Why Are Leading AI Companies Losing Billions?

TL;DR

Leading AI labs remain deeply unprofitable because training and serving large language models require enormous, variable compute spending that revenue has not covered. Ed Zitron argues that OpenAI’s reported $13.07 billion in 2025 revenue came against roughly $34 billion in costs, while planned data-center capacity would require several additional customers as large as OpenAI or Anthropic.

Transcript

Every AI company's unprofitable. The largest customers of AI compute are unprofitable. Their customers also unprofitable. OpenAI spent $34 billion to make 13.07 billion. Their net loss is just a little under 21 billion. The amount of revenue to substantiate and justify the data centers that are allegedly under construction means that we need two or... Read More

Key Insights

• OpenAI’s reported 2025 economics show $13.07 billion in revenue against roughly $34 billion in total costs and expenses, producing an operating loss of about $20.9 billion according to the financial figures discussed in the interview.
• AI compute demand is concentrated among OpenAI and Anthropic, which Zitron says account for approximately 89% of revenue among the leading companies while remaining heavily unprofitable and dependent on continued outside financing.
• The planned data-center expansion lacks enough identified customers, according to Zitron, because justifying more than 100 gigawatts of proposed capacity would require hundreds of billions of dollars in compute demand beyond OpenAI and Anthropic.
• Inference creates significant variable costs because greater customer usage requires more GPUs and additional compute capacity. OpenAI reportedly experienced worsening margins after purchasing last-minute compute, which is more expensive than capacity secured in advance.
• Historical comparisons with Amazon and Uber do not resolve the scale problem, Zitron argues. He estimates inflation-adjusted Amazon capital expenditure from 2003 through 2017 at roughly $53 billion to $55 billion, while OpenAI could raise $122 billion within six months.
• Hyperscalers receive AI-related growth partly because OpenAI and Anthropic spend heavily on their compute infrastructure. Zitron characterizes this as capital circulating through cloud providers before flowing to Nvidia, Broadcom, and Taiwanese server manufacturers such as Hon Hai and Foxconn.
• The AI investment thesis depends on finding a new source of technology-sector hypergrowth. Zitron argues that Microsoft, Google, and Amazon viewed AI APIs, enterprise products, consumer software, and GPUs as possible successors to earlier growth engines such as smartphones, cloud computing, and AWS.
• Cheaper model operation has not arrived despite the use of Blackwell hardware and Anthropic’s years of TPU usage, according to Zitron. He argues that neither hardware advances nor alternative accelerators have yet produced a credible path toward sustainable operating costs.
• More videos from this channel:
  • How Could Insurance Trigger a Credit Crisis?
  • How Did the 1873 Financial Crisis Reshape Money?
  • How Can the Fed Safely Shrink Its Balance Sheet?
• More from this channel: The Monetary Matters Network

Questions & Answers

Q: Why are leading AI companies still unprofitable?

Leading AI companies remain unprofitable because their training, research, sales, and inference costs exceed the revenue customers generate. OpenAI reportedly earned $13.07 billion in 2025 but incurred about $34 billion in total costs and expenses. Zitron also argues that rising usage does not automatically solve the problem because serving additional requests requires more compute and can increase variable costs.

Q: How much did OpenAI reportedly lose in 2025?

OpenAI reportedly recorded an operating loss of about $20.9 billion in 2025, based on $13.07 billion in revenue and approximately $34 billion in total costs and expenses. Zitron rejects a lower loss figure of $8 billion mentioned in other commentary, arguing that the detailed expense figures support a loss of roughly $21 billion rather than the smaller estimate.

Q: What made up OpenAI’s reported 2025 expenses?

OpenAI’s reported 2025 expenses included $7.5 billion categorized as cost of revenue, $19.1 billion for research and development, $5.73 billion for sales and marketing, and $1.57 billion for general and administrative costs. Together, these categories produced approximately $34 billion in total costs and expenses, compared with reported revenue of $13.07 billion.

Q: Why does AI inference create a difficult cost structure?

AI inference is compute-intensive, so increased customer usage requires additional processing capacity rather than producing nearly cost-free incremental revenue. Zitron says unexpected usage can force a company to buy last-minute compute at higher prices than capacity purchased in advance. The compute seller has leverage when the buyer urgently needs a large amount, which can cause margins to worsen as demand increases.

Q: How do OpenAI’s losses compare with Amazon and Uber?

Zitron argues that the comparisons are not equivalent. He says Uber burned roughly $32 billion, with substantial spending devoted to marketing and subsidized rides. He estimates Amazon’s inflation-adjusted capital expenditure from 2003 through 2017 at about $53 billion to $55 billion across the entire company. OpenAI, by contrast, could raise $122 billion within a six-month period if promised funding closes.

Q: Why might planned AI data centers lack enough demand?

Zitron says more than 100 gigawatts of data centers are planned, with costs of approximately $12.5 million to $15 million per megawatt. Supporting that construction would require hundreds of billions of dollars in AI compute demand beyond OpenAI and Anthropic. He argues that two or three additional customers of comparable size would be needed, but no such customers currently appear in the discussion.

Q: Why are major technology companies investing so heavily in AI?

Zitron’s theory is that major technology companies lack another clear source of hypergrowth comparable to AWS, the iPhone, smartphones, or cloud computing. AI was expected to support new startups, enterprise software additions, consumer products, and a GPU-based infrastructure business. Microsoft, Google, and Amazon therefore treated AI as a possible next growth engine, even though its underlying economics remain unsettled.

Q: What evidence would challenge Zitron’s AI bubble argument?

Zitron says his position would be challenged by the appearance of hundreds of billions of dollars in sustainable compute demand outside OpenAI and Anthropic. That demand would need to justify the proposed data-center capacity while supporting profitable operations. He also points to the need for genuinely cheaper inference and a credible path showing how revenue can grow faster than compute and operating expenses.

Summary & Key Takeaways

• Ed Zitron argues that the AI industry’s economics remain unproven despite more than a trillion dollars in spending. Most AI companies generate limited revenue, while OpenAI and Anthropic dominate demand and continue losing billions. He sees the continuing debate over return on investment as evidence that projected demand has not materialized.

• OpenAI reportedly generated $13.07 billion in 2025 revenue while recording approximately $34 billion in total costs and expenses. Those expenses included $7.5 billion in cost of revenue, $19.1 billion in research and development, $5.73 billion in sales and marketing, and $1.57 billion in general and administrative spending.

• Zitron says AI differs from traditional software because inference creates substantial variable costs that can rise with usage. Unexpected demand may require expensive last-minute compute purchases, weakening margins. Meanwhile, planned data centers would need hundreds of billions of dollars in compute demand beyond OpenAI and Anthropic, but he sees no such customers.