GM. This is Milk Road AI, the newsletter trying to stay one bottleneck ahead.
The GPU shortage isn't the biggest opportunity anymore. A new set of bottlenecks is emerging that could produce the next generation of AI winners. We’ve outlined a few of them in today’s edition.
Case in point, today's partner Circle is tackling a bottleneck you haven't thought about yet.
Your agent can code, browse, and find the best deals, but it’s flat broke. Circle Agent Stack is a programmable money layer for the agentic economy that lets agents hold USDC, discover and pay for services, and get paid.
THREE NEW TRADES FORMING WITH MEMORY, OPTICS AND POWER ⚡
A year ago, the trade was simple. GPUs were scarce and Nvidia was the obvious winner.
But now, the trade has moved past the GPU scarcity.
The shortage now sits in everything that has to surround a GPU to make it useful: memory, packaging, optics and power delivery.
To understand this, imagine a GPU as the filling in a very expensive sandwich:
- Memory feeds it data from one side.
- Advanced packaging holds everything together.
- Optical wiring connects it to tens of thousands of other GPUs in the same cluster.
- Power delivery keeps the whole thing alive.
Every one of these layers is becoming more constrained and that's where our analyst Vincent sees the next opportunities.
Squeeze #1: Memory
Only three companies make high-bandwidth memory (HBM) in volume while SK Hynix supplies most of it.
Demand is so strong that HBM4, the next generation of AI memory, is already sold out through mid-2027.
What’s even more interesting is that three memory makers are posting higher operating margins than Nvidia:
- Micron: 80.4%.
- SK Hynix: 71.5%.
- Sandisk: 69%.
Squeeze #2: Optical connectivity
Once the GPUs are assembled, tens of thousands of GPUs are wired together so they can train the same model in parallel.
This is where NVIDIA’s investment in Corning fits. The partnership is designed to expand U.S. manufacturing of the optical connectivity needed to link increasingly large AI clusters.
Squeeze #3: Power
Three years ago a single GPU rack consumed around 40kW.
But now, NVIDIA's Vera Rubin racks will hit between 600kW and 1MW.
That creates a huge opportunity for companies supplying power infrastructure like Infineon:
- Announced its second price hike of 2026 on July 1 following an earlier hike in April.
- Expects AI data center revenue to grow from €1.5B to €2.5B next year.
- Committed an incremental €500M to expand capacity.
Here's the broader point.
The AI trade is broadening beyond the GPU.
The biggest opportunities are shifting to the companies that feed, connect, package, and power AI infrastructure.
That's where the next layer of the trade lives.
All the information in this edition comes from a small section of Vincent’s latest PRO report. You can go check out the entire report here.
THE FIRST FINANCIAL STACK BUILT FOR AGENTS
Your AI agent can browse the web, write code, and schedule your meetings.
Hand it a bill, though? Uh-oh spaghetti-o.
Until now, money was the missing piece. Circle Agent Stack fixes that.
Circle Agent Stack is the financial-infrastructure layer built specifically for autonomous agents:
- Agent Wallets
- Agent Marketplace
- Nanopayments powered by Gateway
- Circle CLI
- Circle Skills
All powered by USDC.
Agents are evolving from assistants into economic actors. The agents that run on purpose-built financial infrastructure will be the ones that actually work.
