Miju LabsEdison Scientific is a FutureHouse spinout that raised a $70M seed in December 2025 to build Kosmos (Endpoints News). The reason it belongs in a specialist register rather than a product register is buried in a benchmark methodology section: *LAB-Bench 2 tasks were generated by "contracted domain experts who hold or are in the process of obtaining PhDs in biology", working through "a purpose-built web-based platform" with multiple review rounds* (arXiv 2604.09554).
That is the expert-data business, built in-house, by a company that is not selling expert data. It contracts the PhDs, it built the platform, and it owns the resulting benchmark. Anyone entering Life-science wet lab with a "Mercor for biology" plan is proposing to sell Edison something Edison already makes.
LAB-Bench 2 as a moat, not a leaderboard
1,900+ tasks across literature retrieval, database access, protocol troubleshooting, molecular-biology assistance and experiment planning, with model performance dropping 26–46% versus the original LAB-Bench, and the largest remaining gaps in retrieval, database access and sequence manipulation (arXiv 2604.09554; FutureHouse).
Unsaturated, credible, and owned by a $70M-funded participant in the market it measures. That is the same position Irregular holds in Offensive security and Vals AI holds in law and finance — and it is the reason the publish-a-benchmark opening does not work everywhere. In biology the reference already exists and the company that owns it has more capital than you will raise.
And the horizontal already bought the vertical
Mercor acquired Sepal AI in February 2026 (Orrick). Sepal was the science-environment specialist — the exact vertical-specialist shape this atlas studies — and it is now inside the largest horizontal. Mercor also acquired Deeptune in July 2026 (RL List).
Read those two facts together and the strategic picture in wet lab is unusually clear. The incumbent horizontal has bought the vertical specialist in this domain; the well-funded application company has built the benchmark and contracted the supply. The remaining independent position is narrow.
For the copycat, the Sepal outcome is also the base case rather than the failure case. A vertical specialist that gets acquired by Mercor has succeeded — it is just a different exit shape from the one the "own a domain" thesis implies, and it prices accordingly. See What the public market pays for labour.
The capital that surrounds it
| Company | Raised | What it is |
|---|---|---|
| Edison Scientific | $70M seed | FutureHouse spinout, Kosmos, LAB-Bench 2 |
| Lila Sciences | $350M Series A (Flagship Pioneering) | Autonomous labs |
| Periodic Labs | $300M seed | Physical lab for closed-loop RL |
| Sepal AI | — | Science environments, acquired by Mercor Feb 2026 |
Sources: Endpoints; FinSMEs; Maginative; Orrick.
Roughly $720M of capital sits in this domain, and the layer where an independent would be genuinely unique — closed-loop physical experimentation — is the one Periodic Labs raised $300M to build. SemiAnalysis states the cost line plainly: "a single biological experiment can cost hundreds to thousands of dollars and take hours to complete, compared to a coding task" (SemiAnalysis).
The supply economics still work
The arbitrage is real even if the competitive position is not. BLS puts biological technicians at $57,510/yr ($27.65/hr) across 74,400 US jobs — a low floor — while Mercor lists Biology Expert (PhD) and a Biology Expert (PhD) — AI Safety variant at $65–$70/hr for 15–25 hrs/week, requiring a PhD in molecular biology, microbiology, virology, genetics, biochemistry, bioinformatics, synthetic biology or immunology plus "sound judgment regarding biosecurity and dual-use information handling" (Mercor).
A 2.3x multiple on the technician floor makes recruiting easy. It is just that the listing is Mercor's, and Mercor now owns Sepal.
Revenue, customers and pricing at Edison — none disclosed. Valuation at the $70M seed. Headquarters and headcount. What the contracted PhD biologists are paid, which is the single number a copycat would most want and which appears nowhere in the LAB-Bench 2 methodology. Whether Edison sells data to third parties at all, or builds it purely for Kosmos. The terms of the Mercor–Sepal acquisition, including price.
What to learn from it: when the best-capitalised company in your domain contracts its own experts and publishes its own benchmark, the specialist position has already been taken from the inside — check who built the reference benchmark before assuming the vertical is empty.