How much!?!?

How much!?!?

How much do the artificial intelligence (AI) hyperscalers’ businesses need to make from selling their AI tools in order to generate a decent return?

The debate surrounding the Return on Invested Capital (ROIC) for big tech’s massive AI infrastructure expansion revolves around a core question: How much end-user revenue is required to cover hardware depreciation, energy costs, and capital costs?

With hyperscaler capital expenditure (capex) on track to surpass US$600 billion- $800 billion annually, several prominent analysts, venture capital firms, and investment banks have published frameworks and scenarios to quantify the revenue needed for a reasonable return.

Here are a handful. You can decide whether the targets are likely to be met.

1. David Cahn (Sequoia Capital)

Cahn built a bottom-up model working backward from chip manufacturer revenues (predominantly Nvidia) to calculate how much end-user revenue the AI industry must generate to make its massive data centre investments profitable.

He starts with hardware spend – specifically Nvidia’s data centre chip revenue – and applies two 2x multipliers. 

The first 2x multiplier is for Data Centre Total Cost of Ownership (TCO). Specialised AI chips (GPUs) only represent about half the total cost of building an AI data centre. The other half goes toward physical infrastructure, including land, specialised cooling systems, power, and backup generators. Therefore, Cahn doubles chip spending to estimate total data centre construction and operating costs. 

The second step is another 2x multiplier for Business Margins. Cloud providers (like Microsoft, AWS, or Google) and software startups buying these Graphic Processing Units (GPUs) cannot just break even on compute costs; they need to earn roughly a 50 per cent gross margin to run a viable business.

Doubling the total infrastructure cost again reflects the actual revenue end-users must pay for AI services. 

Cahn’s approach suggests that for every US$1 spent on AI chips, the market must generate roughly US$4 in end-user revenue. If tech companies spend $150 billion on GPUs in a year, the broader economy needs to spend roughly US$600 billion on AI products and subscriptions to make those capital expenditures pay off.

At the time of the analysis, actual end-user revenue (from ChatGPT subscriptions, API usage, and cloud AI add-ons) was estimated at US$15 billion–$20 billion annually. Cahn argues that the gap won’t be bridged unless AI moves beyond ‘speculative point solutions’ into mission-critical corporate workflows.

2. Bain & Co

Modelling a world where AI compute demand grows at twice the rate of Moore’s Law, Bain projected that global AI capacity would require roughly 200 gigawatts of power, demanding US$500 billion in annual capital expenditures (capex) to cover short 3-to-5-year GPU depreciation cycles, power grid expansion, and facility operations.

Bain then applied historical cloud provider business models – where capex typically represents about 25 per cent of top-line revenue to maintain healthy profit margins – yielding a 4x multiplier.

Therefore, sustaining US$500 billion in annual capex requires US$2 trillion in yearly AI revenue across the ecosystem by 2030. This is a number larger than the current combined revenue of Apple, Microsoft, Alphabet, Meta, Amazon, and Nvidia.

Bain then modelled a hyper-optimistic scenario where 100 per cent of legacy enterprise on-premises IT budgets switch to AI cloud infrastructure, and 100 per cent of productivity savings are ploughed back into buying AI tools.

Even under these assumptions, Bain found a circa US$800 billion annual revenue shortfall. They concluded that software subscriptions alone cannot fill the gap; consumer monetisation and new AI-driven business models will have to emerge.

3. Goldman Sachs

Goldman Sachs estimates that Amazon, Microsoft, Alphabet and Meta alone will invest around US$5.3 trillion in AI infrastructure between 2025 and 2030. That figure includes data centres, GPUs, networking, power infrastructure and the enormous ecosystems needed to support them.

Assuming US$5.3 trillion of cumulative investment, an average asset life of eight years, a required return on capital of 12 per cent and a 40 per cent cash margin.

Under those assumptions, which are, of course, open to debate, the AI ecosystem would ultimately need to generate approximately US$1 trillion of annual operating cash flow. At a 40 per cent operating cash flow margin, that implies around US$2.7 trillion of incremental annual revenue, equivalent to roughly 2-3 per cent of global GDP or the entire GDP of the UK or India.

Currently, the global software industry generates around US$800 billion to US$1 trillion of annual revenue, global advertising around US$1 trillion, and global pharmaceuticals approximately US$1.7 trillion.

4. Morgan Stanley, Bernstein & LPL Financial

These models assess the return on invested capital (ROIC) for an ‘average hyperscaler’ by evaluating how many paid enterprise seats and API calls are required to hit target ROIC hurdles of 12-18 per cent.

In one scenario referred to as Enterprise Seat Penetration, specialised AI copilots and productivity tools at a price of US$30-$50 per user per month would suggest 30-40 per cent of all global enterprise knowledge workers (about one billion workers globally) must become paying, daily active users.

5. Morgan Stanley Equity Research

In Morgan Stanley Equity Research’s Token Economics & API Return Model, rather than asking whether millions of consumers will pay US$20–US$30 monthly for co-pilot subscriptions, the model evaluates data centres as digital factories that convert raw electrical power and specialised chips (GPUs) directly into billable output – measured in “tokens” (units of processed text, code, or images).

Hyperscalers (Microsoft, Amazon, Google, Meta) do not need every enterprise worker to buy a standalone AI app to generate strong returns. By embedding AI capabilities into large, high-frequency platforms – such as search engines, automated ad creation, e-commerce recommendations, and background developer workflows – the idea is the hyperscalers can continuously monetise GPU compute behind the scenes through automated, high-volume usage.

Under their assumptions, Morgan Stanley believes hyperscaler AI capital expenditures can generate attractive ROICs of 25-50 per cent, which of course is well above typical corporate cost-of-capital thresholds.

In Path A (GPU Rental / Infrastructure as a Service), where cloud providers lease raw hardware – such as Nvidia GB300 chips – at $8.50/hour with 75 per cent facility utilisation, incremental ROIC reaches approximately 31 per cent with Earnings Before Interest and Taxes (EBIT) margins near 67 per cent.

Perhaps surprisingly, Path B (Proprietary Model APIs / Inference) is even more lucrative. By turning hardware into raw token throughput (2,750 tokens/sec/GPU at US$1.75 per million tokens with 65 per cent capacity dedicated to inference), hyperscalers capture value from both the hardware and the AI intelligence itself, driving EBIT margins above 70 per cent and ROIC into the 40–50 per cent range.

But while Morgan Stanley might conclude building proprietary AI infrastructure allows cloud giants to extract significantly higher economic rent than merely acting as digital landlords renting raw server rack space, the returns are based on assumptions about unit pricing and capacity utilisation.

Because hardware depreciation and data centre power account for billions in fixed overhead per gigawatt, a drop in token pricing below US$1.00 per million (caused by open-source competition or aggressive price wars) or a decline in GPU utilisation below 70 per cent would cause operating margins and ROIC to evaporate rapidly.

Their bull case relies heavily on ‘Jevons Paradox’, which we wrote about on the blog here ‘Why Cheaper AI doesn’t mean easy profits’ – the economic principle that as token costs fall and throughput speeds increase, total market demand will explode exponentially through continuous AI agents, long-context analysis, and automated enterprise workflows. These are still in the realm of a few enterprises and hobbyists.

Investors will have to monitor whether hyperscalers maintain pricing power and high utilisation. If token prices crash faster than efficiency grows, today’s multi-trillion-dollar data centre buildouts will quickly suffer from margin compression.

6. Plato

Plato asks; how much revenue must AI infrastructure ultimately generate to justify the investment?

Plato adopts the same assumptions as Goldman Sachs in Example 3., and reaches the same conclusion; US$5.3 trillion of cumulative investment, an average asset life of eight years, a 12 per cent annual return on capital and a 40 per cent cash margin.

“Under those assumptions, which are, of course, open to debate, the AI ecosystem would ultimately need to generate approximately US$1 trillion of annual operating cash flow. At a 40 per cent operating cash flow margin, that implies around US$2.7 trillion of incremental annual revenue, equivalent to roughly 2–3 per cent of global GDP or the entire GDP of the UK or India.

“To put that figure in perspective, the global software industry generates around US$800 billion to US$1 trillion of annual revenue, global advertising around US$1 trillion, and global pharmaceuticals approximately US$1.7 trillion.”

Plato concludes thus: “The question is no longer whether AI will change the world. It’s whether the world’s businesses will ultimately spend an extra ~US$2.5 trillion every year using it.”

7. Jim Covello (Head of Global Equity Research, Goldman Sachs)

Covello evaluated AI capex by comparing it to historical tech transformations (such as the shift to cloud, internet, or PCs). Historically, new technology was immediately cheaper than what it replaced.

With cumulative global AI spend projected to reach $1 trillion–$1.8 trillion in the near term, Covello calculated that for companies to realise an acceptable ROIC, AI tools cannot merely automate simple, low-wage tasks or rearrange existing software budgets. Instead, AI has to solve complex, multi-trillion-dollar problems or directly replace massive corporate labour payrolls.

Covello noted that if executing tasks via AI remains significantly more expensive than traditional manual or software methods, enterprise demand will plateau once initial experimentation budgets dry up.

8. Tom O’Malley and Ross Sandler (Barclays Equity Research)

Barclays looked at the financial capacity of hyperscalers (Microsoft, Alphabet, Meta, Amazon, Oracle) to sustain current capex trends by analysing their ratio of Capital Expenditure to Operating Cash Flow (Capex / OCF).

As capex approaches 97 per cent of operating cash flow, and hyperscaler AI Capex reaches approximately US$1.16 trillion annually by 2028, hyperscalers can no longer fund infrastructure purely from organic operations without sacrificing dividend programs, share buybacks, or debt ratings.

Barclays suggests returns will only normalise under one of two conditions:

1. Recursive Self-Improvement (RSI): AI models reach a threshold at which they begin to improve and optimise their own code, drastically reducing the compute required for future training iterations.
2. Inference Offloading: Standardised training hardware from earlier years transitions effectively to handle high-margin inference workloads, deferring additional chip capex.

Hyperscalers will have to earn their return not just through consumer subscriptions, but as the foundational utility layer renting compute to labs whose annual Application Programming Interface (API) and subscription revenues must scale into the hundreds of billions.

MORE BY RogerINVEST WITH MONTGOMERY

Roger Montgomery is the Founder and Chairman of Montgomery Investment Management. Roger has over three decades of experience in funds management and related activities, including equities analysis, equity and derivatives strategy, trading and stockbroking. Prior to establishing Montgomery, Roger held positions at Ord Minnett Jardine Fleming, BT (Australia) Limited and Merrill Lynch.

He is also author of best-selling investment guide-book for the stock market, Value.able – how to value the best stocks and buy them for less than they are worth.

Roger appears regularly on television and radio, and in the press, including ABC radio and TV, The Australian and Ausbiz. View upcoming media appearances. 

This post was contributed by a representative of Montgomery Investment Management Pty Limited (AFSL No. 354564). The principal purpose of this post is to provide factual information and not provide financial product advice. Additionally, the information provided is not intended to provide any recommendation or opinion about any financial product. Any commentary and statements of opinion however may contain general advice only that is prepared without taking into account your personal objectives, financial circumstances or needs. Because of this, before acting on any of the information provided, you should always consider its appropriateness in light of your personal objectives, financial circumstances and needs and should consider seeking independent advice from a financial advisor if necessary before making any decisions. This post specifically excludes personal advice.