A Fortune 500 CFO told me about the procurement question that stalled her team’s AI rollout for two months. The vendor’s product was excellent. The vendor’s finances were not: her analyst had pulled the numbers, and the company she was about to sign a three-year contract with was losing more money per year than her own company’s entire market capitalization. “Nobody at the vendor would discuss it,” she said. “Every answer was about the growth story.”
She is asking the question everyone in this market eventually asks, and the numbers behind it have become extraordinary. OpenAI’s own internal forecast projects a $14 billion loss for 2026. Anthropic’s draft IPO prospectus, reported by the Wall Street Journal in October, shows an $80.7 billion net loss for the first half of the year alone. HSBC calculates that the industry’s flagship company cannot fund its operations from revenue at any point before 2030. The Financial Times’ Alphaville described the situation with a phrase that deserves immortality: “a money pit with a website on top.”
To be fair to every company on this list, losses at infrastructure scale are not automatically failure. Amazon lost money for years while building something worth trillions, and the revenue growth rates here are the fastest in tech history. But honesty cuts both ways, and these eight companies are burning cash at rates their marketing departments would prefer you not examine too closely. Here are the real numbers, what is driving them, and what it means if any of these names sits in your vendor stack.
The 8 at a glance
| Company | The loss | The driver |
| OpenAI | $14 billion projected for 2026 | Compute commitments revenue cannot cover |
| Anthropic | $80.7 billion net loss in H1 2026 | $72 billion in non-cash charges, plus real burn |
| xAI | Reported $1 billion-plus per month | Colossus infrastructure and the talent war |
| CoreWeave | Debt-funded expansion, 41% October sell-off | Borrowing against depreciating GPUs |
| Zhipu | Listed lab down 25% in October | China’s price war meets public markets |
| MiniMax | Listed lab down 36% in October | Global competition at unfavorable economics |
| Safe Superintelligence | $32 billion valuation, no product | Pure research burn |
| Oracle | Credit rating under pressure | $300 billion Stargate bet on one customer |
1. OpenAI: the $14 billion year

The scale of OpenAI’s position is difficult to overstate, in both directions. Revenue passed $20 billion annualized at the end of 2025, reaching that mark faster than Google or Facebook did. And the company’s own documents, reported by The Information, project a $14 billion loss for 2026, roughly triple the prior year’s estimate, on the way to cumulative losses of $44 billion before a projected 2029 profit. HSBC’s semiconductor team went further in November: modeling $792 billion in infrastructure costs through 2030 against revenue reaching $213 billion, the bank concluded OpenAI cannot fund operations from revenue at any point in the forecast window, leaving a $207 billion funding shortfall.
The pressures compound each other. ChatGPT’s web traffic share fell from 86.7% to 64.5% in twelve months as Gemini surged, enterprise share dropped to 27% against Anthropic’s 40% per Menlo Ventures, and December’s internal “Code Red” delayed revenue products to defend the core product. OpenAI filed a confidential S-1 in June and is reportedly seeking up to $100 billion in new funding at an $830 billion valuation. The bull case is real: 92% of the Fortune 500 uses ChatGPT. So is the math. Both can be true, and buyers should price both. We tracked the valuation side of this story in our report on OpenAI’s $852 billion valuation.
2. Anthropic: the $80 billion loss that is mostly accounting
Anthropic’s draft IPO prospectus produced the year’s most alarming headline number: an $80.7 billion net loss for the first half of 2026, up from $13 billion a year earlier and, by the Wall Street Journal’s accounting, more than all other major companies combined. The detail that matters is buried in the second paragraph of every responsible report: $72 billion of that loss is non-cash accounting charges from convertible notes repricing as the company’s valuation soared. Strip those out and Anthropic is still burning heavily, on compute and the talent war, but its revenue trajectory (from $1 billion ARR at the end of 2024 to projections of $20 to 26 billion for 2026) is the strongest in the industry, and its enterprise momentum is what triggered OpenAI’s Code Red.
The honest verdict: Anthropic is losing money faster than its fans admit and slower than its headline loss suggests. Its IPO, expected within months, will be the first real test of how public markets price that combination.
3. xAI: the billion-dollar month
Elon Musk’s AI venture has never published financials, which makes every number a reported estimate, and the estimates are severe: reporting through 2025 put the burn above $1 billion per month, driven by the Colossus data center buildout in Memphis and compensation packages designed to poach from OpenAI and Google. The company’s absorption into SpaceX’s structure ahead of the June IPO changed the question from “can xAI fund itself” to “can SpaceX’s public shareholders tolerate the burn.” The June IPO’s success suggests they can, for now. The Cursor acquisition that followed suggests the spending is accelerating, not stabilizing.
4. CoreWeave: the neocloud debt machine
CoreWeave’s business model is simple to describe and stressful to model: borrow billions against NVIDIA GPUs, rent the compute to AI labs, and hope the chips hold value longer than the loans. When October’s AI sell-off hit, CoreWeave, Nebius, and IREN fell between 35% and 41% in weeks, a public verdict on the most leveraged corner of the AI economy. The company is not failing; its revenue growth is real and its customer list is elite. But it is the clearest example of the industry’s foundational bet: that demand for compute will outlast the depreciation schedule of the hardware providing it. If inference prices keep falling at their current rate, that bet gets harder every quarter.
5. Zhipu: China’s AI champion meets the public market

Zhipu AI, the first major Chinese lab to list publicly, dropped 25% in October’s sell-off. Its underlying problem is structural: China’s AI price war, where DeepSeek-level models are effectively free and inference margins approach zero, punishes any lab trying to charge real money for tokens. Zhipu remains technically strong and government-supported, which in China is its own form of runway. But its public listing did something private funding never did: it put a daily price on the distance between China’s AI capabilities and China’s AI economics.
6. MiniMax: global ambitions, unfavorable math
MiniMax, the second listed Chinese lab, fell 36% in the same sell-off, the market’s judgment on the hardest position in AI: competing globally from a domestic market where prices have collapsed, against American labs with ten times the capital. Its Hailuo video models are genuinely competitive, and its consumer products have real traction. The losses reflect the cost of staying in a game where the ante rises every quarter. Of the eight companies here, MiniMax faces the purest version of the industry’s existential question: what happens to the lab that is very good but not the best, in a market with no prize for second place?
7. Safe Superintelligence: the $32 billion research project
Ilya Sutskever’s post-OpenAI venture represents the industry’s most honest form of loss: a company with no product and no revenue, valued at a reported $32 billion, employing some of the world’s best researchers to pursue safe superintelligence directly. Every dollar it spends is a loss by definition, and its investors know it. SSI belongs on this list not because it is failing but because it is the purest expression of the market’s logic: capital is so convinced AI will be worth trillions that it will fund a company whose entire business plan is “we will tell you when we get there.” If that bet is right, SSI is the bargain of the century. If the timeline stretches, it is the most expensive research sabbatical ever funded.
8. Oracle: the enabler betting its credit rating

Oracle is the only profitable giant on this list, and it belongs here because of what its balance sheet is absorbing. The company’s $300 billion Stargate commitment makes it the landlord of OpenAI’s ambitions, funded substantially with debt, and the strain is showing in public: Barclays downgraded the stock to underweight with a warning that continued spending could push its credit rating toward junk, credit default swaps on Oracle debt spiked sharply enough to alarm Morgan Stanley’s wealth management CIO, and January’s revelation that 45% of Microsoft’s $625 billion cloud backlog is tied to OpenAI briefly erased $440 billion of Microsoft’s market value. The circularity is the story: Microsoft invests in OpenAI, OpenAI commits to Oracle’s cloud, Oracle borrows to build it, and everyone books the same dollar as good news. When the music slows, the companies holding the debt, not the equity, absorb the first loss.
What the burn rates actually mean
Three honest conclusions survive the numbers. First, this is what infrastructure revolutions look like from inside: railroads, electrification, and fiber optics all destroyed more capital than they returned in their first decade, and the companies building them were called money pits by the Alphavilles of their day. The losses are real, and so is the historical pattern.
Second, the risk is concentrated and circular. A small number of companies (OpenAI at the center, Microsoft, Oracle, NVIDIA, and SoftBank around it) have cross-invested, cross-committed, and cross-sold to each other to a degree that makes independent failure unlikely and correlated failure the actual scenario to model. Third, for buyers the lesson is vendor diligence, not avoidance: ask the question the CFO from our opening asked, demand contractual protections for vendor disruption, and favor tools that let you switch models, because the companies most likely to raise prices in 2028 are the ones losing money fastest today. The AI boom is real. So is the bill. The professionals who thrive in this market are the ones who read both.
Frequently asked questions
Is OpenAI profitable?
No. OpenAI’s own internal forecast projects a $14 billion loss for 2026 and cumulative losses of $44 billion before a projected 2029 profit. HSBC models a $207 billion funding shortfall through 2030, concluding the company cannot fund operations from revenue at any point in its forecast window, despite revenue projected to exceed $213 billion by 2030.
Why did Anthropic lose $80 billion in six months?
Anthropic’s draft IPO prospectus shows an $80.7 billion net loss for the first half of 2026, but $72 billion of that is non-cash accounting charges from convertible notes repricing as its valuation rose. The underlying operating burn is real but far smaller, and its revenue is growing faster than any competitor’s, from $1 billion ARR in late 2024 to a projected $20 to 26 billion for 2026.
Which AI company has the worst finances?
By absolute burn, OpenAI’s projected $14 billion 2026 loss leads. By leverage, CoreWeave’s debt-funded GPU expansion carries the most structural risk. By dependence on a single customer’s promises, Oracle’s $300 billion Stargate commitment is the most concentrated bet in the market.
Should businesses avoid vendors that are losing money?
Not avoid, but diligence them. Ask about funding runway and path to profitability, negotiate contractual protections for acquisition or shutdown, and favor tools with model flexibility so a vendor’s financial stress cannot become your operational crisis. The biggest losers are also, in several cases, the best products.
Will AI companies ever become profitable?
History says some will: infrastructure buildouts from railroads to cloud computing produced enormous early losses followed by durable profits for the survivors. The open question is which tier captures the value, models, infrastructure, or applications, and current pricing trends (inference costs falling 280x for equivalent performance) pressure the model layer hardest.











-white.png)
-white.png)

