One AI that does it all.
Aether is a single foundation model that designs molecules, simulates engineering systems, lays out chips, runs autonomous wet-labs and writes software. One model. Every discipline. The seven-figure stacks it retires are named below.
The frontier labs describe the world. We simulate it.
Apex is the foundation-model lab built on a different axis of the frontier. The chat-trained labs read text and reason about reality. Aether is trained on reality — petabytes of solver trajectories, instrument traces and design history — and runs forward in time across the physical world the way reality does. Same scale of model. Same investment in evaluation and safety. A different ceiling.
Trained on simulation, not text
Our pretraining corpus is solver trajectories — CFD, FEA, MD, DFT, RTL — instrument runs and shipped-design archives. The chat labs train on the internet. Aether trains on what the internet describes.
Does the work, not the description
Aether integrates the equations forward in time, drives the instrument, lays out the chip, lands the PR. It produces the artefacts a senior engineer used to produce. The chat surface stops at words; Aether doesn't.
One model, every physical discipline
Engineering, semiconductors, drug discovery, autonomous labs, materials, fusion, software — the same Aether weights, the same context, the same safety story. Cross-domain transfer is the empirical result, not the marketing.
The AI that does the work.
Apex is a research and product company. We are training one foundation model that thinks in physics, chemistry, biology and silicon — and we are shipping it into the disciplines that have to face reality.
AI scientists, not chatbots
Aether does the work. It plans the study, picks the method, runs the agents, validates against measured reality, writes the memo. The work a senior engineer used to do by hand — done by the model.
One AI for every discipline
The same foundation model serves docking, structural FEA, chip signoff and wet-lab campaigns. Specialisation lives in the agents and the workloads — not in five sibling models with five sets of blind spots.
AGI through physical reality
We train on simulation trajectories, instrument data and design history. Reality is the regulariser. The model gets better the more reality it touches — and reality is what tells it when it's wrong.
Every workload. One model.
A twelve-panel slice of what Aether already runs — from docked ligands at −9.2 kcal/mol to tokamak plasma at 18 MW/m² divertor flux to ΛCDM N-body cosmology. The same foundation model behind every panel, with more workloads shipping every release.
- Drug discovery
Docked ligand · ΔG = −9.2 kcal/mol · FEP-validated
- Materials discovery
DFT + ML potential · screening 10⁶ candidates
- Quantum
VQE on 64 qubits · ground-state E₀ = −74.83 Ha
- Cosmology
ΛCDM N-body · 1024³ particles · σ₈ = 0.811
- Launch vehicles
Mach 2.4 ascent · 1.8% drag error vs flight
- Nuclear fusion
Tokamak MHD · 18 MW/m² divertor flux
- Automotive
Cd 0.214 · ΔCd 4.2% vs wind-tunnel
- Drones & eVTOL
Rotor wake · −4.1 dB ground-effect noise
- Robotics
6-DOF planning · 142 Nm shoulder torque
- Hypersonics
Mach 8 real-gas · 3.2 MW/m² stagnation
- Battery & energy
Pack thermal · 54 °C T_max at 4C
- Spacecraft
Sun-pointing · ±0.04° attitude knowledge
We are building AI scientists.
You write the brief in one sentence. Aether plans the study, picks the right solvers, runs the agents, validates each artefact against measured reality, and hands you back a memo signed by the model version that produced it.
“Reduce drag on this wing by 8% while keeping structural margin above 1.5 and respecting our manufacturing constraints. Hand me a CAE memo at the end.”
- 01Prompt
One sentence from an engineer
“Reduce drag on this wing by 8% while keeping structural margin above 1.5 and respecting our manufacturing constraints. Hand me a CAE memo at the end.”
- 02Plan
Aether decomposes the brief
The planner posts a study tree — geometry, mesh, CFD, FEA, optimiser, manufacturability check — for your approval before any compute runs.
- 03Execute
Specialist agents run the work
Mesh agents, CFD agents, FEA agents, optimiser agents — each a real production module — execute the plan in parallel.
- 04Validate
Every artefact against a benchmark
Cross-checked against the regression suite for that discipline. Drift halts the run and surfaces a citation, not a silent miscalibration.
- 05Report
A signed CAE memo with the receipts
Pressure contours, mode shapes, stress fields, optimisation history, model-version hash. Ready for design review.
Aether is the AI doing the work.
We aren't wrapping a chat model around your existing tools. We are training a foundation model that does the work — plans the study, picks the method, runs the agents, validates the artefact, writes the memo. Six things every senior practitioner used to do by hand. Now Aether does them.
- 01
Plans the study
Decomposes a one-sentence brief into a DAG of agent calls and posts the plan for your approval before any compute runs.
- 02
Picks the right method
Knows when RANS is enough and when DDES is required. Knows which solver pairs with which physics. The judgement of a senior practitioner, encoded.
- 03
Runs the work
Specialist agents execute the plan in parallel. Each is a real production module — not orchestration JSON glue.
- 04
Validates against reality
Cross-checks every artefact against the regression suite for that discipline: published benchmark, customer-validated study, your historical data. Drift halts the run.
- 05
Writes the report
Signed memos with figures, tables, margins-of-safety, optimisation history, convergence log and the model-version hash that produced each.
- 06
Improves itself
Failed cases enter the eval corpus. Successful pilots become regression tests. The model that solves your next study is better than the one that solved this one.
A foundation model that does the work, not just the chat about it — the substrate for general physical intelligence.
Aether is built to think in physics, chemistry, biology and silicon — the substrate of the physical world. We are moving toward general intelligence through the disciplines that have to face reality. The model that designs a rocket today is the model that runs a lab tomorrow.
“We expected an AI that would write Python around our existing tools. We got something that replaces them. The renewal we were dreading is the renewal we retired.”
The incumbent stack, named.
We don't sell intelligence as a service. We ship one AI that retires specific seven-figure software stacks — and we'll tell you which one is up first.
- DisciplineEngineering & scienceStack todayCommercial structural FEACommercial CFDHigh-frequency electromagneticsExplicit-dynamics crash codesMultiphysics suitesAetherAether for engineeringCFD, FEA, electromagnetics, multiphysics, acoustics, thermal — under one model.
- DisciplineSemiconductorStack todayLogic synthesisPlace-and-routeStatic timing analysisFunctional verificationMultiphysics signoffAetherAether for semiconductorsRTL through signoff, 3DIC architecture, multiphysics analysis, manufacturing optimisation.
- DisciplineDrug discoveryStack todayComputational chemistry suitesDocking and FEP toolsMolecular dynamics packagesADMET prediction pipelinesPharmacophore softwareAetherAether for drug discoveryDocking, FEP, MD, QM/DFT, ADMET, pharmacophore, generative chemistry, biologics.
- DisciplineWet-lab automationStack todayVendor lab schedulersScientific-data lake platformsCustom LIMS automationAetherAether for autonomous labsActive-learning campaigns driving liquid handlers and plate readers — with audit-grade lineage.
- DisciplineSoftware engineeringStack todayEditor-embedded coding assistantsLong-horizon coding agentsPull-request automationAetherAether for softwareWhole-repo edits, performance profiling, CVE-aware review, on-call drafts.
Aether 1.0 is here.
One AI, every discipline.
Specialisation lives in the agents and the workloads — not in different models. The same Aether weights serve a drug-discovery team in Boston, a chip-design house in Hsinchu and a hypersonics group in Huntsville.
Retire the seven-figure stack.
Pilots scoped to one workload your team actually cares about. Three to eight weeks. Win condition defined together. We walk away if it doesn't land.
See solutionsCall the model directly.
Python, TypeScript and OpenAPI. Streaming trajectories, tool use, memory, evaluation. Per-seat pricing — no per-token throttling on the work.
Read the APIMonths of R&D, compressed to hours.
The slowest part of R&D isn't the simulation — it's the hand-offs around it. Setting up the case, queuing the solve, re-meshing, writing the memo. Aether collapses each of those into the autonomous agent flow.
The shape of the thing.
Aether is one AI on the outside and a lot of careful engineering on the inside. The numbers below are the ones we measure ourselves against — and the ones customers measure us against.
Built for the regulated parts of the work.
The same workloads that make this useful — pharma, aerospace, semiconductor, public sector — are the ones with auditors, regulators and standing data boundaries. The platform is designed for that, not retrofitted for it.
One foundation model
Aether is one model, owned end-to-end. We train it ourselves on simulation, instruments and design history. We don't wrap a third-party chat API.
Your boundary, your data
Managed cloud, your VPC, on-prem, or air-gapped. ITAR-clean, GxP-grade, export-control-aware. Air-gapped deployments produce zero outbound traffic.
Signed, replayable artefacts
Every output is signed by the model version that produced it. Promotion gated by an eval suite. Old versions reproducible by hash.
Safety inside the model
Biosecurity refusals, controlled-pathogen lockouts, dual-use gating — at the model layer, not in a wrapper anyone can disable. Versioned red-team corpus.
Notes from the laboratory.
Architectures, evaluations, case studies, and the mistakes worth admitting. Reproduce the work before you bet a roadmap on it.
Where Aether is in production.
The same model serves a drug-discovery team in Boston, a chip-design house in Hsinchu and a hypersonics group in Huntsville. Every industry today — specialisation lives in the agents, not in different models.
Drug discovery
Docking, FEP, ADMET, virtual screening.
Chip design
RTL, P&R, timing, signoff.
Aerospace & energy
CFD, FEA, thermal, fatigue.
Materials & quantum
DFT, MD, multi-scale coupling.
Autonomous labs
Liquid handling, assay scheduling, EBR.
Software & SRE
Coding, review, migration, on-call.
Manufacturing
Additive, drying, process control.
Public sector
Sovereign deploys, export-controlled work.
Robotics
Sim-to-real, world models, policy training.
Cybersecurity
Vuln research, IR, detection, digital-twin defense.
Cyber-physical
OT/ICS, IT/OT boundary, consequence simulation.
Digital twins
Forward-rolled, multiphysics, synced to reality.
5G & telecom
RF propagation, RAN optimization, network twins.
Studio
Prompt-to-app — agents, sims and UIs.
Cloud
Compute, GPUs, data — managed to air-gapped.
Generative design
Topology optimization, simulation in the loop.
- Northwind Aerospace
- LYREBIRD
- Helix Silicon
- FERN/PLATFORM
- Continental Labs
"A foundation model trained on simulation, not text — with the receipts to back it up."
- MIT Technology Review
- The Information
- Stratechery
- IEEE Spectrum
- Nature Methods
- Semiconductor Engineering
A foundation model that does the work, not just the chat about it.
Aether is built to think in physics, chemistry, biology and silicon — the substrate of the physical world. We are moving toward general intelligence through the disciplines that have to face reality. The model that designs a rocket today is the model that runs a lab tomorrow.
Tell us which stack to retire.
Most pilots start with one line item — the renewal you'd rather not sign, or the license server that costs more than the engineer running it. We go from there.