Ornn Helps Simular Scale Capacity as Fast as Its Research Moves
Overview
Simular, Inc. is building the autonomous computer: AI agents that operate desktop applications by seeing the screen, moving the cursor, and typing, the way a person does. Founded by former DeepMind researchers and headquartered in Palo Alto, Simular develops both an open agentic framework, Agent S, and consumer products built on top of it.
In December 2025, Agent S scored 72.6% on OSWorld, the leading benchmark for multimodal agents performing real computer tasks, edging past the benchmark's human baseline of 72.36%. Behind that work, Simular runs GPU-accelerated training and evaluation on dedicated NVIDIA H100 nodes sourced through Ornn Compute.
Agent Infrastructure
Dedicated capacity for training
Challenge
Simular's compute needs grow with its research, not on a schedule anyone could commit to a year in advance. A training run and an evaluation sweep are followed by a gap, and the size of the next requirement is not known until the last one finishes.
At one end sit multi-year contracts, written for buyers who know their demand years ahead. At the other, spot capacity priced by the hour, reclaimable in the middle of a run. A research team buying nodes rather than clusters falls between them, too small for the first and too exposed on the second. What Simular needed was dedicated capacity on terms short enough to match a research cycle.
“Ornn gave us the flexibility to train ambitiously without locking up our capital, at half the price of a hyperscaler.”
Ang Li, CEO
Solution
Ornn Compute aggregates dedicated capacity from operators across regions, so a buyer specifies what the work requires instead of choosing from what one operator happens to hold.
Simular specified region, hardware, and term. Ornn returned capacity matching all three: H100 80GB SXM in Singapore on a month-to-month term, signed and running that week. Ornn had already done the operator diligence, so Simular contracted with one counterparty, on terms it recognized, for infrastructure on the other side of the world.
That structure is what made the decision easy. When the work went well and Simular wanted more nodes, the capacity came from the same platform under the same agreement. Buying a second block from a second operator would have meant a new contract, a new security review, a new credential system, and a new counterparty to manage. Instead, Simular was able to double their footprint inside a relationship it already had.
Results
Roughly half the cost of hyperscalers. Dedicated capacity through Ornn Compute prices at approximately 48% of comparable hyperscale on-demand.
Capacity doubled without added procurement. As the work went smoothly, Simular doubled their footprint with no new onboarding and no second credential system to manage.
Terms that match the research cycle. Simular reserved capacity on a short term and extended it as the work justified, rather than committing to a year up front.