Artificial intelligence companies are usually compared by model size, benchmark scores, response speed and the number of tokens their systems can process.

Perplexity CEO Aravind Srinivas believes the longer-term competition may be determined by a broader efficiency metric: token value per watt per user.

In a June 2026 CNBC interview, Srinivas argued that the eventual winner of the AI race will be the company that produces the greatest economic value from the electricity its systems consume while balancing accuracy, latency, cost, privacy and intelligence.

The idea represents the next stage in a much older history of measuring computational efficiency. Traditional processors were compared through operations per second, while supercomputers were ranked by floating-point operations per second, or FLOPS.

As power consumption became a major engineering constraint, the Green500 project began ranking supercomputers according to FLOPS per watt.

Its first rankings were published in 2007, when an IBM Blue Gene/P system achieved 357 megaflops per watt.

By 2015, the leading system had reached more than 7,000 megaflops per watt.

AI infrastructure requires a more complicated measurement because completing more calculations does not necessarily produce a better answer.

A system may generate thousands of tokens, use extensive reasoning steps and consume substantial energy while still returning an inaccurate or unhelpful response. Token throughput per watt therefore measures output volume, but it does not measure the usefulness, accuracy or economic importance of that output.

Srinivas’s proposed metric attempts to place value above raw production.

A short AI response that correctly diagnoses a machine failure, identifies a security vulnerability or completes a difficult engineering task could be more valuable than a much longer answer generated with greater computational effort.

The best AI system would therefore not be the one that merely generates the most tokens. It would generate the most useful and reliable work from each unit of electricity available to it.

This also means that “more compute per token” is not automatically evidence that AI is improving.

Additional inference compute can improve reasoning on difficult problems, but only when the resulting increase in answer quality exceeds the added energy, latency and financial cost.

An AI model that uses twice as much compute for a one-percent improvement may be less efficient than a smaller model that completes the same practical task with slightly lower accuracy.

MLCommons has already documented this tradeoff, including benchmark results in which systems sacrificed as much as 50 percent of their energy efficiency to move from 99 percent to 99.9 percent inference accuracy.

The AI industry is beginning to formalize these measurements. MLPerf Power evaluates machine-learning performance and power consumption across systems ranging from tiny embedded devices to megawatt-scale data centers.

Its methodology considers not only processors and accelerators but also memory, storage, networking, interconnects and cooling because the complete system contributes to the energy cost of an AI workload.

Researchers have also proposed an academic metric called Intelligence per Watt, or IPW. A 2025 study defined IPW as task accuracy divided by power consumption and evaluated more than 20 local language models, eight hardware accelerators and one million real-world queries.

The researchers reported that intelligence-per-watt performance improved approximately 5.3 times between 2023 and 2025, showing that advances in smaller models and local hardware are making useful AI possible within increasingly limited power envelopes.

There is one important engineering limitation in the terminology. A watt measures the instantaneous rate at which energy is being consumed, while a joule or watt-hour measures the total energy required to complete a task.

For individual AI requests, metrics such as joules per query, joules per useful answer or task value per watt-hour may ultimately be more precise than value per watt. Srinivas’s phrase should therefore be understood as a strategic framework rather than an established scientific unit.

A practical version of the metric might be expressed as:

AI Efficiency = Useful Task Value ÷ Total Energy Consumed

Useful task value could incorporate answer accuracy, completion rate, user satisfaction, economic output and time saved.

Total energy would include model inference, memory, networking, storage and the supporting data-center infrastructure.

The measurement would also need to specify the workload because coding, image generation, search, translation and scientific reasoning have very different computational requirements.

The metric becomes especially important as AI’s electricity requirements expand. The International Energy Agency estimated that data centers consumed approximately 415 terawatt-hours of electricity in 2024 and projected consumption of roughly 945 terawatt-hours by 2030.

Accelerated servers used primarily for AI are expected to account for almost half of the net increase. Under these conditions, improvements in useful output per unit of energy can directly affect operating costs, grid capacity and the pace at which new AI infrastructure can be deployed.

Token value per watt also supports technologies such as model routing. A simple request does not always require the largest available model.

An intelligent routing system could send basic queries to small local models, difficult reasoning tasks to frontier systems and private workloads to hardware controlled by the user. The result would be a hybrid AI environment designed to allocate the right amount of intelligence, hardware and energy to each individual task.

Ultimately, Srinivas’s metric reframes the AI race. The most important question may no longer be which company owns the largest model or the biggest data center.

It may be which system can produce the most accurate, useful and economically meaningful result for every unit of electricity it consumes.

In that framework, genuine AI progress is not simply more tokens, more parameters or more compute. It is more useful intelligence with less computational waste.