From people to GPUs: when will AI actually pay for itself?

Does machine intelligence create enough value to justify its massive costs? Dr. Ruth Wandhöfer breaks down the true human and economic impact of AI.

I think by now every business has somehow gotten hooked into the AI revolution. Incredible capabilities opened up with the arrival of Large Language Models (LLMs). At the same time, the arrival of AI has created a strange economic paradox. Companies are reducing their workforces in the expectation of AI-driven productivity increase while at the same time increasing spending on the infrastructure required to make AI work.

The expense line is clearly moving. Fewer people. More GPUs.

But the question behind this is whether that actually translates into lower costs or not.

For years, technology investment was justified by a relatively straightforward proposition: automate a process, remove manual work, reduce the cost base, and improve productivity. AI promises something much bigger. It can write, analyse, code, investigate, advise, detect, predict, and increasingly act.

But the economics are not yet so straightforward.

Replacing a person with AI does not necessarily eliminate the cost of doing the work. It can change the nature of the cost - from salaries and benefits to model licences, tokens, cloud infrastructure, GPUs, data, integration, cybersecurity, and human oversight (there is still an HR cost!).

The real question is therefore not whether AI replaces people. It is whether the total cost of delivering a hopefully better outcome falls.

The expense line is moving

On the one hand, it is understandable that businesses want to create more efficiencies through automation, whether that is in the context of customer service, software development, research, compliance reviews, or back-office processing. Using AI to do so should theoretically achieve more with fewer people.

But headcount is only one part of the equation.

Imagine removing 1,000 employees while simultaneously increasing spending on cloud computing, GPUs, model licences, data engineering, cybersecurity, and AI specialists. The organisation may look dramatically more efficient from an HR perspective while its technology cost base is rising.

That is not necessarily bad. If AI allows the company to process ten times more transactions, launch new products, or provide significantly better service, the investment may be entirely justified.
But in practice it is not simply cost reduction. The expense has moved.

Nobody knows the economics yet

One of the defining characteristics of the current AI cycle is experimentation.

People are using AI because it is available to them. Employees ask models to write reports, analyse documents, produce presentations, summarise meetings, generate code, and answer questions. Businesses are embedding models into products and workflows before they have necessarily established where AI creates sustainable economic value.

Experimentation is healthy until experimentation quietly becomes infrastructure.

A team starts with a few thousand tokens. Then thousands become millions. A model becomes embedded in a workflow. The workflow becomes business-critical. Suddenly the organisation has a substantial recurring bill for something that was never originally budgeted as strategic infrastructure.

This is where shadow AI becomes more than a cybersecurity issue, because now it is also a cost-accounting problem.

If employees independently adopt multiple AI tools, the organisation may have no consolidated view of how much AI it is consuming, which models it is using, what data is being shared, or what contractual dependencies are being created. At the same time, in the absence of a clear AI policy and adequate enforcement, there is no perspective on which models actually work well for the intended purposes.

The company may believe it has an AI strategy, whilst in reality it has hundreds of individual AI habits.

AI can become addictive

AI has another unusual economic characteristic: it is so very easy to consume. Unlike a traditional software licence that tends to have relatively predictable economics, AI is different in that the more useful it becomes, the more people end up using it. Isn’t that a good thing?

An employee who previously spent an hour researching something may now ask a model ten questions in five minutes. Developers can generate more code. Analysts can run more scenarios. Marketing teams can produce more content.

Productivity rises but, on the flipside, so does compute consumption.

This creates an AI consumption loop: better AI drives more use; more use drives more compute; more compute drives infrastructure costs; and the increased availability of AI encourages still more use.

The critical question then becomes whether productivity grows faster than consumption.

Payments: the perfect AI test case

There is one sector where the AI proposition should naturally be compelling: payments.

Payments generate enormous volumes of data, operate continuously, demand decisions in milliseconds, and sit directly at the intersection of fraud, customer experience, risk, and operational efficiency.

The economics of AI in payments can be compelling.

A fraud model that prevents losses, reduces false positives, and allows more legitimate transactions through can create measurable value on every transaction. An AI system that automates customer-service interactions can reduce operating costs. A model that improves payment routing or authorisation rates can increase revenue without increasing transaction volumes. But payments also expose the other side of the equation.

AI has to operate at enormous scale and at very low marginal cost. It is not enough for a fraud model to be accurate. It must make millions of decisions in milliseconds without adding an uneconomic cost to each transaction.

This is why payments may become one of the best real-world tests of AI economics. The question is not: ‘Does the AI work?’

It is: ‘Does the AI improve the economics of the payment?’

The fintech problem

The calculation becomes harder for the fintech sector.

A large financial institution can potentially spread AI infrastructure costs across enormous transaction volumes. A smaller fintech may need AI simultaneously for fraud, customer service, compliance, software development, and financial crime.

The technology stack can quickly become: = Cloud + GPUs + foundation models + specialised models + data infrastructure + cybersecurity + integration + human oversight

Fintechs tend to have fewer employees than a traditional bank, and logically they tend to run a higher technology cost per employee. So, can a fintech achieve AI-driven operating leverage before AI becomes another structural cost base? 

One answer is specialisation. A payments company does not necessarily need the world's largest language model to determine whether a transaction is suspicious. A smaller model trained for a specific financial task may be faster, cheaper, and more effective.

The future is all about choice. One of the smallest models capable of doing the job reliably could be a better option than simply the largest model available.

The AI arms race in fraud

As banks and fintech are investing in AI to reduce fraud, criminals are investing in AI to increase it.

Synthetic identities, deepfakes, automated social engineering, convincing phishing messages, and voice cloning are making attacks faster and more scalable.

This creates an uncomfortable dynamic - AI investment may become less about gaining competitive advantage and more about maintaining the existing level of security.

Or simply put, we shift from asking ourselves ‘Do we need AI?’ to: ‘What happens if we don't have it?’

From real-time payments to agentic payments

The next development could make the economics even more interesting. Real-time payments have already made money move at machine speed. The next step ahead is making payment decisions move at machine speed. An AI agent could search for a product, compare prices, select a supplier, and initiate a payment within parameters set by its owner. As soon as an AI agent is in the mix, we have to ask the following questions:

• Who authorised the payment?
• Was the agent acting within its mandate?
• What happens if the agent is manipulated?
• What if it misunderstands the customer's intent?
• Who is liable?

These questions become even more important because real-time payments are difficult to reverse. The faster the money moves, the further upstream fraud prevention has to move.

And the volume could explode

Like the early days of APIs and payment initiation and account information services appearing in Europe - which, for some banks, felt more like DDoS attacks with the explosion of API queries on account balances - agentic commerce could also dramatically increase volumes of activity.

Imagine millions of agents continuously:

• comparing prices
• making purchases
• replenishing inventory
• booking services
• paying APIs
• buying data
• purchasing compute
• paying other agents.

Some transactions could be tiny - fractions of a cent - particularly where machines pay other machines for data, processing, or access to digital services.

This triggers another economic challenge. What happens when the cost of making the payment approaches the value of the payment itself?

The payments infrastructure of an agentic economy may need fundamentally different economics from today's card- and account-based systems.

And there is a deeper implication. The more autonomous payments become, the more closely AI governance becomes payment governance and vice versa.

The GPU problem

Behind all of this sits the physical infrastructure. AI companies, banks, fintech, pharmaceutical companies, consultancies, cybersecurity firms, and technology companies are competing for the same underlying resources. GPUs cannot simply be manufactured overnight. Capacity, advanced memory, networking, and data-centre infrastructure are all subject to supply constraints. Whereas we have traditionally thought about labour, capital, energy, and real estate as fundamental inputs into business, AI adds another: compute.

And the economics are increasingly being measured in cost per token and useful output per unit of compute, rather than simply the price of a chip. Current industry analysis increasingly treats inference - the ongoing process of generating outputs after a model has been trained - as a central driver of AI infrastructure economics¹.

Even organisations that train their own models do not escape this problem. They may reduce dependence on a third-party model provider, but they still need GPUs, energy, data-centre capacity, and specialist infrastructure. You can reduce one dependency, but you will not really get out of the other.

Eventually, AI should get cheaper

There is, however, a reason not to become too pessimistic. Technology markets tend to commoditise. More models will enter the market. Open models will improve. Specialised models will proliferate. Inference will become more efficient. Hardware will improve. Alternative architectures will compete.

The cost of generating a unit of intelligence should thus fall over time - so a ‘new AI Moore’s Law’ of some form should still hold.

The economics are already moving in that direction: the industry is increasingly focused on reducing cost per token through better hardware, software optimisation, and more efficient models.

But falling unit costs do not necessarily mean falling total costs.

If AI becomes dramatically cheaper and organisations consequently consume dramatically more of it, the overall technology bill can still logically continue to rise.

The real AI strategy is capital allocation

The AI debate should quickly move beyond the simplistic question of whether AI will replace people. A new fundamental question becomes the following:

What happens to the economics of an organisation when intelligence becomes a variable technology cost?

For payments, the opportunity is enormous. AI can reduce fraud, improve payment acceptance, automate operations, strengthen compliance, and enable new forms of commerce. But it can also introduce model dependency, compute concentration, cybersecurity risk, and unpredictable consumption costs. The strongest businesses will not necessarily be those that deploy AI everywhere.

They will be those that know where AI creates measurable economic value - and where it is simply an expensive way of doing something that could be done more simply. Replacing people with AI is not the objective. Creating more value at a lower total cost, however, is. The companies that ultimately win the AI race may thus not be those using the most models, buying the most GPUs, or cutting the most jobs. They may be the companies that understand exactly where and what type of intelligence creates how much value, what that intelligence precisely costs, and which infrastructure they are becoming dependent upon.

Ultimately, the real AI revolution is not about replacing the human cost of intelligence with the machine cost of intelligence. It is about discovering whether machine intelligence can create enough additional value to justify both. 

 

¹Nividia: https://perspectives.nvidia.com/ai-infrastructure/total-cost-of-ownership/task/faq/revenue-per-rack-ai-inference-datacenter-economics/?utm_source=chatgpt.com (last updated 17.06.2026)

 

About author
Dr. Ruth Wandhöfer
Adviser | Author | Speaker | Board Director | Professor

Dr. Ruth is a leading authority at the crossroads of finance, technology, cybersecurity, and regulation. A former senior Citi executive, Ruth combines roles as an independent Non-Executive Director, Head of European Markets at Blackwired Cybersecurity, Author, Adviser, and Investor.

She also runs her own business, Leximar Advisory, which supports financial institutions, tech, and fintech/cyber security businesses. She is a global keynote speaker, strategist, and Visiting Professor at Bayes Business School.