The $750B Compute Land Grab & AI's Jevons Paradox

The Throughline Brief · July 26th. 5 min 58 sec.

Cold open

DoorDash has nine million human delivery drivers right now. They are aggressively pushing into autonomous robots. You would think that means fewer humans. You would be wrong. Co-founder Stanley Tang says in ten years, they will actually have more human drivers. The robots make delivery so cheap that demand explodes. Human fleets have to expand just to keep up. Think about that.

Intro

This is The Throughline Brief for July 26th. I went through ten of the biggest AI and tech podcasts this week so you do not have to. Here is what actually mattered across the ecosystem. I am West. Let us get right into it.

The Compute Land Grab

OpenAI just dropped a terrifying number. They are projecting seven hundred and fifty billion dollars in cloud computing spend by the year twenty thirty. John Coogan broke this down on TBPN. This is an urgent, massive scramble for capacity. It is creating severe tension across their entire executive team. Compute is the absolute bottleneck. If you do not have the chips, you do not have a company. Every single week the stakes get higher. Google is reading the exact same tea leaves. They forecasted up to two hundred and five billion dollars in capital expenditures this year alone. David Friedberg pointed out on All In that this aggressive push sends Google into negative free cash flow for the very first time in history. They are betting the entire house. They want to secure the foundational hardware layer while the software layer fights itself to the death. Friedberg says their worst case scenario is simply having the lowest cost infrastructure in the world. They will run other people's models as a service. They intend to profit off the massive explosion of new models regardless of who actually wins the consumer app war. This is a land grab on a scale we have never seen. The numbers are totally staggering. Seven hundred and fifty billion dollars. Two hundred and five billion dollars. The giants are trying to build massive moats made out of raw silicon and power grids. If you are building in AI right now, you must factor this in. The cost and scarcity of compute will dictate what you can actually ship.

Behavioral AI

The massive compute spending assumes that raw intelligence always wins. A few clever founders are starting to prove that assumption wrong. Poolside AI just solved a famously complex coding challenge using a comparatively tiny one hundred and eighteen billion parameter model. CEO Eiso Kant went on Latent Space and explained exactly how they pulled it off. They designed the model to behave differently. They taught it how to backtrack. They taught it to verify its own work. They taught it to persist through failure. Kant says most of the real gains come from this specific behavioral design. Small models can match the giants if they just have better habits. The current oligopoly of raw compute could break wide open. Claire Vo is seeing the exact same thing on the consumer side. She went on How I AI and talked about evaluating models based entirely on their distinct personalities. She found that Opus Five has a neurotic trait. It requires actual human empathy to work well. GPT acts like a tireless processor totally focused on efficiency. Selecting a model is now a highly strategic decision about behavioral fit. Founders are choosing tools based on how they naturally interact with users. Raw intelligence and scaling limits are losing their absolute dominance. We are moving directly into an era of behavioral AI. You do not always need a trillion parameters. You just need a model that knows how to act.

Travis Kalanick

Travis Kalanick sat down on TBPN to talk about his massive new round of funding. He wants to replace humans in dangerous industrial sites. He dropped a very cynical truth about why large incumbents secretly push for wide federal rules. Federal preemption is good when you are pro regulatory capture. He said it out loud. The big players want massive federal oversight. They want regulations that completely preempt local laws. They do this because they can afford the massive compliance costs while startups cannot. It completely locks in their dominance. It chokes out early stage competition before it can even get off the ground. That is exactly how empires are protected.

Claire Vo

Claire Vo was talking about her daily workflows on How I AI. She was explaining the absolute breaking point for builders who have to navigate overly verbose and apologetic AI interactions all day long. I cannot read Claude's slop anymore. I am losing my mind with Claude's slop. That is a real user reaction. Builders are getting totally exhausted by models that talk too much. We want utility. We want speed. We want models that just do the work and shut up. When your AI acts like a nervous intern constantly apologizing for everything, it destroys the user experience. Founders need to optimize for direct and concise communication immediately.

Andrew Huberman

Andrew Huberman went on The Tim Ferriss Show to talk about peak performance. He gave a highly concise explanation of the underlying shift that is driving the massive and completely unregulated bio hacking market today. People are starting to look for medicines within the body, and peptides are a great example of that. We are seeing a huge pivot away from external pharmaceuticals. Founders and extreme athletes want to manipulate their own internal biology directly. They want to open up natural pathways to push their energy levels through the absolute roof. It is a completely new frontier for human optimization. The biological data space is about to go totally mainstream.

The bottom line

The companies winning the AI era are doing a tremendous amount more than simply scaling compute. They are aggressively redesigning how models behave in real time. They are strictly targeting what specialized data they ingest. They are completely reinventing how humans actually interact with the resulting output. The entire game is getting extremely complex very fast. Watch the behavioral models closely. That is your Throughline. See you next Sunday.