When and how are AI firms going to recoup their immense and ongoing investment in AI models and factories?
The Financial Times reports Sam Altman, CEO of OpenAI and Jensen Huang, CEO NVIDIA will accompany President Trump on his unprecedented and profoundly unpopular state visit to the UK* this week. Most of it will be spent beyond the public gaze at Windsor Castle (pictured) and the Prime Minister’s grace-and-favour residence, Chequers. The heads of the companies are apparently to pledge support for the development of new AI data centres in the UK which “could be worth billions”. Oh really?
Field of dreams?
Well maybe in terms of the cost of building, kitting out (including chips, other hardware and software) and powering the premises, and the people required to do all those things then run and upgrade them. Whether the new data centres/Ai factories produce billions of pounds-worth of value from the AI models they train and operate is another matter.
Investors in AI firms like Anthropic and OpenAI by hyperscalers in particular and others, as well as Google’s own Gemini model, is predicated on the notion that their large language models (LLMs) will develop an unassailable lead – like Google did with search, say.
According to The Guardian in August, “The US’s largest companies [Meta, Microsoft, Amazon, and Alphabet] have spent 2025 locked in a competition to spend more money than one another, lavishing $155bn on the development of artificial intelligence, more than the US government has spent on education, training, employment and social services in the 2025 fiscal year so far.”
How are they going to get the return on their investment from AI? There are a number of reasons to think they might not.
Untested business model
Firstly, some commentators, like John Mihaljevic of MOI GLOBAL, an investor community, argue that “by ramping up AI-related capital spending to unprecedented levels, [hyperscalers] have set themselves on a perilous path, away from high-margin, capital-light models toward a capital-intensive future in which their return-on-capital and margin profiles are highly uncertain.
“This AI-driven capex frenzy is eerily similar to the telecom bubble of the late 1990s and early 2000s. Extravagant spending on fiber optics and network infrastructure promised growth but delivered catastrophic oversupply and collapsing prices. Today’s hyperscalers may be repeating history’s costly mistakes.”
In this scenario, it might be that consumers of AI – individuals and all kinds of businesses and industries – could be the winners?
In the meantime, there is the matter of feasibility: “It can be so expensive to implement the use cases,” Nilmar Seccomandi David, Director of Autonomous Network and Infrastructure in the Global CTIO Unit of Telefónica told Mobile Europe in a recent interview for a new research report – see banner below. “In some cases, you can spend €1 million to implement a use case, but the benefit might be just €50,000. We can’t justify that.”
Telefónica has implemented about 400 autonomous network use cases, many of them using AI, as part of its Autonomous Network Journey (ANJ) programme, launched in 2021. And yes, the cost of the tech will come down, as it always does, but the timing and degree of that fall is unknown and many other things develop in parallel.
Is an unassailable lead indefensible?
Which brings us to question of whether an unassailable lead really achievable? China’s DeepSeek-R1 paused the stock market band wagon briefly last January, appearing from nowhere yet apparently rivalling the performance of some celebrated LLMs at fraction of the cost. Its open source approach uses relatively little computing power and so consumes less electricity and needs fewer and less sophisticated chips. Surely open source could challenge the unassailable leader approach to the market?
Certainly, there are numerous examples of communities of interest doing their own thing. For example, in May Sweden launched its largest enterprise AI supercomputer (aka an AI factory) in partnership with NVIDIA and AstraZeneca, Ericsson, Saab, SEB and Wallenberg Investments. Clearly the plan is for the country’s pharma, automotive and financial sectors to devise and run models specific to their needs, keeping the models and training data within their own territory.
In June some of Europe’s largest and most forward-thinking operators – Orange, Fastweb (of Italy), Swisscom, Telefónica and Telenor – announced they would collaborate collectively with NVIDIA to develop and expand sovereign AI factories and edge infrastructure across Europe, bringing scale – and therefore the potential of less cost – into the equation.
Will LLMs deliver on their promise?
Some pundits say that GenAI is too general and inexact. “Packaging [the models] into services seems premature and hard to bring to market just now, especially when it comes to the financial sector, healthcare, and other critical industries that won’t allow for the high degree of variability in end results that AI currently allows for,” as Techzine’s Erik van Klinken so neatly puts it.
He suggests that deals such as that announced earlier this month between the Dutch chipmaker ASML and France’s Mistral AI with the former leading a €1.7 billion C round of funding might be a more successful blueprint for investing in AI than of Microsoft splashing the cash between OpenAI, Anthropic and others. Or OpenAI agreeing to pay Oracle Cloud $300 billion over five years, starting in 2027, for providing five gigawatts of compute capacity, as reported in the Wall Street Journal.
This might have helped make Larry Ellison, Oracle’s founder, richer than Elon Musk temporarily, but elsewhere it has been greeted with scepticism – and the suggestion that as things stand, OpenAI is unlikely to have the funds to pay for it.
Microsoft has invested heavily in OpenAI, to the tune of more than $10 billion, but relations have been strained over the last two years, with dramas over its not-for-profit status which at one point led to CEO Sam Altman being fired. Now it seems the the two are trying to realign, having signed a non-binding agreement that could allow OpenAI to change its modus operandi and eventually float on the stock market.
Not meeting expectations
Then, as the investment advice always goes, “past performance is no guarantee of future success”. ChatGPT5’s launch in August was as welcome as a wet summer holiday. Underwhelmed users voiced their disappointment and dissatisfaction. There were accusations of ChatGPT5 being about saving money rather than pushing the boundaries of the tech and demands to reinstate the older version.
ChatGPT5 had been hyped: it was widely expected it be a big step beyond GenerativeAI to Artificial general intelligence (AGI) which, according to Google, “refers to the hypothetical intelligence of a machine that possesses the ability to understand or learn any intellectual task that a human being can”. On other words, it can mimic the human brain which means it can transfer knowledge and skills learned in one domain to another and adapt to new situations.
Also, it is predicted that AGI will have “common sense” – knowledge about the world beyond facts to include relationships and social norms, which it would factor into its reasoning and conclusions. Common sense is a slippery concept. For one thing, there is no universal definition of common sense and you don’t have to look far to see that people make a lot of stupid decisions – it’s part of the human condition. On the other hand, good decisions are sometimes illogical – depending on the desired outcome.
Economics for telcos?
Back to this week’s state visit by the US President and AI ‘royalty’ to the UK. The support for new AI factories in the UK by OpenAI and NVIDIA add to the slew of deals this year that the two have signed this year with countries that are US’ allies (or perhaps more accurately countries that since World War II assumed they were allies) as they scramble to develop sovereign AI infrastructure.
The drive for tech sovereignty is in large part driven by fear of over-reliance on US hyperscalers and tech – although both OpenAI and NVIDIA are US companies, which is rather underlined by them riding on the coattails of the US President’s state visit.
Chipmaker NVIDIA supplies about 80% of the world’s AI chips, although some estimates put it as high as 95%.
OpenAI developed ChatGPT which burst onto the market just three years ago. In February, Brad Lightcap, OpenAI’s Chief Operating Officer, said ChatGPT had 400 million users a week, which according to the BOND report, had increased 800 million weekly users by April. Other sources suggest this might have been an overestimate but either way, OpenAI expects to attract a billion users by the end of the year. Semrush said chatgpt.com had 5.38 billion visits and was the world’s fifth most visited website.
* A state visit is a royal event using assets of UK’s Civil Service, the Royal Household and the Household Division.
