The Math Doesn't Work (Yet): Inside the AI Profitability Problem
Why Scaling Doesn't Lead To Profitability
The Math Doesn’t Work (Yet): Inside the AI Profitability Problem
OpenAI’s own projections show losses getting bigger as revenue gets bigger. Leaked investor documents reported across WSJ, Fortune, and The Information put the company at roughly $74 billion in operating losses in 2028 — on roughly $100 billion in projected revenue. That pairing is the headline. Not a smaller loss as scale arrives. A loss that grows faster than the top line.
That single relationship is the whole story. We tend to model AI companies as software businesses that will eventually grow into their cost structure the way SaaS companies did before them. The numbers say otherwise. These are capital-intensive infrastructure plays wearing software-company clothing, and the unit economics underneath them run in the opposite direction from the SaaS playbook most of us internalized over the last fifteen years.
A caveat before the numbers, because it matters for what’s below: neither OpenAI nor Anthropic publishes audited financials. Most figures here come from leaked investor decks, run-rate annualizations the companies announce in funding rounds, or SEC filings made by their cloud partners. The uncertainty is part of the analysis, not a footnote to it. When a specific timeline or quarterly figure couldn’t survive cross-checking against primary sources, I left it out.
TL;DR
OpenAI’s leaked projections show operating losses widening as revenue grows — roughly $74B in losses on ~$100B revenue projected for 2028, with cumulative cash burn near $115B through 2029.
In 2025 OpenAI spent about $1.69 for every dollar of revenue (~$9B net loss on ~$13B revenue), per leaked documents confirmed by multiple outlets.
Anthropic’s revenue trajectory is steep — roughly $1B annualized at the end of 2024 to a figure announced in the tens of billions by mid-2026 — with Claude Code alone reported at $2.5B annualized by February 2026.
Inference token prices fell about 75% in a year. Selling more AI makes the per-unit economics cheaper, which makes revenue growth harder, not easier.
AI-native gross margins sit near 45% versus 75–85% for mature SaaS — a structural gap of 23–33 points no company has yet closed.
No company has a verified path to profitability. Every specific breakeven-by-year claim I tried to confirm fell apart under scrutiny.
Two Companies, Two Shapes
OpenAI ended 2025 at roughly $20 billion in annualized revenue, a figure CFO Sarah Friar has stated directly. That is a large business by any normal measure. It is also a business that, in the same year, spent about $1.69 for every dollar it took in — somewhere around a $9 billion net loss on roughly $13 billion in recognized revenue, according to leaked documents that WSJ, Fortune, and The Information each reported. The company is majority funded by Microsoft, runs its compute primarily on Azure, and its strategy is scale-first: build the largest models, capture the most usage, and trust that revenue follows the curve.
Anthropic’s shape is different. Its revenue trajectory is steeper and more concentrated. The company grew from roughly $1 billion annualized at the end of 2024 to a figure it announced in the tens of billions by mid-2026 — the number it cited in its Series H materials. I’m deliberately not pinning an exact figure to a month here, because the company has grown several-fold inside a single five-month window and any precise number is stale by the time you read it. What’s verifiable is the slope, and the slope is steep.
The more interesting detail is the concentration of value. Claude Code, one product, was reported at roughly $2.5 billion annualized by February 2026. A single coding tool driving that much of a company’s run rate tells you something about where the margin-bearing demand actually lives. Anthropic is backed by Amazon (over $8 billion invested) and Google (over $2 billion), and it has committed to spend more than $100 billion on AWS over ten years, with roughly 1 GW of Trainium capacity targeted by the end of 2026. Two large companies funding it; one of them also selling it the silicon it runs on.
Why Compute Is the Problem
What separates these companies from every SaaS business you’ve evaluated: they don’t own their infrastructure. They rent it, at hyperscaler rates, from the same companies that fund them.
OpenAI’s Azure spend reportedly ran around $3.7 billion in 2024 and roughly $8.7 billion across the first three quarters of 2025. Treat those numbers as medium-confidence — they come from leaked documents, and Microsoft pushed back that the figures “aren’t quite right.” But the direction is consistent with everything else: compute cost is the dominant line item, it’s largely fixed, and it grows with usage.
Anthropic’s arrangement produced one of the stranger details in modern enterprise finance I’ve read in some time. The company signed a deal for compute from Colossus 1 — Elon Musk’s Memphis data center, operated by xAI — at roughly $1.25 billion per month for 300 MW of capacity, running through May 2029. That’s not from a leak. It surfaced in SpaceX’s S-1 SEC filing and was confirmed by CNBC, Axios, and Data Center Dynamics, with a potential total value above $40 billion. There’s a 90-day mutual cancellation clause, so the headline total overstates the firm commitment. Still: Anthropic — funded by Google and Amazon — is paying Elon Musk’s company more than a billion dollars a month for compute. The AI capital world is stranger from the inside than the press releases suggest.
Zoom out and the renter problem gets sharper. Hyperscaler capex for 2026 is projected at $660–690 billion. Against that, OpenAI’s $20 billion ARR is roughly 3% of a single year’s data-center buildout by its suppliers. The companies selling AI applications are small tenants in an infrastructure market they don’t control and can’t currently price against.
The Unit Economics Trap
This story inverts an instinct most of us trust.
In normal software, even in the Cloud, scale is your friend. Marginal cost trends toward zero, gross margin climbs as you grow, and a mature SaaS business lands at 75–85% gross margin because serving the millionth customer costs almost nothing. Volume is the cure.
Inference doesn’t behave that way. Every token generated costs compute — real, metered, non-zero compute — so the marginal cost of serving usage stays stubbornly positive. And the price you can charge for that token is collapsing. Enterprise transaction data from Ramp shows inference prices falling roughly 75% in a single year, from around $10 per million tokens to around $2.50. Capability per dollar is improving fast, which is good for buyers and brutal for sellers, because it means the revenue you booked at last year’s prices reprices downward while your compute bill does not.
Put the two forces together and you get a squeeze that worsens with success. The better you are at selling inference, the more usage you drive; the more usage you drive, the more the per-unit price falls; the more it falls, the harder it is to grow revenue against a compute bill that scales with that same usage. Volume isn’t the cure here. Under these dynamics it’s part of the disease.
The survey data puts a number on how far this world sits from SaaS. ICONIQ Capital polled about 300 software executives and pegged AI-native gross margins at 41% in 2024, 45% in 2025, and a projected 52% in 2026. Improving — but starting from a base 30-plus points below mature SaaS, and closing the distance slowly. A 52% gross margin is a respectable hardware business. It is a structurally difficult software business, especially one still spending heavily to grow.
Two Different Bets
OpenAI and Anthropic are running different experiments on how you eventually close that gap. Neither has been validated.
OpenAI’s bet is scale and breadth. Build the broadest platform, capture consumer and enterprise and API demand simultaneously, and assume that at sufficient scale you gain pricing power over compute, model-efficiency gains compound, and the revenue base grows fast enough to absorb the fixed cost. The leaked projections embody the risk in this bet: they show losses widening through 2028 even as revenue approaches $100 billion, with cumulative cash burn near $115 billion through 2029. The theory requires the curve to bend after the window we can currently see.
Anthropic’s bet is narrower and more product-led. Find a wedge where the work is valuable enough that buyers tolerate real prices, prove the margin there, and expand outward. Claude Code is that wedge made concrete — $2.5 billion annualized from developers who pay because the output is worth more than the inference under it. Coding, agents, and enterprise automation are higher-value work than chat, and higher-value work supports prices that don’t immediately erode under token deflation. The risk: revenue concentration in a single product line, and an infrastructure bill — AWS commitments plus the xAI deal — that’s enormous relative to a company still proving the model.
Two theories of the same problem. Scale your way past the margin gap, or find work valuable enough that the gap doesn’t bind. We don’t yet have the data to say either works.
What Would It Actually Take
I’ll skip the timeline speculation — every specific breakeven-by-year claim I tried to verify died on contact with the sources. The structural requirements are clearer than the dates.
Three things have to move. First, gross margins have to climb from the mid-40s toward something defensible — call it 60-plus — and stay there while volume grows. That means model-efficiency gains (cheaper inference per unit of capability) have to outrun price deflation, rather than getting passed straight through to buyers as lower prices.
Second, these companies need pricing power over compute, which today they don’t have. At current scale they’re tenants. The open question is at what ARR a vendor becomes large enough to negotiate compute like a partner instead of a customer — or to build its own. Anthropic’s Trainium commitment and OpenAI’s various infrastructure moves are bets that vertical integration eventually changes the cost equation. That’s unproven, and it’s expensive in the interim.
Third, the product mix has to keep shifting toward work that resists deflation — enterprise agents, coding tools, automation that’s measured against labor cost rather than against the falling price of a token. Claude Code is the cleanest evidence that this category exists and that buyers will pay. Whether it’s a large enough share of total volume to lift blended margins across a company is the question that decides the whole thing.
The Open Questions
What we don’t know outweighs what we do.
We don’t know the actual, current gross margins at either company — only the AI-native sector estimate of roughly 45%. Neither company publishes the number that would settle the argument. We don’t know whether Anthropic’s stack of compute commitments, the decade-long AWS deal alongside the month-by-month xAI arrangement, creates structural tension or healthy redundancy. A company hedging across three infrastructure providers is either diversifying supply or revealing that no single supplier can meet its demand. We don’t know the ARR threshold at which compute pricing becomes negotiable, which is the hinge the entire margin story turns on.
And there’s the strategic risk that has no clean precedent: your infrastructure supplier is also your competitor. Microsoft ships Copilot. Amazon and Google both build models that compete with Anthropic’s. xAI builds Grok. Every dollar these companies pay for compute partly funds a rival’s model program. In normal software you don’t hand your gross margin to the company trying to beat you. Here it’s the default arrangement.
So the real question isn’t whether the AI labs are growing. They obviously are, faster than almost any companies in history. The question is whether revenue growth and margin improvement are the same trend or opposing ones. The SaaS era trained a generation of operators to believe that scale fixes economics. The leaked numbers describe a business where scale, so far, makes the loss bigger. Until one of these companies publishes a gross margin that shows the curve bending, that’s the math we have. And the math doesn’t work yet.
Bob Matsuoka is CTO of Duetto and writes about AI-powered engineering at HyperDev.
Related reading:
The First 70% Era — Where agentic AI delivers value and where it stops, and why the higher-value work resists token deflation
AI and the Rise of the Hyperdev — Why developers pay real money for AI tooling, the demand side of the margin story
AI Power Ranking — Tool comparisons and benchmarks for AI practitioners
LinkedIn Newsletter — Strategic AI insights for CTOs and engineering leaders