A venture capitalist argues that the feverish build-out of data centers and AI models isn’t just a typical tech bubble—it’s a unique historical moment where a collapsing price commodity, government policy, loose credit, and pure technological exuberance are all colliding. According to Paul Kodrosky, this “rational bubble” is creating a massive, overbuilt fabric of computing infrastructure, much like the railroad or dot-com fiber booms. While this will inevitably lead to a painful bust for many current players, including giants like OpenAI and NVIDIA, the resulting crash in the cost of “cognition” will unlock a new wave of innovation and productivity.

Kodrosky describes the current AI economy as an “enthusiasm” replaying a classic pattern: enormous capital expenditure goes into building foundational infrastructure, which rarely pays off for the builders themselves but creates a platform for future winners. The critical, unusual twist this time is that the core unit being sold—AI “tokens,” or units of computational output—is plummeting in price by 50-70% annually. This collapsing price undermines the business models of frontier AI companies, which must grow at astronomical, near-impossible rates just to maintain revenues. Meanwhile, the boom has metastasized deep into the wider economy, drawing in insurers, private credit, regional governments offering subsidies, and construction firms, meaning the coming correction will cause broad economic pain.

Despite predicting a shakeout, Kodrosky is profoundly optimistic about the underlying technology’s long-term impact. The rapid decline in token prices means cognition will become cheap and ubiquitous, acting as a “cheat code” for progress in fields like medicine, materials science, and robotics. This will help offset powerful demographic headwinds, such as aging populations in developed nations, by boosting productivity. The future beneficiaries won’t be today’s AI poster children but companies and sectors we can’t yet envision, which will leverage this cheap, powerful substrate to solve real-world problems and reshape the workforce.

Surprising Insights

• Consumer chat is a tiny, almost irrelevant part of the AI economy. Kodrosky asserts that all global consumer chat inference (e.g., using ChatGPT) could be handled by a single large data center, representing less than 5% of total AI compute demand. The boom is being driven by training runs and internal coding tools, not by end-user applications.
• The bottleneck is shifting from processing to memory. The analogy used is that NVIDIA’s powerful chips are like a super-fast highway tollbooth, but the real constraint is the number of lanes feeding it. As AI moves toward an “inference world,” high-bandwidth memory, not raw processing power, is becoming the critical scarcity.
• Major investors are already treating top AI models as commodities. Microsoft’s Satya Nadella discussing a “multi-model world” signifies that even a massive backer of OpenAI sees little differentiation between frontier models. This commoditization at such an early stage is a bad omen for the business models of pure-play AI companies.
• Demographic decline will partially offset AI-driven productivity gains. While AI could significantly boost productivity, an aging and shrinking workforce in developed nations will act as a strong countervailing force on economic growth, making the AI productivity surge more about compensating for losses than creating hyper-growth.
• The AI bubble uniquely combines four classic bubble ingredients. Kodrosky notes this is the first U.S. “bubble” that sits at the intersection of technology, real estate (data centers), loose credit, and direct government policy/geopolitical anxiety, making it particularly potent and dangerous.

Practical Takeaways

• Skeptically evaluate “new name” AI IPOs. As the bubble progresses, a wave of speculative AI companies will go public to soak up retail investor appetite. Their business models will be especially vulnerable to collapsing token economics, so extreme caution is warranted.
• Look for winners beyond the obvious infrastructure builders. History shows the fortunes are made by those who build on the new infrastructure (e.g., Netflix after the telecom build-out), not by those who lay the pipes. Focus on applications in robotics, enterprise software, and scientific research.
• Consider vocational paths less susceptible to AI displacement. With AI poised to transform white-collar cognitive work, skilled trades and vocational jobs with high demand, good pay, and physical components may offer more stable, long-term career paths.
• Recognize that “demand” can be a misleading signal. Much of the current insatiable demand for AI compute is for training and internal coding—activities that are costly, wasteful, and often done by the AI companies themselves. This is not necessarily validation of a sustainable external market.