CuspAI screened 300 trillion possible molecular structures and returned 20 validated candidates for the Finnish chemicals firm Kemira. That search had taken Kemira years. CuspAI says it ran it in six months.
Hold onto that number—20—because it is the only number in this week's roughly $850 million of funding that actually settles anything.
Two rounds closed within days of each other. CuspAI raised $450m at a $2.6bn valuation, led by Kleiner Perkins and NEA with Bezos Expeditions writing a check. Chai Discovery reportedly raised $400m at $3.8bn, led by Index Ventures.
One hunts for new materials—semiconductors, carbon capture, cleaner water. The other designs antibodies and proteins from scratch for drug targets that have beaten traditional discovery.
Strip away the materials-versus-drugs difference and both rounds are buying the identical thing. And it isn't the AI model.
Last issue this column tracked capital flowing into data infrastructure and robotics—AI pointed at the digital and the semi-physical. This week the money crosses fully into hard science: materials labs and pharma pipelines, backed by sovereign funds and the largest drugmakers on earth. That is the delta.
Start with the shared story, because both companies tell the same one. CuspAI's platform, MIRA, is pitched as a "search engine" for materials that don't exist yet—specify the properties you need, and it generates synthesisable candidates "up to 10x faster." Chai treats "medicine design as an engineering problem, one that can be solved with precision, speed, and scale," per General Catalyst.
The verb both companies keep using is "design," set against the older world of "discover." Chai's line: most drug discovery screens vast libraries of existing molecules; Chai instead designs new ones "tuned to a disease target before they reach a lab."
Compress years into months. That is the compression narrative, and it is genuinely attracting money faster than the market itself grows. Analysts put the AI drug-discovery market at about $2.35bn in 2025, growing to $13.7bn by 2033—yet Chai's valuation tripled in seven months, from $1.3bn to $3.8bn. Investors are moving faster than their own forecasts.
Chai's valuation tripled in seven months while the AI drug-discovery market barely grows. Investors are moving faster than their own forecasts.
Here is the turn. Read the fine print in both pitches, and notice that neither company is actually selling you the model anymore.
CuspAI didn't just raise money—it launched the AI Materials Foundry, a coalition of more than 45 organisations: Nvidia supplying compute, Meta's FAIR lab contributing an atomistic chemistry model, plus Samsung, Hyundai, Applied Materials, Tokyo Electron and Lam Research. It also announced a multi-year partnership with A*STAR, Singapore's public R&D agency, for autonomous synthesis.
CuspAI itself lays out why. Software-led materials discovery, it says, requires four things: training data at scale, compute to screen at molecular resolution, synthesis infrastructure to move from digital design to physical reality, and domain expertise to interpret what the machine finds.
Notice what that list is. Only one of the four items is the AI. The other three are the validation stack—labs, synthesis, people who know a real result from a hallucinated one.
The Foundry exists precisely because generation alone proves nothing. A model that dreams up 300 trillion structures is impressive; a model that returns 20 you can actually make is a business. The gap between those two is everything CuspAI just spent $450m building out.
Josh Coyne of Kleiner Perkins says the quiet part directly. What sets the founders apart, he argues, is that "they design for materials that can actually be built, not just ones a model can dream up."
Read that sentence as a confession. The differentiator isn't the search engine—search engines are becoming cheap, and Meta is contributing a frontier atomistic model to the coalition for free. The differentiator is buildability, which is a claim about the lab, not the model.
Chai's version of the validation stack is its customer roster. Its models are deployed at Eli Lilly, Pfizer, and Novartis, running the latest Chai-3 against targets that defeated traditional methods, with a new argenx collaboration added this week.
TNW names why that matters: "Big drugmakers rarely trust unproven software, so paying customers at that scale set Chai apart from most AI startups chasing the same problem." The moat isn't Chai-2 posting double-digit success rates in de novo antibody design. The moat is that Lilly is willing to run it.
That is the same move CuspAI made. Both companies have quietly conceded that the generation step is the commodity, and that the expensive, unproven part is the physical loop—the thing that turns a prediction into something you can synthesize, manufacture, or push through the FDA.
Which is where the valuations get uncomfortable. The $2.6bn and the $3.8bn are pricing the generation step—the speed, the compression, the search engine. But the risk lives entirely downstream, in the loop the models haven't closed yet.
TNW's closing line on CuspAI is the whole thesis in one clause: the Foundry "now has to prove the models can find materials that survive contact with a real lab." Survive contact with a real lab. Everything before that is a prediction.
The Kemira result is the closest thing either company has to a posted score, and even it is unfinished. Twenty "validated candidates" is validation in the screening sense—candidates that survived simulation and initial checks. It is not twenty materials in production replacing iridium and ruthenium in chips, which is the actual prize CuspAI is chasing.
Chai's scoreboard is even further out. As TNW puts it: a model can generate candidate molecules in minutes, but "those molecules still face years of lab work, clinical trials, and regulators before they reach patients."
And then the sentence that should be printed on the term sheet: "Whether that belief survives the first wave of clinical failures is what a $3.8bn price tag is really testing."
That is the real test for both. Not the demo, not the coalition logos, not the compression story. The first wave of clinical failures for Chai. The first materials that do or don't survive a real lab for CuspAI.
This extends a question this column raised in July around Helsing—can software scale into hardware?—but shifts it from manufacturing throughput to something slower and less forgiving: lab-and-clinic validation. Defense hardware ships when it works. A drug candidate can look perfect in silico and die in Phase II, and a material can screen beautifully and crumble on synthesis.
The bull case is that the compounding loop CuspAI describes—breakthroughs inside the network accelerating discovery across it—is real, and that owning the validation infrastructure is exactly the durable moat generation alone can't be. If the AI is the commodity, then the labs, the exclusive chemistry databases, the pharma trust, and the synthesis partners are the actual asset. That is a defensible read.
The bear case is that they've priced the cheap part and deferred the expensive one. The market has decided the models can find candidates. It has not yet seen whether those candidates become materials or medicines.
Both scoreboards read the same right now: zero final scores posted. Twenty validated candidates and one first wave of clinical failures still to come.
The money has already voted. The lab and the clinic haven't.