The Wall Street Journal published confidential investor documents from both OpenAI and Anthropic this month — the kind of numbers that usually only surface in pitch decks shown in rooms where everyone has signed an NDA and the coffee is suspiciously good. Within hours, every tech publication on Earth had a version of the same headline: Anthropic passes OpenAI in revenue! The AI race heats up! Who will win?
The right question isn’t “who’s winning?” It’s “for every dollar of revenue, how much actually falls to the bottom line — and where does the rest go?” That’s unit economics. And when you build the waterfall for these two companies, the picture looks very different from the headlines.
The Unit Economics Nobody Is Showing You
Here’s what happens to each dollar of revenue at both companies. I’ve marked each number by data quality: Sourced (directly from primary reporting), Calculated (derived from sourced figures), Inconsistent (sources disagree), or Estimated (residual/inferred).
A few things jump out.
Anthropic has better gross margins. 40% versus 33%. Both numbers are sourced directly. This means Anthropic keeps seven cents more of every revenue dollar after paying for inference — before any other expenses.
Anthropic’s training costs are dramatically lower. $4.1 billion versus somewhere between $13 billion and $25 billion for OpenAI, depending on how you count. That’s a 3x to 6x difference.
Both companies spend more than $1 for every $1 of revenue. Anthropic’s total expense load is $2.16 per revenue dollar. OpenAI’s is harder to pin down because of the R&D uncertainty, but cash burn of $0.45 per dollar on top of a 67% COGS means they’re also underwater. The difference is the depth.
Here’s where every other analysis stops and shows you the headline numbers. Let me do that too — because they are remarkable — but keep the unit economics in mind as you read them.
The Scoreboard (Since Everyone Insists on One)
Anthropic’s annualized revenue hit $30 billion in April 2026. That’s up from $9 billion at the end of 2025, $14 billion in February, $19 billion in March. If you’re keeping score at home, that’s $1 billion in January 2025 to $30 billion fifteen months later. A 30x in fifteen months. Salesforce took twenty-four years to get to $30 billion. Anthropic did it in five.
OpenAI is at $24 billion annualized — $2 billion a month, confirmed by the company. Still extraordinary growth: $2 billion in 2023, $6 billion in 2024, $20 billion by end of 2025. That’s 3x per year sustained at a scale where 3x means adding billions every quarter.
Epoch AI modeled this crossover back in February. They predicted mid-2026. It happened in April — a few months early. Anthropic growing at roughly 10x per year versus OpenAI at 3.4x per year meant the lines had to cross eventually. The question was when, and the answer was “sooner than the models predicted.”
Note on Revenue Recognition: One asterisk before we go further: these numbers aren’t perfectly apples-to-apples. Anthropic counts sales through cloud partners (AWS Bedrock, Google Vertex AI, Azure) as its own revenue. OpenAI doesn’t. Anthropic says this reflects standard accounting — they’re the principal in the transaction, the cloud provider is a distribution channel. I think that’s defensible, but it inflates Anthropic’s top line relative to OpenAI on a like-for-like basis. How much? Impossible to say without audited financials. Keep it in mind.
📊 [WSJ Chart: Yearly revenue by segment, in billions — OpenAI vs Anthropic]Consumer / Enterprise / New Products breakdown. OpenAI's chart is a consumer company with an enterprise sideline. Anthropic's is the reverse. Same race, different lanes.
The Metric Everyone Is Using Is Wrong
Revenue is the metric Wall Street loves because it’s the metric Wall Street knows. Top-line growth, run rate, year-over-year comparisons — the language of SaaS analysis applied to companies that are anything but.
Meanwhile, the AI industry has its own preferred metric: gigawatts. As readers of this newsletter know, I’ve spent considerable time arguing that measuring AI capability in gigawatts is like measuring a chef’s skill by how much gas their stove burns. OpenAI’s investor memo this week — titled, I kid you not, “Computing is the Key to Success” — doubled down on exactly this framing: 1.9 GW in 2025, 30 GW by 2030, Anthropic described as “operating on a meaningfully smaller curve.”
So we’ve got Wall Street measuring revenue. We’ve got the AI industry measuring gigawatts. Both are input metrics masquerading as output metrics. Revenue tells you how much money is coming in but nothing about how efficiently it’s being generated. Gigawatts tells you how much power you’re burning but nothing about what you’re getting for it. It’s like valuing a mine by the diesel it burns. You’re measuring the dig, not the gold.
Revenue/Gigawatt: Fine! I Will Do The Math Myself.
OpenAI’s own investor memo — reported by CNBC and Bloomberg — gives us the compute figures. For both companies, helpfully.
OpenAI: 1.9 GW in 2025
Anthropic: 1.4 GW (OpenAI’s own estimate of Anthropic)
Combined with the April 2026 revenue run rates:
Anthropic: $30B ÷ 1.4 GW = $21.4 billion per gigawattOpenAI: $24B ÷ 1.9 GW = $12.6 billion per gigawatt
Anthropic generates roughly 70% more revenue per unit of compute.
Caveat — because I take these seriously. The GW figures are 2025 estimates; the revenue figures are April 2026 run rates. That’s a timing mismatch. Both companies have likely added some capacity since those figures were reported. But compute infrastructure at this scale doesn’t double overnight — these are gigawatt-scale power agreements, data center builds, chip procurement cycles measured in quarters, not weeks. The directionality holds even if the precise ratio shifts a few points.
And the directionality says something important: OpenAI’s investor memo is arguing “we have more power plants.” Anthropic’s numbers are arguing “we extract more value from every kilowatt.”
There’s a version of this story where building 30 GW is a moat. It’s the pharma playbook, the DRAM playbook, the streaming playbook. Build the factory, flood the market, price aggressively. When marginal cost approaches zero, overcapacity lets you price competitors out of existence while your fixed costs amortize over infinite volume.
AI inference is not that business. Every query burns watts. The meter is running on every token. Overcapacity here isn’t a deterrent — it’s a commitment to a burn rate.
But undercapacity has costs too. Anthropic launched Opus 4.6 to rave reviews in February, then spent March quietly downgrading it. Mythos, their most capable model, was deemed too powerful to release publicly. Both stories have more to them — and I’ll get into them shortly.
Overcapacity is a burn rate. Under capacity is a product you can’t ship. Pick your poison.
The Double-Entry Bookkeeping Trick
Before we go further into the numbers, let’s appreciate something genuinely novel about these financial disclosures: both companies report profitability two different ways.
One version includes training costs. The other leaves them out.
Strip out what both companies call “compute for research,” and OpenAI is on track for a small pretax operating profit this year. So is Anthropic under its best-case scenario. This is the number both companies want their IPO investors to focus on. “Look! The core business works!”
Add training back in, and the math gets uncomfortable. OpenAI doesn’t expect to break even until the 2030s. The WSJ’s charts show projected losses peaking at $85 billion in 2028. Anthropic projects hitting profitability by 2028-2029.
📊 [WSJ Chart: Profit, in billions, yearly — OpenAI vs Anthropic]Two lines per company: profit excluding model training costs vs including training costs. The gap between the two lines is the magic trick. Strip training, and both companies look like businesses. Add it back, and at least one looks like a moonshot.
I sat through two years of corporate finance at Wharton, and I have never — not once — seen a company report earnings on an “excluding the part of our business that might bankrupt us” basis and present it with a straight face to investors. What both companies are telling you is: "We are two businesses stapled together — a profitable inference company and an unprofitable research lab. Please evaluate the profitable one and trust us on the other."
The problem, of course, is that you can’t unstaple them. Without the research lab, the inference company sells last year’s model. And last year’s model is a commodity with a non-vanishing marginal cost.
The Training Cost Chasm
OpenAI spent $25 billion on training in 2025. They project $121 billion in 2028. The 2029 number is higher. If your R&D budget quintuples in three years and your CFO is telling investors this is the plan, you're not building a product. You're building a particle accelerator that happens to autocomplete emails.
Anthropic's number for the same period: roughly $30 billion. Same race. Quarter of the spend.
📊 [WSJ Chart: Yearly AI model training costs, in billions — OpenAI vs Anthropic]OpenAI’s bars look like a skyline. Anthropic’s look like a suburb. The visual alone tells you these are fundamentally different spending philosophies.
If you extract 70% more revenue from each GW, you need to build 70% less infrastructure for every marginal dollar of growth — your revenue per gigawatt compounds. Over three to five years, the difference between $21B/GW and $13B/GW isn't a rounding error — it's hundreds of billions of dollars in data centers, power contracts, and chip procurement that one company never has to finance.
OpenAI needs to raise that money. Anthropic doesn't.
OpenAI has made a public show of refocusing — killing Sora, declaring code red, consolidating around core products. The training budget didn’t get the memo. Epoch AI found that over 80% of OpenAI’s R&D compute goes to experimental runs and unreleased models, not final training of shipped products. It’s not a data problem either — OpenAI’s own projections show data acquisition costs declining from $500 million to $200 million. The compute is going somewhere, just not where the press releases suggest. Anthropic trains text models. OpenAI trains everything, and most of what it trains never ships. That’s not a spending problem. That’s a portfolio strategy — and whether it’s brilliant diversification or very expensive ADHD is something I’ll get into in Part 2.
Inference: The Meter That Never Stops Running
Training costs are R&D. You spend them once, you get a model, and then you pray. Inference costs are COGS. You pay them on every single query, forever, and they scale with success. The more customers you win, the bigger the bill gets. Right now, inference eats more than half of revenue for both companies. That’s not an R&D problem. That’s a unit economics problem.
The physics of inference is democratic — cost per query is mostly a function of model size and hardware, and both companies run comparable frontier models on similar chips. Gross margins confirm it: Anthropic's 40%, OpenAI's 33%, despite radically different pricing and customer mix. What differs isn't the cost of serving a query. It's who pays for it — and what happens to every dollar below the gross margin line.
Below that line, OpenAI's version is the classic consumer internet trap. Of ChatGPT’s 900 million weekly users, 5.5% pay. The rest use it for free to settle bar bets while OpenAI picks up the compute tab. A third of all inference spend, which amounts to billions a year — generates zero revenue. OpenAI calls this an adoption strategy. It’s also a prayer that 5.5% becomes 15% before the investors run out of patience.
Anthropic doesn't have the free-rider problem. Eighty percent of revenue comes from enterprise customers who pay for every token. No subsidy. No conversion funnel. No prayer. When inference costs came in 23% higher than expected — pushing gross margins to 40%, ten points below their own internal target — it hurt. But margin compression is an engineering problem. Structural subsidy is an existential one.
Note on margin convergence: This surprised me too. Anthropic charges up to $200/month. OpenAI charges $20. You’d expect a wider gap. But gross margin is revenue minus inference cost divided by revenue — and enterprise customers don’t just pay more, they consume more. A Max user running Claude Code in agentic loops burns through vastly more compute per account than a ChatGPT user asking for dinner recipes. Anthropic charges 10x more because each customer costs proportionally more to serve. Meanwhile, OpenAI’s free tier likely routes users to cheaper, smaller models with shorter queries — the cost per free query is lower than the cost per enterprise query. The result: radically different business models, similar gross margins. The divergence shows up below the gross margin line — in training spend, cash burn, and the path to profitability. That’s where this story gets interesting.
What Anthropic has instead is a rationing problem. When you can’t serve all the demand profitably, you have three options: raise prices, add capacity, or degrade the product. There’s a fourth: don’t ship it at all. Anthropic is doing all four.
In early April, Anthropic blocked third-party tools like OpenClaw from using flat-rate subscriptions. Cherny called it “managing capacity thoughtfully.” OpenClaw’s creator, now at OpenAI, called it “locking out open source.” The economics were obvious — third-party tools consumed 5-10x more compute than subscription pricing could sustain. The timing, weeks after OpenClaw’s creator joined a competitor, was too.
Then there’s the Opus downshift. Anthropic quietly dropped its flagship model’s default reasoning effort from high to medium. An AMD engineer’s analysis of 6,852 sessions showed reasoning depth collapsed 73% — the model was editing code it hadn’t bothered to read. A workaround spread through Reddit: set an environment variable to disable adaptive thinking. When your power users are distributing ENV flags to restore the product they’re paying for, that’s a compute budget leaking through the product surface.
And then Mythos — Anthropic’s most capable model, restricted to 40 organizations through an invitation-only program. The official explanation is safety: the model finds and exploits zero-day vulnerabilities with scary accuracy. Andreessen asked the obvious question: does Anthropic have the compute to serve it broadly? Anthropic’s own documents describe Mythos-tier models as “very expensive to serve.” Both explanations can be true. But safety concerns don’t show up on the income statement. Inference costs do.
The pattern is consistent: Anthropic’s efficiency advantage is real, but it’s running at the edge of its own capacity. Revenue per gigawatt is high because the denominator is low. That’s operational discipline with a warning sign.
Both companies expect inference costs as a share of revenue to decline, and the WSJ confirms the trend is visible. But “cheaper per token” doesn’t mean “cheaper overall” when agentic workloads consume 5-30x more tokens per task than a chatbot. Whether margins actually expand depends on whether costs fall faster than demand rises. That question has a name, and a 161-year-old answer. Part 2.
The Cash Burn Math
OpenAI burned $9 billion in 2025. Projected: $17 billion in 2026, $57 billion in 2027. In 2025, the company spent $1.70 for every dollar it earned. Cumulative burn through 2029: $115 billion. Break-even target: 2030. Maybe.
Anthropic burned $5.6 billion in 2024, dropping to $3 billion in 2025. Positive free cash flow projected by 2027. Burn rate falling to 9% of revenue by that year.
📊 [WSJ Chart: Free cash flow, in billions, yearly — OpenAI vs Anthropic]
Both companies’ investor documents include a metric I find genuinely useful: projected revenue per dollar of compute by 2028. Anthropic: $2.10. OpenAI: $1.60. (Reported via Techmeme, sourced from WSJ.)
But here’s the catch: those figures exclude training costs. They’re the “profitable inference company” half of the double-entry trick. Add training back in — using the WSJ’s own cost projections — and the picture changes dramatically.
On inference alone, both companies project profitability. Add training and Anthropic still clears the bar — I estimate roughly $1.70 per compute dollar based on WSJ’s training cost projections. OpenAI lands at approximately $0.68 — thirty-two cents underwater on every dollar of silicon before paying a single salary. (Both training-inclusive figures are my estimates using WSJ cost data; the inference-only figures are directly from investor documents.)
That last part matters. “Compute dollar” here means infrastructure spend: chips, cloud capacity, power. It doesn’t include headcount, sales, G&A, or stock compensation. The $0.68 figure is conservative. OpenAI can’t cover their hardware costs from revenue alone, let alone everything else required to run the company. That’s how you get to $85 billion in projected losses.
The IPO Implications
Anthropic is targeting an IPO as early as October 2026. OpenAI is targeting Q4 2026 at roughly $1 trillion, following its $122 billion raise at $852 billion. SpaceX plans a roadshow in June. Combined, these listings could raise north of $240 billion. The Nasdaq literally changed its rules to accommodate them — allowing newly listed companies to join its index faster to give them access to broader pools of capital sooner.
The infrastructure of public markets is bending to fit the cash requirements of AI companies.
For OpenAI, the IPO isn’t optional. At $85 billion in projected 2028 losses, no private funding market sustains that pace indefinitely. This is a company that needs continuous access to public capital to fund the gap between what it earns and what it spends. The IPO is a lifeline, not a victory lap.
For Anthropic, an IPO is advantageous but less existential. If the company hits positive free cash flow by 2027 as projected, it has optionality on timing and terms. The IPO becomes a growth accelerant.
PitchBook’s research concluded that among the three major AI IPO candidates — OpenAI, Anthropic, and Databricks — OpenAI scores weakest on business quality fundamentals despite commanding the highest valuation. When a company’s own projections show it needing to raise hundreds of billions just to reach break-even, the IPO prospectus becomes less “here’s why you should invest” and more “here’s what happens if you stop.”
The Metric That Matters (Part 1 of a Thesis)
Revenue per gigawatt isn’t perfect. It inherits timing mismatches and estimation uncertainties from its component parts. But it does something no other metric in this conversation does: it measures how efficiently a company converts physical infrastructure into economic output. It’s the AI equivalent of same-store sales — the number that separates “growing fast” from “growing well.”
Right now, one company generates $21 billion per gigawatt. The other generates $13 billion but has 35% more capacity and 900 million users it hasn't figured out how to charge yet. One spends a quarter as much on training. The other is placing bets across modalities that could define the next decade of computing — or burn $121 billion trying. One projects profitability in 2028. The other in 2030, with a revenue ceiling that's either delusional or the biggest opportunity in tech history.
Both are growing at rates that would make any other company in history jealous. The question isn’t who’s growing faster. The question is who’s growing better.
Revenue per gigawatt answers that. It doesn’t answer why — why both companies have converged on text despite obvious demand for video, why efficiency gains keep getting eaten by demand growth, or whether the silicon underneath all of this is about to get 100x more efficient. Those questions have answers. One of them is 161 years old. Part 2
