Artificial intelligence (AI) investment is getting bigger, more complex and harder to judge from headline numbers alone. Nvidia is helping to establish financing platforms targeting more than $500bn of third-party capital, Microsoft has signed $329bn of leases that are yet to begin, and cheaper AI models are raising questions about future pricing.
These are genuine risks, but the underlying data paints a more balanced picture than the biggest figures suggest.
This article isn’t personal advice. If you’re not sure an investment is right for you, seek advice. Investments and any income from them will rise and fall in value, so you could get back less than you invest. Ratios also shouldn’t be looked at on their own.
Investing in an individual company isn’t right for everyone because if that company fails, you could lose your whole investment. If you cannot afford this, investing in a single company might not be right for you. You should make sure you understand the companies you’re investing in and their specific risks. You should also make sure any shares you own are part of a diversified portfolio.
Is AI spending becoming too reliant on leases?
One concern is that reported capital spending no longer captures the full scale of the AI buildout. Hyperscalers are increasingly using leases and long-term agreements to secure data centres, power and equipment. These arrangements provide flexibility but make the timing and true cost of investment harder to judge.
Microsoft’s $329bn of signed leases that have not yet commenced looks alarming at first glance, but the detail is less dramatic. The facilities are due to come online between financial years 2027 and 2033, with contracts lasting from one to 20 years.
Based on reasonable assumptions, annual lease payments could rise from $13.9bn to around $24.9bn-$28.8bn by financial year 2031. That is significant but spread over time it implies annual growth of roughly 12-16%, broadly in line with historical trends. The total commitment is huge, but it’s not a near-term bill.
Nor is the infrastructure solely for AI.
It can support Microsoft’s wider cloud business, while purchase commitments beyond the next financial year remain relatively limited. That leaves scope to adjust how quickly new facilities are filled with expensive equipment.
Leasing does not make the cost disappear, and this level of investment deserves close attention. But long-dated, reusable infrastructure should not be confused with immediate, inflexible spending. Microsoft’s numbers warrant caution, but do not suggest the reckless overbuilding of the headline figure.
Is Nvidia becoming the bank of AI?
Nvidia’s new financing initiative brings a second concern into focus: circular financing.
The company is working with six major financial institutions to establish independent platforms with the ambition of mobilising more than $500bn for AI infrastructure.
That does not mean Nvidia or its partners have already committed $500bn. It is a potential pool of mostly third-party capital, with each project assessed separately and no guarantee the full amount will be deployed.
Nvidia may provide residual-value support of up to 25% on some projects. This is not an upfront commitment to fund a quarter of every deal, but a potential backstop if the underlying assets prove less valuable than expected.
Support will be considered case by case and is intended to complement, rather than replace, independent underwriting.
The attraction is clear.
More capital should support demand for Nvidia’s systems, while independent investors must test the commercial case before putting their own money at risk. Nvidia could also earn usage-linked or revenue-sharing income from some projects, although both the scale and structure remain uncertain.
The circularity concern still matters, but this agreement helps us get more comfortable, not less. These are independent projects, each with its own investor pool, due diligence, and financing. This is not Nvidia throwing cash at a customer so they can go out and buy its chips. Instead, it’s using its platform to position itself as the orchestrator of this buildout.
Are cheaper AI models bad for hyperscalers?
The rise of cheaper, open-source models raise a third concern. If companies can’t charge as much for each AI token, it might seem logical that returns on their expensive data centres must fall.
But that misses the other half of the equation.
Revenue depends on both the price per token and the number of tokens a data centre can process. More efficient models may pressure pricing, but they could drive better output from the data centre itself and make AI affordable for a wider range of tasks.
The offsetting nature of these two inputs works to a cloud giant's advantage, as we model below. Case 1 lowers the token price (Price per Million Tokens) and raises the output (Tokens/Second/GPU).
Revenue: | Reference | Case 1 |
|---|---|---|
AI Chips in 1GW Data Centre | 410,256 | 410,256 |
(X) % of Capacity Used for Inference | 65% | 65% |
(=) Revenue Generating Chips | 266,666 | 266,666 |
(X) Tokens/Second/GPU | 2,750 | 3,025 |
(X) Seconds per Year (k) | 31,536 | 31,536 |
(X) Utilization | 75% | 75% |
(=) Tokens per Year (bn) | 17,345 | 19,079 |
(X) Price per Million Tokens | $1.50 | $1.35 |
(=) Revenue/GW | $26 | $26 |
This is just an illustration, but all else being equal, if token prices fall by 10% and that leads to an equal improvement in output, revenue would be broadly flat. So, better efficiency would offset the pricing pressure.
The real-world outcome could be better still. Lower prices can unlock more usage, while more intelligent models may command a premium by delivering better results from each token.
Hyperscalers have other levers too.
Newer chips can deliver more for the same energy use, software can process workloads more efficiently, and simpler tasks can be routed to smaller, cheaper models.
The most bullish lens would suggest that these advances could increase a data centre's value over time.
But the main takeaway is that cheaper models are not automatically bad news for those who own their own infrastructure.
What does it mean for investors?
The next stage of the AI trade will be harder to assess.
Conventional capital spending captures less of the overall buildout, while financing risk is spreading across more participants.
That calls for scrutiny, not a rush to the exit. Investors need to look beyond the largest headline numbers and focus on when commitments become cash costs, who carries the risk, and whether efficiency gains can keep pace with lower prices.
The winners will not simply be those willing to spend the most. They will be the companies that use scarce power most effectively, grow output faster than costs and retain enough flexibility to respond as demand develops.
On today’s evidence, we think the leading tech companies remain well placed as AI infrastructure becomes a competitive advantage.
The author holds shares in Nvidia and Microsoft.
This article is original Hargreaves Lansdown content, published by Hargreaves Lansdown. It was correct as at the date of publication, and our views may have changed since then. Unless otherwise stated estimates, including prospective yields, are a consensus of analyst forecasts provided by LSEG. These estimates are not a reliable indicator of future performance. Past performance is not a guide to the future. Investments rise and fall in value so investors could make a loss. Yields are variable and not guaranteed.
This article is not advice or a recommendation to buy, sell or hold any investment. No view is given on the present or future value or price of any investment, and investors should form their own view on any proposed investment.