Who Will Pay the $2.6 Trillion AI Infrastructure Bill?

We are at the start of the largest infrastructure build-out in human history. As AI investment increasingly contributes to both GDP growth and market returns, concerns about overbuilding and irrational exuberance are growing. Recent analyses using different methodologies flash the same warnings: to remain economically viable, the AI ecosystem will require annual revenues measured in the trillions of dollars. According to our model, roughly $2.6 trillion of end-market spending will be needed per year by 2030. (See the exhibit.) But we believe that this is feasible.

$2.6 trillion is an enormous number. But it’s important to put it in context. Industry forecasters project that enterprise customers will spend more than $7 trillion on IT in 2030, excluding AI. Consumer spending on AI is growing, and AI-assisted advertising and e-commerce are in their infancy, as are other nascent technologies such as robotics and physical AI. Open-weight models, software giants and startups engaged in building AI apps, agentic commerce, and other new use cases will all contribute to AI revenues. We believe that demand for intelligence is highly elastic: as token prices fall, usage should grow fast enough to raise total spending. And unlike past infrastructure booms, this one is funded largely from hyperscalers’ cash flow rather than debt, giving demand time to catch up.

Although there are broader debates about AI, our analysis focuses purely on the financial viability of the infrastructure build-out. This article follows the economic chain, building from AI data center costs (including their cost of capital), to the AI labs’ costs (including a hypothetical operating margin), and then distributes the costs among potential payers, and explains why—despite the challenges—we’re optimistic.

Section 1 · Scale of the Build-Out

An Unprecedented Build-Out

The five hyperscalers driving the bulk of the build-out (Amazon, Alphabet, Meta, Microsoft, and Oracle) are expecting to commit approximately $4.6 trillion in capital expenditure for AI between 2022 and 2030—nearly $3.9 trillion of it after 2025. The $4.6 trillion represents 60% to 70% of AI capital expenditure.

In terms of scale, only the development of the US railroad system and the early internet are comparable. The industry structure is much the same in all three cases: heavy upfront capital, a risk of supply scaling faster than demand, and early investors taking significant operating losses even as the infrastructure proves valuable to the wider economy. What differs is the financial consequence. In those earlier cycles, losses came from low near-term demand elasticity combined with heavy debt financing, which created untenable economics while demand caught up. Hyperscalers have funded the current build largely from internally generated cash, and we believe that demand for intelligence is highly elastic. Together, these suggest that this build-out may be more resilient than prior booms.

Hyperscaler Capital Expenditure How the AI Build-Out Compares Why Prior Booms Destroyed Equity Hyperscaler Capex and Gross Debt Issuance

• Hyperscaler Capital Expenditure
• How the AI Build-Out Compares
• Why Prior Booms Destroyed Equity
• Hyperscaler Capex and Gross Debt Issuance

Section 2 · Data Center Economics

Economics of a Data Center Build-Out

In our model, a gigawatt (GW) measures critical IT load: the power delivered to servers and racks (excluding cooling, power conversion losses, and other facility overhead). We use this as the standard capacity metric for AI data centers—in place of direct compute measures like floating-point operations per second (FLOPS)—because expenses (both capital expenditure and operating expenses) scale roughly linearly with power draw, whereas the cost per unit of compute goes down over time. We estimate that installed AI capacity will grow from 17.8 GW in 2025 to 120.3 GW in 2030, which is 14% more than Brazil’s highest peak energy demand and only slightly less than Japan’s.

The cost of data centers is spread across the data center shell, power infrastructure, land, and—the most significant cost—the powerful computer chips that process AI requests. Our model assumes an average five-year lifespan for these chips, in line with hyperscaler accounting practices.

Once a gigawatt of compute capacity is online in 2026, our model estimates that a data center will need to earn $15 billion per gigawatt to cover its costs, including both the cost of debt and an acceptable ROI for equity investors.

AI Capacity Build-Out Upfront Capital Needed per GW Clearing the Cost of Capital on a Single GW Total Annual Data Center Requirement

• AI Capacity Build-Out
• Upfront Capital Needed per GW
• Clearing the Cost of Capital on a Single GW
• Total Annual Data Center Requirement

Section 3 · AI Lab Economics

Economics of the Frontier Labs

Frontier labs pay a substantial share of the hyperscalers’ bills, so their economics matter to the whole value chain. The key distinction is between training and inference. An AI model’s training investment is paid back when gross profit from subsequent inference recovers its training cost. A lab can clear that hurdle and still report negative free cash flow, because the next training cycle is already under way.

More compute may appear to make the ecosystem more fragile, since it adds fixed costs that must be paid for. But if the additional compute comes from increasing the ratio of inference to training, it makes the system more resilient. This is because there is more compute going toward earning revenue than toward developing the product. A significant proportion of inference does not require frontier hardware and can run on older or cheaper chips, extending their useful life and lowering the cost of goods sold (COGS). Labs can then reduce prices and accept lower gross margins while still recovering their training costs within a fixed payback period. A broader inference market also spans a wider range of workloads from frontier reasoning to coding, advertising optimization, consumer assistants, commerce agents, embeddings, and batch processing.

The key question is whether demand for intelligence is elastic enough to cover the cost of previous training runs. Here AI differs from railroads and the early internet: the distribution infrastructure is already in place through networks, Wi-Fi, mobile devices, and cloud access, giving the roughly 5 billion internet users and nearly every global business the ability to consume more intelligence.

Training vs. Inference Frontier Lab Cost Mix Illustrating Payback of a Training Run

• Training vs. Inference
• Frontier Lab Cost Mix
• Illustrating Payback of a Training Run

Who Will Pay the Bill?

Although we have modeled the required AI revenue to be $2.3 trillion in 2030, our total system or end-market economic requirement in 2030 is $2.6 trillion. That’s because AI providers are businesses and need a reasonable profit margin to remain sustainable. There is no reliable way to forecast the steady-state operating margins of AI providers, so our analysis uses a 10% margin on the portion paid by external customers.

This is intended as an illustrative assumption rather than a prediction. Although $2.6 trillion is a massive sum, there are many routes to generating it. The path to hyperscaler revenue flows through enterprise, consumer, internal use, and other nascent innovations—with multiple commercialization options inside both the enterprise and consumer segments. If demand at this scale doesn’t materialize, we expect to see price wars and economic losses at both the model and the data center layer.

Section 5 · Enterprises

Can Enterprises Pay Their Share of the Bill?

In our model, enterprises are responsible for $1.8 trillion of the $2.6 trillion total bill. The International Data Corporation (IDC) forecasts that companies will spend $7.3 trillion on IT in 2030, excluding China and AI. Of this amount, $2.2 trillion is for non-AI software and $1.6 trillion is for traditional technology services. Given the current trajectory of enterprise IT spending, we expect a significant portion of that same budget to be reallocated to AI-infused software and services as these products evolve, which means that innovating with AI will be crucial for traditional software and tech services providers if they are to maintain growth. Therefore, we expect that enterprise AI can be funded through a mix of net-new budget and reallocation from existing technology categories, rising from 5% of enterprise tech spending today to 25% in 2030.

In our illustrative example, we see $857 billion reallocated to AI. The remaining $937 billion of the $1.8 trillion requirement would need to be added to IT budgets. Although significant, that is just 13% of the existing IDC IT spending forecast.

The spending per worker is also large but not implausible. We estimate annual AI spending at about $21,200 for an onshore tech worker at a tech company—the heaviest users today—falling to about $900 for an offshore knowledge worker at a nontech company.

IT Spending and Enterprise AI Requirements Funding AI Through IT Budgets Token Allocation Multipliers Worker Personas

• IT Spending and Enterprise AI Requirements
• Funding AI Through IT Budgets
• Token Allocation Multipliers
• Worker Personas

Section 6 · Consumers

Can Consumer Markets Pay Their Share of the Bill?

The modeled end-market expenditure reaches approximately $366 billion in 2030. Historically, consumers have been reluctant to pay large sums for software. But the consumer bill can also be paid via two far larger markets: digital advertising and e-commerce, where tech companies have made fortunes sitting between consumers and their purchases. Our illustrative scenario assumes that AI companies will play an increasingly critical role in both.

Digital advertising can pay for AI through advertisements inside consumer AI apps and AI-driven on existing digital platforms. This doesn’t mean diverting existing advertising budgets. If AI improves advertising personalization, relevance, click-through, and conversion, the same ad inventory will become more valuable, and the overall pool can grow faster than current forecasts predict.

E-commerce can pay for AI through the fees generated by agentic-commerce intermediaries between the consumer and retailers, brands, or service providers. As consumers delegate research, shopping, and ordering to personal agents, the AI provider can take a referral fee or commission. Many investors see this as the largest consumer opportunity. If the interface shifts from apps and websites to agents, existing intermediaries may have to rethink their model or cede part of their fees. This outcome is highly uncertain, as it’s unclear how far consumers will shift their spending to agents and whether AI companies can win a material share of fees from incumbents. For example, in September 2026, Amazon blocked Meta’s Muse shopping agent from its e-commerce sites. In our example, however, the share of the total AI build-out paid in this way is small relative to global e-commerce value.

Direct purchases make up an even smaller share. But a consumer AI app market of approximately $37 billion would be in line with existing paid-app categories, and account for roughly 1.4% of the total AI system revenue requirement. Underperformance would not meaningfully change the overall economics.

The Consumer Funding Split Advertising-Funded AI E-Commerce-Funded AI Direct Consumer Market Benchmark

• The Consumer Funding Split
• Advertising-Funded AI
• E-Commerce-Funded AI
• Direct Consumer Market Benchmark

This analysis is not an AI revenue forecast. It does suggest that existing enterprise and consumer markets contain spending pools large enough to support the AI value chain, particularly if AI providers displace major digital incumbents. The nine indicators below show what needs to remain true to support this analysis, and what could change the picture.

The infrastructure is soundly financed and productively used:

1. Funding stays grounded in cash flow. The build-out is led by hyperscalers with large operating cash flows and balance sheets and by frontier labs that are moving toward public markets that will test their revenue streams and margins. A sustained shift toward debt and structured financing—or toward repeated IPO delays and ever-larger private rounds—would reduce our confidence in business resilience.

2. Deployed AI chips remain highly utilized. There is little evidence today of compute going unused, but growing idle capacity would be an early warning.

3. Chips keep a reasonable useful life. Our model assumes a useful chip life of five years, and broadening inference workloads let older chips cascade into less demanding work. If chips wear out or become obsolete faster, more revenue would be needed to fund replacement.

The model layer turns compute into profit:

4. Compute shifts from training to monetized inference. A rising inference share makes the model layer more profitable and resilient. Prolonged heavy training would add costs.

5. Demand for intelligence remains highly elastic. If cost is the primary constraint on usage, labs can cut prices and margins to unlock volume while still covering their costs.

6. Multiple channels draw on the same compute. Frontier-model APIs, open-weight deployments, and applied AI products all fund the same infrastructure, so the costs aren’t tied to the performance of a single product.

Customers can pay the bill:

7. The enterprise bill remains manageable. It can be absorbed through IT budget growth and reallocation, and it is modest relative to compensation at the per-worker level. A sustained slowdown in IT budgets would put it at risk.

8. Consumer AI draws on markets larger than subscriptions. Direct purchases, advertising, and e-commerce provide three distinct funding pools.

9. New use cases keep emerging. Applied AI continues to broaden beyond coding and knowledge work, with agentic commerce and physical AI as the major additional opportunities to watch.

To remain economically viable, the AI infrastructure rollout will require enormous revenues. But our analysis suggests that generating these sums is feasible, given the potential rewards and the deep pockets of the groups that must foot the bill. As a result, leaders can be more confident in the stability of the AI ecosystem and can monitor key indicators for advance warning of any impending changes.