AI's $3 trillion question: how to budget cloud and GPU spend against a $1.5T buildout

Summary. AI infrastructure spending for 2026 is projected at about $1.5 trillion, and by Sequoia partner David Cahn's arithmetic the industry needs to earn roughly $3 trillion to justify it. He reached that figure by taking the $1.5 trillion of projected 2026 infrastructure spend, adding operating costs, then adding the return operators and investors expect on that capital. Frontier lab revenue does not currently reach it: the combined run rate sits in the low hundreds of billions, with Anthropic estimated near $60 billion ARR as of July 2026 and OpenAI reporting about $13 billion of 2025 revenue. Alphabet guided to $180–190 billion of 2026 capital expenditure, Microsoft to roughly $190 billion, Amazon to roughly $200 billion, and Meta to $125–145 billion. Google, Meta, Microsoft and Amazon all predict sharp free-cash-flow acceleration in 2028. That 2028 date is the load-bearing assumption in the whole structure, and it is not one you can influence. What you can influence is whether your own AI budget survives either answer.

The arithmetic

The $3 trillion number is not a forecast of doom. It is a payback calculation, and the inputs are public.

Cahn's own caveat matters: $3 trillion is probably an underestimate, because rising memory costs and the increasing use of exotic or inference-specific chips push the number up.

Put the rows next to each other and the gap is not subtle. A bill in the trillions against frontier revenue in the low hundreds of billions is an order of magnitude, not a rounding error. The industry's answer is that the payback arrives in 2028.

The case that it is already working

The bear reading is not the only defensible one, and the counter-evidence is specific.

Bloomberg reported on June 25, 2026 that global AI sales excluding China reached $25 billion in the first quarter of 2026, exceeding the industry's estimated $21 billion in depreciation costs tied to data centre and chip investments. That was the second consecutive quarter it cleared the bar.

Read carefully, though. Depreciation charges still consume more than two thirds of revenue, which leaves a thin buffer for power, labour and financing. Clearing depreciation is not the same as earning a return on capital, and the $3 trillion figure is a return-on-capital number. Both facts are true: AI sales now cover the depreciation line, and they do not come close to the payback arithmetic.

The demand mix supports the bull case in one specific way. Inference has already passed 50% of total AI compute. That marks the shift from building models to running them, and inference revenue is recurring in a way that training spend is not.

Google Cloud growing 63% year over year is not the growth rate of a business whose demand is imaginary. Neither is Azure at 40%. The question was never whether anyone wants this. It is whether they want $3 trillion of it by 2028.

Why this reaches your budget even if you never buy a GPU

The macro argument feels distant until you notice it already moved prices you pay.

AWS raised EC2 Capacity Block prices roughly 20% effective July 1, 2026, the second increase in six months after a 15% rise on January 4. That is what a supply crunch looks like from the buyer's seat, and we covered the detail in our piece on the AWS GPU Capacity Blocks price rise. Providers spending $200 billion a year have to recover it somewhere.

The token line has moved in the opposite direction, which is where budgets get fooled.

Unit price fell 67%. Consumption per developer rose about 18.6x in nine months. Those two facts produce a bill that grows while every vendor truthfully advertises a price cut. That is the mechanism behind the FinOps Foundation finding that 73% of enterprises blew through their AI projections, and it is not a procurement failure. It is an arithmetic one.

The agentic shift makes it sharper. A simple linear workflow cost about $0.04 per interaction in 2023. In 2026, an orchestrated system with tools, reasoning and iterative loops runs about $1.20 per interaction, roughly 30 times more. Nobody raised a price to do that. The workload changed shape.

One healthcare enterprise consumed a trillion tokens over six months, producing more than $6 million in unplanned costs before the finance team understood what was driving it. Deloitte published a CFO guide on AI token economics in April 2026, on a topic that was not on the finance radar eighteen months earlier.

Falling unit prices are not a budget. They are an invitation to use more.

Five moves that survive either answer

The honest position is that nobody reading this can forecast whether 2028 payback arrives. So build a budget that does not need the forecast.

1. Budget consumption, not unit price

Vendor price cuts are real and they will not protect you. Model the tokens you expect to consume and the direction of that curve. If per-developer consumption is growing 18.6x in nine months across the industry, a flat consumption assumption in your model is the single largest error in it.

2. Tier your models on purpose

The median blended cost for a tiered architecture is $2.31 per million tokens. Routing everything to frontier models costs $18.40. That is roughly an 8x spread available to any team willing to decide which requests actually need the expensive model. It is the highest-return FinOps work in AI right now and it requires no vendor negotiation.

3. Re-plan quarterly

AWS repriced twice in six months. Blended token costs fell 67% in a year while consumption rose. An annual budget set in January describes a market that no longer exists by July. Move AI spend to a quarterly cycle and treat the annual number as a range.

4. Hedge the contract, not the thesis

If the 2028 payback slips, vendors under-earning against a $3 trillion bill have one obvious lever, and it is your renewal price. Prefer shorter commitments and price-protection clauses over long lock-ins that assume today's rate card. If payback arrives instead, competition holds prices down and you lost little by staying flexible.

5. Instrument attribution before you scale

The healthcare team that spent $6 million unplanned did not lack discipline. It lacked attribution. Tokens have to map to a team, a feature and a unit of business value before consumption grows, because retrofitting that after the bill arrives is how a cost problem turns into an argument.

India-specific considerations

For Indian teams, the $3 trillion question has a sharper edge, because the currency risk and the subsidised alternative both cut the other way.

Hyperscaler pricing is dollar-denominated while your revenue may not be. A 20% GPU price rise is a 20% rise plus whatever the rupee did. Teams modelling AI spend in rupees against dollar rate cards need the exchange assumption written down and reviewed, not implied.

The subsidised route changes the arithmetic more than any optimisation. The IndiaAI Mission makes 34,000 GPUs available at ₹115–150 per GPU-hour, roughly 42% below market rates, targeting 100,000 GPUs by the end of 2026. No amount of tiering or rightsizing on a hyperscaler bill matches a 42% subsidy for teams that qualify. Check eligibility before you optimise.

Domestic capacity is now a real option rather than a compromise. Providers host A100, L40S and H100 GPUs in Indian data centres with rupee billing and DPDP-aligned infrastructure, which removes cross-border transfer questions under the Digital Personal Data Protection Act 2023 at the same time as it cuts the rate. Our India FinOps cloud cost playbook and guide to cutting cloud spend for Indian teams work through the options.

The pyramid works against you on tokens. If per-developer consumption is the driver, a large delivery organisation has a larger exposure than headcount alone suggests. Attribution per team is not optional at that size.

What to watch

Three dates, in order of how much they should change your plan.

Your own quarterly consumption curve. It is the only number here you control and the only one that reliably predicts your next bill.

The hyperscalers' 2028 free-cash-flow guidance. Google, Meta, Microsoft and Amazon all predict acceleration that year. Watch whether that date moves in their guidance, because a slip is the first honest signal that the payback thesis is stretching.

The depreciation-coverage line. Global AI sales excluding China cleared depreciation for two consecutive quarters at $25 billion against $21 billion in Q1 2026. If that reverses, the thin buffer disappears before the $3 trillion question is anywhere near answered.

FAQ

How eCorpIT can help

eCorpIT is a CMMI Level 5 certified technology organisation in Gurugram, and our senior engineering teams work across AWS, Microsoft and Google platforms on AI cost and architecture. We build token attribution before consumption scales, design tiered model routing that keeps frontier calls for the requests that need them, and model hyperscaler spend against domestic and IndiaAI Mission routes with DPDP considerations included. If your AI budget is moving faster than your forecast, contact us for a review.

References

1. Can AI answer the $3 trillion question? — TechCrunch

1. The $3 Trillion Question: Can AI Actually Earn Back the Chips It Bought? — Business Model Analyst

1. The $3 Trillion AI Question: Can The Industry Justify Its Infrastructure Spending? — Bitcoin World

1. AI Sales Start to Justify Data-Center Spending Boom, Report Says — Bloomberg

1. AI Demand Begins to Justify Massive Cost of Data-Center Buildout — Yahoo Finance

1. Google Cloud, AWS, Microsoft Azure: The AI vertical integration race — Constellation Research

1. Meta layoffs 2026: 8,000 jobs cut in AI restructuring — Quartz

1. China Plans $295 Billion Investment to Build Nationwide AI Data Centers — Bloomberg

1. AWS hikes prices for EC2 Capacity Blocks amid soaring GPU demand — Network World

1. AI Token Costs: Why Enterprise AI Bills Keep Rising in 2026 — Optimum Partners

1. Enterprise AI Budgeting in 2026: Benchmarks, Cost Breakdown, and CFO-Ready Planning — StackAI

1. Agentic AI Enterprise Token Costs — EY

1. How Enterprises Can Control AI Token Costs — BCG

1. AI Token Cost Enterprise: Stop Budget Blowouts in 2026 — elvex

1. AI Data Center Build Advances at Full Speed: Five Things to Know — BloombergNEF

1. IndiaAI Mission: 34,000 GPUs at Rs 150/Hour for startups — Abhishek Gautam

_Last updated: July 15, 2026._