OpenAI’s $17 Billion Burn Rate: The Unit Economics Don’t Work
I’m sitting in a coffee shop in Dublin, scrolling through OpenAI’s leaked financial documents, when a number stops me mid-sip:
$17 billion.
That’s how much cash OpenAI will burn through in 2026. Not revenue. Not valuation. Pure, undiluted cash going up in smoke.
For context, that’s more than NASA’s entire annual budget. It’s enough to buy Twitter (the old one, not X). It’s the GDP of Iceland.
And here’s the kicker: OpenAI is making money hand over fist. Revenue hit $20+ billion in 2025, tripling from the year before. They’re adding customers faster than any software company in history. ChatGPT has 810 million weekly active users.
So how is a company with $20 billion in revenue burning $17 billion in cash?
The answer reveals something uncomfortable about the AI gold rush everyone’s betting on: the unit economics are fundamentally broken.
The Revenue Mirage
Let me start with the good news.
OpenAI’s growth is genuinely historic. Revenue went from:
2023: $2 billion
2024: $6 billion
2025: $20+ billion
That’s a 10x increase in two years. It took Google seven years to hit $12 billion. It took Facebook six years. OpenAI did it in under three.
The company hit $1 billion in monthly revenue in July 2025—a number they’re now beating every month. Annualized revenue run rate exceeded $20 billion by year-end.
CFO Sarah Friar calls this “never-before-seen growth at such scale.”
But here’s what the PR doesn’t tell you: revenue is completely decoupled from profitability.
The $17 Billion Hole
OpenAI’s 2026 cash burn breaks down like this:
Compute costs (inference): ~$10-11 billion
Running ChatGPT queries costs money. A lot of money.
With 810M weekly users generating 2.5 billion queries per day, the math is brutal
GPT-5 consumes 18.9 watt-hours per query on average (University of Rhode Island estimate)
At current rates: 47.2 gigawatt-hours of energy daily just to keep the lights on
That’s enough to power 1.64 million US households for a year
R&D and training new models: ~$3-4 billion
Training GPT-5 alone likely cost hundreds of millions
They’re already working on GPT-6, o3, and other next-gen models
Unlike traditional software where you code once and deploy, AI requires constant retraining
Sales, marketing, and headcount: ~$2-2.5 billion
Competing with Anthropic, Google, and Meta for top AI talent
Engineer salaries at OpenAI now routinely exceed $500K-$1M base + equity
Marketing push to maintain consumer brand dominance
Infrastructure and data centers: ~$1-1.5 billion
Building out compute capacity grew from 0.6 gigawatts in 2024 to 1.9 gigawatts in 2025
Planning for 10x more by 2030
Long-term contracts with Microsoft Azure, CoreWeave, others
Total projected burn through 2029: $115 billion.
Let me repeat that. One hundred and fifteen billion dollars.
The Manhattan Project cost $30 billion in today’s dollars. The Apollo program cost $288 billion over 13 years. OpenAI will burn through $115 billion in five years (2025-2029) trying to reach profitability.
The Unit Economics Problem
Here’s where it gets uncomfortable.
Every dollar of revenue costs OpenAI more than a dollar to generate.
Let’s break it down using GPT-5’s actual numbers from its August-December 2025 run (per Epoch AI analysis):
Revenue (4 months): $6.1 billion
Inference costs: $3.2 billion (52% of revenue)
Gross profit: $2.9 billion (48% gross margin)
Operating expenses: $3.6 billion
Operating loss: -$700 million (-11% operating margin)
So GPT-5 had a 48% gross margin but an -11% operating margin. Positive gross profit, negative operating profit.
This is the core problem: AI models cost more to operate and improve than they generate in revenue.
According to Microsoft’s leaked revenue share data, OpenAI still burns $2 for every $1 earned on inference alone. That’s before R&D, before sales and marketing, before anything else.
Compare this to traditional SaaS companies:
Salesforce gross margin: 76%
Adobe gross margin: 88%
Microsoft 365 gross margin: ~70%
OpenAI’s 48% gross margin (after inference costs) is half what mature software companies achieve. And remember, that’s gross margin—before all the other expenses.
The Anthropic Reality Check
Want to know how bad OpenAI’s unit economics are?
Look at Anthropic (the company behind Claude).
Anthropic has similar technology, similar challenges, similar customers. But their financial trajectory is wildly different:
Anthropic 2025-2026 numbers:
Revenue: $9 billion ARR (end of 2025), targeting $18-26B in 2026
Cash burn: $3 billion in 2025
Burn rate as % of revenue: ~33% in 2025 → 9% in 2027
Projected to break even in 2028
OpenAI 2025-2026 numbers:
Revenue: $20+ billion (end of 2025)
Cash burn: $17 billion in 2026
Burn rate as % of revenue: 85% in 2026, staying at 57% through 2027
Not profitable until 2029-2030
Anthropic will stop burning cash in 2028. OpenAI won’t stop until 2030.
How is this possible when OpenAI has 2x the revenue?
Because Anthropic focused on enterprise from day one, where margins are better. OpenAI chased consumer scale, where margins are terrible.
Anthropic’s gross margin trajectory:
2024: -94% (yes, negative)
2025: 50%
Projected 2028: 77%
They’re going from losing money on every query to making 77 cents on the dollar in four years.
OpenAI’s path to profitability is... less clear.
The Circular Money Problem
Here’s where it gets weird.
OpenAI’s largest investor is Microsoft. Microsoft has committed $13+ billion and owns significant profit-sharing rights in OpenAI.
But OpenAI runs on Microsoft Azure. Every dollar of compute OpenAI spends goes back to... Microsoft.
Similarly:
Nvidia committed $100 billion to OpenAI
That money will “go back to Nvidia” (per OpenAI CFO) in GPU purchases
Nvidia is also a major investor in CoreWeave, which provides cloud capacity to OpenAI
CoreWeave spends billions buying... Nvidia chips
This is the AI industry’s dirty secret: much of the “revenue” and “investment” is just money circulating between the same handful of companies.
When Microsoft reports that 45% of its $625 billion cloud backlog is tied to OpenAI, investors get nervous. Microsoft’s stock dropped 12% in one day (January 29, 2026), wiping out $440 billion in market cap, after disclosing this dependency.
Oracle’s credit rating is approaching junk status because of AI infrastructure spending under the Stargate project.
The question investors are asking: Is this a sustainable business model, or a game of financial musical chairs?
🔒 The rest of this analysis is for Future Digest premium subscribers
What you’ll discover in the full article:
✓ The make-or-break moment: Why 2026 is OpenAI’s inflection point—and the 3 scenarios that determine if they survive
✓ The advertising gambit: OpenAI just announced ads in ChatGPT after saying it was a “last resort”—what this reveals about their desperation
✓ The profitability playbook: The 4 levers OpenAI is pulling to reach profitability by 2029 (and why 2 of them won’t work)
✓ The competition bloodbath: How Google DeepMind, Anthropic, and Chinese labs are forcing OpenAI into an unsustainable spending war
✓ The investor exodus risk: Why some VCs are quietly getting nervous about the $500B-$830B valuations being discussed
✓ What happens if they fail: The 3 endgame scenarios—from Microsoft acquisition to catastrophic collapse
✓ What this means for YOU: Should you use ChatGPT knowing it might not exist in 3 years? What about building a business on OpenAI’s API?