The story, in depth

Everything behind the demo, in ten parts.

Start with the job, go as deep as you need: how the gate decides, how it validates against experiment, what it cannot do yet, and what the company around it looks like. The slides and the full-control demo are at the end.

Part 1

The job: decisions a chemist makes every week without simulation.

Each situation below is a real recorded run in the demo, wrapped in the decision it supports. Click one to replay it.

Process chemist · C4 skeletal isomerisation (MTBE / butyl rubber feed)

Which isomer should the plant design for?

Butene feed over an acid catalyst; the equilibrium isomer sets the isobutene yield. A cheap screen picks cyclobutane. Wrong by 11 kcal/mol.

Stake: a pilot campaign on the wrong product

Stability chemist · forced degradation

Which degradant to synthesise first?

A new peak at m/z 58. Seven C₃H₆O isomers; acetone is 8.7 kcal/mol clear of the open-chain isomers, so those stay at Tier 1. Only the two rings are checked at DFT, because calibration says a cheap method is wrong by 12 kcal/mol on ring-count changes. 3 of 7 escalated, the rest of the DFT saved honestly.

Stake: 2–3 weeks on the wrong reference standard

Medicinal chemist · lead optimisation

Which tautomer to dock and register?

2-pyridone vs 2-hydroxypyridine in water. Solvent-aware at both tiers; the lactam wins by 5 kcal/mol, matching K_T ≈ 10³.

Stake: misleading SAR for a whole series

Process chemist · epimerisation step

Will equilibration give a clean diastereomer?

cis vs trans 1,2-dimethylcyclohexane: trans by 1.2 kcal/mol, an 88:12 mixture. Honest verdict: budget for a separation.

Stake: an unplanned unit operation at scale

Process safety · scale-up to 500 L

Can the reactor take the heat?

Diels–Alder enthalpy → adiabatic temperature rise ~290 K at 20 wt%. Tier 1 overstated it by 40%. The T2 Laboratories runaway (2007, four killed) is why this number is checked before the reactor is booked.

Stake: thermal runaway · CSB T2 report

Process / catalysis chemist · oxo alcohols

Is our n/iso selectivity thermodynamic or kinetic?

Propylene hydroformylation feeds ~2 Mt/yr of 2-ethylhexanol. Thermodynamics favours the branched aldehyde by ~1 kcal/mol (NIST), so linear selectivity is kinetic and must be protected as such.

Stake: n/iso drift on a process excursion

Electrolyte formulator · Li-ion cells

Which carbonate survives the high-voltage hold?

EC, PC, DMC and the sacrificial additive VC, ranked by HOMO–LUMO gap. Tier 1 compresses all four into noise; DFT separates VC cleanly as the one that reacts first.

Stake: weeks of cell builds on a failing formulation

Process chemist · thermal rearrangement step (flavour & fragrance)

What temperature does the rearrangement need?

Claisen rearrangement of allyl vinyl ether. The xTB scan says 24 kcal/mol, DFT on that geometry says 42; the gate escalates a second time to a DFT transition-state optimisation and lands within 3 kcal/mol of the measured 30.6. The cheap number would have sized the reactor wrong either way.

Stake: a reactor sized for the wrong temperature · Ea Schuler & Murphy 1950

Synthetic chemist · building a norbornene scaffold

Endo or exo: which adduct forms?

Cyclopentadiene + maleic anhydride, the textbook Alder endo rule. A kinetic question needs two transition states. Maxwell builds a guess per pathway from the product side, is forced to DFT, then optimises and Hessian-verifies both saddle points and returns the ratio. GFN2-xTB had put both guesses below the reactants. Days of a student's TS hunting, in minutes of CPU and zero human steps.

Stake: a week of synthesis on the wrong diastereomer · endo exclusive at RT (Woodward & Baer 1948)

Chemical biologist · photo-affinity probe design

Which carbene gives the clean crosslink?

Methylene, CF₂, CCl₂, phenylcarbene: singlet or triplet? GFN2-xTB gets the sign wrong for two of four. Open-shell DFT (restricted singlets, unrestricted triplets) gets all four ground states right.

Stake: a probe that scrambles instead of inserting · exp ΔE(S−T) from photoelectron / kinetic data

Where the order of magnitude actually is.

Not in compute. DFT on a handful of small molecules costs seconds, and the demo says so in a footnote. The gains that matter are in the chemist's own time, in the mistakes that no longer happen, and in work that DFT alone cannot do.

100–1000×chemist time, routine thermochemistry

Which isomer, tautomer, conformer or spin state: a careful manual workflow is 4–13 hours of a chemist's day (the demo lists the steps and takes your estimate). Maxwell answers in 20–80 seconds with σ attached. The factor is computed from your own number, in the room.

measured wall time vs audience estimate

days → minutestransition states, selectivity

Locating and verifying one saddle point is where students lose days to guesses that fall off the saddle; a selectivity question needs several. Maxwell builds the guesses, optimises each saddle point, checks the imaginary mode and quotes the ratio. DFT does not enumerate pathways; people do, slowly.

Claisen: 24 min CPU · endo/exo: two verified saddle points

wrong → rightthe protocol mistakes

All from building this prototype: a linear carbene start that puts the singlet in an excited state (41 kcal/mol, wrong ground state); a single point on an xTB path maximum 13 kcal/mol above the true saddle (10⁶ in rate); a cheap method putting two transition states below the reactants. Each one is caught, and the number it would have cost is shown.

real numbers from the recorded runs

What Maxwell does not claim: it runs the same B3LYP-D3(BJ)/def2-SVP a careful expert would, so it is not more accurate than DFT; a 1 kcal/mol question stays marginal and the verdict says so. It is the expert's protocol judgment applied every time, available to a bench chemist who has no computational group to queue for.

Simulation is gated behind a PhD in theory.

Quantum chemistry can tell a chemist which molecule to make before they spend three weeks making the wrong one. Almost nobody at the bench can use it.

2–3Mchemists in R&D

Pharma, agro, materials, specialty chemicals. Fewer than 5% can run a quantum-chemistry calculation unaided.

founder estimate

1 : 20comp chem to med chem

At a typical pharma site the computational team is a queue, not a tool. Answers arrive after the compound has been made.

founder estimate, range 1:15 to 1:30

€5–20Kper wrong guess

Functional, basis set, dispersion, solvation, convergence: Gaussian, ORCA and Q-Chem assume you already know. So the decision is made on intuition instead.

1–3 weeks of synthesis, reagents and instrument time

Part 2

Three cost curves crossed in 2025/26.

The electronic-structure engine, the incumbents' core asset, is being commoditised from two directions at once. The value moves to whoever owns the decision of what to run.

GPU-native DFT

QC Ware's Promethium runs 100× faster than legacy codes on cloud GPUs; GPU4PySCF does it in open source. DFT drops from hours to minutes.

aws.amazon.com/blogs/hpc

Learned physics

Meta's OMol25 (100M+ DFT calculations) and UMA, Rowan's Egret-1, MACE-OMol25. Transition-state searches with 94–96% fewer DFT gradients. Microsoft's Skala functional ships as a pip package.

arxiv 2604.00405 · arxiv 2506.14665

Agents that plan

Frontier models can read a chemistry question, decompose it into a workflow, call deterministic tools and reason about the result. The chemistry checks wrap the model, not the other way round.

net effect: ~1,000× lower cost per trustworthy answer

Part 3

Maxwell's demon, for molecules.

The thought-experiment gatekeeper sorts fast molecules from slow ones. Ours sorts molecules that need expensive physics from molecules that don't. That gating decision is the product.

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

~0.3 s / molecule

→

The gateUncertainty = method prior for this kind of comparison + disagreement between cheap methods + conformer spread. Escalate only if the decision margin is smaller than k × σ. Hard rules override statistics where cheap methods are known to fail.

escalate iff margin < k·σ

→

Tier 2 · expensive where it mattersDFT (B3LYP-D3(BJ)/def2-SVP today, GPU def2-TZVP in production). Only the molecules the answer depends on.

~3 s / molecule on a laptop CPU · ~50× faster 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 training-data flywheel. Same artefact, two jobs.

The gate, precisely

For a ranking question, let best be the Tier 1 minimum. For each candidate i:

margin_i   = E1(i) − E1(best)                                   [kcal/mol]
σ_i        = sqrt( prior(class(i, best))² + disagreement_i² + (0.25 · spread)² )
disagree_i = | ΔE1_GFN2(i, best) − ΔE1_GFN1(i, best) |
escalate_i = margin_i < k · σ_i          (default k = 1.5)

class(i, best) is the kind of comparison, determined from the structures: conformers (prior 1.5), stereoisomers (1.5), tautomers (3.0), constitutional isomers (4.0), ring-count change (12.0). The priors are calibrated against 131 published reference energies (see Validation, "σ calibration"). The best candidate is always escalated when any contender is, because it is the DFT reference. If nothing is inside the band, no DFT runs and the Tier 1 ranking is reported with σ attached. A budget caps the number of escalations in balanced mode.

Hard rules override the statistics: reaction energies always go to DFT (bond types change), property questions always go to DFT (semi-empirical orbital gaps are qualitative), and elements outside the validated set force DFT. Verdicts: Confident if the final top-two margin exceeds 2 × combined σ, Likely if it exceeds 1 × σ, otherwise Unresolved with a concrete next step.

Why disagreement alone is not enough

Building the prototype produced the clearest lesson of the project: GFN1 and GFN2 share their systematic errors. On the C₄H₈ isomers they agree with each other within 0.1 kcal/mol while both are wrong by ~11 kcal/mol on cyclobutane. A disagreement-only gate would have reported the wrong answer as Confident. The comparison-class prior is the independent signal, and it is exactly the kind of knowledge a computational chemist carries in their head and never writes down. The product encodes it, then calibrates it from the ledger.

Part 4

Proof, not slides.

Ten questions recorded from the prototype engine, from isomer stability to a reaction barrier and open-shell spin states. Default settings (gate on, k = 1.5). Compare against published experimental values. Click any row to replay it.

Question	Tier 1 said	Gate	Final answer	vs experiment	DFT spent

Part 5

Validation against experiment

Every number below is read live from the recorded runs (gate on, k = 1.5), compared with gas-phase Δ_fH°(298 K) from the NIST Chemistry WebBook unless stated. Electronic ΔE vs ΔH at 298 K: expect ~1 kcal/mol of drift from ZPE and thermal terms on isomerisations, more on reactions.

	Molecule	Tier 1 ΔE	Final ΔE	Method	Experiment	Δ(final − exp)

σ calibration: do the uncertainty bands mean what they say?

The gate's σ is a per-class prior combined with the GFN2/GFN1 disagreement. We ran the full Tier 1 → Tier 2 protocol on four GMTKN55 reference sets (131 reactions, CCSD(T)-quality references) and asked how often the real GFN2-xTB error sits inside the band the gate would have drawn. Read live from calibration.json.

Reference set	Class · prior	n	GFN2-xTB MAD	DFT//xTB MAD	sign errors (GFN2)	inside 1σ	inside 1.5σ (k)

Where cheap methods fail, and how the gate handles each

Ring strain and isomerisation across ring counts. GFN-xTB underestimates ring strain by 10–20 kcal/mol (cyclobutane, oxetane, cyclopropanol in the demos). Handled by the ring-count prior, which calibration pushed from a hand-set 6 to a measured 12 kcal/mol (GFN2-xTB MAD 11.7 on the ISO34 ring-count reactions). This is the single most common way a bench chemist would be misled by a semi-empirical screen.
Reaction energies with bond-type changes. Ethene hydration: xTB −33 vs experiment −11 kcal/mol. Diels–Alder: xTB −67 vs DFT −48 vs experiment −40. Handled by a hard rule, not a threshold.
Orbital gaps. GFN2-xTB gaps are compressed by a factor of ~2 (benzene 3.2 vs 6.9 eV at B3LYP). Hard rule.
Tautomers in solution. ALPB and ddCOSMO both reproduce the 2-pyridone preference in water; GFN1/GFN2 disagree by ~6 kcal/mol on this pair, which the disagreement term catches on its own.
Stereoisomers. Small margins, small priors. cis/trans-1,2-dimethylcyclohexane: 1.4 kcal/mol at xTB, 1.2 at DFT, 1.9 experiment. The verdict is honestly Unresolved at def2-SVP electronic energies; def2-TZVP plus thermal corrections is the stated next step.

Part 6

Protocol (prototype)

Tier 0 — structures

SMILES or name → RDKit. ETKDGv3 conformer embedding (5–30 conformers depending on the task), MMFF94 pre-optimisation, heavy-atom RMSD de-duplication at 0.25 Å. Scope checks here: element set, heavy-atom count, open-shell detection, formula consistency for stability rankings (non-isomers are dropped with a warning rather than ranked).

Tier 1 — semi-empirical

Every unique conformer is optimised at GFN2-xTB (tblite 0.7.0, ASE L-BFGS, f_max = 0.02 eV/Å). ALPB implicit solvation when a solvent is specified. The lowest conformer is kept. GFN1-xTB single points on the same geometry provide the second opinion used in the uncertainty estimate.

Tier 2 — DFT

B3LYP with D3(BJ) dispersion (simple-dftd3), def2-SVP, PySCF 2.14, ddCOSMO when a solvent is specified. By default a single point on the Tier 1 geometry, so reported DFT numbers are a B3LYP-D3(BJ)/def2-SVP//GFN2-xTB protocol. Full DFT geometry optimisation (geomeTRIC) is available behind a flag and is not used in the recorded demos.

All relative energies are electronic energies. No zero-point, thermal or entropic corrections. The generated Methods paragraph states this explicitly on every report.

Limitations you should hold us to

Electronic energies only. No ZPE, enthalpy or free-energy corrections yet; conformational entropy is ignored.
def2-SVP is a small basis. Basis-set incompleteness inflates reaction energies (hydration: −23 at def2-SVP vs −11 experiment, of which a few kcal/mol is basis and a few is thermal).
Priors are calibrated on 131 reference reactions from four GMTKN55 subsets (below), which is a start, not a benchmark: one stereoisomer case, no charged species, no transition metals, nothing above 26 atoms. The first engineering milestone is a ≥300-case set with DLPNO-CCSD(T) references per question type, recalibrated weekly from the ledger.
DFT single points on xTB geometries do not improve conformer energies: on 79 alkane and diol conformers the DFT//xTB error (MAD 2.0 kcal/mol) is worse than GFN2-xTB itself (1.0). Conformer and stereo questions need DFT geometries at Tier 2; the verdicts say so, and the protocol will change.
Closed-shell main-group organics only. Transition metals, radicals and excited states are out of scope and are refused rather than guessed.
Single conformer at Tier 2. Conformer ensembles are only sampled at Tier 1.
The rule-based planner handles the demo grammar. Production uses a schema-constrained LLM planner with the same output type; it is in the code but not in the recorded runs.
CPU PySCF in the prototype. Production targets GPU4PySCF or a partner GPU engine, which is a ~50× speed-up at identical numbers.

Methods roadmap

Tier 1b: learned potentials. MACE-OMol25 / UMA / Egret-1 via ASE as a second independent cheap tier. Disagreement between an xTB family member and an MLIP is far more informative than GFN1 vs GFN2.
Tier 2 defaults. def2-TZVP single points on GPU; ωB97X-D or r²SCAN-3c as alternates per question type; Skala where it benchmarks better.
Thermochemistry. Harmonic frequencies at Tier 1 (xTB) for ZPE and thermal corrections; quasi-RRHO entropies; Boltzmann-weighted conformer ensembles.
Solvation. SMD at Tier 2; CPCM-X / ALPB consistency at Tier 1; explicit-solvent microsolvation for ionisable sites.
New question types. Reaction barriers (NEB / FSM with MLIP pre-screen), pKa and redox potentials, crystal structure prediction and polymorph ranking.
Calibration. Weekly recalibration of per-class priors from escalation triples; published calibration curves per question type.

Ledger schema (what every answer carries)

{ "kind": "engine", "molecule": "isobutene",
  "payload": { "engine": "B3LYP-D3(BJ)/def2-SVP", "version": "pyscf 2.14.0",
               "tier": "tier2", "energy_hartree": -157.12512, "wall_time_s": 3.87,
               "converged": true, "geometry_source": "GFN2-xTB opt", "notes": "" } }

{ "kind": "gating", "molecule": "isobutene",
  "payload": { "margin_kcal": 6.39, "uncertainty_kcal": 6.00, "threshold_kcal": 9.00,
               "escalate": true, "reason": "margin 6.39 < 1.5x uncertainty 6.00 [ring-count change]" } }

{ "kind": "warning", "payload": { "kind": "tier1_overturned",
  "message": "Tier 1 (GFN2-xTB) would have picked cyclobutane; DFT overturned it in favour of isobutene." } }

Any report can be regenerated from its ledger plus pinned engine versions. The tier1_overturned entries are the training signal.

Part 7

Why this wins.

Open models are our friend

Meta, Microsoft and NVIDIA give away general potentials and functionals. We consume them as tiers. Our cost of goods falls every time they ship.

Proprietary data where it matters

Every escalation produces a (cheap prediction, DFT truth) pair in exactly the chemical space customers care about: transition metals, charged species, solvated tautomers, the long tail where open datasets are thin.

Validated workflows per question

"Is this the stable polymorph?" and "will this catalyst cycle close?" need different protocols. We encode and benchmark each once, then sell it a thousand times.

Trust is the interface

Chemists stop using a tool after one wrong answer they didn't see coming. Maxwell tells them when it isn't sure, and shows which method produced every number.

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 (QC Ware) · Rowan

Bench chemist

Skala · UMA / OMol25 (free building blocks)

maxwell.ai

Schrödinger took 30 years to reach $200M of software revenue selling to the expert tier. Bottom-up: 200K industrial chemists at €2,400/yr is a €480M pool before materials. The credible seed story is 20 pharma and agro sites, 50–200 bench chemists each, within 36 months.

Business model.

Individual / academic

€0 → €49/mo

Tier 1 free, capped
DFT escalation credits
Purpose: distribution and the next cohort's defaults

Team

€200 / seat / mo + compute

Shared projects, ELN export
Audit trail, Methods paragraphs
Metered compute at a margin

Enterprise

€150–500K / yr + compute

Deployed inside the customer's cloud account
Structures never leave their VPC
SSO, validated workflow packs, support

Gross margin target 70–75% blended. Compute stays at 15–25% of revenue because the gate keeps DFT spend to the handful of molecules that need it.

Part 8

Roadmap.

Q1 2027MVP: tautomer / conformer / stability agent; xTB + MLIP + GPU DFT tiers. 10 academic groups live.

Q2 2027Two pharma design partners in VPC. Benchmark published: accuracy vs cost vs Gaussian defaults.

Q3 2027Reaction energetics and barriers. First paid enterprise contract. 1,000 weekly actives.

Q4 2027Ledger at 1M labelled triples. First in-house tier beats open MLIP on customer chemistry.

H1 2028pKa / redox / polymorph packs. €1.5M ARR. Series A.

The ask.

Raising €3.5M seed 18–20 months

Team55%

Compute & data25%

Go-to-market10%

Buffer10%

Milestones this round buys: two paid pharma contracts, a 1M-triple escalation ledger, and a published benchmark showing Maxwell matches expert-chosen DFT protocols at under 10% of the compute.

The founding team we are assembling

CEO / chemistry: a practising computational chemist who has been the queue inside a pharma or agro company.

CTO / agents: has shipped an agentic product where an LLM planner drives deterministic tools, with evals.

Founding scientist / ML potentials: trained or fine-tuned MLIPs, credible in the OMol25 community.

What we want from design partners

Two groups, academic or industrial, who will run their real tautomer, conformer and stability questions through Maxwell for three months, tell us when it is wrong, and let us keep the anonymised (structure, cheap prediction, DFT truth) triples. In return: free compute, direct access to the team, and your protocol encoded as a validated workflow pack with your name on the benchmark.

Try the demo

Part 9

The deck, in 15 slides.

Arrow keys to navigate inside the frame. Open full screen ↗

Part 10

Advanced demo: all controls, live engine.

Every recorded run with replay speed, accuracy mode and k exposed, plus a live mode that talks to a running maxwell-web server for questions of your own. Open full screen ↗