Our last report identified a $5.5 trillion capex supercycle and closed it by naming three constraints limiting the machine.

Capital is committed. What decides who gets paid is which supply constraint is connected first, and the shift from training models to agents operating is reshaping all three.

Memory is where it first appears, and its blowout quarter is proof of that.

This week’s TechEdge covers:

• CPUs: The most underappreciated twist
• Memory: Battlefield, what else? Micron Just told us
• Networking: Our Most Faithful Calling
• The bottom line: What this means for investors.

The ratio broke down because training was a GPU story. Dozens of GPUs off a single CPU.

Agents behave like people: each requires its own computation, its own memory, and constant communication with other machines. It overturns the math of old hardware and pushes the big demand into three types of training that are treated as afterthoughts.

We believe that the agent cycle will be about 3 times longer than the hardware training cycle in the next two to three years.

CPUs: The most underappreciated inflection

The agent ratio tends to be one CPU per GPU, as each agent requires its own orchestration.

We see the CPU market growing from ~$35-40 billion today to $200 billion plus by 2030, higher than its own ~$120 billion estimate and the ~$170 billion sell-side consensus.

In terms of scale, Cloudflare requires ~10 million CPUs to serve 100 million knowledge workers and ~1 billion globally. AMDand all have flagged it, and 2025 is the first year of inflection. Stocks have moved, but we think they only reflect the first leg.

Memory: Battlefield, and what Micron just told us

In this quarter, Memory ceased to be a thesis and became a print.

Micron beat every segment, and the beat was almost all price, not volume: Adjusted gross margin nearly doubled to ~80%, for a business that historically made 30-50% in good times and turned cash flow negative in bad times.

Prices are rising ~7x from the bottom of the cycle, going straight to Micron’s bottom line, and straight out of the budgets of NVIDIA, Alphabet and Microsoft.The driver is structured. Agentic AI adds a second, separate demand stream on top of HBM: agents require generic DRAM and NAND, the same memory that powers a typical PC.

Meanwhile, each HBM wafer uses three to four conventional DRAM wafers, so producers are shifting their capacity as demand for the commodity increases.

Two demand curves that used to move independently collide in a finite supply base. Micron’s CEO knows when supplies will arrive (especially at HBM) and we think the squeeze will last until 2028.The real change is the contract. Micron locked in 16 strategic customer agreements at fixed prices worth at least $100 billion through 2030, each for ~3 years.

Memory is always sold on spot, which is why it traded at a constant discount to cyclicality, so the multi-year fixed price is exactly what can reduce the downside, and why the market is debating re-ratings.

We won’t underwrite it yet: the proof comes only from generating cash through a full down cycle, which is still one cycle away. The contracts cut both ways (customers could walk if construction slows), and Chinese entry is a real long-term threat.

So we’d rather own a pick and shovel: capital equipment names like , and which sell tools to increase capacity, not the commodity itself.

One to watch: Micron flags humanoid robots, which require ~10x the memory of today’s AI, as a second wave that could extend this cycle into the next decade.

Networking: Our highest assurance call

Agents talk constantly, and traffic has to move.

“Copper vs. Optical” framing is pretty simple. Both of them are on different scale points. Therefore, we prefer technology-agnostic operators as much as any single medium.

The strain is already visible: optical lead times on some products have increased to 12 months, and fiber costs have risen by 50 percent since January. Industry estimates put the final optical market at over $150 billion today, ~9x.

Networking has overtaken both memory and hyperscalers this cycle, and we still see the highest consensus earnings here, along with semi-cap devices.

The bottom line: What this means for investors.

The hurdles now decide the winners.

• CPUs are affected and owned.
• The recall is real and is already being re-priced, but there is much more to the price, and the re-pricing is still to be achieved through the down cycle, which is why we want to own goods rather than commodities.
• Networking is our most trusted calling.

In the agentic period, returns accrue to the controller of scarce inputs, and scarce inputs are simply exchanged.