The AI boom is becoming a voracious claimant on the global pool of capital. Companies developing AI are seeking enormous funding for multi-year investments, with planned capital expenditure running into trillions of dollars. The first source of funding was the healthy free cash flow generated by hyperscalers' pre-AI businesses. This was followed by substantial debt raising through innovative structures backed by GPUs, and increasingly equity markets, as Alphabet, Meta, SpaceX, SK Hynix and Anthropic seek fresh capital.

This competition for capital is beginning to affect borrowing costs elsewhere, notably for governments. Bond yields have risen in several countries, with some of the pressure attributable to AI's enormous financing requirements. As governments refinance debt at higher rates, their fiscal room could shrink. More expensive capital may eventually force selective spending cuts, including reductions in public welfare programmes such as Medicaid and Food Stamps in the US.

Meanwhile, the progression from machine learning to generative AI has demonstrated how greater computing power, combined with increasingly sophisticated models and larger parameter sets, can improve capabilities. The results are remarkable. With frontier AI labs openly pursuing artificial general intelligence (AGI), however, it is difficult to see where the race for more computing power, memory and electricity will end. More fundamentally, how will we know when AGI has actually been achieved?

The response to every constraint appears to be another, more capital-intensive solution. Rising energy requirements invite investment in nuclear power. High land and rental costs prompt proposals to move data centres into space. Computing shortages drive successive generations of chips, making earlier investments obsolete. When the objective is to build ever-smarter models, there is little incentive for competing labs to stop investing. But this arms race requires ever more capital, potentially destabilising other claims on it.

There are two ways in which these effects could be moderated. The first is the discipline imposed by capital markets. Markets may tolerate excesses and irrational exuberance for a while, but eventually demand their pound of flesh.

Capital Discipline
In credit markets, the cost of borrowing should ultimately reflect the risk associated with the debt. The question is whether the expected returns from AI investments can justify the financing costs. Some bonds issued by special purpose vehicles associated with neo-cloud company CoreWeave already yield around 15%, as the debt maturity extends beyond the economic life of the GPUs being financed. That is a clear indication of the risks involved.

For equities, business risk is reflected in the cost of capital used to value a company. This determines the discount applied to its future free cash flows to arrive at their present value. The greater the perceived risk and capital requirements, the higher the cost of capital and the lower the valuation.

Shareholders have little reason to welcome a decline in the value of their holdings. Ideally, they should resist business risks expanding to the point where they materially erode value. Anthropic, the maker of the Claude family of models, has already seen its estimated valuation fall from around US$2 trillion to closer to US$1 trillion, as investors reassess the risks involved.

But before markets can sensibly adjudicate, a more fundamental question needs answering: do capital allocators understand the risk-return trade-off associated with AI? Even a highly risky business can be worth investing in if its expected returns justify that risk. Equally, the difficulty of quantifying AI's benefits today does not mean those benefits do not exist.

If AI investments generate annual returns of 25%, CoreWeave's bonds could be cheap and Anthropic at its lower valuation a bargain. If returns are unlikely to exceed 10%, both could remain expensive despite their reduced prices.

As this author has previously argued in an earlier article; measuring returns from AI remains a challenge. AI is expected to boost productivity, yet US productivity growth in the years since the arrival of ChatGPT, Claude and other models has not shown a clear departure from the preceding trend. None of the large hyperscalers separately discloses AI revenue, while many organisations remain unable to quantify the productivity gains from deploying these tools. Frontier model companies and their investors, meanwhile, maintain that the returns from AGI could be priceless.

Until a credible way of measuring AI's economic benefits emerges, significant capital is likely to keep flowing into the sector, with the associated pressure on government borrowing and other uses of capital continuing.

The consequences extend beyond financial markets. Anecdotal evidence suggests that AI displacement is already affecting employment, particularly at the entry level. If this trend gathers pace, it could mean lower consumption, weaker GDP and reduced tax revenues, alongside higher government spending on unemployment support.

Governments would then face a double whammy: higher borrowing costs as AI competes for capital, and weaker revenues as automation displaces workers. If AI is indeed contributing to job losses, some of the profits accruing to AI companies effectively represent a transfer of wealth from labour and governments to capital.

A second way to moderate these effects is to tax the economic benefits generated by AI, thereby replenishing government revenues and creating fiscal room to address the disruption.

But this raises the same measurement problem that confronts capital markets. How can incremental AI benefits be quantified to determine tax liability? Should the tax apply only to companies such as Anthropic and OpenAI? That would be difficult to justify when it is the corporate users of their models who presumably capture productivity gains and make the associated employment decisions. The question is even more complicated for countries such as India, where the employment effects of AI are being felt despite the absence of prominent domestic AI companies. India is increasingly being viewed as an anti-AI trade.

There are no easy answers to these questions. Market mechanisms will impose some discipline on capital allocation, as the CoreWeave and Anthropic examples suggest. That should limit the extent to which AI investment destabilises government borrowing costs, although it cannot eliminate the risk.

Governments, meanwhile, must begin developing credible ways to quantify the benefits businesses derive from AI and consider how some of those gains can be shared with society. Otherwise, AI-related profits could become little more than a redistribution of wealth from labour and the public exchequer to technology companies.

Measurement may be difficult, but that is no reason to avoid the exercise. If AI is to deliver sustainable economic progress, its gains cannot accrue exclusively to those who own the technology. Making AI equitable will require both market discipline and a framework that ensures society shares in its rewards.