The AI took 64 hours. The lab took two months and three times the money.
22 Oct 2026 · News
An investment research firm used Claude to design proteins against scorpion venom. The AI part took under three days and about $50,000. Testing the results took about two months and three times the money. Data centers are showing the same shape.
The model is the cheap part now. In the cleanest receipt we've seen, the AI step cost a third of the lab step and finished in a sliver of the time, and the tests that would show whether it worked haven't even been run. The money and the waiting have moved to the physical step after the AI: the wet lab, the trial calendar, the power hookup. The AI labs have noticed, and they're buying their way toward the bench.
On Tuesday Anthropic published a project it ran with Citrini Research, a firm that normally writes about stocks. An internal Claude research model spent 64 hours designing small proteins meant to latch onto scorpion neurotoxins. It generated 318,484 designs and picked 180, and the team added 31 more. Two contract labs, Adaptyv Bio and Twist Bioscience, made and tested all 211. Seventeen stuck to their target well enough to count. None has been tested to see whether it actually stops the venom, and Anthropic says plainly it hasn't made an antivenom.
Citrini's own write-up has the invoices. Compute came to $49,997, on a little more than 10,000 hours of Nvidia H100 time. The lab bill for making the DNA and proteins and measuring the binding came in "just shy of $150,000." Then the clock: "Over two days, we watched Claude work through problems at a speed that was difficult to comprehend. For the next two months, we spent much of the time becoming familiar with laboratory turnaround times." Their shorter version: Claude "may have compressed a year's worth of work into a couple days, but biology works at the speed of life."
And the AI side keeps getting cheaper. In a footnote, Anthropic says it has since made these models about four times faster on average, and that a simpler setup, one agent on one GPU for 24 hours, reached comparable scores on the computer. The lab step got no such footnote. One of the outside scientists Anthropic asked to review the work put it flatly: "binding itself is no longer the main bottleneck."
The same day, OpenAI bought Orchestra Bio, and look at what it does. It doesn't design molecules; OpenAI already has GPT-Rosalind for the science. Orchestra's agents watch the schedule. They track studies and campaigns, and when, in OpenAI's example, "a failed manufacturing batch" threatens to "delay the start of a clinical trial," they work out which timelines are hit and how to respond. OpenAI puts the road from target discovery to US approval at "roughly 10 to 15 years on average." Terms weren't disclosed, and Orchestra had raised about $20 million, TechCrunch reported. When a lab spends acquisition money on the calendar, it's telling you where it thinks the time goes.
Anthropic got to the bench first. In September it confirmed a wet lab in the Bay Area, with its head of life sciences telling Reuters that "the final test is still, and will be for a while, in real lab work." For now, per Anthropic, "all of the lab work is performed by human scientists."
Chips have the same problem, at a much bigger scale. On Tuesday Business Insider reported, citing internal emails from August, that Amazon set up an "Early Rack Receipt" process to take about $5 billion of Nvidia GB300 server racks before final deployment, with roughly $3 billion more a month expected "for the foreseeable future." Some AWS sites had room for the racks but not enough power to turn them on, so finished equipment waits in warehouses. Amazon told BI that staging hardware near sites is "best business practice" and disputed that it lacked a path to install it. Either way, the order has flipped: buy the brains first, then wait for the plug.
Last November that was a confession. Microsoft's Satya Nadella said he might have "a bunch of chips sitting in inventory that I can't plug in. In fact, that is my problem today," DCD reported. Eleven months later it's a procurement process with a name.
We've watched this queue move for two weeks. On 11 October the slow part of AI bug hunting was someone saying "yes, that one is real." On 16 October it was getting the fix installed. On 18 October, inside Anthropic, it was a person's time to check the work. Each time the AI got faster and the line formed one step further down. This week the line walked out of the building, into contract labs, trial calendars and power hookups.
Our call: by 30 June 2027, OpenAI follows Anthropic onto the bench. It buys, builds or signs an exclusive capacity deal with a wet lab or contract research lab, because agents that manage the schedule still wait on someone else's turnaround. And the brag changes: the headline number in the labs' science posts moves from how fast the model designed something to how fast a lab confirmed it, the "confirmed in the lab within weeks" line Anthropic already slipped into its own write-up. One project's invoice isn't an industry, but it points the same way as $5 billion of racks waiting on power. The model is the part that keeps getting cheaper. Whoever owns the slow part sets the pace.