AI Quarterly Revenue Surpasses Infrastructure Depreciation for the First Time, Annualized Revenue Reaches $175 Billion
N.R. Finch
Global generative-AI annualized revenue has reached $175 billion, and Q1 2026 marks the first quarter income exceeds infrastructure depreciation — AI can now cover its own hardware costs, but proving a return on $2 trillion in cumulative capex remains a distant goal.
Revenue covers depreciation — what does that actually mean?
In Q1 2026, global generative-AI revenue (excluding China) exceeded infrastructure depreciation for the first time. This means → the cash AI generates now covers the accounting cost of servers, GPUs, and data centers wearing down.
In plain terms = previously, AI earned less each quarter than its hardware "aged" on the books. Now it has just broken even — like a new shop finally making enough to pay rent.
But this is only the first threshold. By end-2026, cumulative AI-related capex by hyperscalers and AI cloud platforms will reach roughly $2 trillion, with AI-driven incremental spending at about $535 billion. Cumulative revenue is still far from covering that historical outlay.
Where do these numbers come from — are they reliable?
The data comes from Exponential View's *State of AI Economics* report. The team spent six months disaggregating disclosures, financials, and cloud-procurement records from over 1,000 companies.
The method is bottom-up: strip out double-counting across the supply chain, then build an independent revenue model. This means → it is not a simple sum of each company's reported AI revenue, but an attempt to remove overlaps and size the real market.
Founder Azeem Azhar told Bloomberg TV: "The supply side was almost in plain view, but the demand side was shrouded in fog."
How fast is this growing?
Generative-AI revenue is growing at roughly 200% year-on-year — about three times the pace of any prior IT platform shift, outrunning the early trajectories of the internet, cloud computing, and smartphones.
In plain terms = in 2023, adding $1 billion in cumulative AI revenue took about 180 days. Today it takes fewer than two.
Azhar said his team expected growth to cool at the start of the year. "Instead, Anthropic's explosive growth kept the industry near 200% year-on-year." This reflects a competitive race still accelerating, not slowing.
What is the "compute super-cycle"?
Surging inference demand is driving compute into a new super-cycle. Since 1971, global compute has grown at roughly 66% compound annually; in the AI era that has risen to 80%.
Large AI data centers have scaled roughly 50× in four years. Chip costs as a share of data-center spending are projected to rise from about 40% in 2021 to 60% in 2026.
HBM — high-bandwidth memory that stacks multiple memory layers to boost data throughput — has jumped from roughly 2% to about 18% of that cost. This means → more of the data-center dollar goes to chips, and more of the chip dollar goes to premium memory.
Prices are collapsing — is demand really keeping up?
The cost per million tokens has plunged from about $17 in 2023 to roughly $2, while token consumption has grown approximately 14× year-on-year.
Google, OpenAI, and peers observe a consistent pattern: for every 10% drop in token price, demand rises 12–18% — elasticity exceeds the price cut itself. In plain terms = cut $1, and the extra volume is worth more than $1. The market expands faster than prices fall.
The report draws a parallel to internet advertising: just as Google's cost-per-click model spawned a vast digital-ad ecosystem, token pricing is becoming the unit of value for the AI era.
Where is profit migrating — and who should worry?
AI profit is shifting from upstream to the application layer. Over the past year, application-layer revenue share rose from about 7% to 11%, the model layer slipped from 11% to roughly 9%, and cloud infrastructure fell from about 82% to below 80%.
This reflects a structural reality: frontier models have a limited pricing window. As open-source models close the capability gap, the most advanced models commoditize rapidly after release.
This means → AI labs that want to protect margins must extend into vertical applications — legal, coding, and similar domains — rather than relying on API fees alone. Whether falling costs can keep catalyzing market expansion will determine if this trillion-dollar compute bet pays off.
Content is for reference only, not financial advice.