AI's Circular Money Machine: Bubble or Breakthrough?
Summary of the video “Is AI’s Circular Financing Inflating a Bubble?” by Patrick Boyle.
Tech giants are investing billions in each other's AI companies and infrastructure, creating circular financing loops reminiscent of Japan's post-war keiretsu. While fundamentals are stronger than past bubbles, the system depends on massive electricity buildout, uncertain monetization, and real demand that may never materialize.
The Circular Financing Loop
Nvidia's Self-Reinforcing Ecosystem
Nvidia invests in its customers (like Coreweave), who buy Nvidia chips and rent compute to other companies that also buy Nvidia chips. Nvidia pledged up to $100 billion to OpenAI, which in turn buys millions of Nvidia AI graphics cards. This creates a closed loop where the same dollars circulate through the system, making it difficult to distinguish real demand from subsidized purchases.
OpenAI's Multi-Stakeholder Web
OpenAI announced a $300 billion cloud infrastructure agreement with Oracle, a $10 billion custom chip partnership with Broadcom, and strategic alliances with memory suppliers. According to UBS analysts, OpenAI's memory commitments alone account for half of the world's current capacity. These deals position OpenAI as a central hub where suppliers have vested interests in its success.
AMD's Stock-as-Payment Model
OpenAI agreed to buy tens of billions in AMD chips and received the right to buy 10% of AMD stock for 1 cent per share, contingent on AMD hitting share price targets. When OpenAI announces partnerships, the stock price rises, effectively reimbursing OpenAI for its chip purchases. This structure allows companies to pay for infrastructure with capital gains rather than cash.
Amazon-Anthropic Closed Loop
Amazon invested over $8 billion in Anthropic, which committed to using Amazon as its primary cloud provider. Anthropic trains and runs models on Amazon's custom AI chips, rents compute from AWS, and integrates Claude into Amazon Bedrock. Amazon is effectively funding a company that will use its own infrastructure and help sell its services.
Google's Anthropic Entanglement
Google invested $3 billion in Anthropic and now provides access to up to 1 million TPUs worth tens of billions, bringing over a gigawatt of compute online by 2026. Google is simultaneously a major investor, infrastructure provider, and chip supplier. Anthropic claims it's pursuing a multicloud strategy, but it is deeply entangled with all three largest US cloud providers, each with financial interest in its success.
Elon Musk's Corporate Incest
XAI acquired Twitter (now the everything app) which supplies real-time data to Grok. Tesla uses Grok in its cars and robots. Musk owns a majority stake in XAI and a minority stake in Tesla, and wants Tesla shareholders to invest in XAI. The companies are financially and operationally interdependent, creating conflicts of interest and circular value flows.
Historical Parallels: Keiretsu and Chaebol
Japan's Keiretsu Model
Post-war Japan built large industrial groups around banks and trading houses where companies took stakes in each other and coordinated supply chains. These structures were designed for survival in capital-constrained economies but were criticized for obscuring financial risk, misallocating capital, and propping up uncompetitive firms. When Japan's asset bubble burst in the 1990s, the tangled web of crossholdings made it nearly impossible to unwind bad bets.
South Korea's Chaebol System
South Korea's chaebol followed a similar pattern to Japan's keiretsu but with families in control rather than banks. Like keiretsu, they created tight financial relationships between businesses to secure supply chains, but suffered from the same problems of obscured risk and capital misallocation.
AI Giants Building Similar Structures
Today's AI companies are assembling webs of mutual dependence similar to keiretsu and chaebol. The question is whether they are building a fragile structure that looks stable from outside but depends on constant new capital influx to keep operating. Unlike Japan and Korea, AI giants are not capital-constrained, but they are creating similar interdependencies.
The Electricity Constraint
Stargate's Massive Power Demand
OpenAI's Stargate project announced in January is a $500 billion plan to build 10 gigawatts of AI data center capacity across the US. A typical nuclear power plant produces 1 gigawatt and powers approximately 1 million US households. Therefore, 10 gigawatts would power about 26 million average Americans (2.6 people per household).
OpenAI's Full Infrastructure Buildout
OpenAI has committed to building 23 gigawatts of new data center capacity at a cost of well over $1 trillion. This would require 23 nuclear power stations to power. The 6 gigawatt deal with AMD alone is enough energy to power Singapore for a year.
Industry-Wide Power Requirements
McKinsey estimates $5.2 trillion in capex will be needed by 2030 for chips, data centers, and energy for AI workloads alone. Traditional IT data centers require an additional $1.5 trillion, totaling nearly $7 trillion in projected data center spending over 5 years. All major US tech firms plus companies in China and elsewhere are building out AI capability simultaneously.
Grid Connection Bottleneck
In Texas, where several Stargate sites are planned, electricity demand is rising so quickly that operators are installing on-site gas turbines and exploring nuclear partnerships to avoid waiting for grid hookups. The XAI data center in South Memphis is running gas turbines with no emissions controls and no permits, creating enough pollution that the area leads Tennessee in asthma hospitalizations.
Nuclear Permitting Crisis
The last new nuclear reactor in the US took more than a decade to complete and came online in 2024. There are no new nuclear sites currently under construction. Permitting for solar and wind has been tightened and tariffs have raised costs. Even fast-tracked projects face multi-year delays. The chips might arrive on schedule, but the electricity probably won't.
Revenue vs. Spending Reality
The Profitability Gap
McKinsey forecasts $5.2 trillion in capex for chips, data centers, and energy over 5 years. Bain estimates we need $2 trillion in annual revenue from AI companies to justify that spending. OpenAI has about $13 billion in revenues today and is essentially a money furnace. Anthropic is a smaller money furnace. Nvidia is profitable but not $100 billion profitable. The question becomes who will pay for all of this.
OpenAI's Debt Financing Shift
OpenAI secured a $4 billion revolving credit line from a consortium of banks. This is unusual for high-growth tech firms, which historically raised capital through equity. The shift from equity to debt and from public listing to private investment is happening across the sector. Data center providers are borrowing against assets like GPU racks that might quickly become obsolete.
Low AI Adoption Success Rate
OpenAI claims 700 million weekly users, but only 5% are paying customers. Most revenue comes from enterprise contracts, not individual subscriptions. McKinsey reports that the success rate of AI pilot projects among business users is less than 15%. We are not seeing the mass AI-driven layoffs predicted years ago; labor data shows no clear relationship between AI deployment and employment trends except for freelance graphic designers and copywriters.
GPU Rental Market Stress
Collapsing Chip Rental Prices
The price to rent Nvidia's B200 chip dropped from $3.20 per hour to $2.80 per hour in just a few months. Older A100 chips are available for as little as 40 cents per hour, below break-even for many operators. A cluster of eight chips costing $200,000 five years ago with a 5-year useful life needed to generate $4 per hour in rental fees to break even. A100 average rental fell from $2.40 per hour in 2020 to $1.65 per hour today.
Risk of Stranded Infrastructure
If demand for AI infrastructure does not materialize as expected, data centers built for 5 years of peak usage might sit half empty. Many pandemic-era GPUs could end up in liquidation without ever earning back their cost. There is historical precedent: telecom firms in the early 2000s built fiber optic networks that were never used, and 19th-century railways laid track to nowhere.
Systemic Risk Beyond Startups
If demand for infrastructure does not materialize, the fallout will not be limited to a few startups. It will hit lenders, landlords, and public utilities that signed up to support the boom without necessarily understanding the bet they were making.
Demand Quality and Valuation Questions
Nvidia's Valuation Depends on Real Demand
Nvidia's stock market valuation is based on the idea that demand for its chips is massive and will keep rising indefinitely. However, if Nvidia's biggest customers are also its investment targets and those customers are using Nvidia's money to buy Nvidia's products, then the margins may not be what they seem. The circularity makes it difficult to assess the quality of revenues.
OpenAI-Nvidia Deal Scale
The OpenAI-Nvidia deal should account for around 13% of Nvidia's expected 2026 revenue according to UBS, assuming the full gigawatt deployment goes ahead. That would mean $50-60 billion in total capital investment with Nvidia receiving $35 billion back. Nvidia says it might reinvest $10 billion into OpenAI, but only if monetization keeps pace, a performance-based approach that gives room to back out.
AI Output Quality Issues
For a technology supposed to make scientific breakthroughs and cure diseases, a surprising amount of AI output looks like slop. OpenAI's Sora can generate realistic video, but viral clips have been deepfakes of Taylor Swift and SpongeBob in Breaking Bad. Elon Musk's XAI built an anime girlfriend chatbot. However, real breakthroughs exist: the 2024 Nobel Prize for Chemistry went to Google DeepMind researchers for AI-powered protein folding, which is already being used to combat cancers.
Bubble Assessment and Competitive Dynamics
Stronger Fundamentals Than Past Bubbles
The mega-cap US tech firms are expected to generate over $200 billion in free cash flow next year alone, even after capex. That is enough to fund infrastructure buildout without leaning hard on debt or requiring new financing. Balance sheets are strong and earnings are real. This is not the same as the telecom bubble. Valuations are elevated but not absurd: in the late 1990s, internet stocks traded at 60 times forward earnings; today's AI leaders are closer to 35 times.
Winner-Take-All vs. Competitive Market
High private market valuations for OpenAI, XAI, and Anthropic only make sense if one dominates the space. However, DeepSeek earlier this year showed models can be replicated quickly and cheaply. Elon Musk's rapid Grok deployment showed the same. If models are all roughly the same, the market may not reward any one player. Instead of a big winner and monopoly, we might see a very competitive market where none have pricing power.
Uncertainty About Who Profits
It might not be the model builders who profit. It might be the businesses that use the models. AI could end up boosting productivity across the economy while the labs themselves struggle to monetize. The outcome is still really uncertain, and it is not clear who wins or if anyone does.
Notable quotes
An extension cord plugged into itself. It won't power your home appliances. You just need outside energy to get things going. — Patrick Boyle
The same dollars are circulating through the system, possibly inflating purchase orders and revenue projections. It's hard to tell where the demand ends and the subsidy begins. — Patrick Boyle
The chips might arrive on schedule. The electricity probably won't. — Patrick Boyle