Seed · Q4 2026 · confidential

maxwell.ai

Ask a chemistry question. Get a simulated answer you can trust, with the receipts.

The computational chemist every bench chemist can afford.

Problem

Simulation is gated behind a PhD in theory.

2–3M

chemists in R&D

Fewer than 5% can run a quantum-chemistry calculation unaided.

founder estimate

1 : 20

comp chem : med chem

The computational team is a queue, not a tool. Answers arrive after the compound is made.

founder estimate, 1:15–1:30

€5–20K

per wrong guess

1–3 weeks of synthesis. Decisions are made on intuition because Gaussian assumes you already know the method.

founder estimate

Why now

Three cost curves crossed in 2025/26.

GPU-native DFT

Promethium: 100× faster than legacy codes on cloud GPUs. GPU4PySCF in open source. Hours → minutes.

Learned physics

OMol25 (100M+ DFT calcs), UMA, Egret-1, MACE. TS searches with 94–96% fewer DFT gradients. Skala ships as a pip package.

Agents that plan

Frontier models decompose a chemistry question into a workflow and drive deterministic tools.

Net effect: ~1,000× lower cost per trustworthy answer. DFT stops being the product and becomes the oracle you call only when the cheap model is unsure.

Product

Maxwell answers the question, not the input file.

Input

"Which of these six analogs is the stable tautomer in water?" plus structures from the ELN, ChemDraw or SMILES.

Output

A ranked answer, a confidence verdict, every number tagged with its method, and a ready-to-paste Methods paragraph. Minutes, for a few euros of compute.

Plans the workflow (method, conformers, solvation, what "stable" means here).
Runs the cheap tier on everything.
Estimates uncertainty per result.
Escalates to DFT only where the decision margin is smaller than the uncertainty.
Returns the answer with receipts.

How it works

The demon in the middle.

Tier 1 · cheap on everythingGFN2-xTB or a learned potential. A second cheap method gives an independent opinion.

~0.3 s / molecule

→

The gateσ = prior for this kind of comparison + disagreement + conformer spread. Escalate iff margin < k·σ. Hard rules where cheap methods are known to fail.

escalate iff margin < k·σ

→

Tier 2 · expensive where it mattersDFT, only for the molecules the answer depends on.

~3 s / molecule CPU · ~50× on GPU

Every escalation is a labelled example: (molecule, cheap prediction, DFT truth) on chemistry a real user cared about. The provenance ledger is both the trust mechanism and the flywheel.

Demo · recorded from the real engine

The gate catches the cheap method being wrong.

Question	Tier 1 said	Gate	Answer	vs experiment	DFT spent
Most stable C₄H₈ isomer	cyclobutane ✗	4/5 escalated (ring-count prior)	isobutene	✓	4/5
Most stable of 7 C₃H₆O isomers	acetone	open-chain isomers clear of σ; rings checked (ring prior 12)	acetone, Confident	✓	3/7
2-pyridone vs 2-hydroxypyridine, water	2-pyridone	2/2 (tautomers, σ 7)	2-pyridone by 5.2	✓ (exp ≈ 4)	2/2
Diels–Alder ΔE	−67 kcal/mol	hard rule	−48 kcal/mol	exp ΔH −40	3/3
Claisen barrier ΔE‡	24 (scan) ✗	hard rule → DFT 42 → escalates again: DFT TS search	29.2, saddle verified	✓ exp Ea 30.6	3/3 + TS
Diels–Alder endo vs exo	both TSs below reactants ✗	hard rule → DFT → two TS searches + Hessians	endo, verified ratio	✓ endo exclusive	4/4 + 2 TS
Carbene spin states (4)	all singlet ✗ 2/4	hard rule, open-shell DFT	T · S · S · T	✓ 4/4	8/8

Run it live →all ten demos

Moat

The moat is the escalation ledger.

Open models are our friend. Meta, Microsoft, NVIDIA give away general potentials. We consume them as tiers; our COGS falls every time they ship.
Proprietary data where it matters. (cheap prediction, DFT truth) pairs in exactly the chemical space customers care about, where open datasets are thin.
Validated workflows per question type. Encode and benchmark once, sell a thousand times.
Trust is the interface. Uncertainty and provenance on every number. Chemists stop using a tool after one silent wrong answer.

Market

Expand the buyer from the expert to the experimentalist.

Sells the engine

Sells the answer

Expert user

Gaussian · ORCA · Q-Chem · VASP · GPU4PySCF

Schrödinger · Promethium · Rowan

Bench chemist

Skala · UMA / OMol25 (free building blocks)

maxwell.ai

Bottom-up: 200K industrial chemists × €2,400/yr = €480M before materials. Schrödinger took 30 years to $200M selling to the expert tier only.

Business model

Free editor-class distribution, enterprise revenue.

Individual / academic

€0 → €49/mo

Tier 1 free, capped. DFT credits. Sets the next cohort's defaults.

Team

€200/seat/mo

Plus metered compute. Shared projects, ELN export, audit trail.

Enterprise

€150–500K/yr

In the customer's cloud account. Structures never leave the VPC. Workflow packs, SSO, support.

Gross margin 70–75% blended; compute 15–25% of revenue because the gate keeps DFT to the molecules that need it.

Go-to-market

Wedge with one expensive question.

Months 0–12: tautomer, conformer and relative-stability questions for medicinal chemistry. High frequency, clear right answer, DFT-sized compute, buyer has budget.
Land: free academic tier seeded through 10 synthetic-chemistry groups; PhD students carry tools into industry. Two pharma design partners via founders' network.
Months 12–24: reaction energetics and barriers, pKa and redox, then polymorph and crystal-structure questions.
24+: materials, electrolytes, catalysis via the same agent with periodic tiers.

Roadmap

Eighteen months.

Quarter	Milestone
Q1 2027	MVP agent for tautomer / conformer / stability; xTB + MLIP + GPU DFT tiers. 10 academic groups live.
Q2 2027	Two pharma design partners in VPC. Benchmark published vs Gaussian defaults.
Q3 2027	Reaction energetics and barriers. First paid enterprise contract. 1,000 WAU.
Q4 2027	Ledger at 1M labelled triples. First in-house tier beats open MLIP on customer chemistry.
H1 2028	pKa / redox / polymorph packs. €1.5M ARR. Series A.

Team

What the founding team must look like.

CEO / chemistry

Practising computational chemist, 5+ years as "the queue" inside pharma or agro. Knows which 5% of method choices matter.

CTO / agents

Has shipped an agentic product where an LLM planner drives deterministic tools, with evals. Chemistry checks wrap the model.

Founding scientist / MLIPs

Has trained or fine-tuned MACE / UMA-lineage potentials; credible in the OMol25 community.

Advisors: one ex-Schrödinger or Chemaxon commercial lead; one pharma head of computational chemistry.

The ask

Raising €3.5M seed · 18–20 months runway.

Team (6 eng/sci + founders)55%

Compute & dataset generation25%

Go-to-market10%

Buffer10%

What this round buys

Two paid pharma contracts. A 1M-triple escalation ledger. A published benchmark showing Maxwell matches expert-chosen DFT protocols at under 10% of the compute.

Appendix · unit economics per answer (illustrative)

Step	Compute	Cost
Conformers, 5 molecules × 50	CPU, seconds	< €0.01
Tier 1 on 250 structures	GPU, ~10 s	€0.02
Gate: 2 of 5 escalate	—	—
Tier 2 DFT, 2 × 3 conformers	GPU, ~4 min	€0.60
LLM planning + report	API	€0.05
Total		~€0.70

Priced at €3–5 per escalated answer inside a seat. "DFT on everything" would cost €1.50–2.50 and take 2.5× longer.

Appendix · risks

Risk	Mitigation
Silent accuracy cliffs (transition metals, radicals, excited states)	Comparison-class priors + disagreement across independent cheap methods; hard rules force DFT for flagged classes; publish failure cases.
Big labs bundle free tooling, compress pricing	Value is the gating, protocols and ledger, not the models. Free models lower our COGS.
Pharma procurement cycles	In-VPC from day one; academic distribution so champions arrive inside pharma already trained.
Chemists distrust black boxes	Provenance ledger and Methods paragraph on every answer.
Hiring MLIP talent	Partner for Tier 1 models while building in-house capability on the ledger.

← demo · story · methods & validation