Pushing electrons thru silicon at scale doesn't scale ... efficiently

 Angular Momentum

How to gin up the IPO. You got to fund us to protect you from US. The Twilight Zone applies.

The Securities and Exchange Commission‘s publication on Monday of Anthropic’s S-1 form—ahead of a heavily anticipated initial public offering—has led to plenty of discussion of whether the company’s quickly ballooning revenues will be able to overcome its even higher costs. But amid the numbers and forward-looking optimism about AI “rapidly transform[ing] every sector of the global economy in ways greater than the impact of industrialization, electricity, semiconductors, and the internet,” Anthropic also spends an inordinate amount of time in its S-1 laying out the myriad risk factors that could bring the company down.

Across 93 pages and over 70,000 words, Anthropic’s S-1 “Risk Factors” section discusses literally hundreds of potential problems that could negatively impact its business in the coming years. Many of these potential risks are the kind of standard worries about competition, costs, customer acquisition, and potential regulation that almost any business must worry about. But some risks highlight interesting potential points of failure for Anthropic specifically and the hype-filled AI bubble more generally.

The smell of desperation applies when you factor in an 80 billion loss against 16 billion in profits Think 64 billion as net loss, and combine this with a possible 750 billion needed to keep the system running does not portend a happy ending because the bond guys are starting to get twitchy and for good reason. 

It's the ultimate Silicon Valley protection racket, and it is Dr. Strangelove writ large. Anthropic's built a computational Doomsday Machine, and now are holding a multi-billion-dollar funding round to figure out how to build the off switch but they don't have one. Rod Serling couldn’t have scripted a better episode.

But if we cut through the IPO hype and look at this through a purely architectural lens,
this 93-page apocalyptic disclaimer is actually a confession.

Why are the creators of this technology terrified of it? Because they have absolutely no structural control over it. Anthropic, much like DeepMind and OpenAI, has bet the house on probabilistic brute force. Their models are statistical guessing engines, floating in a digital void with no physical constraints, no geometric boundaries, and no anchoring in physical reality. When you build a system that relies entirely on hallucinating the right answer, of course you lose sleep worrying about what it might hallucinate next.

The emperor has no clothes. The problem is the fact uninformed investors don't know this but they will. Physics doesn't care. Pushing electrons through silicon at scale doesn't scale ... efficiently.

As proof ...

A Google data center in Nevada.
Credit...Bizuayehu Tesfaye/Las Vegas Review-Journal/Tribune News Service, via Getty Images

Tech companies are struggling to find enough electricity for all the new data centers they need for artificial intelligence, and lately they’ve explored ideas such as restarting old nuclear plants and stringing together off-grid gas engines.

Now Google is trying a different approach: paying, at least indirectly, for homes and businesses to use less electricity at times to help free up power for data centers.

On Thursday, Google announced a deal with a company called Voltus to set up what’s known as a “virtual power plant” in the large Mid-Atlantic grid known as PJM. Voltus will provide Google with up to 1,070 megawatts of electric capacity when needed — roughly as much as the output from a large nuclear reactor.

That electricity won’t come from new power plants. Instead, Voltus will sign up homes and businesses and give them financial incentives to curb their usage when local power demand is high or the grid is strained. Many of these reductions could occur automatically, such as through smart thermostats that dial back air-conditioning during peak summer hours or batteries that can provide temporary on-site power.

Again, pushing electrons through silicon at scale doesn't scale ... efficiently.