Investor materials
Competitive landscape
Where Cerebel sits among robot foundation models, enterprise automation vendors, and systems integrators — and why only one of them competes with us.
ACompetitive landscape
Four different businesses — only one competes with us
Layer 1 · Model
Robot Foundation Models
The model is the product. Sold to robot makers and enterprises.
Skild AI
~$2B raised · $14B valuation
“Omni-bodied brain” via cloud API. Any robot, any task. Hardware-agnostic by design.
Physical Intelligence
$5.6B valuation · pre-revenue
π-series models to control any robot body. Deliberately sells nothing yet.
RLWRLD
~$41M seed · Seoul
Industrial foundation model trained inside partner sites — CJ Logistics, Lotte. Korea/Japan conglomerate coalition.
Relationship: supplier, not competitor. We already build on GR00T and π baselines. A better general model makes our deployment better, not obsolete — because the model was never the thing we sell.
Layer 2 · System
Enterprise Warehouse Automation
A system sold to large operators. Capital-intensive, long cycles.
Nimble
$221M raised · $1B valuation
Vertically integrated robotic 3PL — FedEx-backed. Runs its own fulfilment centres rather than deploying into yours.
Dexterity · Covariant
Hundreds of $M each
High-throughput picking and truck loading for enterprise logistics. Bespoke integration per site.
RightHand · Osaro · Plus One
Piece-picking incumbents
Cell-level picking sold as capex through integrators. Enterprise buyers, engineered installs.
Relationship: adjacent, different buyer. All of them sell to operators who can absorb a six-to-seven-figure capex and an integration project. At a $60–70K robot cost, mid-market simply doesn't clear — which is why it stays unserved.
Layer 3 · Labour
Systems Integrators
Engineering hours sold against bespoke requirements.
Thousands of regional firms. The unit of learning is a site — this machine, this fixture, this part — and it never recurs. Job 50 starts where job 1 did, because the knowledge lives in an engineer's head rather than in an asset.
Relationship: the failure mode we're designed to avoid. This is why the market looks fragmented, and why fragmentation is a property of the business model rather than of the demand.
Layer 4 · Task
Cerebel
We sell the task, not the robot, not the model, not the hours.
One task — kitting — deployed as a service on hardware we build ourselves at a $20K BOM, into mid-market 3PLs and fulfilment centres. We retain the asset, the model, and the data by construction. The unit of learning is the task, which recurs in every building.
BWhere the lines actually fall
Same words, different businesses
| Dimension | Foundation models | Enterprise systems | Integrators | Cerebel |
|---|---|---|---|---|
| What is sold | A model, via API | A system, as capex | Engineering hours | A completed task, monthly |
| Who buys | Robot makers, enterprises | Large operators | Anyone with a bespoke need | Mid-market 3PLs & fulfilment |
| Hardware | None — agnostic | In-house or partner | Third-party | In-house, $20K BOM |
| Data source | Web, sim, partner sites | Own deployments | None retained | Capture instrument + deployments |
| Contact data | Inferred or absent | Largely absent | None | Measured — joint angle & force |
| Unit of learning | Everything, at once | A deployment | A site — never recurs | A task — recurs everywhere |
| Capital to prove it | Billions | Hundreds of millions | n/a | $1M for the first pilot |
CThe honest version
Where they are ahead of us
What we concede
- Capital and talent. Skild and Physical Intelligence outspend us by three orders of magnitude and employ many of the best robot-learning researchers alive. We will not win a general-model race and are not entering one.
- Real deployments. RLWRLD has MOUs and joint deployments with CJ Logistics and Lotte. Nimble has FedEx and a live facility. We have engaged conversations and no signed pilot yet.
- Corporate access. RLWRLD's strategic investors hand it factory floors. That is a genuine advantage in data collection, and it is the axis where we most need a partner.
What does not transfer to them
- The cost basis. None of them can serve a mid-market 3PL at $5K/month. Not because they lack the model — because the arithmetic on a $60–70K robot never clears.
- The instrument. Site access gives you vision and telemetry. It does not give you absolute finger-joint angle under occlusion, or contact force. Those require hardware built to measure them.
- The ownership. Under RaaS the asset, the model and the data stay ours. A vendor who sells hardware ships the data-generating asset to the customer and receives nothing back.
The well-funded labs are building the brain. We are building the business that puts one to work in buildings nobody else can afford to serve.
If a general model arrives and is excellent, we adopt it and our deployments get better overnight. The thing we are accumulating — tactile-paired demonstration data from real floors, and a cost basis that reaches the mid-market — is not on anyone else's roadmap, because it is not what their business needs.
Funding figures are latest disclosed as of August 2026; positions summarised from public statements.