Contrary to the belief that Big Tech's AI spending is a rational bet on the future, the data suggests otherwise. Tether's CEO recently warned of four fissures in the infrastructure boom. I have spent 45 years—most of them dissecting complex systems—and I see the same patterns that led to the Tezos formal verification debacle in 2017 and the Terra collapse in 2022. The proof is in the logic, not the promise. This is not a market cycle; it is a capital allocation error dressed in algorithmic hype.
Context: The Four Mismatches
The warning, delivered by Paolo Ardoino, identifies four structural cracks: the mismatch between hardware lifespan and capital commitment, the chasm between cost and revenue, the timing gap between capital expenditure and profit realization, and the erosion of pricing power via open-source alternatives. These are not minor frictions. They are the foundational flaws of an industry that has spent over a trillion dollars on infrastructure before proving a single sustainable business model.
Let me ground this. The AI boom is unprecedented in scale. Morgan Stanley projects $500 billion in capital expenditure by 2028. JP Morgan sees cumulative spending exceeding $1 trillion by 2030. Yet revenue growth from AI services remains anemic. The market is paying for a future that may never materialize. I modeled similar dynamics in 2020 when I audited Yearn Finance's vault strategies. The code was elegant, but the assumption of constant liquidity depth was mathematically unsound. The same assumption underlies current AI investments: that demand will grow exponentially to absorb the capacity. It will not.
Core: Systematic Teardown of the Four Cracks
Crack One: Capital Commitment vs. Hardware Lifecycle
AI chips—GPUs, TPUs, and custom ASICs—have a useful life of three to five years for cutting-edge training. Yet the financial structures supporting them (debt, leases, depreciation schedules) assume a ten-year recovery window. This is a mismatch baked into the balance sheets. In a worst-case scenario—which I always model—a single architecture shift (like the transition from transformer-based to state-space models) could render $200 billion worth of hardware obsolete overnight. Complexity is the camouflage for incompetence. The industry hides this risk behind rosy depreciation assumptions.
I saw this exact pattern in 2021 when I analyzed the Bored Ape Yacht Club metadata storage. The IPFS pinning service was a single point of failure disguised as decentralization. The community called me a bot for pointing it out. Today, the same dynamic plays out at macro scale: the AI infrastructure is centralized on a few chip architectures and cloud providers, and any disruption cascades.
Crack Two: Cost vs. Revenue
The unit economics of AI are broken. Major providers are subsidizing inference costs to capture market share. Each ChatGPT query costs more than the revenue it generates. This is not a temporary promotional period; it is a structural deficit. In my 2024 analysis of EigenLayer's restaking slashing conditions, I identified a vector where adversarial actors could exploit latency to double-slash validators. The team called it low probability. I called it inevitable entropy. The same principle applies here: when the subsidy ends, demand will contract, and the cost-revenue gap will widen.
The financial numbers confirm this. The average AI API price has dropped 80% in two years, driven by open-source competition. Yet the cost of compute has only declined 20%. The wedge is being filled by venture capital and corporate cross-subsidies. This is unsustainable. Yields are just risk wearing a tuxedo.
Crack Three: Capital Expenditure Timing vs. Profit Realization
Capital expenditure for AI is front-loaded. Data centers, power infrastructure, and chip purchases require massive outflows before any revenue materializes. The typical construction cycle is 18 to 36 months, and then another 12 months for hardware installation. By the time the facility is operational, the hardware may already be obsolete. This is a classic asset-liability mismatch. The 2022 Terra collapse taught me that systems requiring infinite growth to maintain stability are mathematically doomed. The AI capex cycle assumes continuous exponential demand growth. It is a seigniorage feedback loop with real assets.
Central banks are noticing. The Bank of England has warned that AI valuations are approaching dot-com bubble levels. The Bank for International Settlements flagged the risk of a sudden stop in investment. The signs are there, but the market is in euphoria. Assume malice, verify everything, trust nothing.
Crack Four: Open-Source Threat to Pricing Power
Open-source models like Llama are closing the performance gap with proprietary models at a fraction of the training cost. This erodes the pricing power of closed-source providers. In a commodity market, only the lowest-cost producer survives. But the lowest-cost producer in AI may be a decentralized community, not a corporation. I saw this dynamic in 2020 when Yearn's vault strategies were copied fork-by-fork. The ability to sustain a premium requires a moat. Open source is a moat destroyer.
The implications are severe. If AI becomes a commodity, the capital deployed for proprietary moats is wasted. The industry will face a wave of asset impairment not seen since the dot-com bust. Static analysis reveals what marketing hides.
Contrarian: The Bull Case—and Why It Falls Flat
Bulls argue that Big Tech's diversified revenue streams can absorb the capex. They point to strong cash flows from advertising and cloud services as buffers. They also note that AI could unlock productivity gains that justify the investment over a longer horizon. This is not entirely wrong. Microsoft Azure, Amazon Web Services, and Google Cloud have deep pockets. A delayed payoff is not the same as a default.
But the scale matters. The capital being deployed is not marginal; it is existential. A 20% reduction in AI capex would still leave the world with oversupply. And cross-subsidies work only as long as the core businesses remain healthy. If AI fails to deliver, the drag will pull down the entire conglomerate. I modeled this in 2017 for Tezos: the transition from centralized foundation to on-chain governance was theoretically sound but practically fragile. The same fragility applies to AI budgets within giant corporations. A recession or antitrust action could trigger a reallocation, and AI spending would be first to be trimmed.
Moreover, the open-source threat is not a short-term blip. It is a structural shift in the competitive landscape. The bull case underestimates how quickly a commodity market can erode pricing. Just ask the telecom carriers who built fiber networks in the 1990s.
Takeaway: A Call for Accountability
The four cracks are not surface-level anomalies. They are rooted in basic arithmetic. The industry is gambling that future demand will rescue present oversupply. History suggests otherwise. The dot-com bubble collapsed because the infrastructure exceeded the usage. The same script is being replayed with faster hardware and bigger numbers.
The question is not whether a correction will happen, but when and how violently the market will reconcile the math. Until that day, the only rational position is skepticism. The proof is in the logic, not the promise.