Hey
The past few months have been a blur.
I’ve gotten a little too familiar with the “AI psychosis” people joke about. Sleep has been rough. I wake up early and spend an unreasonable amount of time with Claude Code. I’ve started referring to it as “we.” (to my wife’s chagrin)
That’s because we’ve been building something, and it’s finally live enough to share!
The short version: it’s an intelligence system for the AI infrastructure supply chain, compute, memory, and power. It tracks three indices across 80+ sources and updates daily. It shows when something is tightening before the market fully catches on, and it shows the evidence behind the signal.
The core question it answers is simple: is the AI build-out tightening, or is the trend starting to break?
Hyperscalers are deploying roughly $700 billion in capex this year. I'm convinced that the companies sitting on the binding constraints in that supply chain could become the defining infrastructure winners of the decade. The hard part is figuring out which layer becomes the constraint before it shows up in earnings.
Our system does. It's called Tessara.
Here's what two years of compute supply-demand looks like:
The thing most people missed was.. because DeepSeek proved you could do more with less compute…the market assumed GPU demand would ease. Instead it accelerated.
Because efficiency made entirely new workloads viable. Specifically, agents. Which we’ve been talking about here a lot. The same efficiency gains that let you run inference 10x cheaper are what made it practical to chain dozens of LLM calls into autonomous workflows. Every agent run is a multiplier on compute demand.
We're now in the tightest compute regime since the original H100 rush. The difference is that this time the demand is not training runs. It's millions of agents running inference 24/7. And the supply chain is still catching up. That's Jevons paradox playing out in real time.
Today, I’m making our regime dashboard public. No signup, no paywall. You can see the scores, what moved, and which layer looks like the binding constraint:
I’m also running a private beta for the full terminal: the evidence chains behind the scores, the complete signal feed, regime shift alerts and everything you need to stay at the frontier. Early results look strong.
If you're an investor, analyst, or operator tracking AI infrastructure and want in, reply to this email. I'm onboarding people in small batches.
— Teng