The AI Power Race: Who Wins, Who Pays?

You ask an AI assistant to help with a letter, a spreadsheet or a piece of work. The answer arrives in seconds. The payment may arrive monthly. Behind that small exchange sits a much larger bargain: investors financing computing infrastructure, workers being asked to adapt, and communities supplying land, electricity and sometimes scarce water.

The AI power race connects those worlds. Its participants promise systems that could accelerate science, make expertise more accessible and eventually achieve artificial general intelligence, or AGI. They are also competing to own the services through which those benefits would reach us. A useful technology and a struggle for power can be the same story.

This investigation follows the chain from corporate ambition to the household bill. It separates observed results from forecasts, pricing from value, and safeguards against dangerous AI from control exercised through AI.

Evidence checked: 7 September 2026. Prices are a dated snapshot. Company statements establish what companies say; independent research tests particular claims.

AI is moving from something people try to something institutions build into everyday work. Stanford's 2026 AI Index reports AI use in 88% of surveyed organisations, with generative AI used in at least one business function by 70%. These are survey findings, not a census of every business, and adoption does not establish that an organisation has made a profit from using it. ^[1]

That distinction matters to anyone wondering whether they are being left behind. A company can purchase a tool, announce an AI strategy and run pilots without proving the resulting work is better. An individual can use a chatbot every day without needing its most expensive plan.

The pressure to adopt nevertheless changes the choices available. If an employer expects AI-assisted output, a supplier makes its service AI-first or a public body automates a decision, the consequences extend to people who never bought a subscription. The relevant test becomes practical: what changed for the person receiving the service, doing the work or paying the bill?

Artificial intelligence is a broad category. A system that predicts a protein structure and a chatbot that writes an email are both AI, but their abilities and purposes differ. AGI usually refers to broadly capable intelligence able to perform across many cognitive tasks, rather than excelling only at a narrow speciality. Superintelligence goes further, implying capabilities substantially beyond human performance.

There is no single, universally accepted finish-line test. Google DeepMind's Levels of AGI framework distinguishes breadth of capability, performance and autonomy. Those dimensions should not be collapsed into a marketing claim that one impressive demonstration means the race is over. ^[2]

An assistant may generate excellent prose and still need checking. A system may solve difficult technical problems without being authorised to act independently. Intelligence, reliability and permission are different things.

Being first would not necessarily mean owning the future. A lead might be temporary; rivals might offer cheaper systems; useful applications might matter more than the original breakthrough. Distribution, energy access, customer trust and the ability to operate economically could determine commercial success.

The public should therefore ask two questions whenever a company claims an AGI milestone: what exactly was demonstrated, and who independently tested it? A definition that changes to suit the announcement is not a reliable measure of progress.

The competitors do not all describe the destination in the same way. OpenAI explicitly links its infrastructure ambitions to AGI. Google DeepMind develops both advanced systems and frameworks for judging their progress. Anthropic describes powerful AI as a potential source of scientific and economic gains. Meta has framed its ambition as personal superintelligence. These are stated objectives, not proof that the promised future has arrived. ^{[2] [3] [5] [6]}

ChatGPT, Claude, Gemini and Grok are the familiar consumer faces of a wider contest. The other participants include cloud providers, chip suppliers, manufacturers, energy companies and governments. Treating this as a contest between chatbot personalities misses the businesses supplying the underlying capacity. Our AI Compute investigation examines that infrastructure in detail.

In January 2025, OpenAI announced that Stargate intended to invest $500 billion over four years in US AI infrastructure. That number describes an announced intention. It should not be repeated as $500 billion already spent, earned or returned to investors. ^[3]

The potential prize has several forms: recurring subscriptions; enterprise contracts; paid access for software developers; and services that become difficult for customers to replace. A company need not monopolise intelligence itself to benefit from becoming the route through which millions of people access it.

But revenue is not profit. Construction, financing, equipment replacement, research and daily operation all have to be paid for. A spectacular funding round is evidence that investors are willing to take a bet. It is not evidence that the bet will succeed or that its returns will be shared widely.

Greed belongs in this story. So do curiosity, status, fear of failure and the desire to produce something consequential. Declaring that every AI executive is motivated only by greed would make the argument easier and the evidence weaker. We cannot read private motives from a subscription price.

We can examine the choices organisations make when their interests conflict with those of others. Do contracts make leaving unnecessarily difficult? Can customers challenge changes? Are public subsidies matched by public benefits? Are risks disclosed before people become dependent? These are observable tests of power.

The FTC's January 2025 study of major cloud-provider and AI-developer partnerships identified potential restrictions on access to important inputs, higher switching costs and privileged access to sensitive business information. The report concerns specific arrangements and competition risks; it is not a finding that every partnership is unlawful. ^[7]

The mechanism is more revealing than a villain's portrait. A developer needs computing capacity. A cloud provider supplies it, invests in the developer and may gain a long-term customer. The arrangement can finance innovation while also making independence harder. Both consequences deserve scrutiny.

Competitive fear can intensify the problem. If each company believes slowing down will hand an advantage to its rivals, restraint becomes commercially costly. That possibility does not excuse harmful decisions. It explains why relying only on personal goodwill would be a fragile way to govern a powerful industry.

The dividing line is accountability. Earning money from useful work is not proof of exploitation. Seeking returns while concealing material risks, obstructing meaningful choice or passing avoidable costs to others would warrant a different judgement. The question is what the record establishes, not which label produces the strongest reaction.

AI's strongest defence is a task completed better. In the published research Generative AI at Work, access to an assistant increased customer-support productivity, measured as issues resolved per hour, by an average of about 15% in the studied setting. The gains varied across workers. This is evidence for a particular deployment, not a guaranteed return on every chatbot subscription. ^[8]

Scientific work provides another example. The AlphaFold 3 paper demonstrated improved predictions of structures involving different biomolecules. Such tools can help researchers investigate interactions that would otherwise be difficult to study. A predicted molecular structure is not an approved medicine, and success in this specialist system does not establish the clinical reliability of a general chatbot. ^[11]

For everyday users, the useful possibilities are less theatrical: understanding a difficult document, receiving explanations at a comfortable pace, adapting text into a more accessible form or getting help to start a task. Their value should be assessed through whether the person actually understands more or completes useful work. The benefit need not be income; saved effort, confidence and access can matter too.

We should be equally careful with negative results. METR's early-2025 trial found that 16 experienced open-source developers took 19% longer with AI on 246 tasks in projects they knew well. Its February 2026 update said newer tools likely helped more, but selection effects prevented a reliable estimate of the current speedup. Quoting the old slowdown as a verdict on all present-day AI would misrepresent the research. ^{[9] [10]}

The lesson is to measure the whole task: instructions, waiting, checking, correction and final delivery. An answer arriving quickly is only one part of the job. Useful output matters more than impressive output.

The ILO and NASK's 2025 index estimated that one in four jobs worldwide was potentially exposed to generative AI. It explicitly distinguished exposure from actual job losses and judged transformation more likely than wholesale replacement. A task within an occupation can be affected without the occupation disappearing. ^[12]

That does not make workers' concerns imaginary. If software removes part of a job, an employer has choices: reduce drudgery, shorten hours, improve service, increase output targets or reduce staffing. The technology does not itself decide how the gain is divided.

A worker may become more productive without receiving more pay. A business may lower its costs without lowering its prices. Conversely, a small operator may gain capabilities previously affordable only to a larger company. All three outcomes are possible; the article's test is which one occurs in practice.

The overlooked question is what happens to the route into skilled work. If entry-level tasks are automated, institutions must still provide ways to learn judgement and acquire experience. Telling people simply to learn AI does not explain who pays for the transition, what training is available or how a displaced worker gets the next job.

For the person at home, the AI race can become another recurring expense. A low entry price is different from the cost of sustaining a demanding workflow. The following comparison records selected advertised US-dollar web prices checked on 7 September 2026. It is not a ranking of quality or a quotation for UK checkout.

Local currency, applicable tax and purchase platform can change what appears on a bill. Annual discounts involve different commitments: Claude also advertises Pro at $200 paid upfront for a year. Free tiers exist, and these examples are not an exhaustive list of plans. ^{[13] [15] [16]}

Paying monthly for the three everyday plans above would total $70 a month, or $840 a year, before applicable additions. A $200 monthly plan alone annualises to $2,400. These are arithmetic examples, not a claim that anyone needs multiple subscriptions or should buy the most expensive tier.

The comparison must also ask what happens at the limit. OpenAI documents variable caps on Plus and separate model allowances on Pro. Anthropic explains that usage depends on factors including conversation length, task complexity, model and effort, and offers optional paid usage credits. Paying more does not make the meaning of an allowance self-evident. ^{[13] [14] [17] [18]}

A fair consumer standard would make the price, allowance, reset period and extra-cost options understandable before purchase. The crucial measure is the cost of completing the user's work, not the number of features on the pricing page.

A token is a unit a model processes, often a word fragment or punctuation mark. It is not a word in every case, a unit of truth or a guarantee of useful work. Providers may distinguish input, generated output, cached input and reasoning usage. ^[25]

Consumer subscriptions and API billing must be kept separate. A ChatGPT Plus subscription does not include separately billed API use. If someone builds an automated bot using an API, its spending depends on that API's model rates and any other chargeable services, not simply on the chatbot subscription. ^[13]

Consider an automated task that reads a long document, generates a response, checks its work and tries again. Depending on the service and configuration, more than the final answer contributes to metered usage. Large inputs, repeated context and multiple attempts can therefore matter as much as the visible response. Cached processing may have a different rate; tool use can have its own charges.

It is possible to consume an allowance while producing something the customer cannot use. That does not establish deliberate waste by the provider. It does expose a mismatch between what is sold—processing and access—and what the customer wants: a reliable outcome.

For a business, a sensible comparison includes subscription or API charges, connected services and the human time required to check the result. For a household, the same principle applies without pretending every spare hour has a cash value. A cheaper plan that consistently solves the actual problem can offer more value than a premium plan bought through fear of falling behind.

Controlling an AI system means setting boundaries on what it may do and testing whether those boundaries hold. Controlling access to AI means deciding who can use a service, at what price and under which rules. Controlling people through AI means using systems to influence, monitor or make consequential decisions about them. These questions overlap, but they are not interchangeable.

The case for safeguards is substantial: a system with access to tools or important information can cause harm if it is unreliable or misused. The International AI Safety Report 2026 surveys current reliability and misuse risks as well as possible future severe loss of control. It records disagreement over the likelihood of those future scenarios and distinguishes them from today's unwanted outputs. ^[24]

The counterargument is about authority. If a small group defines acceptable answers, permissible uses and access conditions, its decisions deserve scrutiny. A restriction might protect someone from harm; it might also be poorly designed or block legitimate work. Labelling every restriction censorship avoids the first problem. Labelling every restriction safety avoids the second.

Nor would releasing every powerful capability without restrictions guarantee public freedom. It could also give harmful actors more capacity. The useful questions are specific: what threat is addressed, how proportionate is the response, who evaluates it and what remedy exists when the system gets it wrong?

For the ordinary person, control includes mundane rights: knowing when a decision was automated, reaching a responsible human, correcting information and leaving a provider without losing essential work. Our related files examine AI safety versus AI control and AI in benefits, hiring, housing and insurance.

AI has a physical operating cost. Training builds a model; inference runs it to answer requests. Beyond processors, a working facility needs networking, storage and cooling. Making a service popular creates an ongoing need to serve it.

The IEA's updated 2026 central projection puts global data-centre electricity demand at roughly 950 TWh in 2030, up from 485 TWh in 2025, reaching about 3% of global electricity demand. These are all-data-centre figures, not an AI-only total. A terawatt-hour is one billion kilowatt-hours. ^[19]

Efficiency and growth can happen simultaneously. The IEA reports improving energy efficiency per AI task alongside growing use and more energy-intensive applications. A short text request and an extended reasoning or video-generation task should not be assigned the same footprint. ^[19]

The carbon consequence depends on the electricity supplied. The IEA's supply analysis describes contributions from both renewables and fossil fuels. Energy use and emissions are related, but they are not identical measures: the same quantity of electricity can have a different emissions footprint in different grids. ^[20]

The choice for society is not simply whether to use electricity. Hospitals, transport and homes use it too. The question is whether a particular use produces enough public value, whether cleaner supply is actually delivered and whether alternatives could achieve the same result with fewer resources.

Water is where a global technology becomes a local argument. Berkeley Lab estimated that US data centres directly consumed about 66 billion litres in 2023. That is a historical estimate for US data centres as a whole. It is neither a worldwide total nor water attributable solely to AI. ^[21]

The terms matter. Withdrawal is water taken from a source. Consumption is the part not promptly returned for use in the same system, including evaporation. Direct use occurs at the facility; indirect use can arise from electricity generation. Water moving around a sealed cooling circuit is not necessarily being consumed at the same rate. ^[21]

OpenAI says its Abilene site uses closed-loop cooling rather than traditional evaporative cooling towers. That is relevant evidence of the company's stated design. It is not an independently verified account of every litre used across the site's operation, construction and electricity supply. ^[4]

Air-based, evaporative and liquid-cooling arrangements involve different trade-offs. The phrase liquid cooling alone does not establish the facility's net water consumption; the method used to reject heat outside the building matters. ^[21]

For a resident, the meaningful questions are the local water source, seasonal demand, drought conditions, competing users and enforceable limits. A modest share of national consumption can still matter in a stressed catchment. A large national figure, by itself, cannot prove that a named facility caused a household shortage.

We therefore do not attach a universal bottle-of-water figure to every chatbot prompt. Such a number needs a specified model, task, location, cooling system and accounting boundary. Without them, apparent precision obscures the information that would let a community judge the proposal.

Companies are not ignoring the environmental challenge. Google reports contracting more than 12 GW of new clean energy in 2025 and says its water projects replenished roughly 78% of that year's total freshwater consumption. Its disclosures also explain that actual generation can differ from contracted amounts. ^[22]

Those efforts deserve assessment on their merits. They do not mean the underlying operations have no footprint. A contract is not continuous electricity delivered at every operating hour, and replenishment is not proof that a withdrawal caused no local pressure.

Google also publishes estimates of emissions avoided through its initiatives and products. Such figures compare reality with a modelled alternative. They should not be confused with an independently observed fall of the same size in the company's total emissions. ^[22]

The fair standard applies to critics too. A credible assessment states its boundaries: which activities, locations and years it includes; whether it counts manufacture and construction; and which assumptions shape the comparison. Environmental benefits from using AI and environmental costs from operating it both belong in the account. Neither should be used to make the other disappear.

More investment cannot instantly produce every physical input. The IEA's 2026 analysis identifies tightened supply bottlenecks, including high-bandwidth memory used in AI systems, with shortages expected to persist through at least the end of 2027. That is a specific forecast about a particular constraint, not proof that every consumer computer part is unavailable. ^[19]

A server order also needs manufacturing, packaging, networking, electrical equipment and an operational site. Scarcity at one step can delay the whole project. The leverage belongs partly to organisations controlling those steps, even if they never operate a consumer chatbot.

At the other end of the equipment's life is disposal. A 2024 study in Nature Computational Science modelled cumulative generative-AI-related electronic waste of 1.2–5.0 million tonnes over 2020–2030 under different development scenarios. This was a projected cumulative range, not measured annual waste or an inevitable outcome. ^[23]

Hardware lifetime, reuse and refurbishment therefore deserve a place beside model performance. An efficiency improvement in a new machine must be considered alongside the resources required to make it and the fate of the equipment it replaces. Rapid replacement can create an environmental cost even when the replacement runs a task more efficiently.

Not automatically. The IEA says the effect depends on supply, demand, investment and policy choices. New load can put pressure on a constrained system; different conditions can allow existing assets to be used more efficiently. The allocation of new infrastructure costs is crucial. ^[19]

For a proposed development, ask who pays for the connection, substation and additional generation; what demand the operator commits to; and what happens if the expected load never arrives. A promise of jobs does not answer those questions. Neither does a developer paying its own electricity bill necessarily establish that all wider costs are covered.

Local benefits also need to be defined. Construction work, permanent jobs, tax receipts and improvements available to the wider community are different things. A council should be able to explain which benefits are contracted, which are estimates and what remedies apply if commitments are missed.

This is where the household and the AGI race meet. The person who never opens ChatGPT can still have an interest in the infrastructure built nearby. The appropriate response is transparent evidence about the actual project, rather than assuming every data centre is either a public gift or a local disaster.

The answer is conditional: AI benefits people when it helps them do something worthwhile at an acceptable total cost, and when the wider gains are shared rather than merely promised.

For an individual, compare the service with the best realistic alternative, including an existing tool or a free tier. Record useful tasks completed, time spent checking and whether the result would have been good enough without assistance. The point is not to turn leisure into a productivity contest; it is to find out whether the purchase serves the person.

For a worker, ask whether increased output brings higher pay, better conditions, less drudgery or more demanding targets. For a customer, ask whether the service becomes cheaper, more accessible and easier to correct. For a citizen, ask whether institutions remain accountable when something goes wrong.

A person can benefit from AI without subscribing to anything: through useful research, improved tools or a service that works better. Another person can pay for several subscriptions and receive little lasting value. Spending and benefit are separate measurements.

The most revealing outcome would be greater independence. If AI helps someone understand a problem, acquire a skill or build something they can maintain, that is a durable gain. If every task becomes dependent on an expensive service they cannot question or leave, convenience has come with a different bargain.

A defensible AI economy would make its benefits observable. Providers would explain prices and limits clearly, support usable exports and allow meaningful correction. Organisations would evaluate systems on completed work and consequences, rather than deployment announcements alone.

Workers would have a say in how automation changes their jobs. Public bodies would retain accountable humans for consequential decisions and offer realistic routes to challenge mistakes. Safeguards would have an identified purpose, evidence of effectiveness and proportionate limits.

Infrastructure approvals would publish local water and electricity assumptions, distinguish projected from delivered benefits and make cost allocation explicit. Environmental reporting would show absolute impacts alongside efficiency gains, with transparent methods for claimed savings.

These are standards proposed by this investigation, not a claim that one checklist resolves every technical risk. Their value is that they can be tested. They make the promise of public benefit answerable to someone outside the organisation making it.

The next evidence to watch is similarly concrete: independently assessed capabilities, actual operating costs, changes to subscription allowances, completed infrastructure, local resource data and measured effects on workers and service users. A revised forecast or a new product name is less decisive than a result that survives scrutiny.

What is the difference between AI and AGI?

AI covers many systems with different capabilities. AGI refers to broadly general cognitive capability. There is no universally accepted test that turns a single impressive result into proof of AGI. ^[2]

Does reaching AGI first guarantee a monopoly?

No. Capability, cost, adoption, infrastructure and competition would still matter. A permanent monopoly is a possible concern to investigate, not an established consequence of being first.

Why can a paid AI subscription still have limits?

The subscription purchases access under a defined plan. OpenAI and Anthropic document usage allowances; some limits depend on model and workload. API use and optional extra usage can have separate charges. ^{[13] [14] [17] [18]}

Does every AI query use the same electricity and water?

No. Task, model, location and infrastructure affect the footprint. All-data-centre totals cannot be assigned entirely to AI or divided into a universal per-prompt figure.

Will AI take everyone's job?

The evidence does not establish that. The ILO measures exposure to changes in tasks, not a forecast that all exposed jobs disappear. The distribution of gains and losses remains a policy and workplace question. ^[12]

Is paying for AI worth it?

It can be when useful outcomes exceed the total cost for that user. Include checking and correction time. The highest-priced plan is not automatically the best choice, and free access may be sufficient.

Overall assessment: Unresolved on the public bargain. The evidence supports useful applications, paid access with limits, infrastructure pressure and mechanisms through which power can concentrate. It does not support a predetermined story in which AI either saves everyone or inevitably subjugates them.

Greed is a legitimate subject of investigation. So is the possibility that ambitious people build valuable things. The test is what happens when public benefit and private advantage diverge: who can inspect the decision, challenge it and change the outcome?

The strongest case for AI is that it can put useful capability into more hands. The strongest warning is that the means of delivering that capability can make people dependent on a small number of organisations. Both deserve to be taken seriously.

The AI race should be judged by whether people become better off, more capable and more independent—not by how much intelligence they are persuaded to rent.