The compute inversion, agentic commerce, and broken math

The compute inversion, agentic commerce, and broken math

The Throughline Brief · October 11th. 6 min 18 sec.

Cold open

The standard AI budget just flipped completely upside down. The days of throwing endless money at pre-training are over. Foundation model builders are suddenly pouring everything into reinforcement learning. Tens of thousands of GPUs running for weeks on post-training alone. The low-hanging fruit of web scraping is totally gone. The real money is moving fast.

Intro

This is The Throughline Brief for October 11th. We went through every big AI and tech podcast this week so you don't have to. We listened to all the noise, pulled out the exact moments that shift the space, and here's what actually mattered.

The Compute Inversion

Let's look at the absolute total inversion of how artificial intelligence gets funded. The pre-training budget is suddenly shrinking. Misha Laskin went on No Priors and broke down exactly how ReflectionAI built Beam. They used 6 thousand GPUs to pre-train. But their reinforcement learning took a little over 10 thousand GB300s running for four solid weeks. That's a massive shift. The spending flipped. Frontier labs are reallocating billions of dollars right now. Nathan from Air Street Capital told The Boys Product Network that foundation pre-training compute plummeted. It dropped from 50 percent of total GPU workloads down to just 10 percent. The reason is simple. The web scraping era is maxed out. You end up spending less money on exploratory work because the solution space is known. And that brute-force compute is bleeding into the hard sciences. David Friedberg mentioned on the All-In Podcast that OpenAI quietly dumped 370 completed math proofs into the public domain. An unreleased model compressed centuries of theoretical labor into a single afternoon. Friedberg called it probably the biggest day of discovery in human history. The sheer volume is completely unmanageable for human reviewers. Justin Solomon told Odd Lots that university math departments are facing a massive backlog. We left proof scarcity behind. Now we have proof abundance, and academics have to figure out an entirely new way to do their jobs. If you build models today, your competitive advantage shifted entirely to post-training pipelines. You have to adapt fast.

Agentic Commerce

Autonomous agents are fundamentally rewiring how software gets bought and sold. This is happening right now in enterprise sales. Jason Lemkin sat down on 20VC with Harry Stebbings and dropped a staggering number about Vercel. They hit 600 million dollars in annual recurring revenue. Half of their new business is initiated autonomously by machines. Think about that. Human buyers are stepping back. Coding agents are the ones inspecting APIs, evaluating the tools, and setting up the accounts. Lemkin pointed out that agents make up 50 percent of new business today. That number was only 3 percent at the start of the year. The entire enterprise funnel just changed. And this exact behavior is coming straight for consumer retail. Kevin Rose was on The Tim Ferriss Show warning about the next wave of proactive assistants. These bots run inside text messages and threaten to dethrone Amazon as the top search destination. Consumers are happily handing their credentials over to models like Instinct. Then agent to agent coordination handles the logistics. The transaction happens entirely outside the browser window. No human in the loop. The takeaway for founders here's massive. You need to build your product discovery for machines first. If your API is difficult for an autonomous agent to read, you'll lose the sale before a human even knows your company exists. It changes every assumption about go-to-market strategies.

Chamath on All-In

Chamath Palihapitiya went on the All-In Podcast to discuss the exact moment an AI agent decompiled thirty years of Adobe proprietary code into open source Rust. He told the crew, I think that IP and software was effectively rendered worthless two days ago. That's a terrifying reality check for legacy tech companies. Building a walled garden of proprietary code used to guarantee dominance. Now a sufficiently advanced agent can look at your product, reverse engineer the entire thing, and rebuild it from scratch over the weekend. Defensibility is entirely about speed and distribution now. If your moat relies on closed source code, you're completely vulnerable.

Woodward on Lenny's

The skills required to thrive in a rapid AI transition are totally counterintuitive. Josh Woodward leads Google Labs and he told Lenny on Lenny's Podcast about the single most mispriced skill in tech hiring right now. He said, I'm actually looking a lot right now for people, what's their unlearning rate, like how fast can they learn something and then walk away from it. This flips traditional hiring on its head. Explosive endurance matters, but walking away from sunk costs matters more. When the underlying models change every six months, clinging to your previous architecture is a death sentence. You have to hire builders who can torch their own work and start over without ego.

Senra on Founders

Moving over to hard business truths, David Senra broke down the brutal reality of corporate acquisitions on the Founders Podcast. He was analyzing Endeavor and their supposedly friendly takeover of William Morris. Senra put it bluntly. He said, there's no such thing as a merger, this is a lie, there's one company. That's the exact clarity founders need when evaluating exit opportunities. We love to use soft language like strategic partnership to protect feelings. The reality is that someone always ends up with the steering wheel. If you sell your company, you hand over the keys completely. Expecting equal footing is just a guaranteed path to severe disappointment.

The bottom line

The era of simply scaling up pre-training to win is officially stalling out. The entire artificial intelligence industry is locked in a brutal race for post-training data and autonomous agent utility. Everything hinges on what your models can actually do in the wild. If you build software today, your main customer is quickly becoming an autonomous agent. That's your Throughline. See you next Sunday.