Follow the Money Down: Why AI’s Revenue Pyramid Has a Paper-Thin Base

This paper should be getting more attention than it is. Like most data-backed findings about AI that don’t confirm the narrative, it arrived and quietly sank. But it tells you something fundamental: there is a massive, growing disconnect between the infrastructure being built for AI and the value that infrastructure is producing. Not in theory. In the numbers.

We’ve written about the 18x revenue gap, the GDP dependency, and the capex-driven layoffs. This piece takes a different approach. Instead of looking at any single layer of the AI economy, we started at the top of the pyramid and traced the money down from the profitable peak through every layer beneath it. Where does it come from? Who keeps it? And how much of the $400 billion in annual AI infrastructure spending produces measurable business value?

The answer is a revenue pyramid with five layers. At the top sits one company making more profit than every downstream layer combined. At the bottom sits a customer base with 60% annual churn and a productivity thesis that doesn’t hold up under controlled measurement. In between, the money flows upward.

Layer 1: The Only Profitable Layer

nvidia’s $120b profit and where it comes from

NVIDIA posted $215.9 billion in total revenue for fiscal year 2026 (ending January 2026). Data center revenue alone was $193.7 billion, or 90% of the company. Net income: $120.1 billion, a 56% net margin. Gross margins stayed in the 71–75% range. By every conventional metric, this is one of the most profitable companies in history.

But look at where the money comes from. The top six customers generate roughly 85% of quarterly data center revenue. In Q2 FY2026, the top two customers alone represented 39% of total revenue: one at 23% (likely Microsoft) and another at 16% (likely Meta or Amazon). AllianceBernstein calculated that NVIDIA captures 30% of all AI data center spending as pure profit.

This is the critical fact about the AI revenue pyramid: NVIDIA’s net profit alone exceeds the total revenue of every model provider combined. OpenAI, Anthropic, xAI, Midjourney, and everyone else together generate roughly $30–35 billion. NVIDIA’s profit is $120 billion. The semiconductor layer captures more value than all downstream layers put together. So where does NVIDIA’s $120 billion come from? One layer down.

Layer 2: $320 Billion In, $30 Billion Out

hyperscaler capex vs. ai-specific revenue

The five hyperscalers spent over $345 billion on AI infrastructure in 2025. Amazon: $128 billion. Google: $91 billion. Microsoft: $88 billion. Meta: $38 billion. Their 2026 guidance is even more aggressive. Goldman Sachs projects $527–650 billion in combined AI capex.

Capital intensity has reached 45–57% of revenue. These companies have never spent at this rate. They raised a record $108 billion in AI-related debt in 2025. J.P. Morgan estimates the sector will need $1.5 trillion in bonds over five years.

Against all that spending, what do hyperscalers actually earn from selling AI to customers? Roughly $25–30 billion. That’s a 7–13% return on annual AI infrastructure spending. The math only works if downstream customers grow massively. And the NBER data says they’re not.

An estimated 40–50% of hyperscaler capex flows directly to NVIDIA and other chip suppliers. The money goes up. The cloud layer is a pass-through. But a pass-through for what? Who’s paying the cloud bills?

Layer 3: The Money Goes Back Up

model providers, cash burn, and the circular flow

OpenAI crossed $20 billion in ARR in January 2026. It burned $8–9 billion in 2025. Leaked Microsoft documents show OpenAI paid $8.67 billion to Azure for inference costs through Q3 2025 alone, plus a 20% revenue share to Microsoft. For every dollar OpenAI earns, approximately 67 cents to a dollar goes back to Microsoft Azure just for inference.

Anthropic’s numbers are starker. The company hit $9 billion in ARR by late 2025, with 85% from enterprise. But its AWS spending through September was $2.66 billion against estimated revenue of $2.55 billion: spending over 100% of its revenue on AWS alone. Morgan Stanley projects AWS will generate $1.28 billion from Anthropic in 2025, rising to $5.6 billion by 2027.

This creates a self-referential loop. Microsoft invests $13 billion in OpenAI. OpenAI spends most of it on Azure. Azure uses that revenue to buy NVIDIA GPUs. NVIDIA invests in OpenAI at a $100 billion valuation. Amazon invests $8 billion in Anthropic. Anthropic spends more than 100% of its revenue on AWS. AWS buys more GPUs. The capital circulates through the same companies, inflating revenue figures at each layer, while the actual new money entering the system from end customers remains thin. Which raises the question the entire pyramid depends on: are those end customers actually getting value?

Layer 4: The ROI Crisis

category by category, the demand unravels

Menlo Ventures surveyed 495 enterprise decision-makers in December 2025. Total enterprise generative AI spend: $37 billion, up from $11.5 billion in 2024, a 3.2x year-over-year jump. Add $12 billion in consumer AI spending and the total addressable revenue is $49 billion. The Sankey below traces every dollar from provider through market segment to end application. Here’s what each category looks like under the hood.

Coding Tools / $4 Billion

Cursor: $1.2 billion ARR, negative 30% gross margin. The fastest B2B SaaS ever to scale, and it loses money on every subscriber. It pays roughly $650 million annually to Anthropic for API access, against approximately $500 million in revenue. GitHub Copilot reached $2 billion ARR with 1.3 million paid subscribers and 90% of the Fortune 100, but was losing an average of $20 per user per month at its $10 price point. Microsoft has since moved to $19–39/month tiers to fix the economics.

Then there’s the productivity question. A July 2025 METR study, a randomized controlled trial with experienced open-source developers, found that developers were 19% slower with AI coding tools. But they believed they were 20% faster. A 39-point perception gap. Faros AI’s telemetry across 10,000 developers confirmed the pattern: 75% use AI tools, yet organizations see no measurable performance gains. Developers merge 98% more pull requests, but review time increases 91%. The gains evaporate at the bottleneck. Sixty percent of enterprise coding tool deployments fail outright.

Copilots & Assistants / $7.2 Billion

Microsoft 365 Copilot has 15 million paid seats out of 450 million commercial Microsoft 365 users. A 3.3% conversion rate after two years of aggressive rollout. A Gartner survey of organizations that piloted Copilot found half gave it “some value, shows promise,” and 50% of tech leaders said it was “too soon to know” whether the ROI justified $30 per user per month. Forrester projected 116% ROI over three years , but the study was commissioned by Microsoft and modeled hypothetical benefits, not measured outcomes.

Salesforce Agentforce crossed $500 million ARR with 18,500 deals closed, but adoption remains modest relative to the installed base. ServiceNow’s Now Assist hit $600 million in annual contract value. The pattern across copilots is consistent: rapid initial adoption by early-mover enterprises, then a long plateau as organizations struggle to prove value beyond isolated use cases. Most remain stuck in perpetual pilot.

Enterprise Infrastructure & Other B2B / $25.8 Billion

42% of companies abandoned most of their AI initiatives before reaching production in 2025, up from 17% the prior year. The largest category by dollar volume: $18 billion in direct model API and infrastructure spend, plus $7.8 billion in vertical AI, agents, and other enterprise applications. And the one where the failure data is starkest. The proportion of organizations reporting positive AI impact fell across every metric: revenue growth, cost management, customer satisfaction. All declining.

Vertical AI ($3.5 billion) shows the closest thing to genuine ROI in healthcare, which leads at $1.5 billion. Enterprises with strong data integration report 3.5x higher success rates. But even here, MIT’s NANDA lab found that purchasing from specialized vendors succeeds roughly two-thirds of the time, while internal builds succeed barely a third as often. The NBER study isn’t alone. The convergence across independent research is the signal:

Run the math. $49 billion in total AI revenue, enterprise and consumer combined. The converging estimate from NBER, McKinsey, and BCG is that roughly 4–5% of implementations produce measurable P&L impact. That yields approximately $2 billion in proven business value out of $49 billion in revenue. Four cents on every dollar. And less than 0.5% of total AI infrastructure spending.

Layer 5: Paper-Thin Base

consumer ai, churn, and the one exception

Consumer Chat / $8.5 Billion

ChatGPT has 800 million weekly active users and captures roughly 70% of the $12 billion consumer AI market. The numbers sound impressive until you look at conversion: only 3–5% of users pay. That’s 20–35 million paid subscribers out of 800 million. ChartMogul’s retention data for AI-native companies shows a median gross revenue retention of 40%, meaning 60% annual churn. For products under $50 per month, retention drops to 23%. Compare that to healthy SaaS at 90–95% retention.

ChatGPT’s market share fell 19 percentage points in 12 months, from 87% to 68%, as Google Gemini surged to 18%. Switching costs are near zero. An OpenAI/Harvard analysis of 1.5 million conversations found that over 70% of consumer messages are non-work related: asking for advice, playing, reflecting. Most users treat it as a search replacement, not a productivity tool. OpenAI is now testing ads for free-tier users, the classic signal that organic monetization isn’t working.

Other Consumer / $3.5 Billion

There is one notable exception to the entire AI profitability crisis: Midjourney. $500 million in revenue, profitable since month two, 107–163 employees, zero external funding. No VC capital inflating the loop. No debt. No dependency on external model APIs because it built its own. Revenue per employee exceeds $5 million. It’s the only major AI company that makes money on its own. The lesson: profitability in AI requires owning your model and having no API dependency.

Elsewhere in consumer B2C, the novelty is wearing off. Barclays tracked traffic to leading vibe coding platforms: Lovable hit $100 million ARR then saw traffic crash 40%. Vercel’s v0 fell 64% since May. Bolt.new dropped 27% since June. Users love quick prototypes but hit walls when the code gets complex. AI startups fail at 90–92% versus 63% for regular tech startups, and they burn cash roughly twice as fast due to compute costs.

The economics cascade upward. A developer pays $20 per month to Cursor. Cursor sends that $20 (and more) to Anthropic. Anthropic sends most of it to AWS. AWS buys NVIDIA GPUs. NVIDIA keeps 56% as net profit. The developer got a code completion tool. The $20 ends up with the chipmaker.

The $600 Billion Question

solow paradox, j-curves, and what comes next

In 1987, Robert Solow made an observation that became one of the most cited lines in economics. Forty years later, Torsten Slok at Apollo updated it for the AI era:

“AI is everywhere except in the incoming macroeconomic data.”– Torsten Slok, Apollo Global Management, February 2026

The bull case is the J-curve. The IT boom of the 1970s produced a similar paradox before productivity surged from 1995 to 2005. But there’s a structural difference. IT companies had monopoly pricing power: Microsoft Windows, Oracle databases, Cisco routers. Once adopted, switching costs were enormous. AI has fierce competition driving prices toward zero. AI will probably create value eventually. But will that value accrue at AI companies, or at the companies using AI? If it’s the latter, the current capital structure doesn’t justify itself. Not at $400 billion in annual infrastructure spending. Not at $108 billion in new debt. Not at 45–57% capex-to-revenue ratios.

The NBER authors were careful to note that their data captures the present, not the future. Executives predict a 1.4% productivity boost and 0.8% output growth over the next three years. Maybe they’re right. But executives also predicted AI-driven employment cuts of 0.7%, while employees at the same firms predicted a 0.5% employment increase. Someone is wrong.

What we can say from the data: $49 billion in AI revenue currently produces approximately $2 billion in proven, measurable business value. Four cents on every dollar. The money flows upward through five layers and concentrates at the top, in a single company that captures 30% of all AI spending as net profit. Coding tools make developers slower. Copilots convert at 3.3%. Forty-two percent of enterprise AI projects get abandoned. Consumer churn runs at 60%. The base of the pyramid is paper-thin.

The technology might be transformative. The capital structure built to fund it might not survive long enough to find out. If you’re evaluating an AI vendor contract, sizing an NVIDIA position, or betting a company on this stack, the numbers say look closer.