Our last report laid out the $5.5 trillion capex supercycle and closed by naming the three bottlenecks that could cap the machine.
The capital is committed. What decides who gets paid is which supply constraint binds first, and the shift from training models to running agents is rewiring all three.
Memory is where that shows up first, and ’s blowout quarter is the proof.
This Week’s TechEdge covers:
- CPUs: the most underappreciated inflection
- Memory: the battleground, and what Micron just told us
- Networking: our highest-conviction call
- The Bottom Line: what this means for investors
The ratios broke because training was a GPU story; dozens of GPUs off a single CPU.
Agents behave like people: each needs its own compute, its own memory, and constant communication with other machines. That inverts the old hardware math and drives outsized demand into three categories training treated as afterthoughts.
We think the agentic era will be roughly 3x the size of the training era in hardware spend over the next two to three years.
CPUs: The Most Underappreciated Inflection
The agentic ratio moves toward one CPU per GPU, because every agent needs its own orchestration.
We see the CPU market growing from ~$35–40 billion today to $200 billion-plus by 2030, above ’s own ~$120 billion estimate and the ~$170 billion sell-side consensus.
For scale, Cloudflare pegs U.S. demand at ~10 million CPUs to serve 100 million knowledge workers, and ~1 billion globally. AMD, , and have all flagged it, and 2025 is the first year of the inflection. The stocks have moved, but we think they reflect only the first leg.
Memory: The Battleground, and What Micron Just Told Us
This quarter, memory stopped being a thesis and became a print.
Micron beat across every segment, and the beat was almost all pricing, not volume: adjusted gross margins nearly doubled to ~80%, unheard of for a business that historically earned 30–50% in good times and went cash-flow negative in bad ones.
Prices are up ~7x from the cycle bottom, flowing straight to Micron’s bottom line, and straight out of the budgets of NVIDIA, Alphabet, and Microsoft.
The driver is structural. Agentic AI adds a second, separate demand stream on top of HBM: agents need ordinary DRAM and NAND, the same memory that powers a normal PC.
Meanwhile each HBM wafer consumes three to four conventional DRAM wafers, so producers are converting capacity just as commodity demand climbs.
Two demand curves that used to move independently are colliding into one constrained supply base. Micron’s CEO sees no line of sight to when supply catches up (particularly in HBM) and we think the squeeze runs into 2028.
The real shift is contractual. Micron locked in 16 strategic customer agreements at fixed prices, ~3 years each, worth a minimum $100 billion combined through 2030.
Memory has always sold at spot, which is the reason it traded at a permanent discount for cyclicality, so multi-year fixed pricing is exactly what could dampen the downside, and why the market is debating a re-rating.
We wouldn’t underwrite it yet: the proof only comes from generating cash through a full down cycle, which is still a cycle away. The agreements cut both ways (customers could walk if the buildout slows), and Chinese entrants are a real long-term risk.
So we’d rather own the picks-and-shovels: capital-equipment names like , and that sell the tools to expand capacity, not the commodity itself.
One tail worth watching: Micron flags humanoid robots, which need ~10x the memory of today’s AI, as a second wave that could stretch the cycle into the next decade.
Networking: Our Highest-Conviction Call
Agents talk constantly, and that traffic has to move.
The “copper vs. optical” framing is too simple. It’s both, at different scale points. So, we prefer technology-agnostic enablers like and over any single medium.
The strain is already visible: optical lead times have stretched to 12 months on some products, and fiber pricing is up 50% since January. Industry estimates put the eventual optical market at $150 billion-plus, ~9x today.
Networking has outperformed both memory and the hyperscalers this cycle, and we still see the largest consensus earnings upside here, alongside semi-cap equipment.
The Bottom Line: What This Means for Investors
The constraints now decide the winners.
- CPUs are inflecting and under-owned.
- Memory is real and already re-pricing, but a lot is in the price, and the re-rating still has to be earned through a down cycle, which is why we’d rather own the equipment than the commodity.
- Networking stays our highest-conviction call.
In the agentic era, returns accrue to whoever controls the scarce input, and the scarce inputs just changed.