Validation infrastructure for AI-designed biology

Design got cheap. Proof didn't.

In 2024 the hard part of AI × Bio was generating a plausible candidate. In 2026 it is finding out whether the candidate is real.

• Design capacity exploded — a dozen-plus protein design foundation models now ship publicly; generating 10,000 candidates is a weekend of GPU time
• Validation capacity did not — the same wet-lab step costs the same weeks it did a decade ago
• Q1 2026 was the inflection — the first wave of AI-designed molecules reached peer-reviewed preclinical validation, and the field discovered its real constraint
• The bottleneck moved — from algorithms to throughput: iterating candidate libraries and proving which survive contact with a bench

The field keeps two ledgers

These two ledgers diverge constantly, and the gap between them is where AI × Bio value is created or destroyed. We sell throughput on the second ledger.

What a scientist actually goes through

• Weeks to a quote — email a CRO, wait for scoping calls before a price exists
• Protocols re-typed by hand — the computational design is flattened into a Word document and re-entered by a technician
• Results arrive as a PDF — unstructured, un-versioned, impossible to feed back into the model that generated the hypothesis
• No provenance — when a result fails to reproduce, there is no signed record of what the instrument actually did
• Academia is priced out entirely — the largest single source of AI-designed candidates cannot transact at biopharma contract rates

Order validation the way you order compute

No scoping call. No account manager. Transparent per-assay pricing at the moment of ordering, and a result that arrives as a typed object rather than a report.

LabOS — the layer between the model and the bench

• Protocol compiler — turns a described experiment into an executable instrument program, not a technician's to-do list
• Feasibility & quote agent — instant answer on whether the fleet can run it and what it costs
• Scheduler — packs orders across instruments to keep utilisation high; utilisation is the entire unit economic story
• QC & anomaly agent — catches a failed run while it is still running, not at delivery
• Provenance signer — every result carries a signed record of the exact instrument, reagent lot, and parameters

Every run emits a training example by construction — the data asset is a by-product of serving the customer.

A cost base competitors cannot refactor into

• Price is already the battleground — Ginkgo launched US-based ADME profiling explicitly to match or beat Chinese vendor quotes. Incumbents are defending on cost, not attacking
• Structural, not promotional — facilities, technical staff, and utilities in India run at a fraction of a Boston cost base; we can price where they cannot follow
• Deep talent pool — India's CRO sector has trained two decades of bench scientists now available to an automation-first employer

Apollo: the substrate no one else can buy

Cost advantages get competed away. Data advantages compound.

• Pan-India clinical substrate — EHR and biobank access across Apollo's hospital network
• South Asian genetic diversity — systematically underrepresented in the reference data every Western validation stack is built on
• Validation against real patients — a candidate can be tested against the population it is meant to treat, not only a cell line
• Compounding — every validation run enriches an atlas that makes the next run more informative. Ginkgo can rebuild our robots; it cannot rebuild this

Where we sit — honestly

We are not first to autonomous cloud labs. We intend to be first to autonomous validation that an academic lab can actually afford, tied to a clinical substrate no incumbent holds.

Land in academia. Expand into industry.

The developer-tools motion applied to biology: the scientist who validated a construct during their PhD specifies us when they join a company.

Three revenue layers

• Per-assay self-serve — transparent published pricing, paid at order. Acquisition engine and the source of utilisation
• Capacity subscription — reserved throughput plus API access for teams running continuous design–build–test loops
• Data products — the India Biomarker Atlas and aggregate validation benchmarks, built from consented, de-identified run data

FILL Insert current pricing per assay class, target gross margin, and payback period per instrument cell.

Sized on validation spend, not lab software

• Beachhead — India preclinical CRO, $191.7M (2024) growing to $549M (2033)
• Serviceable — global outsourced preclinical validation and assay services, where price and turnaround are the deciding criteria
• Category — venture capital now treats autonomous labs as an infrastructure asset class alongside chip fabs and data centres; over $1B has entered the category since 2025

FILL Replace with your own bottom-up TAM/SAM/SOM build once assay pricing is fixed. Top-down analyst figures are directional only.

From first cell to fleet

FILL Attach real dates, current phase, and any live pilots or LOIs. An early-stage VC will ask what is running today — this slide must answer it.