The AI Boom's Dirty Secret: Infrastructure Is Failing

The Breach Nobody Wanted to Talk About

The timing could not be more awkward. Anthropic is reportedly approaching a trillion-dollar pre-IPO valuation, Amazon and Google are writing checks in the tens of billions to back it, and somewhere in that same news cycle, the company confirmed it is investigating unauthorized access to Claude Mythos preview, its most powerful and most restricted model.

According to reporting in the Financial Times, the breach appears to have come through a third-party partner, not a direct intrusion. That detail matters. Anthropic had granted access to Mythos only to around 40 top-tier companies and a handful of additional partners, precisely because the model is considered so capable that exposure to adversarial actors could cause serious harm. Banking, critical infrastructure, and national security are all in the frame here. And now the chain of custody on that model is in question.

The China Dimension

This is where the story sharpens. Analysis from CSIS, the nonpartisan national security think tank, puts China just seven months behind US frontier model performance. That gap has closed through a sustained campaign of distillation attacks, a technique where competitors reverse-engineer a model's behavior by querying it at scale and training against the outputs. Both Anthropic and OpenAI have raised public complaints about this practice.

Layer on top of that the smuggling of US chips into China and a recent decision by the current administration to permit the sale of advanced chips to Chinese buyers, and the picture becomes uncomfortable. A breach of a model as capable as Mythos, regardless of who accessed it, adds fuel to a fire that was already burning.

The Money Is Real. The Infrastructure Is Not.

Here is the counterintuitive part. In response to all of this uncertainty, big tech responded by writing larger checks.

Amazon is putting $5 billion into Anthropic upfront, with up to $20 billion more committed over time. In return, Anthropic has agreed to purchase $100 billion of compute from Amazon. Analysts have flagged the circular nature of this arrangement before. Anthropic owes Amazon that $100 billion regardless of whether its revenue materializes. Google, not to be outdone, is offering $10 billion upfront and up to $30 billion more tied to performance milestones, alongside a five-year deal to bring five additional gigawatts of data center capacity online. The FT puts the potential value of that arrangement at $200 billion.

But here is what the satellite imagery tells a different story about: 40% of US data centers are behind schedule. An FT report with overhead imagery shows that many of the infrastructure commitments announced with great fanfare are not yet underway in any meaningful way. For companies like Anthropic and OpenAI, both of which are eyeing 2026 IPOs, the gap between what has been promised to customers and what can actually be delivered is becoming a structural risk.

Oracle sits at the sharper end of this problem. The company made a major infrastructure commitment to OpenAI last year. Its debt has since drifted toward junk status, and the banks that extended credit are struggling to sell that exposure to other investors. Oracle is reportedly issuing additional corporate debt to stay in the game, while simultaneously carrying a large personal guarantee from Larry Ellison tied to the proposed Paramount and Warner Brothers merger, a deal carrying roughly $90 billion in debt of its own.

The Human Cost of the Build-Out

While capital piles into infrastructure, it is flowing out of payrolls. Meta is cutting 8,000 positions in May 2026 and canceling 6,000 open roles. Microsoft is reducing its global workforce by 7%. The rationale in both cases is the same: AI investment is replacing headcount, not supplementing it.

At a moment when inflation is creeping back up, Federal Reserve leadership is uncertain, and broader economic indicators are deteriorating, large-scale tech layoffs carry contagion risk well beyond Silicon Valley. Neeta makes the case that the mid-2026 period could see this pressure spread further across sectors.

What the JAMA Study Says About Limits

One grounding note amid the hype: a new study in JAMA found that large language models tend to reach medical diagnoses too quickly, before enough data has been gathered, producing inaccurate results. Accuracy improved significantly later in the diagnostic process, when more clinical information was available. It is a small but concrete illustration of where LLMs sit today: capable under the right conditions, unreliable at the front end of ambiguous problems.

The trillion-dollar valuations, the circular debt structures, the infrastructure shortfalls, and now a breach of the most tightly held model in the industry, all of it points to the same underlying tension. The AI supercycle is real. Whether the physical and financial plumbing can keep up with it is a genuinely open question.