11.06.2026

Kismet, robot head made in 1990s: can simulate emotions

Just one big trade

Hype around AI remains so big that essentially all private investments in the US are now in tech hardware and software. But if AI is to succeed for capital, writes Michael Roberts, it will be at the expense of the working class

Goldman Sachs, the mega investment bank, reckons that artificial intelligence is just “one big trade on the US economy”. And the AI investment bubble is getting even larger. The AI model maker, Anthropic, has just announced it was issuing shares to potential investors in (what is called in stock market jargon) an initial public offering (IPO). Anthropic was following the planned IPO of Elon Musk’s SpaceX - a humongous $1.8 trillion. This would value SpaceX in the market at 92 times its annual revenue!

Alphabet, Google’s parent, also plans to raise $85 billion in equity funding - its first stock offering in more than two decades. Together, these three giant IPOs could command a combined valuation of around $4 trillion: that is one-third of all the value of US IPOs since 1980 (inflation-adjusted)! And yet SpaceX, OpenAI and Anthropic are all currently loss-making, while the commercial potential of AI models (in the case of SpaceX going to Mars) remains unknown.

AI is one big trade for the US stock market investors and one big bet on the US economy. That is because the amount of capital investment being made by the companies called the ‘hyperscalers’ into AI models, data centres and other AI equipment is staggering. As a share of US gross domestic product, it is now set to far surpass the 19th-century railroad build-out.

Back in December 1996, then Federal Reserve chair Alan Greenspan characterised the boom in technology, media and telecom stocks as showing signs of “irrational exuberance”. Almost 30 years later, we can say the same with bells on about the AI boom. This is already much larger than the dot.com internet investment of the late 1990s ever was. In 2025, US businesses invested almost $1.5 trillion in information technology (IT) equipment and software. At the peak of the dot.com bubble, it was $466 billion (or $829 billion when adjusted for inflation). The hyperscalers - Microsoft, Alphabet, Amazon, Meta and Oracle - plan to invest hundreds of billions in the next five years in data centres to provide the computing power to run these AI models. Capital investments are expected to rise by 20% a year - a growth rate never seen before in this industry.

US GDP growth is now driven almost exclusively by rising tech spending. If this starts to drop, the US economy will enter recession very quickly, even if tech investments decline only by a little bit - say, 4‑6%, as happened after much smaller tech booms in the 1960s and during the 2009 recession.

As I recently showed, US corporate profits have risen significantly.¹ But, according to Brian Green in a recent post,² around 80% of the increase in US non-financial corporate profits came from Nvidia and hyperscalers. The stock market is increasingly concentrated in a handful of A‑linked stocks, which now account for roughly 40% of the Standard and Poor (S&P) 500’s market capitalisation, according to Bank of America data. Headline profitability is being flattered by a small slice of the economy earning extraordinary returns from the scramble to build AI capacity. The risk, then, is that the economy, the profit cycle and the stock market “are all leaning on the same narrow pillar. If the expected returns on AI infrastructure and platforms are questioned, the fallout may not stop at a few richly valued technology stocks”.³

As I have pointed out previously,⁴ up to now the massive investment in AI has been mostly funded by the profits already being made by the hyperscalers. But, given the impossibility of finding enough additional revenues to self-finance their capital expenditure (capex) plans, hyperscalers and their hardware providers are increasingly using external financing to fund them.

Circular financing

The first game is ‘circular financing’: ie, cross-investments between Microsoft, OpenAI and others. In essence, a cash-rich hyperscaler like Microsoft buys hardware from Nvidia, AMD and other suppliers. Nvidia then uses that revenue to buy a multi-billion-dollar stake in OpenAI. OpenAI then uses this cash to ‘secure compute’ in Microsoft data centres. Microsoft itself also invests in OpenAI and enters into a mutual revenue share, where some of OpenAI’s revenues flow to Microsoft and vice versa, as the two companies use each other’s products. Assuming that Microsoft spends $100 billion to order hardware for data centres, Nvidia, AMD and other suppliers can recognise that as $100 billion in revenue. They then use that cash to invest in OpenAI (for example), which then uses this money to book data centre capacity with Microsoft. Microsoft recognises this OpenAI investment as revenue, thus effectively turning its $100 billion expense into billions of revenue!

Even this is no longer enough, and increasingly, hyperscalers have started to resort to borrowing to raise the cash for investment. The US tech giants are issuing debt all over the world. Google/Alphabet is leading the charge.

So first they invested with their own funds; then in each other; then they borrowed from the banks and so-called private credit funds; and now they are putting the risk of success or failure on investors in the stock market. If all this investment fails to deliver the expected returns, it will hit the financial sector and the wider economy big time.⁵

But don’t worry: say revenues for the AI companies and hyperscalers are expected to grow 15% annually. If we make the heroic assumption that there are no costs, then this additional revenue is the profit these companies are expected to make from their additional investments in AI data centres. Yet, even under these extremely optimistic assumptions, the implied return on investment is highly negative for all except Amazon.

If the hyperscalers need to generate, say, a 10% return on investment, they would have to find an additional $2-5 trillion in revenue a year. That is a tall order for a group of companies that currently generates revenues of just $1.5 trillion per year. The other option is that the planned investment in data centres, computer chips and other areas never materialises - maybe as equity investors turn more cautious on the sector, or if debt funding for data centres becomes harder to get. A JP Morgan analysis found that more than 60% of data centre capacity planned for completion in 2027 is not yet under construction, and another 7% is delayed.⁶ What will happen if these companies announce cutbacks on some of their investment plans?

Will the AI heroes, OpenAI and Anthropic, deliver the returns that the hyperscalers and their investors hope and expect? Corporate CEOs are optimistic. Over the last three years, since OpenAI launched ChatGPT, they claim that cumulative productivity gains have been in the order of 0.3% to 1% per year. For the next three years, they estimate productivity gains will accelerate to 1.4%, with executives in the US and UK far more optimistic than in Germany and Australia.

These productivity gains, they reckon, will be achieved by shedding labour. Business leaders expect headcount in their firms to drop by about 0.7% in the next three years - again with executives in the US and the UK expecting far more pronounced drops in employment than executives in Germany and Australia. In the last three years, the same executives saw no employment impact from AI. So this is all expectation. Moreover, the Business Trends and Outlook Survey of the US Census Bureau shows that companies with 50 employees or more show no more growth in AI use since the second quarter of 2025. Businesses are still unsure how to use AI effectively and are increasingly worried about the drawbacks when they do.

Those drawbacks include ‘hallucinations’ (ie, fictions made up by the AI model), which are inherent in large language models (LLMs). One study found that for a training set of 32,000 words, the average hallucination rate in LLMs was 6.8%. When that was expanded to 128,000 words, the average rose to 10%. That is a lot of correction time and monitoring for human workers!

Another problem is that, because LLMs are designed to be ‘good at everything’, they are not very good at any one thing, compared to specialised apps. One report on using AI in software development found an explosive impact at the start, with coders creating or editing almost 300% more files - but that boost was halved to 150% by the time companies got the number of pieces of work submitted for review, and that in turn shrunk five-fold to a roughly 30% uplift at the point of full software release.

Moreover, when researchers looked at whether AI-assisted increases in software production have led to increased usage by clients, they found little evidence. The marked increase in mobile app releases over the past year has not been accompanied by any increase in downloads - most of the new apps fail to capture even a modest audience.

Meanwhile, OpenAI has burnt through some $6 billion, rising to $17 billion in 2026. By 2028, inference (training) costs alone are expected to grow to $121 billion and losses are projected to be $85 billion. Anthropic’s cash burn is much smaller, but was still $3 billion in 2025. Unless the companies that build LLMs can find large amounts of new revenue in the next couple of years, the losses will increase exponentially, especially given the fact that current price charged per ‘token’ is not the true cost of compute. If AI companies were to charge the actual cost price per token, the losses may decline, but demand for LLMs would decline even more.

Despite this, the hype around AI remains so big that essentially all private investments in the US are now in tech hardware and software. Over the last three years, the average annual growth in IT equipment investments has been 11% and 8% in software. Meanwhile, investments in all other parts of the US economy put together have declined by 1.6% per year.

Two in one

The US economy today really is two economies in one. There is the tech economy and then there is everything else. Over the last four quarters to the end of March 2026, 93% of US GDP growth is due to tech investment alone (although much of the purchases are imports and not produced domestically).

This is a bubble waiting to burst. In the aftermath of the dot.com technology, media and telecom (TMT) bubble, private fixed investment dropped more than 12.7% between 2000 and the end of 2002, as a recession took hold in the US. In the initial year after the TMT bubble burst, tech investments dropped 12%, while fixed investments in general dropped 7.6%.

Gita Gopinath, former chief economist at the International Monetary Fund, has calculated that an AI stock market crash equivalent to that which ended the dot.com boom, would erase some $20 trillion in American household wealth and another $15 trillion abroad - enough to strangle consumer spending and induce a global recession.⁷ This is also the view of the IMF. The IMF fears that AI firms could fail to deliver earnings commensurate with their lofty valuations. The collapse of previous investment booms knocked about 1% on average off US real GDP growth. Even a moderate correction in AI stock valuations would reduce global growth by 0.4%: “Combined with lower-than-expected total factor productivity gains, and a more significant correction in equity markets, global output losses could increase further, concentrated in tech‑heavy regions such as the United States and Asia.”⁸ Another study has found that even a very mild drop in tech investment of just 3% would cut US real GDP growth by 1% - or half the current rate (the impact would be greater in Europe).⁹

None of this is to conclude that AI will not at some point deliver with higher profitability for the companies involved and higher productivity for the US economy as a whole. But that will not happen before there is a bursting of the investment bubble - as there was in the railway mania of the 1870s¹⁰ and in dot.com bubble of the late 1990s. As other studies have shown, it will take a decade or more for AI to become a generalised technology that delivers.

For working people, AI poses a different problem. For capital and the mega media companies, the aim is to make AI a profitable technology, but that can only be done by shedding labour and by stopping any attempt to regulate its applications and use. If AI is to succeed for capital, it will only be at the expense of most working people and their families.

Michael Roberts blogs at thenextrecession.wordpress.com