Pantheon
Data centre, rack-scale, replaces the metered API bill
The Mickai Pantheon is the largest member of the compute family, the tier above the Prometheus edge server. It is a fully integrated, rack-scale sovereign system that plugs into an enterprise data centre and runs the whole Sovereign Intelligence Operating System on hardware the enterprise owns. The building block is the NVIDIA GB200 NVL72 rack, seventy-two Blackwell GPUs and thirty-six Grace CPUs fused into a single NVLink domain with 13.5 TB of HBM3e, liquid-cooled, purpose-built for trillion-parameter inference. Storage runs to multiple petabytes of NVMe for model weights, the retrieval corpus, and the Open Audit Record. The fabric is 800 Gb/s Quantum-X800 InfiniBand, so Pantheon scales from one rack to a multi-row hall without changing the software. Every token is produced inside the perimeter, and nothing leaves. The metered API bill, a nine-figure monthly obligation for a frontier-scale enterprise, is replaced by an estate the company owns: the marginal cost of a token becomes the electricity to produce it, and the data never crosses the boundary. The operator holds the keys, the audit record signs under FIPS 204 ML-DSA-65 at rack scale, and the system runs fully air-gapped where required. The price you pay for the estate is the estate you own. No meter, no per-seat licence, no egress.
The Pantheon was the temple raised to all the gods at once, the whole pantheon under a single dome. The data-centre system carries the name because it brings the entire cooperative of brains under one roof, at the scale of a bank or a state, on hardware the operator owns.
- Building block
- NVIDIA GB200 NVL72 rack, 72 Blackwell GPUs and 36 Grace CPUs in one NVLink domain, or HGX B300 nodes for modular build-out
- HBM per rack
- 13.5 TB HBM3e, the working set for frontier-scale models at high batch concurrency
- System memory
- Multi-terabyte coherent Grace memory per rack for context, retrieval, and orchestration state
- Storage
- 1 to 4 PB NVMe per deployment, model weights, knowledge corpus, and the Open Audit Record ledger, encrypted under operator keys
- Intra-rack fabric
- NVLink 5, 1.8 TB/s per GPU, the rack behaves as a single accelerator
- Cluster fabric
- NVIDIA Quantum-X800 InfiniBand at 800 Gb/s, or Spectrum-X Ethernet, scales one rack to a multi-row hall
- Power and cooling
- About 120 to 130 kW per NVL72 rack, facility liquid cooling, or air for B300 nodes
- Signing
- Open Audit Record signed under FIPS 204 ML-DSA-65 at rack scale, operator keys in an on-prem HSM, optional FPGA at line rate
- Deployment tiers
- Pantheon One, 1 rack, around 10,000 seats. Quad, 4 racks, around 50,000. Octa, 8 racks, around 150,000, the half-billion-a-month case. Hall, 16 to 64 racks, 500,000 and up
- Commissioning
- Pre-commissioned racks, Mickai seals the OAR root into the operator HSM on site and hands over the keys, sovereign from day one, air-gapped where required
Frontier-class inference for hundreds of thousands of operators, signed at rack scale, sized from the customer's own token telemetry.
Beyond the standard Windows and Linux workflow, here is what Pantheon does with the Mickai SIOS running on the box. Concrete capability per category, with the level the hardware can carry.
The whole SIOS for an organisation of five hundred to several hundred thousand operators. Hundreds of thousands of concurrent sessions against a shared brain pool, multi-tenant audit isolation, clearance mapped to the corporate directory, signed audit at rack scale. The cooperative of brains behind the firewall of a bank or a state.
GB200 NVL72 racks serve frontier-class inference at the throughput of a nine-figure monthly API contract, sized from the customer's measured token volume with redundancy and peak headroom. The Mickai frontier model drops onto the identical estate when it ships.
The intelligence backend for the entire organisation. Login federation, role-based clearance, signed-action audit on every operator, integration with on-prem identity, every internal tool served from inside the perimeter.
Agent fleets per developer across the whole engineering org, signed code review and signed deploy pipelines, on-prem inference for every internal developer tool, no source code ever leaving the building.
Render and simulation backend at national-infrastructure scale, CAD, BIM, 3D, and virtual production for hundreds of concurrent artists and engineers.
AI video and post-production backend for multiple concurrent productions at the largest model size, broadcast colour pipelines, real-time virtual production.
- Replaces a nine-figure monthly metered API bill with an estate the enterprise owns
- Right-sized from the customer's own token telemetry, neither short nor wasteful
- Every token produced inside the perimeter, the data never leaves the building
- Operator-held keys, Open Audit Record signed at rack scale, air-gap mode with continuous attestation
- Clearance-gated retrieval mapped to the enterprise directory
- Dead-man succession with sealed envelopes, control survives the operator
- Mickai SDK so the enterprise builds its own applications on the same sovereign substrate
- Payback measured in days of the avoided bill, then a token costs only the electricity to produce it
- MICKAI(TM) on the rack door, built in Britain
Mickai is held privately by its founder, Micky Irons. One hundred and one UK patent applications have been filed against the substrate, with approximately 2,234 claims across them, on the public UK IPO register under GB2607309.8 onwards. The corpus covers the cooperative architecture, the Open Audit Record, the FIPS 204 ML-DSA-65 signing primitive, the clearance-gated retrieval, the audio dual-layer watermark, the silicon root of trust, and the host-acceptance attestation that lets the Mickai bundle migrate from one machine to another without breaking its signed chain.
Poseidon, the Mickai Sovereign AI SoC. In-house silicon that complements the GPU fabric on every SKU. Carries the OAR signing primitive in silicon, holds the hardware identity and cryptographic root of trust, and sits between the network interface and the GPU fabric so every action that crosses the bus can be policy-checked before commit. Optional on every SKU as the silicon lands.
Pricing TBD. The price on day one is the price for life. No subscription for context or usage. Only paid upgrades through the sandboxed channel, optional.
Launch window: 2026 onward, post-funding rollout