Amazon's AI Is Breaking Its Own Warehouses

The Company Breaking Things With AI Wants to Fix Everyone Else With AI

Here is the tension at the center of this week's episode: Amazon's own AI tools have been quietly destroying its warehouse operations for months, wiping out six-figure order counts in single incidents, and forcing the company to deliberately slow down AI deployment across 335 retail systems. At the same time, Amazon founder Jeff Bezos is out pitching sovereign wealth funds on a $100 billion vehicle to buy struggling manufacturers and rebuild them from the inside using AI.

That is either a story about extraordinary confidence, or it is a story about how far ahead of the evidence the AI hype machine has gotten.

Why Amazon's AI Keeps Getting It Wrong

The technical explanation is straightforward, even if the implications are underappreciated. AI models are non-deterministic. Ask the same question twice and you may get two different answers. For most applications, that variability is a feature. For a logistics and fulfillment operation where every order, every inventory count, and every shipping confirmation must be exactly right every time, it is a serious liability.

Neeta flags that the executive who raised the alarm internally noted the trend had been deteriorating since Q3 of 2025. This is not a one-off glitch. It is a systemic pattern, and Amazon's response, introducing what it calls controlled friction, is essentially a human override layer designed to catch AI errors before they cascade. The company is now in the awkward position of deploying AI while also building guardrails against AI.

This matters beyond Amazon. Virtually every large enterprise is somewhere on this same adoption curve, and the honest lesson from Amazon's experience is that non-determinism is a structural problem, not a bug that a future model version will simply patch away.

Bezos's $100 Billion Counterbet

Against that backdrop, the Project Prometheus fund is striking. The strategy, as described, is to acquire manufacturers in chipmaking, defense, and aerospace, sectors that are at least as precision-dependent as warehouse logistics, and then apply AI developed by Prometheus to automate their operations from the inside.

The fund target of $100 billion puts it in the same weight class as SoftBank's Vision Fund. Bezos is reportedly in early talks with sovereign wealth funds to assemble the buyout vehicle. Project Prometheus itself launched last November with $6.2 billion in backing and is currently raising an additional $6 billion, with Bezos serving as co-CEO.

The strategic logic is not hard to follow. Struggling manufacturers are available at distressed valuations. AI-driven automation, if it works, compresses labor costs dramatically. And the sectors being targeted, defense and advanced chipmaking, are precisely where governments worldwide are desperate to rebuild domestic capacity, which means political tailwinds and potentially government contracts.

The risk is also not hard to identify. If Amazon, one of the most sophisticated technology operators on the planet, is having to pump the brakes on AI in its own warehouses, the idea that a buyout fund can successfully transplant that same AI into older, more complex manufacturing environments is a proposition that deserves real scrutiny.

The Broader Chip War Gets Criminal

The week's other major AI story is less philosophical. Three Super Micro executives have been indicted in New York for allegedly smuggling NVIDIA chips to China in violation of US export controls. Two have been placed on administrative leave and a third contractor has been fired.

This is the shadow war for AI hardware becoming a criminal matter. The US government has spent considerable energy trying to understand how China has been accessing top-tier chips despite the ban, and this case appears to be one concrete answer. Neeta also notes that NVIDIA itself faces separate scrutiny over direct assistance it may have provided to Chinese AI development, including to DeepSeek.

The contradiction the episode surfaces is worth sitting with. The US government is aggressively prosecuting chip smuggling while simultaneously appearing open to loosening certain chip export restrictions in other contexts. The line between strategic flexibility and strategic incoherence is thin, and right now it is not obvious which side of it policy is on.

What to Watch

Amazon's controlled friction experiment will be a useful signal. If the guardrail approach stabilizes operations without killing efficiency, it becomes a model for enterprise AI deployment broadly. If the outages continue, it raises harder questions about where non-deterministic AI actually belongs in production environments.

Project Prometheus will take longer to evaluate, but the sovereign wealth fund conversations are the near-term tell. If major funds commit at scale, the $100 billion target becomes credible. If they balk, it may be because they are doing the same math Amazon's warehouse operators are doing right now.