We pre-train physical-AI models for nations, inside their borders.
The nation keeps the data, the compute and the model. We do the pre-training and hand it off.
The nation keeps the data, the compute and the model. We do the pre-training and hand it off.
18 days
June 12 to June 30, 2026. A US export-control directive suspended every foreign national's access to a frontier AI model, worldwide. The switch is real, and someone else holds it.
If Washington or Beijing owns the models, then in a conflict or a tension they can shut them off. So nations are going to need to own their own.
184 sovereign-AI projects in 67 countries. Governments already know this. Physical AI, the models that will run cranes, grids, plants and vehicles, is where owning the model stops being abstract.
59%
of the 184 sovereign-AI projects are infrastructure. 9% are data. More compute projects were announced in Q1 2026 than in all of 2024.
3x
NVIDIA's sovereign-AI business more than tripled to over $30B in FY2026, then grew another 35% last quarter.
Open
Physical-AI base models are open: Cosmos Predict 2.5, Cosmos Reason 2 and GR00T N1.6 shipped on Hugging Face on Jan 5, 2026. Continuing one on a nation's own data is now a 90-day project.
The nation's port, grid and factory data, with a provenance record for every file. Faces and plates blurred.
Deduplicate, mine the events that matter, fix the mixture. Versioned, with a reason for every dropped record.
Continue an open physical-AI base model on that data, on the nation's compute, at a fixed and logged budget.
Held-out cameras and months, matched compute, pre-registered threshold. No threshold, no claim.
Checkpoints, model card, runbooks. The nation's lab re-runs it without us. That is the acceptance test.
Data never leaves the country. We do the pre-training, the half-art, half-science part. The nation does the post-training and truly owns the model.
| Scope | One environment, one task, one metric, agreed in writing before training starts. |
|---|---|
| Data | Stays in-country. Orion works inside the customer's permissioned boundary. |
| Compute | The nation's own, or rented in-region under the nation's account. |
| Deliverables | Dataset manifest, pre-trained checkpoints with model card, evaluation report, runbooks, handoff. |
| Rights | The nation owns the model. Orion keeps the method. |
| Then | A national program: more environments, more tasks, in-country training capability, refreshed every cycle. |
Pilot economics, expected casePrice $250,000. Delivery cost $103,900. Contribution $146,100, 58%. Fully loaded margin 40%.
Compute is the customer's, so margin never depends on renting GPUs.
What compounds: the evaluation standard nations judge vendors by, recipes that travel between nations when data cannot, and lineage that brings every refresh back to us.
The labels are already written. A terminal records what moved, when and where, in its operating system and its crane controllers. Align the camera archive with those logs and you get labels nobody has to draw.
Curation is where the compute saving is. Port video is overwhelmingly redundant: fixed cameras, idle hours. DatologyAI reports 7.7x faster training from curation on text. We expect the same lever on port video, and we will measure it.
Domain data goes in at pre-training. "We look for a lot of domain data to put into pre-training," Alexandr Wang, Meta, Oct 12, 2026. Nations have the domain data. We bring the pre-training.
$30B+
NVIDIA's sovereign-AI revenue in FY2026, more than tripled in a year.
3.3tn won
Korea's public funding for physical-AI ports by 2035, announced Jul 23, 2026. One country, one sector.
$138B
Goldman Sachs' 2035 humanoid market, raised from $38B in September 2026.
Bottom up. 67 countries fund sovereign AI; a handful host a frontier lab. One $250K pilot opens each door. A national program then covers every port, grid and plant the nation owns, and refreshes every cycle.
| Who | What they sell a nation | Why Orion wins |
|---|---|---|
| Frontier labs | OpenAI for Countries, Google, Anthropic sell in-country data centers and model access. The weights stay theirs, and so does the switch. | Ours is handed over. The lab can re-run everything without us. |
| NVIDIA, clouds, neoclouds | Compute and reference stacks. | Our substrate, not our rival. We train on the GPUs the nation already bought. |
| National labs | Research and post-training. | Our hand-off partner. They take over the day pre-training ends. |
| Systems integrators | Deployment and integration. | They don't pre-train. They ship what we hand off. |
| Data-curation startups | DatologyAI curates for labs, text first. | Not physical AI, not in-country, not for nations. |
We only do pre-training, so nobody the nation already pays loses a contract. That is why the door opens for a small team.
Spencer Brown, founder and CEO. Founder of LVL3.AI, AI agent infrastructure, Palo Alto. Harvard master's in finance, in progress. Runs Orion's lab with an always-on agent swarm: nightly builds, 57 tests passing, one pull request per run.
Hiring now. A technical lead for continued pre-training and the evaluation harness. A data and evaluation owner for permissions, splits and the gold set.
Nations with world-class industry and no frontier lab need a partner who will go in, pre-train, hand off and leave the capability behind. We are built to be that partner, and nothing else.
First conversations: Brazil, Argentina, the UAE, Saudi Arabia, Indonesia, Thailand, Cambodia and Central Europe.
What exists today, Oct 21, 2026. The pilot offer and contract terms. A lab scaffold that runs nightly: provenance checks, four-pass curation, matched-compute arms, bootstrap intervals, 57 tests. Thirty target accounts researched, with sources.
What the money buys. December 2026: a controlled result on public maritime data, with intervals. Q1 2027: the first paid pilot, accepted when the nation's lab re-runs the pipeline without us. Three paid pilots within 90 days of the first. Two hires.
The round is a SAFE. Size and terms on request.
No customers, no model results and no revenue are claimed in this deck. Every number carries its source and date.
| Export controls | Al Jazeera, "US asks Anthropic to block global access to top AI models," Jun 14, 2026. Euronews, "US lifts export controls on powerful AI models, Anthropic says," Jul 1, 2026. Access was suspended on Jun 12 and the lifting announced Jun 30, 2026. |
|---|---|
| Sovereign AI index | CNAS, "Sovereign AI's Second Wave Is Coming Into View," Aug 20, 2026: 184 projects in 67 countries through H1 2026; CNAS Sovereign AI Index, updated Aug 18, 2026: 59% infrastructure, 32% models, 9% data; more infrastructure projects announced in Q1 2026 than in all of 2024. |
| NVIDIA | Q4 FY2026 earnings call, Feb 25, 2026: sovereign-AI business "more than tripled year over year to over $30 billion"; physical AI contributed "north of $6 billion" of FY2026 revenue. Q2 FY2027 call, Aug 26, 2026: sovereign AI up 35% sequentially and more than tripled year over year. Press release, Jan 5, 2026: Cosmos Predict 2.5, Cosmos Transfer 2.5, Cosmos Reason 2 and Isaac GR00T N1.6 released as open models on Hugging Face. |
| Physical AI market | Bank of America Institute, "Physical AI, part 2: Humanoid robots," Mar 12, 2026: 20,000 shipments in 2025, 90,000 in 2026, 1.2 million in 2030, 10 million in 2035, an 86% compound annual growth rate. Goldman Sachs humanoid outlook via BigGo Finance, Sept 1, 2026: about 6.5 million units and $138 billion by 2035, up from 1.4 million units and $38 billion. |
| Korea | Asia Economy, "Ports to Become Giant AI Robots," Jul 23, 2026: Ministry of Oceans and Fisheries Physical AI Port Strategy, about 3.3 trillion won of public funds by 2035, Busan, Gwangyang and Jinhae New Port. |
| Pre-training and curation | Alexandr Wang, Cheeky Pint #35, Oct 12, 2026. DatologyAI, curation results on text, Aug 18, 2025: 7.7x faster training. |
| Orion planning figures | Orion AI Command Plan and Operating Workbook, Oct 21, 2026: pilot price, cost cases, milestones and the lab scaffold's test count. Planning figures, not results. |