Ignite Keynote · Fictional Demo Build
Accelerated computing for
everything that thinks.
174 products across silicon, systems, networking, robotics, models, simulation and cloud — one programming model from a $149 module to a 132 kW rack. None of it is real; all of it is here to click on.
174
Products & services
1.4 EF
FP4 per Titan rack
20.7 TB
Coherent memory per rack
11
Cloud regions
The stack
One architecture, three layers, nothing rewritten between them
A container built on a $3,999 Atelier Mini runs unmodified on a 32-rack pod and on a serverless endpoint. That is the whole product strategy.
01 / Silicon
Accelerators, CPUs, DPUs and the fabric between them
Fourteen hardware families from a 5 W edge module to a 132 kW rack that behaves as one machine — plus the busbars, cold plates and optics that make a hall of them work.
93 products →
02 / Software
FLUX, the libraries, the models and the industry platforms
One programming model underneath everything, a free library stack on top of it, open-weight models, and application frameworks for eighteen industries.
67 products →
03 / Cloud
Capacity, endpoints and fleet operations, billed per second
Rent an accelerator for two minutes or a 4,096-way cluster for three years. Same images, same telemetry, same CLI as the hardware on your own floor.
14 products →
Featured
What the keynote was about
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 Titan Rack TR72
72 accelerators behaving as one.
- Accelerators
- 72 × M400
- Coherent memory
- 20.7 TB
Axiom-L 500B
The frontier model, licensed for commercial use.
- Parameters
- 500 B total / 41 B active
- Context
- 1,000,000 tokens
Reserved Clusters
Contiguous capacity with a fabric you can benchmark.
- Size
- 32 – 4,096 accelerators
- Terms
- 1 month – 3 years
AXION Spark N4
Humanoid-class compute in a module.
- AI throughput
- 1,100 TOPS INT8 / 550 TFLOPS FP8
- Memory
- 64 GB LPDDR5X
AXION Loom
One scene graph, every DCC tool.
- Format
- OpenUSD native
- Live editing
- Multi-user, multi-app stage
AXION Kinesis H1
A reference humanoid, sold to be modified.
- Height / mass
- 1.72 m / 61 kg
- Actuated joints
- 41
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
Families
21 product families
Every family carries the same driver ABI, the same telemetry schema and the same container runtime.
AXION Play GPUs
HardwareConsumer graphics cards with hardware ray reconstruction and neural upscaling.
AXION Studio GPUs
HardwareWorkstation graphics for rendering, CAD, simulation and on-desk model training.
Accelerators
HardwareRack-scale training and inference accelerators built on the Meridian and Halide architectures.
Server CPUs
HardwareArm-based datacenter CPUs and coherent CPU+GPU superchips.
DPUs & SuperNICs
HardwareInfrastructure processors that offload networking, storage and security from the host.
Networking
HardwareScale-up and scale-out fabrics: lossless Ethernet, HyperLink, optics and cabling.
Systems & Racks
HardwareReference servers, liquid-cooled racks and turnkey AI factories.
Edge Modules
HardwarePower-efficient system-on-modules and dev kits for robots, cameras and instruments.
Automotive
HardwareFunctional-safety compute for assisted and autonomous driving.
Robot Platforms
HardwareReference humanoids, manipulators, drones and underwater vehicles.
Quantum
HardwareQPUs, control electronics and the classical fabric that keeps them in sync.
Emerging Silicon
HardwareNeuromorphic, photonic, analog in-memory and DNA-storage research silicon.
Power & Cooling
HardwareBusbars, CDUs, immersion tanks and rack telemetry for dense AI halls.
Consumer Devices
HardwareHandhelds, AR glasses, capture hardware and the set-top streaming box.
Core Libraries
SoftwareThe compute stack: kernels, compilers, math, data frames and graph analytics.
Inference & Serving
SoftwareCompilers, servers and microservices that put models into production.
Foundation Models
SoftwareOpen-weight and licensed models for language, vision, speech, world and science.
Industry Platforms
SoftwareApplication frameworks for health, vision, telco, security, science and industry.
Simulation & Digital Twins
SoftwarePhysically accurate worlds for design review, synthetic data and robot training.
Developer Tools
SoftwareProfilers, debuggers, container registries and local workbenches.
AXION Cloud
CloudManaged capacity, orchestration, endpoints and fleet operations, billed per second.
Open weights
The Axiom model family
Frontier, workhorse, small and on-device. Evaluation cards include the tasks each model fails, which is the part usually left out.
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
Try the fake control plane
Provision a cluster that does not exist.
The console runs against a real API in this app: it allocates instances, streams simulated utilization telemetry, meters per-second billing and lets you deploy model endpoints. No hardware, no charges, no account.
Launch the console