The $600K SaaS Replacement, AI's Physical Limits, and Agent Supervisors

The Throughline Brief · July 19th. 6 min 3 sec.

Cold open

A healthcare company just fired a massive warning shot at the software establishment. They canceled a 600,000 dollar annual Salesforce contract. Just completely ripped it out. Why? Because they built a superior internal AI CRM from scratch in exactly two months. Traditional enterprise software is officially bleeding. The shift is happening right now.

Intro

This is The Throughline Brief for July 19th. I went through every big AI and tech podcast this week so you do not have to, here is what actually mattered, I am West. Let us get straight into the signal.

Custom AI vs SaaS

Custom AI agents are slaughtering traditional enterprise software. We are seeing real numbers on this. On 20VC, Curative CEO Fred Turner detailed how his company completely eliminated their 600,000 dollar annual Salesforce contract. They did not transition to a cheaper vendor. They just built their own internal AI CRM. It took them two months. Turner projects they will slash 80 percent of their total software spend this year. They are letting AI handle the heavy lifting of back-office operations. Think about that. The math gets even more brutal when you look at off-the-shelf tools. Jason Calacanis brought this up on the All-In Podcast. He highlighted a healthcare staffing company named Nursa. They saved over 1 million dollars a year by replacing 10 different subscriptions with bespoke applications. They built these through vibe coding. Founders are realizing they do not have to tolerate the friction of pre-made solutions, or siloed data. They can just build exactly what they need. The days of forcing diverse requirements into a single box are ending. This shift is happening all the way down the market. Makun is the CEO of Emergent AI. He told TBPN that almost 80 percent of his users are non-technical business owners. They are skipping traditional subscriptions entirely. Domain experts are sitting down and building custom, production-grade applications using plain natural language. No engineering degree required. Just plain English. Enterprise software companies have relied on lock-in for a decade. That lock-in is evaporating in real time. It is a massive extinction event for lazy software.

Physical constraints

The fear of an AI bubble assumes infinite software growth. Everyone assumes the line just goes up forever. But physical reality is getting in the way. On the All-In podcast, former Intel CEO Pat Gelsinger argued that sheer energy capacity provides a natural, self-correcting brake on speculative computing. We have a hard physical limit on power. Gelsinger believes this guarantees we do not get too far ahead of ourselves. It forces the AI buildout to stretch over multiple decades rather than crashing overnight. Hardware constraints are also capping the boom through sheer inefficiency. Rafa Gomez Bambarelli explained this on Latent Space. In their AI training pipelines, Mean Flop Utilization hovers at a dismal 5 to 6 percent. That means 95 percent of paid computing power is totally wasted. Founders face a strict hardware bottleneck. You cannot simply buy your way to rapid progress by adding more scale. The efficiency just is not there. Even local hardware is forcing difficult trade-offs. Alex Finn broke this down on How I AI. The equipment you choose to run models locally is a foundational constraint. Builders face a fundamental choice. You can use the unified memory in Mac Studios for sheer model size. Or you can use dedicated V RAM in Nvidia GPUs for raw processing speed and bandwidth. There is no perfect machine yet. You have to pick your bottleneck. The physical world is quietly putting a ceiling on the software hype.

Dr. Alan Castel

Dr. Alan Castel delivered a brutal reality check on Huberman Lab this week. He was discussing the biological mechanics of how we recall important events. Memory is really a mental representation of the past and by its very nature it is reconstructive. It is never always accurate. Think about the implications for you as a founder. You walk out of a crucial investor pitch or a defining user interview. You think you remember exactly what they said and how they reacted. You are wrong. Your brain is actively rewriting the tape over time. If you rely on your gut recollection without taking hard notes, you are flying blind.

Elizabeth Stone

Over on Lenny's Podcast, Netflix Chief Product and Technology Officer Elizabeth Stone dropped a massive insight about the future of media. She was breaking down how their product strategy is evolving beyond simple video streaming. Entertainment is not going to be one thing in the future and it is already not one thing now. Static media formats are effectively walking dead. Netflix sees the writing on the wall. They are actively pivoting toward interactive and ambient experiences. If you are building consumer software, your users will expect dynamic, multi-modal engagement. The era of just staring at a flat rectangle of pre-rendered content is ending fast.

Andy Beam

Andy Beam completely reversed the traditional biotech playbook on a recent Latent Space episode. He explained how Lila Sciences is treating their technology stack completely differently than legacy pharmaceutical companies. The model itself is the thing of value at Lyra. This flips the entire industry on its head. Historically, the core asset was a single, heavily patented drug candidate. The physical lab was everything. Now, the foundational AI reasoning system is the actual crown jewel. The physical labs are reduced to mere token generators for the model. Hard science is becoming a software game. The power is moving from the test tube directly to the data center.

The bottom line

The market is obsessed with debating the massive energy costs and training budgets of giant foundational models. Nimble founders are completely ignoring the noise. They are actively using targeted AI agents to rip out expensive legacy enterprise software and completely reshape the corporate org chart. The shift from buying generic subscriptions to building custom intelligence is the defining opportunity of this cycle. That is your Throughline. See you next Sunday.