From Ether to Atoms

About a year ago, I picked up a copy of Abundance, a book written by Ezra Klein and Derek Thompson that examines the reasons why there seems to be a lack of progress for ambitious physical projects in the US. I won't get into specifics, or whether or not I agree with its core arguments (I do, with caveats), but the opening montage stuck with me.

In it, the authors paint a beautiful, albeit fantastical, picture of the future: vertical farms producing our fruits and vegetables, autonomous drones delivering our parcels and packages, and ultra-fast jets getting us from New York to London in under two hours. It left me yearning for this utopia. A car that drives me to work so I can doomscroll for an hour? A drone delivering my DoorDash order while my vegetables continue to sit and rot in my fridge? Sign me up!

Turns out, I'm not the only one left wanting.

In 2022, a16z launched its American Dynamism practice, placing bets on companies solving critical national problems in aerospace, housing, manufacturing, and defense. Today, one in five VCs allocates at least 10% of their investments to US hardware, according to a report by Silicon Valley Bank — more than double the share who said the same in 2022. They're betting that the world wants to see more physical changes, and they're moving their capital to see that it gets done.

The timing isn't a coincidence. In 2022, OpenAI released ChatGPT to the public and the ground moved.

AI, in addition to less government red tape, as the book posits, is what makes the physical world investable again. Hardware used to be venture poison because it couldn't scale like software — but a robot that learns doesn't need to be reprogrammed for every task, and a fleet running one model improves everywhere at once. AI made hardware scalable because it's now just software that shares the floor with humans.

But excited as I am, the annoying insurance voice in me asks: how are we to insure this world? Because once software starts acting physically, errors stop being bugs. They become losses.

In a previous piece, I wrote about the insurability of massive physical structures like data centers on the property side. Inside those structures, though, there are liability risks that still need to be accounted for.

Who's responsible for the autonomous vehicle that miscalculates and causes serious harm to its passengers – the owner's auto policy, or the manufacturer's products liability? Or how do we insure against a fleet of humanoid robots whose software update causes them all to malfunction at once? Is that a product defect, a cyber event, or a general liability claim?

For venture investors backing physical AI, questions about liability and whether their investments can get insurance are important because they affect their portfolio companies' ability to secure two critical pieces: capital and customers.

Capital

Loan agreements require insurance. Specific coverages, specific limits, maintained for the life of the loan, with the lender named as an additional insured. Physical products are enormously capital intensive, which means debt financing is a necessary part of the capital stack. But a robotics company financing a fleet is asking a lender to accept collateral whose liability exposure can't be priced yet — and lenders respond to unpriceable risk by demanding more coverage, not less. The requirement tightens where the capital need is greatest.

Customers

I used to run channel sales at a B2B SaaS company. Whenever we got new customers, the redline song and dance would start. Once in a while, someone would ask us to increase our insurance indemnity limits, and we would oblige knowing that our carrier wouldn't sweat a few extra million in coverage for an innocuous SaaS product. But if I were to sell humanoid robots to be deployed on construction sites, in hospitals, as a part of the customer's new workforce – with the ability to think and act autonomously – getting adequate insurance would be a different proposition entirely. Physical AI touches general liability, workers’ compensation, and cyber coverage all at once, and today's insurance offerings aren't built for a product that multidimensional. That leaves two choices: cobble together policies to close the gap, or build new products that meet the moment.

To be clear, no serious company is uninsured. The business does what the business does and insurance follows. Always. That is why semi-autonomous vehicles sit on personal auto policies; fully driverless ones increasingly fall to the manufacturer's product liability. Industrial robots have been covered under workers' comp and products liability for decades. Until recently, humanoid robot exposure was small enough that carriers had no reason to build products for a market that didn't yet exist.

But that's changing. Fast. Morgan Stanley predicts that up to 13 million humanoid robots could walk among us by 2035. By 2050, that figure is estimated to jump to 1 billion.

1 billion. Only two countries on earth have populations over 1 billion.

While most of these robots are pegged for commercial use, up to 80 million are expected to live in human homes. You can be skeptical about Morgan Stanley's divinations, but the money behind the industry is real. Tesla announced its goal to produce 10 million units a year in its California and Texas plants. China is pouring billions into the race too — Nvidia's most bleeding-edge chip will run in the bots built by Unitree, the country's foremost robot manufacturer.

A billion robots acting autonomously is an exposure current policies aren't equipped to carry. The industry flagged this back in 2017, when Swiss Re published a report detailing exactly how coverage falls short for these machines. Since then, Berkshire Hathaway, Chubb, and Travelers have sought state regulatory approval to exclude AI-related damages from general liability policies outright — and regulators approved more than 80% of those requests, according to PYMNTS. Coverage hasn't improved. If anything, it's grown narrower while the technology expands.

The reason carriers keep pulling back is simple: physical AI breaks the way insurance works.

Insurance works because losses are independent. A thousand drivers on the road make a thousand separate decisions. One may be doing his best impression of an F1 driver and another is as cautious as a first-time father driving home from the hospital. Their decisions don't correlate, which is why insurers can pool them, price them, and stay solvent when the claims roll in.

A thousand robots running the same model don't make a thousand separate decisions. They make one decision, a thousand times.

A single bad software update or one adversarial input and the entire fleet fails the same way. It's an entirely different kind of risk. Whereas human negligence can produce a claim, model negligence can produce a class action.

The industry isn't entirely unfamiliar with this — cyber insurance wrestled with the same vulnerability, and it took years of painful pricing to get right. Today, robot fleets stage their rollouts and canary their updates because operators know the danger exists. But cyber losses are typically data breaches. Physical AI losses can mean bodily injury at scale.

The opportunity

Insurance markets don't form when a risk appears. They form years later, once the premium justifies the work of building policy language, filing rates with fifty states, and convincing reinsurers to stand behind it. It's laborious work, but every technology that scaled needed that layer built first — and building it was itself a massive commercial opportunity, every time.

Consider aviation. Commercial flying took off in the 1950s when risk pools formed specifically to insure commercial jet travel. Insurance gave financiers and passengers alike the confidence to bet on flight. It allowed 31 million people to take to the skies in that decade. Now, there are roughly 4 billion global air travelers annually, and the specialty aviation market that grew alongside them generates billions in annual premium for carriers like AIG and Sompo.

Same pattern in auto, in nuclear power, and more recently in cyber. No insurance, no mass adoption — insurers had to build the product before drivers got behind the wheel, before plants could get financed, before companies could collect and store data at scale. The MGAs that didn't shy away from cyber risks became billion-dollar businesses. In each case, the insurance layer was a precondition for growth.

Physical AI is next. The insurance layer doesn't exist yet in a meaningful way, but the companies that build it will enable physical AI to scale and make money along the way.

So where to invest?

In my data center piece, the core reason for the insurance gap on the property side was that the assets in question have concentrated value and the data needed to underwrite them is scattered across municipality records, inspection reports, and IoT feeds.

With physical AI, the data lives inside the products themselves. Autonomous vehicles and humanoid robots are covered in sensors perceiving the world around them — enough data to make an actuary's head spin. The question now isn't where the data lives. It's whether the manufacturer is willing to share it. A Tesla or a Unitree would be reluctant to hand over complete access to proprietary data just to get insured. So for companies on the edge of insurability, the amount of data they share with carriers gets negotiated right alongside pricing and terms — often through bordereaux reports, where the type and volume of data shared are predefined, deal by deal.

Therein lies the opportunity. If someone can standardize that exchange across the physical AI category, they could own the intelligence layer. And once that layer exists, an MGA can use it to do what ibott did for scooters.

Ibott (insuring businesses of tomorrow, today) was a team within Apollo Group Holdings that first extended coverage to scooter companies when no one else would touch them. In January 2026, Apollo Group was acquired by Skyward Specialty Insurance for $555M. Physical AI will be orders of magnitude larger than scooters, which means the company that plays ibott's role in this category will be worth considerably more than $555 million. Once that MGA proves the risk is priceable, the reinsurance stack deploys behind it at scale. That's how a market gets built.

The proprietary data collector turned MGA playbook works because in insurance, data is what informs capital allocation. Whoever has the edge in knowing when and where to deploy capital ahead of the general market, wins.

The bleeding edge of technology has largely lived in the ether. We cannot feel it, touch it, savor it. We experience it through portals in the form of phones and desktops, but we don't live in it, transport in it, rub elbows with it. Abundance blames government inefficiency for a future the Jetsons would be disappointed in. While I agree in many ways, I'd also add insurance to the list of reasons we're not there yet. Structures don't get built if there's no one covering them. Robots won't get deployed if no one's responsible for them. If that's the future we're yearning for and looking to build, then insurance is the scaffolding. The data is there. The exposure is arriving. The market is forming right now and the open question is who gets the data first.