Ed Zitron: The AI Bubble is Bleeding Cash, Here Are The Receipts

The Unprofitable Landscape of AI Companies [00:00]

• The AI industry is characterized by widespread unprofitability, with even the largest customers of AI compute operating at a loss.
• OpenAI's financial figures show a staggering loss of nearly $21 billion on revenue of $13.07 billion, despite spending $34 billion.
• The projected demand for AI compute to justify planned data centers requires two to three more OpenAI and Anthropic-sized customers, which do not currently exist.
• There's a sentiment that major AI companies lack a clear strategic plan for profitability.

"Every AI company's unprofitable. The largest customers of AI compute are unprofitable. Their customers also unprofitable."

The Debate Over AI's Return on Investment [00:47]

• The ongoing debate about the Return on Investment (ROI) for the enormous sums being spent on AI indicates a fundamental issue with its economic viability.
• If AI were a robust industry with the projected Total Addressable Market (TAM), such a debate would not be necessary.
• Many AI companies struggle to exceed $100 million in annual revenue, with OpenAI and Anthropic dominating the market share (89%) but remaining significantly unprofitable.
• Companies allegedly use "wacky accounting" to reconcile their losses.

"So, trillion plus dollars in the fact we're still debating the ROI kind of says everything."

Financial Discrepancies and Historical Comparisons [02:30]

• Historical comparisons are drawn to companies like Uber and Amazon during their growth phases. Uber burned $32 billion, while Amazon Web Services (AWS) incurred approximately $53-55 billion in total capital expenditure between 2003 and 2017.
• OpenAI's fundraising is immense; in six months, they aimed to raise $122 billion, and Anthropic $95 billion. OpenAI raised $40 billion last year, and Anthropic $16.5 billion.
• These figures dwarf historical comparisons, with OpenAI's single round raising more than half of Amazon's total capex for its retail and AWS operations combined.
• The argument is that these companies do not make economic sense and that comparisons to past tech giants are misleading.

"The answer is incomparable. Like Amazon's capex for its retail operation on top of AWS was less than half of what OpenAI has raised in one round this year."

The "Rockcom Bubble" and Lack of New Ideas [05:06]

• A theory called the "rockcom bubble" suggests the tech industry is out of revolutionary ideas, lacking a new AWS, iPhone, or cloud computing equivalent.
• AI was intended to be the panacea, driving new startups, enterprise bolt-ons, and consumer software.
• Hyperscalers like Microsoft, Google, and Amazon saw AI as an opportunity to create the next generation of growth, as they have no other significant business lines showing comparable potential.
• The growth seen from AI is largely attributed to the massive spending by OpenAI and Anthropic on compute, which primarily feeds back into hardware companies like Nvidia and Broadcom.

"And AI was meant to be the panacea. It was meant to be AI models as in via the API were meant to be the thing that created the next generation of startups that created the next generation of enterprise bolt-ons so that the software industry could grow."

The Exorbitant Costs of AI Compute [09:49]

• The current generation of AI hardware, like Blackwell, is not leading to significant cost reductions, contradicting expectations.
• The costs associated with AI, particularly inference, are extremely compute-intensive and create significant variable costs for both customers and AI companies.
• AI companies are forced to buy compute last minute at higher prices, impacting their margins.
• Unlike software businesses with low variable costs, AI operations are inherently expensive to launch, run, and maintain.

"Where's the cheapness happening? Whatever happened to those OpenAI asex? Anthropic has been using TPUs for years. They're not working out their costs either."

OpenAI's Inflated Sales and Marketing Costs [11:30]

• OpenAI's reported sales and marketing cost of $5.73 billion is unusually high, exceeding that of major consumer brands like Coca-Cola.
• Evidence suggests a significant portion of this cost may be attributed to credits from Microsoft or other non-traditional advertising expenses, rather than direct consumer outreach.
• This inflated cost category raises questions about how OpenAI accounts for its expenses and potentially shifts costs to appear more favorable.

"So that sales and marketing cost is the thing I want people to really pay attention to because it's very very strange."

Questionable Accounting and Cloud Credits [15:38]

• The Financial Times reported alleged losses of $8 billion for OpenAI, a figure disputed by critics who point to undisclosed cloud computing credits from Microsoft that should be accounted for as real costs.
• Documents reviewed do not explicitly mention these credits, but do note substantial stock-based compensation and share-based compensation for compute provided by related parties.
• The practice of trading stock for compute is highlighted as unusual.

"If I get billions of dollars worth of free stuff, I think it it would be honest for me to include that as a cost because it's you can't normally just count that as oh my god, normally yeah, normally I'm going to get tons of free stuff, you know, that just that's just not how it works."

The "Magic" of Accounting and Unrealized Costs [20:18]

• OpenAI's accounting practices are described as bizarre, with substantial costs "magically disappearing" or being attributed to subsidiaries, especially during the conversion from non-profit to for-profit status.
• Specific instances include $3.7 billion in costs vanishing in 2024 and $17.8 billion in 2025, which accountants suggest were merely shifted, not eliminated.
• Warrant liabilities, a non-cash expense, are also noted as contributing to reported net losses, further complicating the financial picture.

"So, the net loss attributable to non-controlling members capital magic away $3.7 billion worth of costs in 2024. Where did they go?"

The unsustainable Demand for Compute [44:39]

• The primary demand for AI compute comes from unprofitable AI companies like OpenAI, Anthropic, and Meta, or hyperscalers providing compute to these entities.
• Even when peeling away these major players, the remaining revenue for AI compute is relatively small, suggesting that the general demand for inference and training is not as high as perceived.
• The argument is that the ROI conversation is pushing back against high costs, and the shift towards open-source models, which use less compute, further questions the need for massive data center expansions.

"Just put aside my bare case for a second, just on very on a very basic level, the demand and natural need for AI compute is not that high. Like it's just not there."

Meta's Disconnected AI Strategy [47:18]

• Meta's AI strategy is described as lacking coherence, with Mark Zuckerberg frequently shifting focus from one area to another, such as from the metaverse to AI.
• While Google and Microsoft have clearer infrastructure and integration strategies, Meta's approach seems to involve acquiring vast amounts of AI hardware without a clear plan for monetization or significant profit increase.
• The company's history of acquiring or copying successful ideas from other platforms suggests a lack of original innovation.

"This is a company that has not had a new idea since Facebook. Instagram, they bought stories, they stole from Snapchat. Reals, they stole from Tik Tok. Every idea they've had, they've stolen. All of this is to say is that Meta doesn't have an AI strategy."

The Limitations of Agentic AI and Future Predictions [38:03]

• Large Language Models (LLMs) are deemed incapable of reliably performing multiple deterministic functions, making them unsuitable for tasks requiring precision, especially those involving money.
• Despite claims of agentic AI capabilities, there's no evidence of actual dollars being spent or transactions completed by these agents.
• The high cost of developing and running agentic AI, combined with its unreliability due to model hallucinations and prompt misinterpretations, makes its widespread adoption unlikely.
• The vision of agentic AI replacing professions like lawyers or accountants is dismissed as unrealistic.

"I think it's wrong I think the everything you're talking about there involves multiple different deterministic functions. Large language models cannot do those."

The Reckoning of the AI Bubble [01:05:15]

• The AI bubble is predicted to pop around 2027, possibly sooner if major tech ventures like SpaceX falter.
• The inflated expectations surrounding Nvidia's performance are a key concern, as their circular financing and ambitious sales targets prop up the market.
• The AI bubble is not driven by real revenue or ROI but by speculative debt and private credit, fueled by a "vibe shift" rather than fundamental economic value.
• The increasing debt burden on hyperscalers and the commoditization of AI businesses are cited as reasons for the market's eventual correction.

"It really comes down to the fact that because not real revenue because real revenues from these companies are not and actual ROI is not what's making the AI bubble inflate. It's going to come down to a vibe shift and it's already begun."

The Misconception of AI as Progress [06:47]

• A significant misconception among AI boosters is the belief that AI represents progress. In reality, it is seen as a "flattening" or "averaging out" of capabilities.
• The industry sells a future potential rather than current demonstrable value, with conversations consistently occurring in the future tense.
• AI boosters often conflate semiconductor sales and debt-fueled AI capex with genuine AI demand, mistaking speculative investment for intrinsic value.
• Companies are accused of treating their boosters with contempt by providing evasive answers and manipulating financial reports, fostering a cult-like mentality rather than genuine progress.

"I actually think the biggest misconception they have is that AI is progress. That AI is a progressive thing when what it actually is is a flattening of everything."