Why Mickai builds the cooperative on NVIDIA Blackwell

The Mickai Sovereign Intelligence Operating System runs frontier-class artificial intelligence on hardware the operator owns, under keys the operator holds, with a complete and cryptographically verifiable record of everything the system does. The cooperative of brains, the Open Audit Record, the FIPS 204 ML-DSA-65 signing primitive, the clearance-gated retrieval, the AudioSeal dual-layer watermark, the host-acceptance attestation, the Poseidon Sovereign AI SoC roadmap, and every other piece of the substrate are Mickai-built, on the UK IPO public register under GB2607309.8 onwards, named inventor Micky Irons.

The inference fabric underneath the cooperative runs on NVIDIA Blackwell silicon. This piece explains why, and what that means for an operator buying a Mickai workstation.

The fabric is the floor, not the work

A useful distinction lives in every Mickai workstation. The fabric is the silicon, the memory, the interconnect, the power and cooling that turn electricity into matrix multiplication at scale. The work is the cooperative of brains, the SIOS layer, the audit substrate, the policy-gated retrieval, the trading bot, the AMT, the brains the operator actually deploys against the work they actually do.

The fabric should be the best available, because the work depends on it. The work should be Mickai, because the operator pays once and keeps it for life.

NVIDIA Blackwell is the best available fabric for frontier inference. The Mickai SIOS is the work.

NVIDIA Blackwell, the short version

NVIDIA announced the Blackwell architecture as the successor to Hopper. The line splits into two surfaces relevant to the Mickai workstation:

Blackwell Ultra B300 and the GB300 platform. The B300 is the data-centre chip and the GB300 NVL72 is the rack-scale platform around it. Each Blackwell Ultra GPU carries 288 GB of HBM3e memory with 8 TB per second of bandwidth, 20,480 CUDA cores, fifth-generation Tensor Cores with NVFP4 support, 15 PFLOPS of dense FP4 compute, a TGP of 1,400 W, mandatory liquid cooling, PCIe 6.0 host connectivity, and NVLink 5 at 50 GB per second per direction across eighteen connections per GPU. Against H200 it delivers roughly 3.5x the FP8 throughput and twice the memory.

RTX PRO 6000 Blackwell Workstation Edition. The professional desktop card. 96 GB of GDDR7, 1.8 TB per second of bandwidth, 24,064 CUDA cores, fifth-generation Tensor Cores, fourth-generation RT Cores, 600 W maximum draw, dual-slot, PCIe 5.0 x16. Launched March 2025. The workstation card replaces the RTX 6000 Ada Generation in the lineup and brings Blackwell capabilities to the desk without the data-centre power and cooling burden.

These two pieces of silicon are what the Mickai workstation fabric is built around.

Where each Blackwell part lands in the Mickai lineup

The Mickai workstation lineup is seven SKUs, named after Greek mythology, all built in Britain by our Birmingham manufacturing partner.

The flagship trading machine, Olympus, carries four Blackwell Ultra B300 NVL GPUs with NVLink, around 1.15 TB of HBM3e in aggregate. Olympus is engineered to deliver under ten milliseconds of end-to-end decision latency for the Mickai Trading Bot, from signal in to quorum-checked, OAR-signed decision out. The B300 is what makes that envelope realistic at the scale a serious trading operator demands.

The enterprise on-prem edge server, Prometheus, carries eight Blackwell Ultra B300 NVL GPUs in HGX configuration, around 2.3 TB of HBM3e in aggregate, dual 400 GbE with kernel-bypass DPDK, hot-swap everything. Five-trillion-parameter inference at organisational scale.

The mid workstation, Hyperion, ships with two RTX PRO 6000 Blackwell GPUs and 192 GB of GDDR7 across them, or a single H200 NVL for the operator who prefers the data-centre card on a regulated desk.

The entry workstation, Hermes, ships with one RTX PRO 6000 Blackwell or one RTX 5090, depending on the operator's budget and use case.

The mini PCs (Castor and Pollux) and the laptop (Daedalus) carry mobile NVIDIA silicon (RTX 4060, 4070, 5070, and 5080 mobile) tuned to fit a quiet chassis and the heat envelope each form factor allows.

In every case the SIOS, the cooperative, the OAR substrate, and the Poseidon SoC roadmap are the same. The GPU configuration is the operator's choice.

How the fabric and the SIOS work together

When an operator runs the AMT, the Trading Bot, or the cooperative of brains on a Mickai workstation, the SIOS routes work across the brains, the brains call into the model weights, and the model weights live in HBM on the GPU fabric. The Chronus orchestration kernel decides which brain takes which call. The Open Audit Record signs every action under FIPS 204 ML-DSA-65 before it commits. Identity and clearance gates are checked in the substrate before the brain runs.

What the operator sees is a single coherent system that thinks fast, signs everything it does, and never sends the work to a third party. What runs underneath is the Mickai SIOS on the operator's hardware, calling into a Blackwell fabric that NVIDIA designed for exactly this kind of load.

The Mickai SIOS is what makes the workstation sovereign. The Blackwell fabric is what makes it fast enough to be useful.

The Poseidon SoC, in-house silicon beside the GPU pool

Mickai is also building its own silicon. Poseidon, the Mickai Sovereign AI SoC, is the in-house chip that complements the NVIDIA GPU pool on every Mickai workstation. The role of Poseidon on the chassis:

• Carries the OAR signing primitive in silicon, so every action can be signed at line rate even on lower SKUs that do not carry an FPGA.
• Holds the hardware identity and the cryptographic root of trust for the operator.
• Sits between the network interface and the GPU fabric so every action that crosses the bus can be policy-checked before commit.

Poseidon is in the patent corpus already. It is an add-on on every SKU as the silicon lands, and it does not replace the NVIDIA fabric. It sits beside it, holding the audit and identity primitives that the SIOS needs to be sovereign.

What we say about NVIDIA on the chassis

Mickai uses NVIDIA silicon. We name it on the spec sheet. We do not put the NVIDIA marque on the chassis at launch, because that requires accreditation through the NVIDIA Partner Network, and we have not yet applied. When NPN OEM status lands, the chassis will carry the partner badge under the standard NVIDIA brand guidelines, in the appropriate location and at the appropriate scale.

Until then, the chassis carries the MICKAI(TM) wordmark, the SKU codename, and a "built in Britain" stamp. The fabric is named on the spec sheet by product (RTX PRO 6000 Blackwell, H200 NVL, Blackwell Ultra B300 NVL) so the operator knows exactly what is in the box.

This is the legally clean and editorially honest way to do it. The operator gets the fabric they need. NVIDIA gets proper accreditation. The Mickai brand stands on its own substrate.

What it means for an investor reading this

A serious investor evaluating a workstation company asks where the fabric comes from. The Mickai fabric comes from NVIDIA Blackwell, which is the best available silicon for the load the cooperative of brains demands. That is the same fabric a Bank of England trading desk, a national defence laboratory, or a national health-service research workstation will demand. We are building the right work on top of the right fabric.

The Mickai patent corpus, fifty-seven applications and approximately 1,535 claims on the UK IPO public register, covers everything above the fabric. The cooperative, the audit substrate, the signing primitive, the clearance gates, the host-acceptance attestation, the Poseidon SoC roadmap. The fabric is rented from NVIDIA by purchase. The work above the fabric is Mickai.

What it means for the operator

If you are an operator buying a Mickai workstation, the practical answer is shorter. You get the best NVIDIA silicon for your use case, you get the Mickai SIOS preinstalled, you get the cooperative of brains running locally, you get the Open Audit Record signing every action under a post-quantum primitive, and you get the price for life. No subscription for context. No subscription for usage. The fabric does the work the cooperative tells it to do. You own the machine.

That is the simple statement of the Mickai workstation, and that is why we build it on NVIDIA Blackwell.

Further reading

• The Mickai workstation lineup at mickai.co.uk/hardware
• The Mickai Workstation Configurator at mickai.co.uk/hardware/configurator
• The Poseidon Sovereign AI SoC at mickai.co.uk/poseidon
• The Mickai patent corpus at mickai.co.uk/patents
• NVIDIA Blackwell Ultra B300 product page at nvidia.com/en-us/data-center/gb300-nvl72/
• NVIDIA RTX PRO 6000 Blackwell Workstation Edition at nvidia.com/en-us/products/workstations/professional-desktop-gpus/rtx-pro-6000/

NVIDIA, Blackwell, RTX, and CUDA are trademarks of NVIDIA Corporation, used by Mickai for product reference and accreditation. Mickai is not a formal partner of NVIDIA at the time of publication, and this article makes no claim to a partnership. The Mickai workstation lineup uses NVIDIA silicon under standard product purchase.