As AI Gets More Powerful, Who Will Be Able to Afford It?

Artificial intelligence is becoming more capable at a remarkable pace. The latest generation of frontier models can reason across increasingly complex problems, write and analyse code, work with large volumes of information and perform tasks that previously required significant human involvement.

For users, that progress can appear relatively simple. A new model becomes available, a subscription is purchased and its capabilities can be accessed through a computer or phone.

Behind that simplicity sits a very different economic reality.

The systems at the leading edge of AI require enormous amounts of computing power. That computing depends on advanced processors, high-performance memory, data centres, networking, cooling and electricity. As AI moves towards more autonomous and computationally intensive workloads, the infrastructure required to support it is becoming an increasingly important part of the story.

This raises an uncomfortable question.

As AI gets more powerful, who will be able to afford it?

The answer may not be as straightforward as assuming that the most advanced AI will eventually become a service reserved for the world’s largest technology companies.

There is another group of companies working on a different problem altogether: how to make increasingly capable AI cheaper, faster and less demanding to operate.

If they succeed, they could play an important role in determining whether the benefits of advanced AI remain concentrated among organisations with enormous computing budgets or become increasingly accessible across the wider economy.

The cost behind artificial intelligence

The price paid by an individual user is only one part of the economics of AI.

Developing and operating advanced models requires infrastructure at several levels. Training requires substantial computing resources, while inference, the process of actually running a trained model to produce an answer or complete a task, creates an ongoing cost.

That distinction becomes increasingly important as AI develops beyond simple question-and-answer systems.

A user asking an AI to summarise a document creates one workload.

An autonomous AI system conducting research, writing software, testing its work, accessing databases and repeating the process until it reaches an acceptable result creates a very different workload.

The more work AI performs, the greater the potential demand for computing.

This creates an unusual economic situation. Improvements in efficiency can make individual AI tasks cheaper, but cheaper AI can also encourage much greater use.

The International Energy Agency expects electricity consumption from data centres to rise sharply this decade, with AI-intensive accelerated servers among the key drivers. The IEA estimates that global data-centre electricity consumption could more than double to around 945 terawatt-hours by 2030.

The precise outcome remains uncertain, but the direction of investment is clear. AI is creating demand for infrastructure well beyond the companies actually developing the models.

The infrastructure problem

The AI industry has become closely associated with GPUs, but the infrastructure challenge is considerably broader.

Advanced processors need high-performance memory. Data has to move between processors and memory at enormous speeds. Servers have to be connected through increasingly sophisticated networking infrastructure. The resulting heat has to be removed, while data centres require reliable access to electricity.

This creates multiple potential constraints.

A company may have access to an advanced AI model but still face limitations in how much AI it can run, how quickly it can run it and how much it costs.

That distinction could become increasingly important as AI adoption grows.

The largest technology companies have the financial resources to invest billions of dollars in computing infrastructure. Smaller companies do not necessarily have the same options.

The question is whether technological innovation can reduce that difference.

There is a growing group of companies attempting to do exactly that.

Making AI inference more efficient

One of the most interesting areas is AI inference.

Training an advanced model can require enormous resources, but once a model has been created, it still requires computing power every time somebody uses it.

At global scale, even relatively small improvements in the cost of inference can become significant.

That has created an opportunity for companies developing specialised hardware and software designed specifically to make AI workloads more efficient.

One example is d-Matrix, a Silicon Valley chip company focused specifically on AI inference.

Its Raptor processors are designed for the execution of AI models rather than primarily for training them. In September 2026, d-Matrix announced that it would integrate its technology with Nvidia’s NVLink Fusion platform, allowing its inference processors to be incorporated into Nvidia-based data-centre systems. The companies expect integrated systems to become available in 2027.

The significance is not simply that another chip company is entering the market.

It illustrates the growing importance of specialised infrastructure designed around the economics of inference.

If AI increasingly becomes something that runs continuously rather than occasionally, reducing the cost of each workload could become strategically important.

Building specialised hardware

Etched is approaching the same broad problem from another direction.

The San Jose-based startup is developing specialised AI processors designed around inference workloads. In August 2026, Reuters reported that the company had raised $700 million at a valuation of approximately $21 billion, more than doubling its valuation in less than a month. The company had also secured more than $1 billion in customer contracts, according to the report.

A valuation of that size does not prove that the technology will ultimately succeed. Semiconductor development is highly competitive and technically demanding, and established companies have substantial resources.

It does, however, demonstrate the level of investor interest in the possibility that specialised hardware could make AI inference faster and more economical.

The broader question is whether the industry needs increasingly general-purpose systems or whether a mixture of specialised processors will ultimately provide a more efficient route to deploying AI.

Using the right chip for the right task

That question sits at the centre of another emerging approach.

British startup Callosum is developing software designed to allow AI workloads to operate across different types of processors rather than relying on a single computing architecture.

The company argues that different AI tasks can benefit from different combinations of models and hardware. Its platform is intended to determine which combination is appropriate for a particular workload.

Callosum announced a $100 million seed round in August 2026 led by Atomico, with participation from the UK Sovereign AI Fund and others.

The concept is significant because AI infrastructure does not necessarily have to be built around the assumption that one type of processor should perform every task.

If a relatively simple AI operation can be completed using less expensive hardware, sending it to a highly specialised and expensive accelerator may not make economic sense.

The potential benefit is not necessarily a more powerful model.

It is better utilisation of the computing resources already available.

That could become increasingly valuable as AI workloads diversify.

Making the models smaller

Hardware is only one way to reduce AI’s infrastructure requirements.

Another is to reduce the size of the models themselves.

Spanish company Multiverse Computing has developed CompactifAI, a model-compression technology designed to reduce the size of large language models while maintaining their performance.

The company announced a $215 million funding round in 2025 and reported that its technology could reduce model size by up to 95%, with claimed reductions in inference costs of between 50% and 80%. These figures are company claims and should not be treated as a guarantee of performance across all models or workloads.

The underlying idea is important regardless of the eventual commercial outcome.

If a model can be made substantially smaller without a corresponding loss in capability, it can potentially require less memory, less computing power and less energy.

That could change where advanced AI can operate.

Instead of requiring access to the largest data-centre infrastructure, sufficiently compressed models could potentially operate on private infrastructure or increasingly powerful edge devices.

Multiverse’s more recent funding plans show how seriously investors are taking the efficiency opportunity. In July 2026, the company announced plans for a $570 million Series C round at a reported pre-money valuation of $1.7 billion, with the stated focus including efficient AI deployment from edge devices to data centres.

Again, the future commercial success of any individual technology remains uncertain.

The larger trend is easier to identify.

Reducing the amount of computing required to deliver useful AI has become a market in its own right.

Could efficiency keep advanced AI accessible?

This is where the question of affordability becomes more interesting.

There are two competing forces.

The first is the increasing cost and complexity of frontier AI infrastructure.

The second is the rapid development of technologies designed to reduce those costs.

If the second advances quickly enough, the cost of accessing sophisticated AI could continue to fall even as the underlying models become more capable.

That has happened repeatedly in technology.

Computing has become more powerful while becoming accessible to a much larger proportion of the population. Storage capacity has increased dramatically while the cost per unit of storage has fallen. Internet access has expanded while the underlying infrastructure has become considerably more sophisticated.

There is no guarantee that AI will follow precisely the same trajectory.

But there is also no reason to assume that today’s infrastructure requirements represent the permanent cost of advanced AI.

The companies working on efficiency are effectively betting on the latter.

The importance of doing more with less

The economic value of AI efficiency could become considerable if AI adoption continues to expand.

Consider a simple example.

If a business can achieve the same result using half the computing resources, it can potentially reduce its costs or use the available infrastructure for additional workloads.

If that improvement is applied across millions of AI interactions, the potential savings become much larger.

At the infrastructure level, the consequences could be even more significant.

• Better model compression can reduce memory requirements.
• More efficient processors can reduce energy requirements.[
• Better orchestration can improve the utilisation of existing hardware.
• More efficient networking can reduce the energy required to move data.
• Each improvement addresses a different part of the same problem.

The objective is not necessarily to make AI less powerful.

It is to make more intelligence possible with the same resources.

The gap between frontier AI and everyone else

This brings us back to the original question.

Does the increasing cost of AI mean that its most powerful capabilities will eventually be available only to the world’s largest companies?

There is currently no basis for saying that with certainty.

The opposite may happen.

If AI efficiency improves rapidly enough, the cost of accessing advanced capabilities could fall considerably. Smaller businesses could benefit from more efficient models, specialised hardware and software that allows them to use computing resources more effectively.

However, there may still be a difference between accessing advanced AI and deploying it at enormous scale.

A small company could have access to an exceptionally capable AI assistant without having the resources to run thousands of AI agents continuously.

A multinational organisation could potentially combine advanced models with proprietary data, large-scale computing infrastructure and extensive automation.

Both organisations would be using AI.

The scale of their deployment could be very different.

That distinction may ultimately matter more than the price of an individual AI subscription.

An emerging investment opportunity

For investors, this creates an important area to watch.

The AI investment story is often dominated by model developers and the largest semiconductor companies.

But the companies attempting to make AI more efficient could become an equally important part of the ecosystem.

d-Matrix is targeting inference hardware.

Etched is developing specialised processors.

Callosum is attempting to optimise how AI workloads are distributed across different types of computing hardware.

Multiverse Computing is attacking the problem at the model level through compression.

They are different approaches to the same broad challenge: getting more useful intelligence from limited computing resources.

None of these companies provides proof that the affordability problem has been solved.

Nor is it clear which technological approach will ultimately prove most successful.

That is precisely why the sector is worth watching.

The economics of intelligence

The next stage of artificial intelligence may therefore depend on two races taking place simultaneously.

One race is to build increasingly capable models.

The other is to make those models economically viable at scale.

The second race may receive less attention, but it could prove just as important.

If AI becomes dramatically more capable but requires proportionally more computing, memory and electricity, the cost of deploying that intelligence could become an increasingly important competitive factor.

If companies can make those systems substantially more efficient, the picture changes.

Advanced AI could become cheaper to operate, more widely deployable and accessible to a much broader range of businesses.

That does not mean the largest technology companies will lose their advantages. They are likely to continue possessing enormous resources, proprietary data and access to infrastructure.

But it could prevent the gap from becoming purely a question of financial scale.

The most important AI companies of the next decade may therefore include not only those building the most intelligent systems, but also those finding ways to make intelligence cheaper to produce.

For investors, that creates a broader opportunity than the race to build the next frontier model.

The companies making AI more efficient could ultimately determine how far that technology spreads.