Tensor Machines has launched an open-source AI hardware benchmark aimed at helping data center operators determine the true economic cost of AI compute. Announced on October 7, 2026 in San Francisco, the benchmark moves beyond GPU specifications and hourly rental prices to connect GPU performance and power consumption directly to the cost of useful AI output. By focusing on output that meets a workload's service requirements, the new benchmark provides a repeatable framework to evaluate the productive capacity of existing AI infrastructure.
GPU specifications and hourly rental prices do not tell an operator how much work a system can sustain. Facility measures such as Power Usage Effectiveness (PUE), describe energy overhead but do not measure the AI output produced by the equipment. Tensor Machines is developing the benchmark around output that meets a workload's service requirements, giving operators a way to evaluate the productive capacity of their infrastructure and understand the energy and cost required to produce useful AI output.
While running their benchmark on several classes of NVIDIA GPUs, Tensor Machines observed differences that a hardware specification or hourly price alone would not reveal. The same GPU model can deliver different amounts of AI output, with the fastest result delivering nearly 15% more tokens per second than the slowest on one inference workload, implying nearly 15% higher cost per reported token for the slower result at equal hourly GPU prices.
More power does not buy proportionately more performance. In a compute comparison, a GPU operating at a higher power cap drew approximately 36% more power while delivering 21% more FP16 compute throughput than its peers. A performance gap can also change with the workload. A group that performed roughly 12% below its peers on one inference workload showed only about a 1% gap on another, proving that a single ranking can miss where hardware is most useful.
"The economic value of a GPU comes from the useful work it delivers," said Muneeb Rasool, founder and CEO of Tensor Machines. "Operators need to know what that output costs and how the answer changes with the workload, power settings and hardware condition. We are opening up the benchmark so they can measure those tradeoffs and make better use of the infrastructure they already have."
Tensor's benchmark recipe applies different loads and records how hardware draws power, heats up, sustains performance and recovers. The benchmark is designed to connect that physical response to the share of requests meeting service requirements, the energy consumed per accepted output token and the cost of that output. Repeating the same workload after a controlled change would let an operator test whether a different power setting, workload placement or cooling condition improves the result.
Tensor Machines' proprietary physics-based models support this work by analyzing the relationships between power, heat, performance and hardware degradation. The models are in private beta with select bare-metal providers and neo-cloud design partners. The broader goal is to help operators sustain productive output and make informed decisions about hardware use and remaining productive life.
"GPUs have the potential to transform computing at the edge, just as they are transforming data centers, but managing GPU health remains a crucial bottleneck," said Dr. Sandip Roy, professor of electrical and computer engineering at Texas A&M University. "The Texas A&M Global Cyber Research Institute (GCRI) has had the good fortune to collaborate with Tensor Machines, to assess the use of their remarkable GPU health telemetry/analytics for edge applications."
Tensor Machines also announced $1.5 million in pre-seed funding, led by Omni VC. Other investors include Reinforced Ventures, Avesta Fund and Draper U Ventures.
Rasool will present the open-source benchmark on October 8 at 11 a.m. Pacific at the NeoCloud Summit at Hotel Kabuki in San Francisco, alongside the president of Infrastructure Masons (iMasons). Tensor Machines invites operators, researchers and developers to use the benchmark and contribute to its development. Benchmark and documentation is available at https://github.com/tensormachines/TensorBench.
The launch of TensorBench marks a shift toward outcome-based evaluation of AI infrastructure, focusing on compute yield and cost per useful token rather than theoretical peak performance. As AI workloads scale across data centers and edge environments, the ability to measure real-world GPU health, energy efficiency, and workload-specific performance will be critical for optimizing AI infrastructure ROI and extending the productive life of GPUs and servers.
About Tensor Machines
Tensor Machines develops physics-based models and benchmarking tools to help AI infrastructure operators understand how power, heat and hardware condition affect useful compute output. Its work focuses on compute yield, energy efficiency and the productive life of GPUs and servers.
