The number looks like a typo – $1.4 trillion. That is the cumulative AI capital expenditure projected for Meta, Amazon, and Google alone by 2028. Morgan Stanley's call is now the most aggressive on Wall Street. The rationale? Unrelenting demand for compute, a supply chain that cannot keep up, and a belief that scaling laws still hold. The market has celebrated. NVDA rallied. Meta broke out. But as someone who spent 2017 auditing ICO whitepapers with flawed tokenomics, I have seen this narrative before: massive infrastructure spending assumed to guarantee leadership, with little empirical evidence linking the two.
Context Morgan Stanley analysts raised their forecast, citing “ongoing demand growth” and “supply chain bottlenecks.” They now expect Meta to spend $250 billion, Amazon $318 billion, and Google $350 billion on capex through 2028 – roughly doubling previous estimates. The thesis is straightforward: AI model training and inference require ever-larger clusters of NVIDIA GPUs, which remain scarce and expensive. The banks maintain “overweight” ratings on Meta and Amazon, implying these investments will eventually generate superior returns. The numbers themselves are not wrong. The ledger of corporate budgets is clear. But ledgers do not lie – the narrative does.
Core As a data detective, I need to examine the assumptions behind these numbers. First, the total addressable demand. The report assumes the current trajectory of scaling – bigger models, more parameters – continues linearly. My own work in 2020 analyzing Uniswap V2 liquidity depth taught me that exponential narratives often break when you model the marginal cost of each additional unit. For AI, the marginal benefit of doubling compute has been declining. Recent papers from DeepMind and Anthropic show diminishing returns on loss reduction beyond a certain scale. If this trend continues, the marginal GPU purchased today will generate less intelligence than the one bought a year ago. The implicit assumption that “more compute = better AI” may be a sunk cost fallacy in disguise.
Second, the supply chain bottleneck is real, but it cuts both ways. When I stress-tested my portfolio during the 2022 bear by tracking whale movements, I saw that concentration of assets in a few addresses magnified exit risks. Here, concentration of demand in four buyers (adding Microsoft) creates a different risk: if any one of them slashes orders, NVIDIA’s revenue falls off a cliff. The capex forecast is essentially a collection of best-case spending plans, not ironclad contracts. The same Morgan Stanley also noted that “key component costs are rising.” That is a polite way of saying that the hardware is getting more expensive per unit of compute, eating into the ROI even before the first model is trained.
Third, commercialization timelines are vague. I have audited protocols that promised revolutionary layer-2 throughput but delivered nothing but gas spent. The AI capex is similar – Amazon’s $318 billion includes both its own model training and inference for AWS customers. But how much of that inference revenue is recurring versus one-time experimentation? In 2026, I led a project analyzing on-chain transaction patterns to detect wash trading. We found that 15% of DEX volume was artificial. I suspect a parallel in AI: a significant portion of current API calls come from developers testing ideas, not building sticky applications. The revenue per GPU deployed may be far lower than assumed.
Contrarian The market reaction treats these capex numbers as a key catalyst. I disagree. The real signal is not the size of the investment, but the correlation between investment and market share stability. In crypto, the protocol with the highest staking ratio does not always win – just ask Ethereum Classic. In AI, Meta’s $250 billion will buy compute, but that compute must be converted into user engagement and ad revenue. If Threads or Llama 4 fails to captivate, that capex becomes stranded. The contrarian view: we are about to witness a “commoditization of inference.” As GPUs become more available over the next 18 months, the marginal cost of inference will drop, compressing margins for all but the best models. The first-mover advantage is in the data moat, not the hardware stack.
Moreover, the concentration of capex creates a single point of failure for the entire ecosystem. A sudden policy change – US export controls on NVIDIA chips to China tightening, or a new energy regulation in EU – could freeze half the planned data centers. I saw this in the Terra collapse: when one node of the ecosystem fails, contagion is rapid because everyone is correlated. The AI infrastructure buildout is highly correlated across companies, geography, and even supply chain. Diversification is an illusion when all your eggs are in NVIDIA’s basket.
Takeaway The next signal to watch is not the next capex upgrade, but the revenue per teraflop. I will be tracking AWS Bedrock and Azure AI revenue growth against the number of new instances added. If those ratios decline, the math breaks. Trust the math, ignore the hype. In the meantime, the smart money is preparing for a rotation from infrastructure plays to application layer survivors. Survival is the ultimate alpha in a bear, and the current bull market is exactly when hidden risks compound.