Is AI profitable?

Four different economic questions

The question “Is AI profitable?” bundles together several different businesses.

For the organisation using AI, the question is whether accepted output, genuinely avoided expenditure or additional profitable business outweighs implementation and operating costs.

For the company serving a model, the question is whether customer revenue covers delivery: hardware depreciation, power, networking, licensing and operations at realistic utilisation. Positive margins at this layer do not automatically cover the cost of developing the model.

For the frontier-model developer, training, research, staff and the rest of the business also need funding. Those costs are real. Describing inference as profitable does not make them disappear.

For the investor, even a useful technology and a profitable business can be a poor purchase at the wrong price.

There is no contradiction in an adopter saving money while its supplier loses money overall. There is also no guarantee that today's favourable customer economics survive future price changes. Both statements belong in the analysis.

But one company's losses do not tell you whether another company's workflow is worthwhile. An enterprise running an available open-weight model does not have to repeat the original research programme; its own deployment still needs to cover licensing where applicable, hardware or hosting, integration and operation.

Economic value, supplier profitability and investment returns are connected. They are not interchangeable.

“A human is still needed” does not mean “the same number of humans is still needed.”

A job is a bundle of tasks. Replacing every responsibility attached to a job title is not a prerequisite for changing how many people are needed to deliver the work.

The International Labour Organization's 2025 occupational-exposure research explicitly examines tasks and distinguishes job transformation from complete automation. Its finding that human involvement remains necessary is not a finding that staffing levels must remain unchanged.³

Consider a deliberately simplified example. A ten-person team performs 1,600 hours of work each month. Automation removes 800 hours, while checking results, managing exceptions and correcting mistakes adds back 160. The remaining workload is 960 hours: six people's worth rather than ten.

That is an illustration, not a forecast. It assumes the remaining work can be redistributed, the necessary skills are available and output requirements stay constant. Under those conditions, four positions could disappear without the occupation disappearing. Alternatively, the business could retain the team and produce more.

This also explains why “AI only assists the worker” is not an economic rebuttal. Assistance can reduce labour requirements, avoid future hiring, increase capacity or improve the service delivered.

None of that establishes that half of all jobs will disappear. Nor does half the task count necessarily mean half the working time. It establishes something more useful: task-level automation can have large economic effects long before whole-job replacement becomes possible.

The evidence already goes beyond impressive demonstrations

A study published in The Quarterly Journal of Economics examined 5,172 customer-support agents and found a 15% average increase in issues resolved per hour with AI assistance. Effects varied across workers. This is evidence of workplace productivity, not an audited calculation of the employer's net AI profit - but it measures delivered work rather than enthusiasm.⁴

There are negative results, too. METR's early-2025 experiment found that experienced open-source developers took 19% longer on the studied tasks when allowed to use AI. That result matters: review, correction and workflow friction can outweigh the assistance.⁵

The update matters as well. In February 2026, METR reported that its subsequent experiment faced substantial selection effects, including developers declining work without AI, and difficulties measuring concurrent agent use. The researchers believed speedups had probably improved but explicitly said their data could not reliably establish the magnitude. That does not invalidate the earlier finding or prove a universal new speedup.⁶

There is also evidence of agents producing operational improvements. Google DeepMind reported that a scheduling heuristic discovered by AlphaEvolve, and deployed in production, recovered 0.7% of Google's worldwide compute resources on average. This is a vendor-reported result, not independently audited profitability. Nevertheless, the claimed output is additional usable infrastructure capacity - not a better-sounding paragraph.⁷

The conclusion is neither “AI always saves money” nor “AI never saves money.” Useful work is measurable. So are the costs that can erase its value.

The unit of value is completed work, not a chatbot subscription

The economic question is not merely:

How many people will pay for a chatbot subscription?

It is:

How much useful work can these systems deliver, and what is that work worth after accounting for compute, integration, supervision and failures?

A practical calculation is:

Net AI value = realised benefits − model and infrastructure costs − integration and maintenance − review and rework − expected failure costs.

Benefits must be counted without double-counting. An hour cannot simultaneously be claimed as eliminated payroll and as an hour an existing employee uses to generate additional revenue. Saved time is not automatically cash saved. Additional revenue is not all profit. Severe risks also need acceptance limits, not just an optimistic average cost.

Suppose, as an illustration, an AI-assisted process genuinely avoids €2,000 of external spending each month. Its model, review and maintenance costs total €700, and implementation costs €3,900. It produces €1,300 of monthly net savings and recovers the initial cost in three months, assuming the savings persist and no material costs have been omitted.

The supplier's company-wide accounts cannot answer that customer-level calculation. Equally, the customer's savings cannot prove the supplier's business is sustainable.

This is why AI should be compared with the cost of the work it delivers - not confined to the price someone will pay for a conversational interface.

Agents make the potential market larger than human prompting time

A tool-using agent can select an action, invoke a tool, examine the result and choose its next step. In software, a compiler or test runner supplies feedback from outside the language model. That does not guarantee correctness, but it is materially different from accepting an unverified answer from a chat window.⁸

Now compare two hypothetical instructions:

“Help me fix this bug.”

“Within this budget, investigate incoming failures, propose fixes, run the required checks and send me the changes that need approval.”

The second can generate work without requiring a person to initiate every investigation. One bounded objective becomes a queue of tasks. The owner still sets priorities and authority; the system performs more of the intermediate execution.

Anthropic's June 2025 engineering report provides a concrete indication of the demand difference: in its data, agents typically used about four times as many tokens as chat interactions, and multi-agent systems about fifteen times as many. These are deployment-specific observations, not universal multipliers or direct measures of GPU consumption. Anthropic also emphasised that the task's value must justify the extra cost.⁹

The point is not that burning more tokens creates value. It is that counting people and subscriptions can miss how much computation a useful objective generates.

An organisation does not necessarily need more employees - or more people staring at chat windows - to create more demand for machine-executed work.

The agent-to-agent economy is an extension of this mechanism

Imagine an operations agent identifying a problem, asking a specialist system to investigate, commissioning a simulation and submitting a verified recommendation for approval. With suitable permissions and budgets, some services could be purchased rather than manually integrated and ordered each time.

This is a scenario, not a claim that fully autonomous commerce is already mature. But its technical ingredients exist. The A2A project released version 1.0 of its agent-communication protocol in March 2026. Google Cloud's April 2026 platform announcement also describes Agent Payments Protocol integration and payment mandates.¹⁰¹¹

Interoperability and payment infrastructure do not establish the size of the resulting economy. They make a particular kind of economy easier to build: one in which software can delegate work to other software across organisational boundaries.

The important expansion is not merely doing today's work more cheaply. It is doing work that was previously too expensive to commission.

Consider a small business that cannot justify a bespoke internal application, daily operational analysis or extensive product experimentation. If reliable delivery becomes sufficiently inexpensive, some of those activities become viable. Those are illustrative possibilities - not additional spending that can be assumed before customers obtain results.

There is a stopping rule: the next unit of computation must be worth more than it costs. Agents repeatedly buying services from one another do not create net value merely by generating transactions. But successful applications can generate income that funds further computation; today's software budgets are not necessarily a permanent ceiling.

Cheaper AI can expand demand rather than exhaust it

A common objection to AI infrastructure investment is that models will become cheaper to run. That is a reason to examine the demand response, not assume it disappears.

Stanford's 2025 AI Index documented a more than 280-fold decline in query prices for models reaching approximately GPT-3.5-level performance on MMLU between November 2022 and October 2024. This was a comparison of API prices at a benchmark threshold - not audited production costs or equivalence across every real-world capability.¹²

The economic mechanism is straightforward. A task that is not worth doing at €100 might be worth doing at €5. Lower prices can attract additional users, increase the frequency of a workflow and make new categories of work viable.

That does not make demand literally infinite. For comparable tasks:

Total compute consumption = completed task volume × compute required per completed task.

If task volume grows twentyfold while compute per task falls fivefold, consumption grows fourfold. Reverse those changes and consumption falls. Efficiency alone does not determine the outcome.

Nor must efficiency savings be spent on doing precisely the same task. They could fund broader searches, more candidate solutions, additional tests or work on harder problems. Whether those are worthwhile has to be measured.

The strongest demand argument is not “models will remain inefficient.” It is “greater efficiency and capability can make more useful work economical.”

Powerful local models could increase demand for frontier cloud models

Local and cloud AI need not be competing end states. A local model can perform routine work while a remote system handles selected difficult parts.

Research presented at ICML 2025 demonstrated one version of that arrangement. On its evaluated long-document tasks, MINIONS reduced cloud inference costs by 5.7 times on average while retaining 97.9% of remote-only benchmark performance. Those figures describe that experiment's cost and performance trade-off - not total ownership costs or a universal promise for every application.¹³

Now consider the wider market:

Suppose local agents become 100 times more numerous, while each requires one-tenth as much cloud assistance. With comparable cloud tasks, total cloud demand still increases tenfold.

This is arithmetic, not a forecast. It shows why “each user needs less cloud” and “the market needs more cloud” can both be true.

A cheap, private, always-available local agent could become a way to initiate work and route difficult parts to more capable remote systems. It could replace individual cloud calls while expanding the number of people and processes using cloud intelligence.

And the frontier does not have to stand still. Even suppose a device in 2030 can perform relevant tasks at the level of a 2025 frontier model. That does not establish parity with the strongest cloud system available in 2030.

Cloud capability can also come from spending more computation on search, verification and revision - not just using larger model weights. DeepMind's February 2026 Aletheia research describes such a loop and improvements from inference-time scaling on its mathematical evaluations.¹⁴

Our expectation is that the strongest cloud systems to retain an advantage on demanding work. The size and commercial value of that advantage are uncertain. A local model can be excellent at ordinary tasks without eliminating the market for more capable systems.

Compute can become a constraint on economic production

If a company can profitably execute additional work with agents, insufficient affordable compute can become a production bottleneck rather than merely an inconvenient subscription limit.

There is evidence of physical demand expanding despite efficiency improvements. The International Energy Agency reported that electricity consumption by AI-focused data centres increased 50% in 2025. Its April 2026 central projection has that consumption tripling between 2025 and 2030, while identifying constraints across electricity infrastructure and other supply chains. Electricity consumption is not the same as paid, useful compute, and a projection is not a guarantee.¹⁵

The implication for Europe is not “buy any GPU at any price.” It is that access to suitable compute deserves the same operational scrutiny as other important production inputs.

A generic availability figure is not enough. A customer may need a particular region, interconnect, delivery date, capacity commitment, data controls and price. Capacity that fails those requirements is not automatically a substitute.

This is also where the investment argument needs discipline. A growing market can still contain overbuilt locations, poorly matched hardware, weak contracts and operators whose financing costs exceed their returns. Useful infrastructure is not necessarily a good investment at every valuation.

Compute could become strategically essential while some compute investments fail. There is no contradiction.

So, is AI profitable - and what should we measure next?

AI can pay its way without eliminating entire jobs, without every provider becoming profitable simultaneously, and without a person manually requesting every piece of work.

The evidence supports measurable benefits in particular workflows, meaningful failure cases and architectures that go beyond one-shot conversation. The larger economic thesis is an inference from those building blocks - not a claim that future demand or returns have already been proved.

Here is the editorial proposition:

If increasingly capable agents continue finding valuable uses for additional computation, efficiency gains can expand demand faster than they relieve capacity constraints. Powerful on-device models may accelerate that process rather than end it.

That proposition can be tested. Do customers renew after experimentation? Does the all-in cost of accepted work fall? Can organisations convert the benefit into avoided spending, better service or profitable output? Does additional computation improve results enough to justify its cost?

Those are stronger questions than whether a chatbot can replace a whole person.

A demonstration still needs a business case. Take a representative workflow. Give a capable system the context, tools and limited authority it needs. Measure the complete result - including review, mistakes and operating costs - against the current alternative.

The market is not simply people paying to chat. It is the work that becomes worth doing when useful intelligence becomes cheaper to deploy.

Sources