Terafab is the Smoking Gun
Posted May 20, 2026
Chris Campbell
The world's most aggressive vertical integrator just made his biggest bet yet.Yesterday, James told you energy was the bottleneck for AI.
Today I'm telling you it's (still) memory.
Both are true. That's the thing about bottlenecks—they stack.
Energy decides how big AI can ultimately get. Memory decides how much useful work the hardware already built can actually do.
Right now, expensive GPUs spend real time waiting on data instead of crunching it.
That's not a rumor. We didn’t hear it through the grapevine.
It's documented in every serious benchmark of large AI models—utilization rates well below theoretical peak, with memory bandwidth the most common culprit.
Because here's the open secret of AI hardware…
A chip can only calculate on data it can reach. If the data can't get there fast enough, it’s a blast furnace fed by a teaspoon.
The Math Everyone Ignores
A modern AI model carries about 400 billion "weights"—the patterns it learned during training. Stored in standard format, that's 800 gigabytes. The equivalent of 200 HD movies.
Just to hold what the model already knows.
Then you start a conversation.
Your prompt. Files you uploaded. The notes the model takes while it thinks. One busy session burns another 50 to 200 gigabytes.
Multiply by every user on the platform at once. You're staring at tens of thousands of gigabytes per data center, just to keep the lights blinking.
That's why the industry obsesses over a specific kind of memory called HBM—high bandwidth memory. HBM is regular memory stacked vertically like a tiny skyscraper, bolted right next to the GPU.
The reason that matters is distance. Memory two inches from the chip feeds it dramatically faster than memory two feet away. HBM is the difference between a country road and a 32-lane expressway running into the factory.
Three companies on Earth make HBM at scale. Micron. Samsung. SK Hynix. That's the entire list.
The memory trade still has legs, but the crowd is catching on.
But let’s consider what’s coming.
The Agent Tax
A chatbot is a tourist. Asks one question, gets an answer, leaves. Memory cost is small and predictable.
An AI agent is an employee. It remembers an objective. Tracks the conversation. Opens files. Calls tools. Branches into sub-tasks. Runs for hours.
Micron's own modeling shows each active agent burns five to ten times more memory than a chatbot session. Per agent. Per user. Running in parallel. Persistently.
The demand math becomes:
More users × more agents per user × more tasks per agent × more memory per task × longer persistence.
Hockey stick stacked on hockey stick.
And then… and then… the second leg arrives—AI leaving the data center.
Physical AI.
Humanoid robots. Self-driving cars. Drones. Smart glasses.
None can act like a data center. A drone can't draw a megawatt. A car can't pause on the memory bus while merging at 70 miles per hour.
Memory becomes a safety issue. It’s a whole new demand layer most analysts still haven't put in their models.
The Smoking Gun
Elon's new project—Terafab—aims to bring logic, memory, packaging, and testing under one roof. The world's most aggressive vertical integrator just decided to bring memory in-house.
If he's pulling memory inside Tesla, it's because memory is a massive gating constraint on robotics, autonomous vehicles, Optimus, and whatever space-based compute SpaceX builds next.
What’s not being talked about enough is Musk’s other vertical integration… 250 miles up. SpaceX is quietly turning into the infrastructure layer for space.
Where those two ambitions meet—AI infrastructure and space infrastructure—is where James has set his focus.
Recently, our team has identified a small public company operating in both lanes at once. And James believes it’s a prime candidate for a SpaceX buyout.
Here's the part I can't put in this letter. The name. The ticker. The buyout math.
None of it goes on a public page.
It lives inside a small beta-test James just launched, and beta-testers see every play like this one first.