SpaceX's $60 billion acquisition of the AI coding agent Cursor reveals where the next phase of artificial intelligence infrastructure demand is headed: away from training and toward persistent, always-on agentic AI systems that will require dramatically more power, cooling, and networking capacity than today's data centers can easily provide. The deal, announced shortly after SpaceX's record-breaking initial public offering, signals that the bottleneck constraining AI growth is shifting from GPU availability to the physical infrastructure required to run AI agents continuously, 24 hours a day.

Why Is Agentic AI So Much More Demanding Than Training?

For the past three years, the AI industry has focused almost entirely on training large language models (LLMs), which are AI systems trained on vast amounts of text data to understand and generate human language. That era is maturing. Now the race is moving toward deployment of agentic AI, which means AI systems that don't just respond to individual questions but instead take autonomous actions, write code, browse the web, and execute multi-step tasks without human intervention.

The computational difference is staggering. Agentic inference workloads run 20 to 50 times more compute-intensive than training-era queries because inference, the process of running an AI model to generate outputs, isn't a one-time event. Instead, it's a persistent, context-heavy, multi-turn process that runs continuously throughout the day and night. A single AI agent might need to maintain awareness of thousands of previous interactions, retrieve relevant information in real time, and generate responses without pausing. That constant operation multiplies power consumption dramatically.

Cursor, which helps software engineers write and debug code faster using AI, is proof that agentic AI coding has become a daily workflow for millions of professional developers. The tool doesn't just answer questions; it continuously assists developers, maintaining context across entire projects and making autonomous suggestions. That persistent usage pattern is exactly what Elon Musk's SpaceX is betting will define the next decade of AI infrastructure.

What Is the Jevons Paradox, and Why Does It Matter for Data Centers?

There's an economic principle called the Jevons Paradox that explains why making something more efficient often leads to using more of it, not less. In the late 1700s, when James Watt improved the steam engine and made it far more efficient, Britain didn't use less coal. Instead, coal consumption roughly quadrupled within a generation because more efficient engines made steam power viable for textile mills, iron foundries, railways, and countless other applications. The efficiency gains unlocked new use cases faster than they reduced overall consumption.

The same dynamic is unfolding in software development right now. AI coding agents like Cursor make building software dramatically cheaper and faster. The intuitive assumption is that this would reduce infrastructure demand: fewer engineer-hours should mean less compute needed, right? Wrong. When software becomes faster and cheaper to build, the world builds vastly more software. More software built by agents leads to more agent usage, which drives more inference compute demand, which requires more GPUs, more networking, more memory, more power, and more cooling running around the clock.

This creates a structural problem for data center operators. They can't simply assume that efficiency gains will reduce their power needs. Instead, they should expect demand to accelerate.

Which Physical Bottlenecks Will Constrain AI Infrastructure Growth?

As agentic AI demand multiplies over the next few years, certain components will become the limiting factors. These are the assets that are hardest to scale, fastest to sell out, and least substitutable. Understanding which bottlenecks will emerge is critical for anyone tracking AI infrastructure investment.

• GPUs and Accelerators: Inference workloads run on the same GPU infrastructure as training, but agentic inference is far more compute-intensive because it runs continuously rather than in discrete bursts. Nvidia remains the dominant supplier, with Broadcom building custom AI chips for Google and 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 and Optical Connectivity: Every token an AI agent generates has to move between memory and processors at extraordinary speeds. When thousands of agents run simultaneously across distributed clusters, the data movement problem rivals the compute problem itself. Arista Networks is the backbone of AI cluster networking, while Corning and Coherent supply the fiber and optical transceivers carrying data between data centers.
• Memory and Storage: Agentic AI is extraordinarily memory-hungry because it requires long context windows, persistent state, and real-time information retrieval. The industry is already structurally undersupplied on high-bandwidth memory (HBM), the specialized memory that AI systems need. Micron is the leading U.S. supplier and has reportedly sold out production under long-term contracts.
• Power and Cooling Infrastructure: 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, making power availability and cooling capacity the ultimate constraint on growth.

How to Assess Data Center Readiness for Agentic AI Workloads

Data center operators and infrastructure investors should evaluate their readiness for the agentic AI era by examining several critical dimensions:

• Power Supply Contracts: Verify that long-term power supply agreements can scale to support 20 to 50 times higher inference loads. Many existing contracts were negotiated for training workloads and may not account for continuous, persistent demand.
• Cooling Capacity: Assess whether current cooling systems can handle the heat output from GPUs running 24/7. Liquid cooling and advanced thermal management will become competitive advantages, not luxuries.
• Network Architecture: Evaluate whether cluster networking can support the extreme data movement requirements of agentic systems. High-bandwidth switching and optical connectivity between racks and data centers will be essential.
• Memory Availability: Confirm access to high-bandwidth memory supplies through long-term contracts or partnerships, since the market is already undersupplied and agentic workloads will intensify that shortage.

The SpaceX acquisition of Cursor isn't just a bet on AI coding tools. It's a $60 billion signal that the infrastructure supporting agentic AI is about to become the most valuable and constrained resource in technology. Companies that secure power, cooling, memory, and networking capacity now will have a decisive advantage over the next decade.