelligence.

Are you not worried about, in the short term, societal unrest, labor displacement, and what that will do to a hollowing out and an increase in inequality in the US?

I think that can be dealt with by significant action, whether or not we do that.

What significant action would you take to mitigate that?

I mean, I think eventually some version of universal basic income.

Really?

Yeah, and I buy that it works. I think we have to build the social structures that give those people purpose and meaning outside of work because I don't think that we're going to have work, and I also don't think that's a bad thing. A lot of these mid-level jobs that are being replaced are awful jobs. They're people sitting at a desk with fluorescent lamps shining in their face, reviewing random paperwork.

When you're little and you say, “What do you want to be when you grow up?” nobody says, “I want to sit in an office and rubber-stamp insurance forms.” It's not a good job.

I would be worried if my child... [laughter]

Right. These are not—it's not like you're taking some super-aspirational thing away from people. I think these are jobs that we'll look back on and say, “God, I can't believe we had people doing that kind of work. That's crazy.”

You know, I walk with my mother a lot, and I always say my job is to invest in the things that we say, “God, I can't believe we used to do it that way.” I said, “Do you remember? I would never put my credit card on the internet, or you'd never find your husband on the internet.”

You'd never get in a stranger's car and have them drive you where you want to go.

What is insane today that will be incredibly obvious in 10 years? Obviously, you have your card online; obviously, you meet your partner online. What is insane today that you think will be obvious in 10 years?

I think empowering agents to do things on your behalf. We've seen internally that Isaac, our CTO and co-founder, and I have trusted the agents faster than most of the team, and we're okay giving the agent authority to do things.

It was a big internal dispute getting the agent to sign these contracts. The agent opens DocuSign and clicks the sign button, and it's legally binding and has my signature on the page. Getting that figured out internally was very difficult. It took a lot of rounds of convincing people that it was okay and that we could do that.

I think it will take time for people to trust these agents with things like giving it your credit card and letting it go book a holiday. Getting people to trust it to act on your behalf, I think, will take longer.

But I'm thrilled that you signed me your house for $12.

Do you worry about the concentration of value when you look at the Magnificent 7 providing 85% of gains year to date in stock markets, and then Anthropic, OpenAI, maybe 1 or 2 more? Do you worry about that concentration of value?

I'm quite bullish now because I think a lot of what's going on in AI is going to massively boost earnings in other areas of the economy that have struggled to grow earnings any other way. If you're health insurance, how do you grow health insurance earnings? You go chase government business and pay a bunch of lobbyists to get the government to overpay for care.

That has all now backfired, and all the government business—Medicare and Medicaid—is now a bad business, and they're all losing money. Everybody has insurance. So unless you're going to increase the total spending, how do you grow earnings?

If you can make it more efficient so you're not spending 9% of your premium on administrative tasks, that's a way you can grow earnings without having to deliver a worse product.

You're in a really good business as well because it's unambiguously not in the path of the model providers.

Yes. Anthropic is not going to start an insurance company. In the past, we've been big investors in Ramp, which is like business banking. Anthropic is not going into business banking.

In Southeast Asia. [laughter]

I would be surprised. I think things that have some regulation around them and are complex industries are going to see the advantages of the models, but they're not going to see competition from Anthropic or OpenAI.

I totally get that. When I listen to you, I'm like, “Jesus, if I were you, I'd also take a chunk of my money and invest it actively into Anthropic.” Can I ask you, have you taken secondaries along the way?

No, no, we haven't sold any secondaries. There was a dividend at the end of COVID. All the investors got 10x their money back before we started the health insurance company, and then they still have their shares today.

Are you serious? They got 10x their money back, and then they kept the shares.

We didn't have that many investors, but yes, they all did well.

That is an amazing deal. [laughter] 10x and then you keep the shares. Yeah.

What?

Well, I think that's why we've seen them double down, right? They made money with us before, and this last round was led by insiders. How big was the last round?

$150 million.

What was the price?

$1.3 billion.

Wow. Nice round, actually. Not too much dilution—enough that it's really impactful cash-wise to come in.

Yeah.

Wow, dude. That's insane. So, can I ask you then personally? I asked this, actually—do you know Josh Browder? He's another Brit in the Valley.

Okay, a phenomenal guy. When you look at your personal allocation today, given our inside access and what we know, is there anything funky that you do with your money outside of Curative?

12. Subcritical: building a fundamentally safer nuclear reactor

Yeah, I invest primarily in companies founded by people that I know, and I do very little investing if I don't know the founders.

Does that work well?

It's had mixed results, but some of them are too early to tell. Some of them are the best investments.

They're all a bit too early to tell. [laughter] Do you have any energy investments?

Yes. There is a company that I co-founded with my wife, Subcritical, that is in the nuclear fission space. This was based on an idea that I had a few years ago: we need more power, and nuclear is a really good way to do this.

It started off as me looking for an investment. This was one of the first times I thought, “We should find a company that's doing nuclear power and try to invest in it and see if we can make it go faster.”

Because I kind of thought I'm pretty good at making things go faster in really regulated spaces. That's kind of what I'm good at.

That's your thing.

Yeah, that's my thing. Everybody's got to have a thing.

Is that your hook on the first date?

Regulated industries make a good first hook on our first date. After our first date, we both shared our genome files with each other, our VCFs. She said she'd done this before, and the guy thought it was really strange. We were both like, “Oh, we should share our genomes and then compare them and check that we were compatible, so it was worth having a second date.” We were both totally into that.

So we knew it was meant to be. We were compatible by genome. We have 2 beautiful kids, so we knew it was meant to be.

I'm sorry. If you're incompatible by genome, you have a ginger child.

Well, that was a concern. My brother is ginger, so I carry the ginger gene.

My brother is ginger, too.

We don't see him anymore. We took him to the woods and said, “Run free.”

Makes sense. [laughter]

I do carry the ginger gene. If she had carried the ginger gene, that would have been a deep concern, but she luckily doesn't. That was one of the key tests.

You progressed to the second date.

Yes, so we made it to the second date.

What does no one know about nuclear that everyone should know about nuclear?

That it is very safe, and that it's not a science or engineering problem. When we started looking at companies to invest in, that was the thing that I was disappointed by: everybody was approaching it as if nuclear is this massive engineering challenge.

Sure, the engineering is hard. It is complicated. But fundamentally, we have built safe nuclear reactors since the '60s. They work great. The technology has not really changed or progressed since then. We know how to build these. That's not the problem.

The problem is that, due to a lot of the anti-nuclear push in the '80s, we have had a regulatory environment that has been incredibly restrictive and difficult to get new nuclear reactors built, particularly in the US, but also worldwide. There's been this push to say, “How do you guarantee that under any possible circumstance, like once-in-a-million-year events, you will never have anything go wrong?”

In traditional nuclear, that is very hard to guarantee. One of the reasons it's difficult is that you're basically balancing on this knife edge. In a reactor, you have what's called criticality, which is where you have to produce enough neutrons each generation that they go off and do exactly 1 more reaction, and it keeps itself going.

If you get too much of that—too many neutrons—it's a bomb, right? It will be a runaway reaction, and it will blow up. That's very bad. That's only ever happened once by accident, which is Chernobyl. All the others have not been criticality events.

If there are not enough, then it just turns off. If you go too far below this exact 1.0 threshold, you get no power out. You're trying to balance perfectly on that knife edge of exactly 1.0, where you can control it. That is a hard problem to guarantee.

This is the fundamental issue with nuclear regulation: how do you guarantee that under no possible circumstances will you deviate from that perfect control? I was initially pretty disheartened. I was like, “Well, we're not going to get new nuclear power. This is not going to work.”

Then I stumbled on this idea of what's called the energy amplifier. It's not a new technology; it's been around since the late '80s, early '90s. It was really pushed by a guy, likely Carlo Rubbia, who used to be the CERN director. He was a Nobel laureate in physics.

The idea is you always operate below that 1.0 threshold. We are designed to operate at 0.97. That means you never have enough neutrons to keep the reaction going. The reaction will always fizzle out. No matter what you do, it's going to fizzle out.

Normally, that would mean you get no power output. What you do in the energy amplifier is point a really powerful particle accelerator at that fuel, and that puts in the extra neutrons to drive the reaction forward. But if you turn that accelerator off, all of your energy output just stops.

You basically have this big on-off switch where you can control fission, and you can guarantee that no matter what you do to it, the fission will never run away. Even if you put in 10 times as much power from the accelerator, it will never run away. There's nothing you can do to cause it to go critical or to have a criticality accident.

It's a fundamentally safer way of doing nuclear fission that is just approaching it from a different angle.

How will the composition of our energy provision change in the next 5 to 10 years? Will nuclear be a demonstrably larger part of energy provision than it is today?

Yes. I think what we're seeing all across the supply chain in nuclear is a push to get more nuclear online. Subcritical is leading the way there with a faster path to market than any of the other players. There's a lot of people working on deploying a lot of new nuclear power.

Which current provision will diminish significantly?

I think any power from coal will mostly go away. I think you're still going to see a lot of gas because, particularly in the US, it's cheap, it works, and it's fast. But I think coal is going to go away. Then you're just going to see more of everything.

What company will be larger, Curative or Subcritical?

Subcritical.

Or Subcritical.

That's a great question. Curative has a larger market opportunity, but I think they're both—

Power generation?

Yeah. The US spends—or US employers spend—$1.5 trillion a year on healthcare, which is our direct TAM every single year.

How much does the US spend on energy that can be addressed through nuclear?

It's a similar order of magnitude.

I mean, you chose good TAMs.

They're both yield optimization. I feel like you've really taken this—

I figured out the TAM thing. [laughter]

No, they're both trillion-dollar opportunities if we execute right.

Yeah.

We're also seeing AI on the nuclear side in the design. Design is traditionally something that's done by a whole bunch of people sitting and doing drawings and mechanical engineering, and the models have gotten really good at that. You can do the design with far fewer people, using AI to optimize a lot of the design parameters.

Historically, you might have needed 100 mechanical engineers to design every single nut and bolt and part. You can do it with 20 really good mechanical engineers who are designing the critical pieces, the important pieces, and overseeing the AI. If you need a little bracket that joins this piece to this piece, that doesn't need a human to design.

I was actually meeting a company the other day that basically said the challenge with hardware engineers is they don't often know what software engineering is. The beauty of today is we've turned hardware engineers into software engineers overnight.

Yeah, and that's amazing. Well, it's another place where we saw code generation as the solution. I think this is one of the bets Anthropic made, and they're totally right on. You can generate really good CAD models by having it write Python to make the CAD model.

It's not necessarily good at 3D-space visualization or outputting a drawing as vectors, but it's really, really good at generating plausible Python code that can draw that part. It is the most exciting time to be alive in many respects.

Yeah. That's why we ended up starting Subcritical. I'm very busy running Curative, but that was an idea that was just too important to pass up, and there was nobody else.

The only one that's under active construction of those systems is in China, based on a US design from the 2010s that the US stopped working on after Fukushima.

How much money do you need to make Subcritical significant?

Each one of our deployments would be about $1 billion of construction cost for a 300-megawatt facility. But it wouldn't be the same—you wouldn't raise that as equity. It would be a mix of equity and debt into the plant, so it's a different kind of financing. It's more infrastructure-build financing.

What do you know now about marriage that you wish you'd known at the beginning? Seriously, it's an amazing thing to build a company with your wife.

It's a challenging thing as well.

Yes. How do you make it work?

I think we're very well matched, which is one of the things. We basically never argue. That's how I knew very early on that it was meant to be: we're always on the same page about things.

It's actually very easy to run a company together because we usually see eye to eye on how something should be done.

Fatherhood. You said you have 2 kids: a 2-and-a-half-year-old and a 6-month-old. Is there anything that you would advise a new father, knowing what you know now?

You should definitely have kids. Don't wait. I think there's too much sentiment of people.

Oh, live your life and wait until you're in your late 30s and then have kids. I think, no, have kids early, when you have the energy and can run around and not sleep. It's one of the best things you'll ever do. You should just get on with it. [laughter]

Okay, we're going to do a quick fire. Sound good?

Yep.

Dude, that was the most twisting and turning conversation ever, from the proliferation of STDs to fatherhood and nuclear. I mean, really, we crushed it. What have you changed your mind on most in the last 12 months?

I think probably a year ago, I would have said that there are workflows that can't be done with today's models. I think today, the current-gen models can do every back-office task we have at Curative. It's just a matter of deploying them, getting them set up, getting them configured, and having the right policies.

A year ago, I thought there was an opportunity and that there were things we could do, but I don't think I would have said you could do every single one of our current back-office flows.

What one change would you make to Europe if I made you president of Europe, in this very strange title, to stay in the race for competitiveness?

You have to have some kind of burden for passing regulation. There needs to be some penalty. Right now, you pass a regulation and it's like, "Okay, you did a good job." The goal is to pass regulation.

There has to be some penalty. If you pass regulation, your country must pay some additional tax for having passed that regulation. Just adding and adding and adding, without refining what you've got today and really digging into how this regulation is affecting things on the ground—just more additive regulation is bad. You need to be looking at the effect of what you've done, refining it, and iterating on it, and not just trying to add some new landmark regulation.

It's very anti-European, Fred. You're not going to [laughter] do anything. You're not going to do very well here for a reason. I mean, you know, I'm a Texan now.

Likely Marc Benioff said he spent $300 million on Anthropic. Equated across the developers that they have, it works out to be about 3.8% of developer salary spent on Anthropic. What do you think the total percentage of developer salary spend will be on Anthropic in 3 years' time?

Between maybe 2 and 5x—2 to 5x the salary. I think that's probably—

2 to 5x is the whole salary.

Yeah. Because I think the way we're driving workflows is that you have 1 senior engineer managing a bunch of downstream agents that are actually doing the work. We're now getting to the point where we have mostly unsupervised agents taking feedback from the team on things, implementing features, and then the engineers are coming in and actually checking that what it built makes sense. So they're becoming more the reviewer and the architect.

Whoa.

Yeah. I think that's what the workflows will be: people are going to be deploying more agents than engineers, and they're going to keep the same number of engineers. We're just going to build a lot more.

You're doing 2 to 5x. I mean, that's not like 3.8% to 20%—50%. If it's 50%, Anthropic's like a $20 trillion company.

I'm going to message my friend to let me into that new Anthropic round. [laughter] Just message Larry. [snorts] There we go. What's the kindest thing anyone's ever done for you?

I think when I first was getting started, there were a lot of people who helped make it possible to move to the US and kind of made a bet on a kid coming from the north of England to come to Silicon Valley. Some of the earliest investors helped with that. Josh Buckley was one of the first guys who invested in us during the YC batch, just because he liked what we were doing and thought it was cool. He was willing to take a bet on a kid.

You know, Josh is like my best friend.

I didn't know that. I haven't seen him in a while.

Yeah, I say hi to him.

I speak to Josh every single night.

Okay, barring, say, Christmas.

All right. Well, he invested in Curative.

That is amazing.

And then STDs. [laughter and snorts]

That's amazing. I didn't know that about Josh.

Yeah. A month into the YC batch, he came by the lab and was super supportive of what we were doing. I think, just coming from the British background, we couldn't even get meetings with investors.

And he was young. I mean—

Yes. But to get—I mean, he'd been through YC and had a successful company, and it was just awesome to have someone like that take a bet on what you're doing. Coming from the UK, where I was used to the cold shoulder and no one was interested in what I was building and no one wanted to take a meeting. [laughter]

That makes me so happy to hear. Okay, final one. What's the best advice that you've been given?

I think one thing that I have learned is to always try and get a lot of different perspectives on a problem. I would historically have approached things from one scientific viewpoint, and sometimes people would say, "No, take a step back and think about that problem more broadly." One of the things I learned during the Curative COVID push is that we had to bring together a bunch of people from very different backgrounds.

We hired a bunch of former military people who were incredible at deployment, but they spoke a different language. Then we were trying to get them to talk to scientists, and we hired a bunch of Silicon Valley developers. They all think about the problem and are all trying to solve the problem, but they all come at it from a completely different perspective. A lot of times, I wouldn't have considered that point of view on doing it.

I think what I found is that the more of those perspectives you can get on a problem, the closer to ground truth you get. You're never going to—no, not one of those people is going to give you the ground truth. But if you hear a lot of perspectives, you can get to that ground truth faster.

Fred, that was the most extraordinary show that I've ever done in breadth, depth, and variance of conversation. Thank you so much for joining me, and it's so great to do it in person.

Yeah, thanks for having me.