The AI Bubble & The Dangerous Hype of The Hyperscalers

The decade of the 2020s will undoubtedly be defined by both the rise of artificial intelligence (AI) and the ever worsening climate crisis. Sadly, at some point, we will all look back and wonder why we invested so heavily in the former to the detriment of the latter.

The proponents of AI have told us that the technology holds all the answers and that humanity-level changes are just beyond the horizon. All they need is more investment and more energy to achieve the goal of developing the “AI for everyone” that will transform the working world and life as we have come to know it.

Yet nearly four years on from OpenAI’s much heralded release of ChatGPT 3.5 in November 2022, and its human-like feedback and conversational presentation of data, there is genuinely little in the way of demonstrable evidence that this technology or other similar AI technologies will be world changing. This is despite the hundreds of billions of dollars spent on training AI systems, creating bigger, more powerful processing infrastructure and the seemingly never ending stream of rhetoric from the world’s AI leaders that the future is almost…almost on its way. It will just take a little more investment.

All the hyperscalers in the AI industry – Open AI, Anthropic, Google, Meta - have followed a similar path in the hope of reaching the holy grail of the AI singularity; that point in time when AI becomes capable of improving itself recursively, triggering a runaway cycle of self-upgrade that results in a superintelligence far exceeding all human capability.

To achieve this, the Large Language Model (LLM) variety of AI has seen the greatest level of investment and hype. LLM is the type of AI system that “learns” by being trained on massive amounts of digital data. To facilitate this, websites have been scraped, databases copied, social media networks repurposed as trustworthy data sources and physical books have been destroyed at a scale which would astonish even those who torched the Library of Alexandria over 2,000 years ago, all in an effort for the AI hyperscalers to reach the AI promised land before their competitors do.

Aside from vast amounts of data, LLMs also need ever-more powerful hardware to be able to deliver on the promises made to investors; these include governments, sovereign wealth funds, asset managers and venture capitalists. The problem is that government and sovereign wealth investment mechanisms use what is essentially public money; public money which is being diverted from essential areas to build and power the AI machine.

If that money was actually achieving the goals the AI hyperscalers had promised, why are they still continuing to seek further, record-breaking investment and involving themselves in questionable circular financing practices with the same technology companies that are already profiting so vastly from supplying the physical infrastructure currently powering today’s AI systems?

The system which feeds on itself

The hardware infrastructure which powers the LLMs of the Western world from inside vast, energy-hungry data centres is mostly supplied by technology companies such as NVIDIA. The American company announced in 2025 its intention to invest extensively in OpenAI to the tune of around $100 billion USD (revised down to $30 billion in 2026) to help OpenAI build the multi-gigawatt processing infrastructure it says it needs to reach its goals. The very same infrastructure NVIDIA develops and manufactures.

Put in simple terms, NVIDIA sells AI companies the hardware they need to function and grow and NVIDIA then injects capital into those AI companies so that they can buy/use more NVIDIA technology going forward. This is known as circular financing and creates an overinflation in the value of the companies involved by generating a level of growth which is not reflected in real terms. It is a profitable wave if you can ride it but a financial tsunami when it inevitably breaks.

For circular financing to function it requires a consistent return on investment over a prolonged period of time. Almost four years on from the public release of first LLM interfaces from the major hyperscalers, such as ChatGPT and Claude, the hundreds of billions invested so far have yielded zero return on investment. Towards the end of 2025 it was reported that OpenAI loses $3 for each $1 it earns.

After burning through investment funding greater than the GDP of many small countries and consuming the equivalent amount of energy of even larger countries than that, this represents an astonishing failure for its business model. Yet, it is a business model which remains in place today and one which is still calling out for (and receiving) ever increasing investment.

Also caught up in this circular financing are the likes of Amazon, Google and Microsoft, who provide data centres, networking, storage and cloud computing services to the AI hyperscalers. They also utilise NVIDIA hardware and they all have, themselves, invested in the AI hyperscalers. This has subsequently boosted the valuation of those companies as they are seen to be investing in the AI promised land which, as we are repeatedly told, is almost here.

In mythology, Ouroboros is the snake which eats its own tail, representing an ever-perpetuating cycle of life, death and rebirth. Yet in the circular economy of AI development, the Ouroboros-like system is consuming itself in order to perpetuate itself; requiring money and resources to continuously circulate within it to survive. 

Factor in the uncomfortable truth that the AI infrastructure will require constant upgrades to remain technically relevant (a necessity for investors), along with the pressures of inflation, rising energy costs and the need to secure enough of the rare-earth mineral resources to make the hardware, and you have a funding model which will potentially devour itself, creating an economic black hole which will have cost well over a trillion dollars to create.

Subscriptions and tokens

Up until recently, the most popular LLMs have been made freely available to users in an effort to establish them as trustworthy AI brands and to utilise those free users’ actions as a training base for the LLM systems. But now that investors are beginning to question when they will begin to see some profits, the hyperscalers have begun rolling-out tiered pricing and tokenisation subscriptions for those that wish to access their AI models going forward.

Tiered pricing gives a user access to more powerful versions of a company’s AI and tokenisation is used to limit the amount of processing a user can take advantage of based on their pricing tier. This is far from a newly fangled method of pricing one’s users but, unsurprisingly with a business model which has so far offered very little in the way of concrete return on investment (ROI), its implementation has caused some embarrassing budget implications for companies such as Uber and Meta who have bet on AI over the value of their own staff, despite learning the hard way that the former now costs more than the latter to do the same job.

Aside from free systems now costing money to access, there has been a growing backlash from the public about how flagrant employers have been about using AI to improve efficiency and reduce headcount costs. This has prompted the major western AI companies to both entice sceptics and hold on to existing user by offering significant reductions in token costs for both private users and businesses. 

It does not take an economic genius to discern that reducing pricing impacts profits. That is not a good look for investors when you have demonstrated no tangible ROI and you are still seeking ever larger amounts of investment to achieve even simple business goals such as breaking even.

The elephant in the room

Meanwhile, China is developing its own AI systems which are, by account, considerably cheaper to develop and operate, and, more fundamentally, can operate “locally”; meaning that the hardware required to run them literally fits in your pocket or sits on your desk, not in a data centre which cost billions to build and just as much to maintain. The parts the Chinese use to build their hardware can also be sourced in-country, removing the impact of tariffs and taxes from profitability calculations and preventing them from being at the mercy of foreign suppliers and supply chains.

Should the Chinese reach the AI promised land before the western hyperscalers, the US government may be forced to take protective measures and ban access to Chinese-developed AI systems and hardware in the United States in an effort to ensure that investors do not pull the plug from the assorted AI companies currently propping up 45% of the S&P 500.

With the aforementioned levels of investment in the companies from governments, sovereign wealth funds, asset managers and venture capitalists already invested with little to show beyond over-inflated stock prices, the potential success of the AI technology being developed in China could pop the AI bubble at a scale which makes the subprime mortgage crisis of 2008 and the dot-com bubble of the early 2000s look like a picnic in comparison. Not even the US government would be able to provide a bailout to prevent that level of collapse.

I do not believe there has ever been a time when so much money has been spent to achieve so little. Billions have been invested to support a self-perpetuating business model which generates no return on investment and ultimately damages the planet through data centre pollution and the pillaging of rare earth minerals, as well as disregarding the value of human purpose and agency through the threat of job replacement. And the companies behind this want trillions more to continue following the same path.

It seems as if, much like the climate crisis, an astonishing lack of corporate accountability and political responsibility has positioned those in power into successfully selling the false promise that today’s problems can always be solved tomorrow.