Within days of its showy IPO, SpaceX (SPCX) has locked in a $60-billion deal to acquire up-and-coming AI coding agent Cursor.
The price tag exceeds what Elon Musk paid for Twitter. In fact, excluding the $1.25 trillion merger between SpaceX and xAI, it’s Musk’s largest acquisition to date.
SpaceX just raised $75 billion in the largest IPO in history. It could have bought almost anything. It bought a coding agent.
That choice tells you everything about where Elon Musk thinks the next phase of AI is headed.
Why SpaceX Needed Cursor: The Software Problem at the Heart of the Musk Industrial Stack
SpaceX’s entire business is centered on rockets, satellites, Starlink terminals, defense systems, autonomous manufacturing lines, humanoid robots, orbital compute infrastructure — and now, through xAI, a large language model.
Every single one of these businesses runs hyper-complex, mission-critical, continuously-iterated software.
If Cursor can make the engineers behind that software more productive, it could compress years of engineering work into months — across rockets, satellites, humanoid robots, and autonomous manufacturing lines simultaneously.
This impact goes deeper still. Cursor isn’t just a productivity tool; it’s a distribution platform. Enterprise developer tools are famously sticky. That means SpaceX just bought daily, persistent, deeply embedded access to the most valuable users in the enterprise software economy: software engineers.
Not to mention — every prompt, every code completion, every debugging session that runs through Cursor? That proprietary usage data is what makes AI models demonstrably better. SpaceX/xAI now owns one of the richest AI training and inference datasets on Earth, packaged inside a tool that users will pay a monthly subscription to provide.
That $60-billion price tag is starting to look much less outlandish.
(While Musk signals his moves publicly, others do it through SEC filings most investors never read. One of those filings just caught my attention.)
The Real Signal: AI Is Moving From the Training Room to Persistent Agentic Deployment
For the past three years, the AI economy has been defined by one thing: training. Who has the most GPUs? Who can build the biggest model?
That’s what moved markets — and it was where the money went.
That era isn’t over, but it is maturing. The frontier labs have their models. The hyperscalers have their infrastructure. Now the race is about deployment; specifically, agentic deployment — AI that doesn’t just respond to prompts but takes actions, writes code, browses the web, executes tasks, and operates autonomously across multi-step workflows.
Cursor is the clearest proof yet that agentic AI coding is a daily workflow for millions of professional developers.
And when the world shifts toward continuously running AI agents, inference demand explodes. We’re talking 20x to 50x the compute from training-era workloads — because inference isn’t a one-and-done query. It’s a persistent, context-heavy, multi-turn process that runs all day, every day.
The SpaceX/Cursor deal is a $60 billion vote of confidence that the agentic shift is happening now, and the infrastructure to support it is worth building — at any price.
The Jevons Paradox Is About to Hit Software — and It’s Bullish for Every Physical Bottleneck
There’s a principle in economics called Jevons Paradox: when a resource becomes more efficient to use, total consumption of that resource goes up.
For example, when James Watt’s improved steam engine made coal-powered machinery dramatically more efficient in the late 18th century, Britain didn’t use less coal — it used exponentially more. More efficient engines made steam power viable for textile mills, iron foundries, flour mills, breweries, railways, and steamships. Applications multiplied faster than efficiency gains could reduce consumption. By the time Jevons wrote his famous treatise in 1865, British coal output had roughly quadrupled in a generation.
The same dynamic is unfolding in software development right now.
AI coding agents like Cursor make software dramatically cheaper and faster to build. The first-order intuition is that this reduces infrastructure demand: fewer engineer-hours means less compute, right? Wrong.
When software becomes faster and cheaper to build, the world builds vastly more software. More software built by agents → more agent usage → more inference compute demand → more GPUs, more networking, more memory, more power, more cooling.
The Cursor acquisition doesn’t just validate agentic AI. It validates the entire AI infrastructure thesis for the next decade.
Where Does the $60 Billion Signal Point? The Physical Bottlenecks of Agentic AI
Nobody got rich from cheaper steam engines. They got rich owning the coal mines, the railroads, and the infrastructure that made the boom possible. The AI version of that trade is right in front of us.
As agentic AI demand multiplies over the next few years, the components that are hardest to scale, fastest to sell out, and least substitutable will capture the most value. Here’s what’s on that list.
GPUs and Accelerators: The First Bottleneck Agentic Inference Pounds
Inference workloads run on the same GPU infrastructure as training — and agentic inference is far more compute-intensive because it runs continuously rather than in discrete bursts.
- Nvidia (NVDA) remains the dominant supplier, with Broadcom (AVGO) building custom AI chips for Google and Meta (META) that handle a growing share of hyperscaler inference.
The GPU shortage is structural, and persistent agentic workloads are about to make it dramatically worse.
Networking: The Least Appreciated Bottleneck in the Agentic Stack
Every token an AI agent generates has to move between memory and processors at extraordinary speeds — and when thousands of agents run simultaneously across distributed clusters, the data movement problem rivals the compute problem.
- Arista Networks (ANET) is the backbone of AI cluster networking, handling the high-speed switching between GPU racks.
- Corning (GLW) and Coherent (COHR) supply the fiber and optical transceivers carrying that data between data centers — the last physical bottleneck before raw compute.
Memory and Storage: Why Agentic AI Is Structurally Undersupplied
Agentic AI is extraordinarily memory-hungry. Long context windows, persistent state, real-time retrieval — all of it demands high-bandwidth memory (HBM) that the industry is already structurally undersupplied on.
- Micron (MU) is the leading U.S. supplier of HBM and has reportedly sold out production under long-term contracts.
- Western Digital (WDC) supplies the storage layer underneath.
- IREN (IREN) operates AI-native data center infrastructure built specifically around these workloads.
Power and Cooling: The Bottleneck That Doesn’t Sleep
Every GPU running inference burns power around the clock — and agentic workloads don’t sleep. A single large AI data center can consume as much electricity as a small city.
- Vertiv (VRT) supplies the power and thermal management systems keeping those racks online.
- Eaton (ETN) provides the electrical infrastructure distributing power at scale.
- Quanta Services (PWR) builds and maintains the physical grid upgrades supporting the entire buildout — a decade-long capex cycle that is just getting started.
The Bottom Line: Own the Bottlenecks the $60 Billion Signal Points To
SpaceX’s latest deal isn’t really about Cursor. It’s about Elon Musk signaling that the next phase of AI is agentic, it runs on inference, and controlling the daily workflow of software engineers is a strategic asset worth $60 billion.
When the smartest, most ruthlessly strategic operator in the technology industry pays 60 billion dollars to make a bet, the right response is to ask what he knows that the market hasn’t priced in yet — and then position accordingly.
The AI economy is shifting from training to inference. From occasional queries to persistent agents. From a few hyperscalers spending capex to the entire software-building world running on AI infrastructure 24/7/365.
The bottlenecks in that world — GPUs, networking, memory, power, cooling — are the assets you want to own.
Those bottlenecks aren’t a secret to everyone.
Peter Thiel recently filed a 13F showing he’d quietly liquidated every share of Nvidia, Apple, Microsoft, and Tesla he owned. Not trimmed — exited entirely. At the same time, his private fund has been deploying capital into exactly the physical bottlenecks this piece describes: energy infrastructure, nuclear power, chip fabrication, and natural resources.
He can’t buy those companies publicly. Most of them aren’t available to retail investors at all.
But I have spent months identifying seven publicly traded stocks that mirror those same private bets — the physical layer of the AI buildout that the billionaires are already funding.
Thiel calls it the shift from “bits” to “atoms.” I call it the Billionaire’s Backdoor.
Here’s the full portfolio — and the thesis behind every position.
