How Much Money Is AI Making? The Real Numbers

The AI industry is generating roughly $60–80 billion a year in software revenue and several hundred billion more in hardware, but almost none of that profit flows to the companies *using* AI. As of the most recent reported figures: Nvidia booked $57.0 billion in a single quarter (Q3 FY2026, ended October 2025), with $51.2 billion of it from data centers. OpenAI exited 2025 at roughly a $20 billion annualized run rate while still losing money. Anthropic went from about $1 billion to $7+ billion annualized inside 2025. Meanwhile, an MIT study found 95% of enterprise generative-AI pilots produced zero measurable P&L return. AI is making enormous money — for chipmakers first, model labs second, and buyers last.

That gap is the actual story, and it's the one most articles on this query skip.

The Four Layers, and Where the Money Actually Lands

"AI revenue" gets reported as one number. It's really four separate businesses with wildly different economics.

Layer 1 — Silicon and infrastructure. Profitable today. Nvidia's fiscal 2025 (ended January 2026) came in at $130.5 billion in revenue, $115.2 billion from data center, with gross margins in the mid-70s. It crossed a $5 trillion market cap in October 2025. This is the only layer reliably converting AI demand into net income at scale. TSMC, Broadcom, SK Hynix, and the power and cooling supply chain sit here too.

Layer 2 — Model labs. Huge revenue, negative profit. OpenAI's ~$20B run rate is real money, and it is dwarfed by compute commitments. Anthropic's climb from ~$1B to ~$7B annualized in twelve months is one of the fastest revenue ramps in software history — and both companies burn cash. Revenue growth here is not a proxy for profitability.

Layer 3 — Application companies. Thin, fast, fragile. Anysphere (Cursor) reportedly hit $500 million ARR by mid-2025 and roughly $1 billion by late 2025. Midjourney reached the low hundreds of millions with fewer than 50 employees. These businesses grow fast because they pay the model layer for their core capability — which means their gross margins are hostage to someone else's pricing.

Layer 4 — Ordinary businesses using AI. Mostly unmeasured. This is where a plumbing company, a med spa, or a law firm lives. The Census Bureau's Business Trends and Outlook Survey has consistently shown under 10% of US firms reporting actual AI use in production, even as adoption headlines suggest near-universal usage. The gap is real: buying a subscription is not the same as changing a P&L.

Nvidia's quarterly data center revenue exceeds the combined annual revenue of every generative-AI application company on earth. The money is being made selling shovels, not digging.

The Number Nobody Publishes: What AI Earns a Single Service Business

Layer 4 is the one that matters if you run a company. Here's the arithmetic, done honestly.

A home services business with $2.4 million in annual revenue and an average job value of $780 closes roughly 3,000 jobs a year. Industry call-tracking data consistently puts missed or abandoned inbound calls at 20–30% for after-hours-heavy trades. If that business fields 6,000 inbound calls and misses 25%, that's 1,500 unanswered calls. At a 30% historical booking rate on answered calls, those missed calls represent ~450 jobs, or roughly $351,000 in lost annual revenue.

An AI voice agent that answers 100% of those calls and books even 40% of what a human would have booked recovers about 180 jobs, or $140,000. Against a fully loaded cost of $800–$2,500 per month for a voice AI plus CRM stack ($9,600–$30,000 per year), that's a defensible return.

Now the same math for a business doing $400,000 a year with 40 inbound calls a month: 25% missed is 10 calls, 3 lost jobs, maybe $2,300 in recoverable revenue annually against a $12,000 tool bill. Identical technology. Catastrophic ROI. You can run your own version of this in our ROI calculator before anyone quotes you.

The variable that decides the outcome isn't the AI. It's call volume, job value, and how leaky the existing process is. AI multiplies an existing system; it does not create one.

Who Should Not Buy AI Marketing (Including From Us)

This is the part that costs us deals, so here it is plainly.

The Failure Modes, Named

What the Revenue Numbers Do and Don't Tell You

Nvidia's earnings tell you how much capital is being *committed* to AI. Combined 2026 capital expenditure guidance across Microsoft, Alphabet, Amazon, and Meta has been reported in the range of $380–500 billion, an increase of well over 50% year over year in some estimates. That is a bet on future demand, not a measurement of realized customer value.

OpenAI's and Anthropic's run rates tell you how much businesses are *spending* on AI. That's closer to useful, but spending is a cost line for the buyer.

The number that would actually answer "how much money is AI making" for you — aggregate net profit improvement at customer companies — does not exist in any reliable form. The best available proxies (the MIT 95% finding, the S&P abandonment rate, the sub-10% Census production-use rate) all point the same direction: the money AI companies make is not evidence that AI makes money for you.

The Honest Version of the Bull Case

None of the above means AI marketing doesn't work. It means the distribution of outcomes is wide, and the inputs that decide your position in that distribution are knowable in advance.

The businesses where we consistently see returns share four traits: inbound lead volume above 75/month, average job or client value above $500, an existing CRM with clean data, and an owner who reviews outcomes weekly. Hit all four and the failure rate drops sharply. Miss two or more and the honest recommendation is to fix operations first.

If you want the underlying adoption and performance data rather than vendor claims, we maintain a sourced set on the AI marketing statistics page, and our pricing is published rather than quoted — partly so the math above is something you can check against a real number instead of a discovery call.

Run the arithmetic on your own call volume and job value before you talk to anyone. If the number doesn't clear the tool cost by 3x, the answer is no — including from us.