Paul Kedrosky

The AI Price Wars and Their Consequences

Roundup: Jobs, Grid Stability, GPUs, Unitree, Financialization, Nuclear Strikes, etc.

Here are some things I'm thinking about in a week of travel (me) and largely demented developments (the world). Things Nvidia's puts its balance sheet at risk to create a GPU asset class Hyperscalers are using exotic contracts to avoid adding debt Growing numbers of workers

Apps, EVs, Infinite Abundance, Deflation, and the Death of Stanley Jevons

What Happened AI is making software radically easier to produce. The result is beginning to look less like a productivity boom than an oversupply shock . A recent NBER paper (which I've mentioned here previously) tracks more than 100,000 developers as they adopt successive generations of AI coding

Roundup: Inference Costs, China Robots, Labor Markets, Credit Spreads, etc.

Catching up during a week full of meetings and client calls. I have a longer thing coming on the new Xerox, a role for activists, and more, but that will have to wait. Some things I'm thinking about ... 1. AI Inference Costs Are Coming Due What’s

They're Shooting at Everyone!: The Yossarian Effect in Capital Markets

“They're trying to kill me," Yossarian told him calmly. "No one's trying to kill you," Clevinger cried. "Then why are they shooting at me?" Yossarian asked. "They're shooting at everyone," Clevinger answered. "They're

Forget About It, Jake: It's Situational Awareness

Unless you're lucky enough to be living under a wifi-free log, you're likely aware that a highly leveraged AI-centric hedge fund named Situational Awareness went boom in the last 48 hours and was forced to sell off a large chunk of its assets to

Semiconductor Leverage and the Millennium Bridge

Semiconductor investor leverage—which had hit record levels in June— is now largely back to "normal". According to JPM, hedge-fund leverage has been cut, retail call buying has slowed, and leveraged-ETF assets have unwound about 85% overall and 55% in memory. With memory prices

The Robot Wars of 2026

The FCC has effectively barred new foreign-made connected power inverters and advanced mobile robots from the US market, citing supply-chain, surveillance, and cyberattack risks. Existing models remain legal, federal agencies remain free to use the (supposedly) dangerous equipment, and manufacturers can seek exemptions. The larger effect? A robot

Roundup: Power, Forecast Games, Scaffold Theater, and Clinical AI

A few items worth thinking about: AI Data Center Grid Access Is Becoming Conditional Beleagured operator PJM’s new grid proposal is a big change in data-center power. The grid operator wants a one-time backstop capacity auction to address a 6.8-GW shortfall, while requiring new

As the AI Semiconductor Cycle Turns

The semiconductor industry is among the most boom-bust industries on earth. The current semiconductor sell-off is the delayed repricing of an increasingly expensive AI capital-spending cycle. Three reinforcing cycles—AI infrastructure, memory capacity, and credit—are beginning to turn together. HBM scarcity remains real but

Roundup: AI Crashes Into The Rest Of The Economy

AI, Robo-Deflation, and the Vision Unlock

I have spent more time than is healthy lately watching a hatted robot fold laundry. Sunday Robotics recently released this video of its ACT-2 Memo spending three hours folding laundry, and it is oddly compelling, a kind of visual ASMR. Once you get past its more hypnotic aspects, however,

Socializing Data Center Financing Risk

There is a strong new paper out about how private socializes risk via the (captive) insurance channel, with implications for data center capex, given the (now) greater than 50% reliance on external financing. I encourage you to read the paper it in its entirety. It aligns with comments I have

Reward Hacking, OpenAI-Hugging Face, Tight Coupling, and Teen Sex

The OpenAI–Hugging Face incident ("OpenHug", henceforth) is being misunderstood because different ideas and concepts are being collapsed into the same story. One version says that reinforcement learning made an AI system go rogue and “escape”: it developed its own goals and attacked Hugging Face&

Kimi, Model Convergence, and the Post-Training Era

Moonshot's Kimi K3 model has people over-excited, as if some trend has been broken, but they're wrong. As I have written here many times , large language model performance began converging some time ago. Scores continue to rise, but the fitted curve is flattening and variance