Boom, Bubble, or Bust? How to Build a Resilient AI Business
Comparisons to the dot-com bust are common but this AI boom rests on short-cycle hardware. Frontier training chases each GPU generation, rendering last year’s chips economically obsolete for training even as they stay serviceable for inference — forcing relentless reinvestment. This dynamic is amplified by a unique, self-referential financial architecture where capital circulates between tech giants and their largest customers, masking true demand and subsidizing the unsustainable economics of computation. This dynamic is amplified by a unique, self-referential financial architecture where capital circulates between tech giants and their largest customers, masking true demand and subsidizing the unsustainable economics of computation.
Support our work by becoming a paid subscriber.
The very excitement over AI’s transformative potential has created a fragile market structure, one that is now likely to face a significant correction. The concentration of risk has spread beyond technology stocks into data center REITs, private credit vehicles, and retirement portfolios, meaning any repricing will ripple far wider than Silicon Valley. To navigate this landscape, it is crucial to understand the distinct pressures at play, from the sector’s fragile physical infrastructure and precarious capital flows to the stark realities of its technical performance in the enterprise and the volatile market dynamics that result.
Financial Architecture and Capital Flows
Beneath the surface of the AI boom lies a precarious financial architecture. Most troubling is the self-referential loop of capital, where tech giants invest in their largest customers, who then use the funds to purchase the investors’ products. This dynamic helps fuel a staggering gap between capital expenditure on infrastructure and the actual revenue generated from AI services. The true cost of computation is further obscured by complex financial arrangements that subsidize current prices, creating an illusion of sustainable unit economics.
These subsidies cannot persist indefinitely. When they end — whether from tightening credit markets or investor demands for profitability — providers may need to reprice services sharply upward, fundamentally altering economics for application builders. More troubling is how this risk has diffused beyond tech stocks. Real estate trusts now hold up to 22% of assets in data centers, while private credit funds increasingly integrated into retirement portfolios carry substantial AI exposure. A downturn in AI utilization or a sudden repricing could trigger credit stress in ostensibly safe investment vehicles, cascading beyond technology stocks into conservative portfolios.
Real demand stands on its own — why is Nvidia underwriting its customer base?
Infrastructure Vulnerabilities
Unlike the durable fiber overbuild of the 2000s, today’s core hardware is short-horizon: GPU platforms see step-function training gains each generation, rendering prior nodes economically outmoded for frontier work even as they’re repurposed for inference. The supply chain is also highly concentrated, with Nvidia controlling roughly 90 percent of the AI accelerator market and facilities clustering in a handful of regions. Perhaps most critically, the industry’s immense appetite for electricity is colliding with the hard limits of regional power grids, with utilities in key markets now offering only 80 percent firm power guarantees and emergency grid responses becoming routine.
This shortened payback window intensifies pressure on data center operators to generate revenue quickly from multi-billion-dollar investments before their hardware becomes economically obsolete. As we explored in a previous article on the physical limits of AI, these infrastructure vulnerabilities represent a fundamental shift. The success of AI applications is no longer just a matter of algorithmic innovation but is now directly tied to navigating real-world constraints, from power grid queues to the global concentration of specialized facilities.
Technical Reality and Performance
The enthusiasm for AI is running far ahead of its practical performance in corporate settings. A persistent gap has emerged between demonstrated potential and real-world reliability, which in turn contributes to disappointing financial returns for most enterprise initiatives. This disconnect is compounded by decelerating progress in frontier models and inflated benchmarks that obscure true capabilities.
The most pressing issue is the gap between capability and reliability. In production environments, models often generate plausible but incorrect outputs that require costly human oversight, a phenomenon that can negate productivity gains and even make experienced employees less efficient. This performance deficit is a primary driver of the sector’s struggle to demonstrate tangible value, with a striking number of enterprise AI projects failing to produce a positive return on investment. The enormous capital flowing into the sector is predicated on transformative productivity gains that, for many, have yet to materialize, raising fundamental questions about whether current valuations are grounded in economic reality.
Market Dynamics and Valuation
The market for AI sits at a crossroads between a durable technology boom and a classic financial bubble. The distinction hinges on whether fundamentals will eventually validate current valuations. A true bubble “pop” would involve more than falling stock prices — it would be marked by a severe, sustained contraction in capital investment, such as a 50 percent cut in hyperscaler spending that would constrain compute availability and alter the economics for application builders. This uncertainty is heightened by rapid commoditization of core technology and the likelihood that returns will concentrate among a very small number of firms.
The threat of margin compression sits at the heart of this risk. When DeepSeek, a Chinese open-source model, demonstrated performance comparable to proprietary systems at far lower cost, it erased a trillion dollars from U.S. AI valuations in a single day. If proprietary models cannot maintain defensible performance advantages, their pricing power will erode, making it harder to justify the enormous capital expenditures on infrastructure. This dynamic — where commoditization undermines margins, which in turn weakens the case for continued investment — represents the mechanism by which a boom could transform into a downturn.
Navigating the Correction: A Playbook for Builders
Our focus is on helping teams put AI and data to work. That mission requires building resilient systems and sustainable businesses that can thrive regardless of market cycles. For teams building with these technologies, surviving a correction and succeeding in the long term will depend on a few core principles.
Architect for Substitution. Design systems with abstraction layers that enable routing between model providers (OpenAI, Anthropic, open-source alternatives) and infrastructure vendors based on cost and performance. Keep prompts, fine-tuning workflows, and retrieval schemas portable to avoid vendor lock-in that could prove catastrophic if pricing models shift or suppliers face distress.
Engineer for Scarcity. Treat computation and energy as expensive, volatile resources rather than assuming infinite availability of cheap GPUs. Implement aggressive caching, model distillation, quantization, and task-appropriate model sizing. Design systems that maintain acceptable performance even under constrained or repriced compute access.
Human oversight to fix ‘plausible’ model errors can wipe out the productivity gains AI promised.
Measure Outcomes, Not Activity. Tie every AI initiative to concrete business metrics — claims processed per employee, support ticket resolution time, days sales outstanding reduced — rather than token consumption or model calls. Calculate total cost of ownership including the “invisible tax” of human verification time, and be ruthless about pausing projects that cannot demonstrate clear returns.
Build Proprietary Moats. As foundation models commoditize, defensible advantages come from domain-specific data assets, deep workflow integrations, unique distribution channels, and user experiences that reduce verification burden. Focus investment on capabilities that create high switching costs and deliver value competitors cannot easily replicate.
Monitor Market Signals. Track leading indicators that provide early warning of deteriorating conditions: hiring patterns at major AI labs, GPU spot pricing trends, hyperscaler capital expenditure guidance, and the quality of AI-related IPOs and acquisitions. Develop pre-planned playbooks for cost optimization and vendor failover that can be executed rapidly when signals weaken.
Design Spike-Aware Budgets. Forecast token consumption per task — including worst cases like asking the model to think through many steps or sample multiple solutions, retrieving lots of background passages, and calling several tools/services at once. Implement circuit breakers, rate limits, and tunable reasoning depth controls that allow shallow processing for routine tasks and deeper analysis only where accuracy justifies additional compute expense.
Develop Power Strategy. For teams managing infrastructure, favor regions with stable electrical capacity and design systems that can gracefully handle load-shedding events. Monitor energy-related clauses in vendor contracts and co-location agreements, as power constraints increasingly drive availability and pricing.
Prepare for Post-Correction Opportunities. Maintain playbooks to acquire discounted capacity, models, datasets, or talent when valuations compress. Teams with strong balance sheets and efficient operations can turn a correction into a competitive catalyst by moving quickly when others retrench.
