Personally, I’m much more interested in new discoveries than new automations. And yet, much of the capital and attention in tech has gone to the latter. But a new generation of startups is taking on scientific discovery head-on, using AI to uncover (then monetize) the hidden secrets of the universe. I couldn’t find a quality piece that explained why this is the moment and who the players are, so I found someone to write it: Arnold Engel. He’s a writer, builder, and former VC-backed founder who writes about business and markets on his blog, Margin Points. For this essay, he got exclusive interviews with multiple players in the space, and I learned a TON reading it.
First though, this piece is brought to you by DeleteMe.
If you read this newsletter, you probably have a public profile. LinkedIn says where you work and what you raised. A podcast or a random blog post will show how you think. Data brokers fill in the rest, by offering your home address, phone number, spouse, kids, property value, up for sale. Put those together and a stranger has everything needed to show up at your door, or to prompt inject your AI agent with a story that checks out. Companies figured this out years ago, which is why over 20% of the Fortune 500 pay to scrub their executives. Founders, investors, and operators carry the same exposure with no fancy security team behind them.
I noticed this problem when the scam attempts I was getting kept getting more personal with my information. DeleteMe removes that information from hundreds of data broker sites and keeps removing it all year, with real privacy advisors you can reach for custom requests. If you are online in any way, I highly recommend you take steps to protect yourself.
I partnered with DeleteMe to get Leverage readers 20% off at with code THELEVERAGE.
And with that, here’s Arnold.
Much of tech looks at AI as labor replacement, which misses the exploration work that is being scaled up right now that will come to dwarf it. I’ve talked with dozens of founders, investors and scientists operating autonomous labs over several months–AI is now driving autonomous science and it’s both misunderstood and underestimated.
The fully autonomous lab has arrived—where AI models have learned each instrument and device’s capabilities and operate experiments without human involvement. When Astra came out, it set a baseline for in-lab capability—operating experiments at roughly a 100x faster clip than PhD grad students.
AI enabling autonomous science, discovery, and exploration is the biggest bull case for AI—bottomless compute demand and next phases of AI build-out—that have precisely nothing to do with chatbots, personal agents, AI lawyers, or coding agents. Every other domain could see underwater economics, slow adoption, and profit skinning competition and it wouldn’t matter.
The motherlode is sitting right in front of us now.
AI isn’t all labor replacement
A common mental model for AI market size distills to: AI can replace some software engineering and legal work and both of those are big markets, so the TAM for AI must be big. Despite some bloviating, there really isn’t demand for 1000x more legal work to be done or 1000x more accounting software—probably not 1000x more investment banking, either? Bearish AI cases typically rotate around human adoption of AI, how much labor can really be replaced, or how much profit will be extracted from labor replacement when considering competition. These first domains have lulled us into discounting the bigger unbounded science market.
The constraint for science isn’t people or ideas, it’s verifying experiments. LLMs on their own can’t do the type of science that survives contact with the real world—they can read every paper and theorize down every rabbit hole but they can’t verify their hypotheses in a physical space. We’re now in the phase of quickly moving to bridge the verification gap.
The science flywheel (Aim, Meet the real world, Learn)
Fans of business flywheels will recognize a powerful one in the sciences. Generate hypothesis, collect data, analyze, then generate a new hypothesis—all running in a loop. Each turn of the flywheel can make the next turn more efficient as the system learns, executes better experiments, and picks targets for subsequent experiments. Traditional LLMs do this with “make a plan, try it, and test if the code works” but the difference here is that the data generated is physical.
As the market shapes up, each competitor picks the parts of the flywheel they want to own and rent. You can own the aiming, you can own the learning, but if you never run a real-life experiment you might well end up like startups who toil with no customer contact or revenue.
The aiming—the research direction—is like deciding where to point a bazooka. Physical Superintelligence ($58M raised) has the scientists aiming and setting direction with the AI agents executing experiments. Of course, the reverse will be tried. AI chooses and runs the experiments—humans help move things around where needed.
Meeting the real world could happen both in massive warehouses converted to autonomous labs or in small distributed garages. The real world verifiability is crucial for learning—it ensures that there isn’t drift for the model and there is always a ground truth to verify. The meeting the real world has implications for the AI-buildout. While autonomous labs could be consumers of a bottomless amount of frontier compute, they also require capex, physical space, power, and at least some human oversight and maintenance.
Warehouses with rows of benches might be only part of the story, and where capex can start to expand beyond what we think of for data centers. In 1850, Foucault built a rotating mirror device to measure the speed of light. In 1924, Michelson built 22 miles of piping in the San Gabriel mountains outside LA to measure it more accurately. Experiments won’t be confined to labs and that changes the profile of capex completely. Already, we see with Mistral and CERN—the capex heavy European particle accelerator—partnering how AI will lead to bigger and bigger experiments.
While the flywheel mechanism underpins a new set of science companies, the power dynamics for AI scientists are playing out and the winners aren’t yet clear.
What you are betting on
There are three tiers of breakthroughs that are accessible now using AI.
The incremental breakthrough happens when a process gets incrementally more efficient. Things like optical power efficiency or battery life are examples of incremental breakthroughs that can be driven by AI. These are primarily driven via efficiency, where thousands of minuscule improvements compound into a significantly improved outcome.
The novel breakthrough is predicated on AI combining or applying technology in novel ways to achieve outcomes that humans haven’t tried before. The breakthroughs are in algorithms or applied technology or experiment design and often combine disparate ideas or approaches that seem surprising to scientists. These have already been observed at two labs I have spoken to.
The fundamental breakthrough has not been observed yet, but would be discovering new physics, for example. Such a breakthrough would likely come from the same combination of efficiency and novelty that leads to other types of breakthroughs. The bull case doesn’t need new science—incremental breakthroughs alone at a 10x clip increase throughput across every domain which looks fundamental from the outside.
As inevitable as this is, as we were all reminded with l’affaire Situational Awareness, you want to be right and be right on timing when highly leveraged. Large amounts of venture capital into a company tend to make the bet timeline compressed.
If you are an AI scientist company, you have agency over what the timeline for breakthroughs will be. Think you can come up with true gamechangers (superconductors)? Invest toward the horizon where you think you’ll be able to achieve it. Don’t have that kind of runway? Scale back to immediate improvements to semiconductors.
The short horizon scientists working on comparatively incremental improvements fear being blown out of the water by scientists discovering moonshots or entirely new domains of exploration. The moonshot hunters fear being subsumed by short horizon scientists racking up fast win after fast win, hoovering up capital and pulling in the best talent and compute.1
In addition to finding amazing breakthroughs and incremental improvements, the new AI scientists have to commercialize them. Novel science can be sold to corporations in the domain who already have the capabilities to bring them to market, or to investors in the form of spinoffs. The latter is more significant now that we are firmly in the era of capital aggregation. An AI scientist that reliably spins off concepts that could take large amounts of capital to develop and provide a return on that capital is something close to the Softbank golden goose.2
Commercialization is easier to conceptualize at massive scales today: GLP-1s showed how massive opportunities can be captured and applied science superiority can be monetized faster than at any point in history. Commercialization in bio has been better ground than other science domains but you’d expect others to move along accordingly with innovations.
Lila: breadth
There’s a bet here that progress in one scientific domain will cross-over and help in other domains. On the surface, this assumption seems reasonable given that LLMs have shown this tendency for cross-domain improvement in non-science areas.
The assumption that a great chemistry model will have advantages in pushing the frontier of biology underpins the thesis for Lila ($550M raised). “Breadth gives depth faster” according to Jon Hennek, CRO and Chief Product Officer at Lila. Jon spoke to me and offered a tie between domains: polymers (materials science) and proteins (biology) are both just large molecules in different contexts. Catalysis chemistry and enzymology are the same underlying physics. Lila is seeing glimpses of cross-domain benefits: training in multiple domains increased holdout performance in areas where no training data was provided.
All AI scientist companies believe that AI can learn from verifiable experiments in the lab, but Lila is the biggest bet on cross-domain work. Lila spun out of Flagship Pioneering, which is severely underappreciated in tech circles, who previously spun out Moderna and who currently has over $10B of assets under management. Flagship provides Lila with a deep bench of scientists across domains and a robust commercialization and partnership arm.
Lila started commercializing ten months ago and they think that the head start combined with Flagship’s GTM prowess give them a massive lead over everyone else in the space. Currently, commercial partners have Lila scientists running campaigns in Lila’s lab but they are working to extend Lila’s model access and AI-enablement into partner labs. Data from experiments is trainable by Lila, but there’s a lot of data from partners that Lila isn’t privy to.
Lila claims to have only trained on 2% of the data it has collected (compute-limited) and boasts a proprietary dataset of ~10 trillion tokens of scientific reasoning based on verified experimental data. They also have over 10,000 questions across different science subdomains for use in model evaluation, training and benchmarking. That question bank alone could be worth $100M if high quality, based on an estimate from friends at two leading data companies. According to Lila’s own testing, their Iris model outperforms frontier models as well as PhD scientists with expertise in the specific domain being tested.
Lila is enthusiastic about flexing its commercial partnerships on big swings where they have active therapeutic programs with high impact if successful. These can be achieved via rapid incremental breakthroughs—they claim 10x faster R&D than prior efforts coming out of Flagship Pioneering—and don’t necessarily rely on novelty or fundamental breakthroughs. The downside for Lila (AI is simply brute force without novelty) already seems to be covered off and likely profitable in and of itself.
Periodic: depth
Periodic Labs ($300M raised) recently released details about their own model, Neon, which was trained on top of an open weight model based on experimental data and reasoning from their own lab. Periodic has been focused on characterization experiments in their real world lab and the model performed at a comparable level to Astra at characterization. Characterization (what is this material?) can be thought of as a building block for work on superconductors, which is the moonshot science that venture capital is underwriting with Periodic.
The early results for Neon can be interpreted in several ways. That Periodic could achieve state-of-the-art equivalency on characterization at a fraction of the training cost as OpenAI incurred for Astra is impressive. However, OpenAI likely had no training data on this type of characterization work and still performed at the frontier on the tasks.
The cost per unit of work for Neon was roughly half that of Astra, which shows the advantage of a model trained on real experiments according to supporters—but the math flatters Neon. The comparison of cost is based on an internal cost for Neon (the model isn’t being served at all except within the company) and compared to the retail pricing for Astra. If we assume OpenAI has 50-70% margins on retail inference, Neon’s cost advantage disappears and might even invert.
Neon opens up more questions than it answers. Would incorporating much more physics experimentation help Neon outperform? Is it necessary to pull experiments from other domains like chemistry to avoid a scientific bitter lesson? More so than others—is this the right time to be spending cash on training, rather than on rapid experimentation? Periodic has a cost advantage over a competitive physics AI scientist in a narrow subdomain but no advantage over OpenAI or Anthropic.
Who owns the loop?
OpenAI did nothing in a specific domain for a year and kept parity with a $300M startup staffed with great talent from OpenAI—which may well breed complacency. Attention is the moat for the science-only players, who are confident that OpenAI and Anthropic (and to some extent Alphabet, Meta, and SpaceX)5 have too many other priorities to dedicate attention. Compute-only players run up against a cornucopia of hypotheses and lack the capabilities to quickly verify them in the lab.
Today, the attention gap is shifting quickly. Anthropic has a wet lab in San Francisco. They want to own the whole flywheel (at least in bio) from aiming, to contact with the real world, to learning. Is Anthropic now directly competitive with Lila? The answer isn’t simple because the market isn’t bounded—Anthropic beating Lila to a specific discovery might have absolutely no impact on the revenue potential of Lila. Lila and Anthropic might be aiming at completely different subdomains for all we know—and for all they know.
We are entering an era of discovery and the only competitive issues are clustering of the exploration. Harvey v. Legora is competitive in the legal AI because neither is advancing new areas of law or developing novel theories. Lila v. Anthropic has no similarity to Harvey v. Legora whatsoever.
Lila is going to be opening up its AI model to customers to use themselves–they shared this with me and I haven’t seen it publicly stated before. The first true frontier science model is a tantalizing possibility that could radically reorient the business of science. 80% of the hardware sitting in Cambridge biotech labs is identical and can be abstracted with a sufficiently strong model that can verify results with tool calls at inference. We may see garage band-like startups in the sciences building billion dollar businesses. There have been two generations of scientists that have left the sciences to pursue other careers, typically on Wall St or in Silicon Valley. AI adoption and tooling might reverse that trend, where experiments become much easier, much cheaper and much faster.7
Bullish
After spending time in the AI and science space, it’s a killshot for AI bearishness. While it’s easy to sound smart decrying AI-slop movies, the beAIux arts, and the end of middle-management, injecting raw intelligence into more science has almost pure upside. AI-enabled labs across the sciences can deploy thunderous compute, a mix of robotics and lab assistants, and human direction to create novel breakthroughs and maybe even moonshots. We don’t know the size of that market. We do know that it’s unconstrained—it’s not labor replacement, it’s not arbitrage, it’s not market efficiency.
AI will push science to fill out the known branches on the tech tree in the next few years. If you find yourself wavering on compute demand projections, close your eyes and imagine warehouses full of lab stations operating 24/7 running tight experimentation loops with robotics. Then, smile, and think to yourself—with hundreds of thousands of PhDs in chemistry, biology, physics in the U.S.—how many stations could a PhD oversee? Could we produce 10x the science we do today?
Lots of compute, big buildouts, world-changing progress, and commercial successes we can now foresee.
The parallel in SaaS-land would be the four year pre-launch period for Figma. It worked for them, but long multiyear pre-launch periods aren’t held as a paragon of smart company building.
Many golden eggs in here.
Early SpaceX employees weren’t highly credentialed. Nassim Taleb explained that surgeons should not look like surgeons.
Underlying open weights models are significant here in timing the thesis. Spending tons of capital early in the lifecycle before you have great data and reasoning traces might leave you vulnerable to being leapfrogged. Being too stingy can also leave you vulnerable to a competitor demonstrating benchmark performance and running away with more capital. Ultimately, companies should focus on science capabilities (of which benchmarks aren’t a perfect indicator in many cases) and those are demonstrated via breakthroughs. Commercialized breakthroughs don’t lie and they should follow the best flywheels with the fastest commercialization capabilities.
Public companies care very much about revenue and once the juice has been squeezed from coding, law, accounting, financial services, and medicine, investors will be looking for an uncapped revenue source totally unbounded from a TAM or employee labor count.
Science isn’t fully ready for this shift—a friend mentioned casually that a mathematics journal has seen 10x the volume of submissions since pre-ChatGPT. Keeping up will have its own set of challenges.
