Yondu Cuts a Five-Day Training Run to 21 Hours
Overview
Yondu AI builds a robot workforce for third-party logistics warehouses. Yondu supplies the model and the software layer that drive off-the-shelf hardware through bin picking, shelf stocking, inventory counts, and order fulfillment.
The whole company is organized around one constraint, brownfield deployment. Most warehouse automation is greenfield, where the facility goes up around the robots with fixed racking and known geometry. Yondu drops into the aisles, lighting, and shelving a 3PL already has, against whatever SKUs come through the door, without a retrofit or a pause in operations.
Getting a robot to work under those conditions takes a model general enough to handle a warehouse it has never seen. Human operators run the robots by teleoperation through Yondu's control platform, every demonstration becomes training data, and Yondu's autonomy stack takes over more of the work as the policy improves. The operator stays in for the cases the model cannot handle yet and hands back control as the model learns them.
Robotics
Model Training
Challenge
Yondu had a customer demo the following week and a model to train before it. The team needed H100 or H200 capacity within a day, and its existing providers could not deliver on that timeline.
Yondu’s requirements were narrow: five to seven days of dedicated bare metal for a single training run, with no commitment past the job. That request falls between the traditional ways compute is usually sold.
Reserved: starts at a month, often runs to a year, arrives at the end of a procurement cycle. Yondu had less than a day.
On-demand: clears faster, but multi-node H100 blocks are rarely sitting unreserved.
Spot: available, but preemptible and reclaimable mid-run.
What remains is calling providers one at a time to ask who has free capacity available, in the right region, on the right hardware. Every one of those calls is also a new vendor relationship to open.
"It's insane that a dataset of this size will only take 21 hours to train 5 epochs, which would otherwise have taken 5 to 6 days on a single H100." — Kiran Kommaraju, Founding Robotics Engineer, Yondu
Solution
Ornn matched Yondu's request to available H100 capacity in US-Central and provisioned it the same evening without a procurement cycle or a monthly minimum to clear first.
Ornn is an omni-channel neocloud, holding relationships across operators and facilities rather than running a single site. The search covered which operators had a matching block free that night, in US-Central, on H100s. Every one of those operators was already under contract, so nothing had to be negotiated before the capacity moved. Ornn holds those relationships in advance so the search runs against inventory rather than against a sales cycle.
Yondu onboarded once and drew on capacity across the operator base from Ornn. It saw the site, the hardware configuration, and the terms attached to the cluster it was assigned, which matters for a training run where interconnect and node topology affect how the job is written.
The term came with an exit. Had the training schedule run shorter, Yondu could have sold the remaining days back into quality offtake on the Ornn secondary market. A buyer signing a dedicated term normally pays for the whole term whether the job uses it or not. Dedicated capacity on Ornn can be transferred or sublet, so unused days can be sold to another buyer.
Results
A term that matched the job. Five days of dedicated bare metal, no auto-renewal, no monthly minimum. Yondu paid for the run and nothing around it. The job finished inside the window and the capacity went back on schedule.
Dedicated capacity for the full run. The capacity was dedicated for all five days, so the run carried none of the preemption risk that comes with spot. Yondu knew on the first day when the job would end.
One relationship for the next request. Yondu onboarded once, and that onboarding covers the operators on the Ornn platform. The next short-notice request starts from available inventory.