Newsletter · · Ashutosh Agarwal

The AI Labs Stop Selling Shovels and Start Digging - AI Drug Discovery - Week of July 9, 2026

AI Drug Discovery newsletter for the week of July 2 to 9, 2026. Anthropic turned Claude Science inward to run its own drug-discovery programs, moving the model builders from selling shovels to owning candidates, while DARPA and CZI reframed the real bottleneck as data volume and diversity rather than model IQ.

AI Drug Discovery

Week of July 9, 2026: The AI Labs Stop Selling Shovels and Start Digging

For a year the frontier AI labs sold pharma the pickaxes and let the miners keep the gold. This week one of them picked up a shovel of its own. Anthropic is now running internal drug-discovery programs on Claude Science, which quietly turns a vendor into a rival, and reframes the whole debate about where value in AI biology actually pools: in proprietary wet-lab data, or in the compute to manufacture data no lab has yet run.

TL;DR

  • The week's defining story is a role change, not a product launch: Anthropic is now using Claude Science to run its own drug-discovery programs for rare, neglected diseases, moving from arms dealer to prospector, and becoming a potential competitor to the pharma clients it sells to (Everyday AI Podcast, Jul 6; Limitless, Jul 3).
  • The "AI-lab-as-drug-company" club now spans Anthropic, Google DeepMind's Isomorphic Labs, and AI-first shops like Insilico, with the model vendors moving down the value chain toward candidate ownership (Everyday AI Podcast, Jul 6).
  • The virtual-cell thesis went mainstream: Mark Zuckerberg and Priscilla Chan laid out CZI's decade-long bet that curing disease depends "as much on better tools as on better treatments," datasets and models first, therapies later (The a16z Show, Jul 9).
  • DARPA reframed the central bottleneck bluntly: biology has both a data-volume and a data-diversity problem, and simulation plus synthetic data is the way through (Free Radicals, Jul 7).
  • Thin week on the listed names: no substantive AI-discovery podcast coverage for RXRX, SDGR, ABCL, CERT, ABSI, RLAY, DNA, TEM, Isomorphic, Xaira, Iambic, or Nvidia's Clara/BioNeMo. LLY's tape strength is obesity-and-rotation, not AI.

What's New

The through-line this week is vertical integration by the model builders. On Limitless (Jul 3), the hosts framed Claude Science as "the way that Claude Code is intended to support engineering: Claude Science is intended to support science," and delivered the load-bearing detail: "Anthropic is actually using it… to pursue their own drug discovery programs for some rare neglected diseases." They cited early productivity proof points, noting the Allen Institute "now completes 100-plus page literature reviews that would previously take up to two years," with some analyses "in a tenth of the time," and a user pairing Claude Science with a Ligand AI MCP tool to identify a molecule's active site "de novo."

Everyday AI (Jul 6) put the strategic point in plain language: "Anthropic is going to make drugs now," describing Claude Science as "one of the most direct attempts by a major AI company to move in[to] drug development rather than just selling AI to other drug makers." The hosts flagged the obvious tension, that Anthropic will be "both a software provider and a potential competitor to its pharma clients," and slotted it alongside Google DeepMind's Isomorphic Labs and Insilico in "a growing race." Anthropic has not disclosed what it would do with any promising candidates it finds.

The bigger-picture vision came via The a16z Show (Jul 9), replaying Zuckerberg and Chan on CZI. Their framing: "most major breakthroughs are basically preceded by the invention of a new tool," so CZI funds the "longer-term, oftentimes more expensive" infrastructure (imaging, Cell Atlas, virtual-cell models) that individual NIH grants won't, "on the order of maybe $100 million to $1 billion over a 10 to 15 year period." Zuckerberg's tell for why the listed toolmakers keep struggling to escape their niche: even "AlphaFold, which is amazing… was built off of a public dataset that had been produced decades ago." The edge, he argued, is pairing "frontier biology with frontier AI" to generate purpose-built datasets, not just running better models on old ones.

The Debate

Last week we framed the constraint as data and GPU allocation, not model IQ. This week that thesis got its most authoritative articulation, from DARPA. On Free Radicals (Jul 7), Dr. Mike Koeris argued biology's data is "vanishingly little… relative to the complexity of the space that is embodied in us," and, critically, that it is also the wrong kind of data: "We get a lot of nucleic acid sequencing data… but we actually had very little metabolomic data. We get very little proteomic data." His resolution mirrors the AlphaFold precedent from the a16z conversation: "One of the reasons AlphaFold was so successful is that they didn't have to worry about the data… it took 30, 40 years to generate all these wonderful structures in the protein databank." DARPA's bet (the NODES and PROSE programs) is to "pay for compute for a year or two to generate a lot of data" synthetically and train on that.

So the swing question sharpens: does the moat accrue to whoever owns proprietary, diverse wet-lab data (CZI's Biohubs, the integrated pharmas), or to whoever can most cheaply manufacture high-quality synthetic data via simulation (the compute-rich AI labs)? Anthropic turning Claude Science inward is a vote for the second camp, that a great harness plus compute can substitute for decades of accumulated experimental data.

Stocks in Play

  • Recursion (RXRX): $3.76, up 1.1% on the day but drifting near the low end of its 52-week range ($2.77–$7.18; market cap ~$1.7B). No AI-discovery podcast or news catalyst this week, a quiet holding pattern into its next pipeline update (FactSet, Jul 9).
  • Schrodinger (SDGR): $16.79, up 4.35% on the day on light volume (52-week $10.95–$23.75; market cap ~$1.3B). No fresh narrative, the move looks technical and rotation-driven rather than event-driven (FactSet, Jul 9).
  • Eli Lilly (LLY): $1,216.95, pinned just under its 52-week high of $1,249.45 (market cap ~$1.15T). A wall of pre-Q2 price-target raises landed: RBC to $1,500 (from $1,250), JPMorgan to $1,400, Truist to $1,370, Cantor to $1,350, Morgan Stanley to $1,347, all constructive, all built on Zepbound/Mounjaro obesity momentum (thefly, Jul 6–9). Honesty check: this is a GLP-1 and rate-rotation story, not an AI-discovery story. Cantor itself cautioned the group is being treated as "a rotation beneficiary from AI names rather than being purely fundamentals-driven."

Read-Throughs

  • Toolmakers face a widening squeeze. If the frontier labs move from selling models to owning candidates, the pure-play computational vendors (SDGR, and privately Iambic and Xaira) increasingly compete against their own suppliers' compute budgets. Watch whether their pitch shifts toward proprietary, hard-to-simulate wet-lab data as the defensible asset, the exact edge Zuckerberg says CZI is building (The a16z Show, Jul 9).
  • Synthetic data is the new contested resource. DARPA funding simulation-generated training data (Free Radicals, Jul 7) validates the compute-heavy path, a directional tailwind for GPU and simulation exposure (Nvidia's bio stack, Schrodinger's physics-based engine) if it works, and a threat to data-scarce single-asset biotechs if it doesn't.
  • "AI biotech" and "obesity biopharma" are trading as one basket but are two different theses. LLY's rip is fund flows into large-cap pharma, not proof that AI is shortening its clinical timelines. Don't let the tape conflate them.

What Changed vs Last Week

  • Last week: Anthropic launched Claude Science and we flagged "watch for pharma partner signings." This week the update is sharper and more disruptive than a partnership: Anthropic is running its own internal drug programs (rare and neglected diseases), positioning as a competitor to pharma rather than a vendor to it.
  • The data-vs-IQ debate we opened last week now has a named, authoritative voice (DARPA's Koeris) and a concrete mechanism (simulation plus synthetic data via the NODES and PROSE programs).
  • New this week: the CZI virtual-cell "tools before treatments" thesis entered the mainstream feed via a widely heard a16z replay.
  • Unchanged and worth noting: still no AI-specific podcast coverage of RXRX, SDGR, ABCL, CERT, ABSI, RLAY, DNA, TEM, Isomorphic, Xaira, Iambic, or Nvidia-bio, a genuinely thin week on the listed names, per dedicated searches.