Catalog
174 products, 21 families
93 hardware SKUs, 67 software products and 14 cloud services. Filter, search and sort — every entry has a detail page with invented specs, pricing and a market analog.
AXION Play GPUs
Consumer graphics cards with hardware ray reconstruction and neural upscaling.
AXION Play X400
Small form factor, no external power.
- Compute cores
- 2,560
- Memory
- 8 GB GDDR7 / 288 GB/s
AXION Play X500
Esports frame rates, 1080p native.
- Compute cores
- 4,608
- Memory
- 10 GB GDDR7 / 448 GB/s
AXION Play X600
High-refresh 1440p on a 180 W budget.
- Compute cores
- 7,168
- Memory
- 12 GB GDDR7 / 672 GB/s
AXION Play X700
The 1440p default.
- Compute cores
- 10,240
- Memory
- 16 GB GDDR7 / 896 GB/s
AXION Play X700 Mobile
Desktop-class 1440p in a 16-inch chassis.
- Compute cores
- 8,192
- Memory
- 16 GB GDDR7
AXION Play X800
Enthusiast 4K without the flagship power bill.
- Compute cores
- 16,384
- Memory
- 24 GB GDDR8 / 1.3 TB/s
AXION Play X900
The halo card. 4K at 240 Hz with neural frame synthesis.
- Compute cores
- 24,576
- Memory
- 32 GB GDDR8 / 1.9 TB/s
AXION Play X950 TITAN
Prosumer memory capacity on a consumer driver stack.
- Compute cores
- 26,112
- Memory
- 48 GB GDDR8 / 2.1 TB/s
AXION Studio GPUs
Workstation graphics for rendering, CAD, simulation and on-desk model training.
AXION Studio Mobile M3000
Certified drivers in a mobile workstation.
- Memory
- 24 GB ECC
- TGP range
- 45–120 W
AXION Studio S2000
Entry certified board, 70 W, single slot.
- Memory
- 16 GB GDDR7
- Board power
- 70 W
AXION Studio S4000
Certified graphics for the mainstream workstation.
- Memory
- 24 GB GDDR7 ECC
- Board power
- 150 W
AXION Studio S6000
The dual-slot rendering workhorse.
- Memory
- 48 GB GDDR8 ECC
- Board power
- 300 W
AXION Studio S9000
96 GB of ECC memory under your desk.
- Memory
- 96 GB GDDR8 ECC
- Board power
- 375 W
AXION Studio Virtual GPU
Slice one board across a floor of thin clients.
- Memory
- 48 GB ECC
- Max sessions
- 64
Accelerators
Rack-scale training and inference accelerators built on the Meridian and Halide architectures.
AXION H1 (Halide Next)
Next architecture. Sampling to launch partners.
- Memory
- 576 GB HBM4e (target)
- FP4 dense
- 45 PFLOPS (target)
AXION M200
Air-cooled training in existing racks.
- Memory
- 144 GB HBM3e / 4.8 TB/s
- FP8 dense
- 4 PFLOPS
AXION M400
The training accelerator the rest of the rack is built around.
- Memory
- 288 GB HBM4 / 8 TB/s
- FP4 dense
- 20 PFLOPS
AXION M400 SuperModule
Eight accelerators, one memory domain.
- Accelerators
- 8 × M400
- Pooled memory
- 2.3 TB HBM4
AXION M400X
Same die, more memory, longer context.
- Memory
- 432 GB HBM4 / 9.2 TB/s
- FP4 dense
- 20 PFLOPS
AXION V20
72 W of inference for edge racks.
- Memory
- 24 GB GDDR7
- FP8 dense
- 310 TFLOPS
AXION V40
Inference, graphics and video on one card.
- Memory
- 48 GB GDDR8 ECC
- FP8 dense
- 1.4 PFLOPS
AXION V90
PCIe inference card for the general fleet.
- Memory
- 96 GB HBM3e / 3.2 TB/s
- FP4 dense
- 3.6 PFLOPS
Server CPUs
Arm-based datacenter CPUs and coherent CPU+GPU superchips.
AXION Fusion GM
One CPU, two accelerators, one address space.
- Composition
- 1 × C1 + 2 × M400
- Unified memory
- 1.0 TB
AXION Vantage C1
72 cores, one NUMA domain, 500 W of memory bandwidth.
- Cores
- 72 @ 3.6 GHz
- Memory
- 480 GB LPDDR6 / 1.1 TB/s
AXION Vantage C1 Edge
24 cores, 85 W, extended temperature.
- Cores
- 24 @ 2.8 GHz
- Temp range
- -40 to 85 °C
AXION Vantage C1X
Two C1 dies on one package for general-purpose racks.
- Cores
- 144 @ 3.3 GHz
- Memory
- 960 GB LPDDR6
AXION Vantage C2
Announced. 192 cores and on-package HBM.
- Cores
- 192 (target)
- Memory tiers
- 96 GB HBM + 1.2 TB LPDDR6
DPUs & SuperNICs
Infrastructure processors that offload networking, storage and security from the host.
AXION LinkPort 400
The volume adapter for inference clusters.
- Throughput
- 400 Gb/s
- Ports
- 1 × OSFP
AXION LinkPort 800 SuperNIC
800 GbE built for all-reduce, not for web traffic.
- Throughput
- 800 Gb/s
- Direct placement
- GPU-to-GPU RDMA
AXION Nexus D3 DPU
Move the infrastructure off the host.
- Ports
- 2 × 400 GbE
- Cores
- 16 Arm + packet pipeline
AXION Nexus D3 Lite
DPU features at a single-port price.
- Ports
- 1 × 200 GbE
- Cores
- 8 Arm
AXION Nexus D4 DPU
1.6 T DPU with a confidential-computing root of trust.
- Ports
- 2 × 800 GbE
- Cores
- 32 Arm
Networking
Scale-up and scale-out fabrics: lossless Ethernet, HyperLink, optics and cabling.
AXION Cascade E-102T
102 Tb/s spine for the 100k-accelerator hall.
- Capacity
- 102.4 Tb/s
- Optics
- CPO + OSFP hybrid
AXION Cascade E-51T
Ethernet that behaves itself under all-reduce.
- Ports
- 64 × 800 GbE
- Capacity
- 51.2 Tb/s
AXION Cascade HL-800
HyperLink fabric switch for lossless training clusters.
- Ports
- 64 × 800 G
- Capacity
- 51.2 Tb/s
AXION DirectLink Cabling
Passive, active and optical, all one part family.
- Types
- DAC / ACC / AOC
- Lengths
- 0.5 m – 100 m
AXION Fabric Director
The appliance that runs the fabric.
- Scale
- 8,192 endpoints
- Telemetry
- 1 s resolution
AXION HyperLink Switch HX72
The scale-up switch inside the rack.
- Endpoints
- 72 accelerators
- Bisection
- 130 TB/s
AXION Photon 800G OSFP
Linear-drive optics, 40% less power per lane.
- Rate
- 800 Gb/s (8 × 100 G)
- Reach
- 50 m to 2 km SKUs
AXION Photon CPO-1.6T
Optics co-packaged with the switch ASIC.
- Rate
- 1.6 Tb/s per engine
- Power
- 4.1 pJ/bit
Systems & Racks
Reference servers, liquid-cooled racks and turnkey AI factories.
AXION AI Factory Pod
A validated 32-rack cluster, cabled and burned in.
- Scale
- 32 racks / 2,304 accelerators
- Storage
- 12 PB flash, 480 GB/s
AXION Atelier A1
A deskside 1 TB-memory dev box on a 15 A circuit.
- Compute
- 2 × Fusion GM
- Unified memory
- 2 TB
AXION Atelier Mini
128 GB of coherent memory in a 1.2 L box.
- Compute
- 1 × Fusion Lite
- Unified memory
- 128 GB
AXION Forge F1
Eight accelerators, air-cooled, 8U.
- Accelerators
- 8 × M200
- CPU
- 2 × Vantage C1X
AXION Forge F1L
The liquid-cooled flagship node.
- Accelerators
- 8 × M400
- Pooled memory
- 2.3 TB HBM4
AXION Outpost E2
Short-depth, ruggedized, fanless-optional edge server.
- Accelerators
- 2 × V20 or 1 × V40
- Depth
- 300 mm
AXION Titan Rack TR144
Two racks, one coherent domain.
- Accelerators
- 144 × M400X
- Coherent memory
- 62 TB
AXION Titan Rack TR72
72 accelerators behaving as one.
- Accelerators
- 72 × M400
- Coherent memory
- 20.7 TB
AXION Vault S1
Storage that can actually feed a training run.
- Capacity
- 1.5 PB raw per 2U
- Read
- 160 GB/s per node
Edge Modules
Power-efficient system-on-modules and dev kits for robots, cameras and instruments.
AXION Spark IX Industrial
Ten-year supply, certified for the factory floor.
- Base module
- Spark N2, ECC
- Certifications
- IEC 61508 SIL-2 elements
AXION Spark N1
5 W of on-device inference for the cheapest bracket.
- AI throughput
- 34 TOPS INT8
- Memory
- 8 GB LPDDR5
AXION Spark N2
The mainstream robotics module.
- AI throughput
- 180 TOPS INT8
- Memory
- 32 GB LPDDR5X
AXION Spark N4
Humanoid-class compute in a module.
- AI throughput
- 1,100 TOPS INT8 / 550 TFLOPS FP8
- Memory
- 64 GB LPDDR5X
AXION Spark N4 Dev Kit
The robot bring-up board.
- Module
- Spark N4
- Fieldbus
- EtherCAT, CAN-FD ×4
AXION Spark N8
Two N4 dies for machines that carry people.
- AI throughput
- 2,200 TOPS INT8
- Memory
- 128 GB LPDDR5X
AXION Spark Nano Dev Kit
$249 to start. Same stack as the racks.
- Module
- Spark N1
- Included
- Camera, PSU, SD image
AXION Spark Vision Module
Sensor, ISP and inference on one 30 mm board.
- Sensor
- 5 MP global shutter
- Output
- Detections + H.265 proxy
Automotive
Functional-safety compute for assisted and autonomous driving.
AXION Autoline A1
Assisted driving compute for volume programs.
- AI throughput
- 320 TOPS
- Safety
- ASIL-D systematic
AXION Autoline A2
Level 4 compute on one board.
- AI throughput
- 2,600 TOPS
- Safety
- ASIL-D, dual channel
AXION Autoline Dev Kit
A bench version of the vehicle computer.
- Compute
- 1 × Autoline A2
- Replay
- 40 Gb/s sensor injection
AXION Autoline Safety Island
The small chip that says no.
- Cores
- 3 × lockstep Cortex-class
- Function
- MRM, watchdog, arbitration
AXION Autoline Twin
Two A2s, cross-checked, for robotaxi duty cycles.
- Boards
- 2 × Autoline A2
- Arbitration
- Hardware comparator
AXION Cockpit C1
Four displays, one SoC, one hypervisor.
- Displays
- 4 × 4K
- Isolation
- Type-1 hypervisor, 6 VMs
Robot Platforms
Reference humanoids, manipulators, drones and underwater vehicles.
AXION Kinesis Arm 7
A 7-DoF arm with torque sensing in every joint.
- DoF / reach
- 7 / 850 mm
- Payload
- 12 kg
AXION Kinesis H1
A reference humanoid, sold to be modified.
- Height / mass
- 1.72 m / 61 kg
- Actuated joints
- 41
AXION Kinesis H1 Torso
The half you actually need for manipulation research.
- Arms
- 2 × 7-DoF
- Payload
- 8 kg per arm
AXION Kite D1
Autonomous drone that flies without GPS.
- Flight time
- 34 min
- Navigation
- Visual-inertial, GPS-optional
AXION Marlin U1
Underwater vehicle with on-board perception.
- Depth rating
- 300 m
- Endurance
- 6 h survey
AXION Porter P2
Autonomous mobile base, 600 kg payload.
- Payload
- 600 kg
- Speed
- 2.2 m/s
AXION Ranger Q4
Quadruped for inspection routes nobody wants to walk.
- Payload
- 14 kg
- Runtime
- 2.5 h / autodock
AXION Scalpel S1 Research Cart
Surgical-grade motion, research-only license.
- Arms
- 3 × 8-DoF
- Precision
- ±0.02 mm
Quantum
QPUs, control electronics and the classical fabric that keeps them in sync.
AXION Frost C1
Cryo-CMOS multiplexing at 4 K.
- Operating temp
- 4 K stage
- Multiplexing
- 64:1
AXION Lumen Q128
128 physical qubits, honest error bars.
- Qubits
- 128 physical
- 2-qubit fidelity
- 99.72% median (measured)
AXION Maestro QC1
Qubit control electronics with a real-time GPU in the loop.
- Channels
- 256 AWG / 128 digitizer
- Timing jitter
- < 400 ps
AXION QLink Bridge
Low-latency link between the QPU and the cluster.
- Round trip
- 3.1 µs
- Topology
- Up to 8 QPUs per host
AXION Syndrome S1
A decoder ASIC for surface codes.
- Code support
- Surface, distance ≤ 15
- Throughput
- 1 M syndromes/s
Emerging Silicon
Neuromorphic, photonic, analog in-memory and DNA-storage research silicon.
AXION Cortex N1
Neuromorphic silicon for always-on sensing.
- Neurons
- 1.2 M / 480 M synapses
- Idle power
- 18 µW
AXION Helix Vault
DNA archival storage. Write once, read in days.
- Density
- ~1 PB per gram
- Write rate
- 24 MB/day
AXION Monolith W1
A wafer as one chip.
- Cores
- 912,000
- On-wafer SRAM
- 40 GB
AXION Prism P1
Photonic matrix engine, research access only.
- Peak
- 760 TOPS-equivalent
- Efficiency
- 0.9 pJ/MAC
AXION Tesseract A1
In-memory analog compute for fixed models.
- Weight capacity
- 64 M analog cells
- Efficiency
- 42 TOPS/W
Power & Cooling
Busbars, CDUs, immersion tanks and rack telemetry for dense AI halls.
AXION Chill C900 CDU
900 kW coolant distribution, in-row.
- Capacity
- 900 kW
- Approach temp
- 3 °C
AXION Cold Plate Kit
The plates, manifolds and quick-disconnects, qualified together.
- Heat flux
- 160 W/cm²
- Disconnects
- Dripless, 5,000 cycles
AXION Deepwell I40
Single-phase immersion, 40 kW per tank.
- Capacity
- 40 kW per tank
- Fluid
- Single-phase dielectric
AXION PowerSpine 1MW
An 800 VDC busbar for the back of the rack.
- Capacity
- 1 MW per rack
- Bus voltage
- 800 VDC
AXION Pulse Rack Telemetry
Know which node is about to lie to you.
- Channels
- 96 per rack
- Rate
- 1 Hz all channels
AXION Volt 33k Power Shelf
33 kW of rectification with N+1 and no drama.
- Output
- 33 kW, 800 VDC
- Redundancy
- N+1 hot-swap
Consumer Devices
Handhelds, AR glasses, capture hardware and the set-top streaming box.
AXION FrameGrab 4K
Capture card with the encoder on board.
- Capture
- 4K 144 Hz / 1440p 240
- Encode
- AV1, H.265 hardware
AXION Play Dock eGPU
An external X600 for laptops with a USB4 port.
- Included GPU
- Play X600
- Host link
- USB4 / Thunderbolt 5
AXION Play Portable
A handheld that streams or runs local.
- Display
- 7" 1200p 120 Hz OLED
- Battery
- 60 Wh, 4–6 h
AXION Play TV
The set-top box for streamed play.
- Output
- 4K 120 Hz HDR
- Latency
- < 28 ms decode path
AXION Studio Cast
Webcam and mic with the cleanup done in hardware.
- Camera
- 4K 60, 1/1.8" sensor
- Audio
- 3-mic beamforming array
AXION Vista One
AR glasses that offload to your phone.
- Field of view
- 46° diagonal
- Mass
- 72 g
Core Libraries
The compute stack: kernels, compilers, math, data frames and graph analytics.
Conduit Data Loader
Stop starving the accelerators.
- Decode
- JPEG, JPEG2000, AV1, H.265, WAV, FLAC
- Pipelines
- Declarative graph, deterministic seeds
FLUX
The parallel programming model everything else is built on.
- Languages
- C++23, Python, Fortran, Rust (community)
- Targets
- All AXION accelerators, x86 + Arm hosts
FLUX Toolkit
Compiler, libraries, headers and samples in one install.
- Install size
- 6.2 GB full / 1.1 GB runtime
- Package managers
- apt, dnf, conda, pip, brew
flux-blas
Dense linear algebra, at the achievable roofline.
- Coverage
- BLAS 1/2/3 + extensions
- Peak efficiency
- 94% of dense roofline
flux-comm
Collectives that know the topology.
- Collectives
- All-reduce, all-gather, reduce-scatter, A2A
- Transports
- HyperLink, RoCE, IB, PCIe, shared mem
flux-dnn
Hand-tuned primitives for neural network layers.
- Primitives
- Attention, conv, norm, GEMM epilogues
- Precisions
- FP4, FP6, FP8, BF16, TF32, FP64
flux-fft
Transforms for radio, imaging and seismic.
- Dimensions
- 1D/2D/3D, batched
- Distributed
- Multi-node slab and pencil
flux-opt
Routing, scheduling and assignment solvers.
- Problems
- VRP, TSP, bin packing, job shop, LAP
- Guarantee
- Anytime: best-so-far at deadline
flux-quantum
Simulate quantum circuits while the hardware grows up.
- Methods
- State vector, tensor network, MPS
- Max qubits
- 42 state-vector on 72-way rack
flux-solve
Direct and iterative solvers for engineering codes.
- Direct
- LU, QR, Cholesky, SVD, eigen
- Iterative
- AMG, CG, GMRES, BiCGStab
flux-sparse
Sparse linear algebra that does not fall off a cliff.
- Formats
- CSR, CSC, COO, BSR, structured 2:4
- Kernels
- SpMV, SpMM, SpGEMM, triangular solve
Torrent DataFrames
pandas and Polars syntax, accelerator execution.
- APIs
- pandas, Polars, Arrow
- Adoption
- Zero-code-change shim
Torrent Graph
Graph analytics at billion-edge scale.
- API
- NetworkX-compatible backend
- Algorithms
- PageRank, Louvain, BFS/SSSP, centrality
Torrent ML
Classical machine learning, still the bulk of production.
- API
- scikit-learn compatible
- Algorithms
- GBDT, k-means, DBSCAN, UMAP, t-SNE, kNN
Inference & Serving
Compilers, servers and microservices that put models into production.
AXION Cells
Pre-built model containers with one API each.
- Catalog
- 310 model containers
- Deploy targets
- K8s, bare metal, edge, cloud
Broker Model Router
Route each request to the cheapest model that can handle it.
- Routing
- Learned difficulty classifier
- Failover
- Cross-provider, per route
Conductor Inference Server
One server, every framework, real batching.
- Backends
- Kiln, PyTorch, ONNX, vLLM, XGBoost
- Batching
- Continuous + prefill/decode split
Kiln Compiler
Take a trained model, get a fast engine.
- Inputs
- PyTorch, ONNX, JAX export, GGUF
- Quantization
- FP8, FP4, INT8, INT4 + accuracy report
Sentinel Guardrails
Programmable rails around a model endpoint.
- Rails
- Topic, safety, PII, grounding, tool-auth
- Latency
- +11 ms typical
Slipstream Distributed Serving
Disaggregated inference across a whole rack.
- Architecture
- Disaggregated prefill/decode
- KV transfer
- HyperLink and RDMA
Foundation Models
Open-weight and licensed models for language, vision, speech, world and science.
Aether Weather
Kilometer-scale forecasts in minutes, not hours.
- Resolution
- 1 km regional / 12 km global
- Lead time
- 14 days, 50-member ensemble
Axiom-Code 34B
Trained on repositories, evaluated on repositories.
- Parameters
- 34 B
- Context
- 512,000 tokens
Axiom-Embed
Embeddings for text, code and images in one space.
- Dimensions
- 128 / 512 / 1,024 / 2,048
- Modalities
- Text, code, image
Axiom-L 500B
The frontier model, licensed for commercial use.
- Parameters
- 500 B total / 41 B active
- Context
- 1,000,000 tokens
Axiom-L 70B
The workhorse. Open weights.
- Parameters
- 70 B dense
- Context
- 256,000 tokens
Axiom-L 8B
Small, fast, and good enough for most routing targets.
- Parameters
- 8 B dense
- Context
- 128,000 tokens
Axiom-Mini 2B
Runs on a phone, a Spark module, or a car.
- Parameters
- 2 B
- Footprint
- 900 MB at INT4
Axiom-Reason 120B
Spends tokens to get the answer right.
- Parameters
- 120 B
- Thinking budget
- Adjustable, 0–64k tokens
Axiom-Rerank
The cheap fix for bad retrieval.
- Parameters
- 1.5 B
- Throughput
- 9,000 pairs/s per V40
Axiom-VL 40B
Documents, charts, video and screens.
- Parameters
- 40 B
- Inputs
- Image, multi-page PDF, video ≤ 2 h
Cadence Speech
Streaming ASR and TTS with word-level timing.
- ASR latency
- 240 ms streaming
- Languages
- 58 ASR / 34 TTS
Kinesis-VLA
A generalist policy for two-armed robots.
- Action space
- Bimanual, 14–41 DoF
- Control rate
- 50 Hz policy / 1 kHz servo
Lattice Chem
Interatomic potentials at DFT accuracy, 10,000× faster.
- Coverage
- 89 elements
- Accuracy
- 4 meV/atom vs DFT (in-domain)
Protean Fold
Structure prediction and binder design.
- Tasks
- Monomer, complex, ligand docking, binder design
- Accuracy
- Published CASP-style report
Reel Video
Video generation with a usable control track.
- Length
- Up to 20 s per shot
- Resolution
- 1080p, 24/30 fps
Verse World Model
Predict the next few seconds of the physical world.
- Horizon
- 6 s at 24 fps
- Conditioning
- Action, text, initial frames
Industry Platforms
Application frameworks for health, vision, telco, security, science and industry.
Aegis Cyber Analytics
Accelerated anomaly detection for defenders.
- Ingest
- 4 M events/s per node
- Sources
- DPU taps, EDR, cloud audit, identity
Atlas Grid Operations
Contingency analysis fast enough to act on.
- Contingencies
- 80,000 screened in 4 min
- Forecasting
- Wind/solar, 15 min – 72 h
Beacon Emergency Response
Situational awareness for incident commanders.
- Mosaicking
- Live, 40 km²/h from Kite D1
- Assessment
- Structure damage classes 1–5
Cadence Studio
Build voice agents that interrupt politely.
- Round-trip
- 620 ms median first audio
- Telephony
- SIP, WebRTC, PSTN partners
Curator Retail Intelligence
Shelf state, demand and pricing in one loop.
- Detection
- Out-of-stock, planogram compliance
- Forecasting
- SKU × store × day
Docket Legal Review
Discovery review with citations or nothing.
- Throughput
- 1.2 M documents/day per node
- Citations
- Mandatory page-level grounding
Foundry Manufacturing QA
Defect detection that learns from twelve examples.
- Cold start
- 12 real defect samples typical
- Synthetic data
- Loom-generated defect library
Harvest Agriculture
Per-plant decisions at tractor speed.
- Discrimination
- Crop vs weed, 42 species
- Speed
- 20 km/h at 2 cm resolution
Ledger Financial Risk
Overnight risk runs that finish before lunch.
- Speedup
- 60× vs CPU grid
- Coverage
- VaR, XVA, FRTB, CCAR-style stress
Mentor Education
Tutoring infrastructure, not a tutoring product.
- Grounding
- District curriculum ingest
- Transparency
- Full transcript to educators
Pitstop Sports Analytics
Skeletal tracking and tactical models from broadcast video.
- Tracking
- 22 players + ball, broadcast feed
- Latency
- Live, 1.4 s behind
Protean Bio Platform
The pipeline around the biology models.
- Workflows
- Screening, docking, ADMET, MD setup
- Screening
- Sequence-of-concern check on all designs
Sable Media Localization
Dub, subtitle and QC a catalog in a week.
- Languages
- 34 dub / 58 subtitle
- Lip sync
- Retimed, viseme-aware
Sightline Vision AI
Video analytics for operations, with policy controls built in.
- Scale
- Up to 1,000 streams per node
- Privacy
- On-device blur, retention policy engine
Signal RAN
A software radio access network on accelerators.
- Standards
- O-RAN 7.2x split, 5G SA, 6G research
- Capacity
- 4 × 100 MHz massive MIMO per V40
Terra Earth Systems
Downscale climate projections to the street.
- Layers
- Flood, fire, heat, wind, drought
- Resolution
- 100 m urban downscaling
Vitals Health Imaging
Imaging AI framework for hospitals and device makers.
- Modalities
- CT, MRI, X-ray, ultrasound, pathology
- Training
- Federated, differential-privacy option
Simulation & Digital Twins
Physically accurate worlds for design review, synthetic data and robot training.
Anvil Physics
Rigid bodies, soft bodies, cloth and fluids — differentiable.
- Solvers
- Rigid, soft, cloth, SPH/FLIP fluid
- Differentiable
- Yes, gradients through contact
AXION Loom
One scene graph, every DCC tool.
- Format
- OpenUSD native
- Live editing
- Multi-user, multi-app stage
Continuum Physics-ML
Surrogate models that respect the governing equations.
- Architectures
- Fourier neural operator, DeepONet, GNN
- Constraints
- PDE residual loss, conservation terms
Loom Kit SDK
Build your own tool on the stage.
- Languages
- Python, C++
- Modes
- Desktop app, headless service
Mint Synthetic Data
Labeled data for the cases your fleet never sees.
- Outputs
- RGB, depth, segmentation, boxes, lidar
- Labels
- Perfect by construction
Proving Ground
Train and test robots where breaking them is free.
- Sensors
- Camera, depth, lidar, IMU, F/T — calibrated
- Throughput
- 40,000 envs on one M400
Developer Tools
Profilers, debuggers, container registries and local workbenches.
AXION CLI
One binary for hardware, cluster and cloud.
- Install
- curl, brew, apt, winget
- Targets
- Local, on-prem cluster, AXION Cloud
AXION Depot
The registry for containers, models and Helm charts.
- Artifacts
- Containers, weights, charts, notebooks
- Supply chain
- Cosign signatures + SBOM per image
AXION Tools for VS Code
Kernel authoring with the profiler inline.
- Language
- FLUX C++ / Python IntelliSense
- Inline hints
- Occupancy, registers, spills
Bench Workbench
Same environment on your laptop and in the cluster.
- Hosts
- macOS, Windows, Linux
- Relocation
- Local → Atelier → Cloud, one command
flux-debug
A real debugger for accelerator code.
- Breakpoints
- Per-warp, conditional
- Checkers
- OOB, races, uninitialized reads
Lens Profiler
Find out which kernel is actually the problem.
- Analysis
- Roofline, occupancy, stall reasons
- Correlation
- Kernel to source line
Lens Systems Trace
The timeline view across CPU, accelerators and fabric.
- Scope
- CPU, accelerator, fabric, storage
- Scale
- Multi-node, 4,096 ranks
Provenance Model Cards
Generate the documentation your auditors will ask for.
- Captures
- Data lineage, hyperparams, evals, limits
- Standards
- Model card + EU AI Act annex templates
AXION Cloud
Managed capacity, orchestration, endpoints and fleet operations, billed per second.
Batch Jobs
Queue a training run and go home.
- Scheduling
- Gang, topology-aware, fair-share
- Resilience
- Auto checkpoint + node replacement
Cloud Insight
Know what every token and every watt cost you.
- Metrics
- MFU, memory pressure, fabric health
- Attribution
- Per team, project, endpoint, model
Cloud Instances
Accelerators by the second, no minimum term.
- SKUs
- M400, M400X, M200, V90, V40, V20
- Billing
- Per second, 2-minute minimum
Fine-tuning Service
Upload examples, get a private adapter.
- Methods
- LoRA, QLoRA, full SFT, DPO
- Data
- Private, not used for base training
Interruptible Capacity
Cheap accelerators with a 90-second warning.
- Discount
- 62–78% off on-demand
- Warning
- 90 s SIGTERM + hook
Loom Cloud
Streamed digital twins, no workstation required.
- Streaming
- WebRTC, 4K 60, 40 ms
- Concurrency
- 500 viewers per stage
Model Gallery
Try any model in the catalog before you deploy it.
- Models
- 410 first-party and partner
- Playground
- Free, rate limited
Outpost Fleet
Manage 10,000 edge devices you will never visit.
- Scale
- 100,000 devices per tenant
- Updates
- Signed, A/B partitions, auto-rollback
Play Stream
Your game library, rendered in a datacenter.
- Tiers
- Free (ads, queue) / Plus / Ultra
- Max quality
- 4K 120 HDR or 1080p 240
Private AI Factory
Our operations team, your building.
- Model
- Customer-owned or leased hardware
- Operations
- 24/7 AXION NOC, on-site SLA
Reserved Clusters
Contiguous capacity with a fabric you can benchmark.
- Size
- 32 – 4,096 accelerators
- Terms
- 1 month – 3 years
Runbook Cluster Manager
Operate your own cluster the way we operate ours.
- Scope
- On-prem, colo, and AXION Cloud
- Health
- Predictive node drain on degradation
Serverless Endpoints
Deploy a model, get a URL, pay per token.
- Models
- Catalog + your own weights
- Cold start
- 8.6 s median (warm cache)
Vector Store
Retrieval that keeps up with the model.
- Scale
- 10 B vectors per namespace
- Latency
- 6 ms p50 at 1 B vectors