AI infrastructure strategy needs a clear account of what capital buys. Spending tells us what has been committed. Installed capacity describes the resources available. Neither, on its own, establishes how much useful work a system delivers or whether customers will pay enough for that work.
A preliminary research framework by Anton Korinek and Patrick McKelvey makes this measurement problem explicit. It is valuable for infrastructure leaders because it challenges a common shortcut: treating growth in dollars, compute, capability, and business value as interchangeable. The investment task is to connect those measures without collapsing them into one headline.
What the research estimates
In Measuring the AI Economy, Korinek and McKelvey estimate U.S. AI production using power and hardware data, rental prices, and adjustments for changing model capability. Their quality adjustment uses API prices at comparable benchmark performance for inference and estimates of algorithmic progress for training. These are constructed estimates, including the spending series, rather than a census of company accounts.
The accompanying policy brief reports preliminary estimates of roughly $250 billion in nominal AI GDP for 2025 and around 2,600 percent annual growth in quality-adjusted AI GDP. That experimental sector measure is distinct from the authors' AI-production index, official U.S. GDP, and measured productivity in businesses adopting AI. It is also distinct from revenue or cash flow available to infrastructure owners.
The working paper identifies substantial uncertainty: limited data on the training–inference split, incomplete visibility into margins, and an uncertain relationship between benchmark gains and economic value. The size of the quality adjustment matters greatly. The paper also discloses Korinek's Anthropic Institute affiliation and states that its conclusions are the authors' own.
The useful insight is methodological. A fixed spending total can purchase a changing level of capability. It does not follow that an enterprise obtains the same increase in productivity, that demand expands at the estimated rate, or that a particular fleet earns an attractive return.
Keep three accounts in the investment case
I would carry three connected accounts into a capacity review. Each answers a different management question.
| Account | What to record | Decision it supports |
|---|---|---|
| Capital and operating inputs | Equipment, facilities, energy, network, software, people, financing, and committed capacity | What must be funded, when it becomes available, and what obligations remain |
| Delivered service | Completed training runs or inference tasks meeting defined quality, latency, and reliability requirements | Whether the system converts those inputs into usable capacity |
| Economic outcomes | Revenue, avoided cost, time saved, or another outcome measured against a stated baseline | Whether using that capacity creates enough value to justify the commitment |
A change in the first account is not automatically a change in the other two. A cheaper model may still require integration work. A faster service may create little value if users do not adopt it. A busy accelerator may be processing retries, redundant context, or work that fails acceptance checks.
This is why I would not use an aggregate AI-growth multiple as the demand forecast in an investment memorandum. It belongs in the scenario discussion. The base case needs evidence about the workloads, customers, and operating conditions the investment will actually serve.
Measure the service the customer needs
For inference, define an accepted task before optimizing cost per task. Specify the quality test, latency requirement, permitted failure rate, and treatment of retries, human review, and tool execution. A lower token price is useful, but it is only one input to that calculation.
For training and model development, define the experiment or evaluation milestone, the time allowed to reach it, and the capacity consumed by failed or repeated runs. Peak throughput alone cannot show whether the research program is moving faster.
This creates a practical distinction between cheaper computation and cheaper completion. If one configuration generates fewer tokens but needs more human correction, its service economics may deteriorate. If another costs more per accelerator-hour but completes the accepted workload sooner, it may release capacity and reduce total cost. Those are hypotheses to test with the same workload and acceptance criteria.
Match the hardware decision to the workload
There is no universal hardware winner implied by the macroeconomic estimates. Compare candidate configurations against demand distributions: model sizes, context lengths, concurrency, batch sizes, latency targets, and the software required to operate them. Include migration effort, network and memory constraints, resilience, and the cost of maintaining another platform.
Industry transactions should not substitute for that evidence. Groq's December 2025 announcement described a non-exclusive inference-technology licensing agreement with NVIDIA, with some Groq personnel joining NVIDIA and Groq continuing as an independent company. It did not establish a general cost advantage for a buyer's workload. Any such advantage still needs a comparable benchmark and a full operating-cost calculation.
The capital decision should therefore include a measured baseline, a credible alternative, and the conditions under which the preferred option changes. Flexibility has value when models and demand are evolving; specialization earns its place when the benefit survives those changes and covers its additional operating burden.
Make uncertainty part of capital governance
A useful investment review should expose the assumptions that can change the decision:
- Demand: Which workloads are contracted, observed, or speculative? What changes if adoption is slower?
- Service quality: Does the cheaper configuration meet the same acceptance and latency requirements?
- Delivery: When will power, cooling, network, equipment, and software be ready together?
- Operating performance: How do queueing, failures, recovery, and support affect useful output?
- Economics: What happens to the case if service prices fall, utilization disappoints, or migration costs rise?
Assign an owner and a review trigger to each material assumption. Stage commitments where possible, so evidence from an initial deployment can change the next capacity decision.
Control of scarce power or differentiated software may create an advantage. Neither guarantees durable returns. The stronger test is whether the organization can repeatedly convert its resources into a service customers value, at an acceptable cost and risk.
AI compute economics becomes useful to an executive when it supports that test. The objective is a traceable connection between capital committed, capacity delivered, and useful work completed.
Originally published 20 May 2026. Revised 13 September 2026 to distinguish preliminary research estimates from measured outcomes, correct the NVIDIA–Groq characterization, and remove unsupported hardware and return claims.