Comma.ai has published a detailed account of the computing facility it operates inside its own office, arguing that organizations with steady machine-learning workloads can gain cost and engineering advantages by owning infrastructure instead of renting cloud capacity. The company estimates that it has spent about $5 million on the data center, compared with more than $25 million for equivalent cloud use.
Those figures are Comma.ai's own estimates, not an independent price comparison, and its setup reflects a specialized workload. The facility runs the company's model training, metrics and data storage and reaches about 450 kilowatts at maximum demand. Electricity is a substantial expense: the company says it spent $540,112 on power during 2025 in San Diego, where its rate exceeds 40 cents per kilowatt-hour.
Cooling relies primarily on outside air rather than a conventional computer-room air-conditioning system. Large intake and exhaust fans move air through the space, while recirculation fans mix warmer exhaust with incoming air to keep humidity below 45 percent. Sensors attached to one server feed a control loop that adjusts the fans for temperature and moisture. Comma.ai says this arrangement consumes only a few dozen kilowatts.
Most of the compute capacity consists of 600 graphics processors in 75 TinyBox Pro machines built in-house. Each system has two CPUs and eight GPUs and can serve either model training or general tasks. The company says its own machines fail at a rate similar to prebuilt alternatives, while internal construction makes repairs faster.
Storage occupies several racks of Dell R630 and R730 servers filled with solid-state drives, providing about four petabytes in total. The main three-petabyte array holds driving data without redundancy because the company says individual records are not critical. It reports aggregate reads near one terabyte per second, allowing training directly from raw data. A separate roughly 300-terabyte array caches processed results, and another redundant system retains trained models and metrics.
Three interconnected 100-gigabit Ethernet switches form the main network. Two InfiniBand switches connect groups of training machines for distributed operations. Ubuntu is installed over the network, Salt manages servers and Slurm schedules compute. Training uses PyTorch's distributed Fully Sharded Data Parallel system, while a lightweight open-source scheduler called Miniray sends other Python tasks to idle machines.
Comma.ai's case is not a general prescription. Owning hardware transfers responsibility for power, cooling, networking, repairs and capacity planning to the operator, and the company accepts reduced redundancy for some services to target 99 percent availability. Its experience instead provides a concrete reference point: when demand is sustained and technically predictable, a small specialized team may be able to trade cloud flexibility for lower claimed costs, direct control and infrastructure optimized around one workload.
