EXECUTIVE SUMMARY

• Last week we argued the cloud is bolted to the ground. This week: what the ground is made of.
• The agentic re-rating of server CPUs is real, well-documented and largely behind us. Intel’s data-centre revenue rose 22 per cent in the first quarter, and the shares rose 23.6 per cent in a day — their best since 1987.
• The CPU-to-GPU ratio is marching from 1:8 towards parity and, in agentic configurations, possibly beyond. That is not a rotation between two semiconductor names; it is a rebuild of the data centre.
• Even when an agent pilot fails, the silicon has already been billed. The application-layer trade is binary; the infrastructure trade is pre-committed.
• Reflation persists. We prefer a basket cut across the stack — memory, packaging, foundry — to the single-name CPU bet whose easy double is already in the price.

The trade everyone saw

It has become respectable, in a way it was not eighteen months ago, to be enthusiastic about a central processing unit. AMD has doubled; ARM likewise. Intel — a company for which leveraged investors were quietly composing eulogies this time last year — has tripled, and in April posted its best single day since 1987 on the strength of a quarter nobody saw coming.

The reason is agents. The simplest version, repeated in roughly every brokerage note since Christmas, runs thus: artificial intelligence used to be a chatbot, which was overwhelmingly a GPU problem. It has become an autonomous worker — booking flights, writing code, opening browsers, clicking through forms — which is much more of a CPU problem. GPUs do the parallel arithmetic; CPUs handle the sequential decisions, the tool calls, the orchestration. The longer the workflow, the further the bottleneck migrates.

We have no quarrel with the story; it is broadly correct. We think it is now widely understood, which is why our attention lies elsewhere.

A quiet rebalance

The best way to grasp the shift is to listen to the people who design the boxes. On Intel’s first-quarter call, Lip-Bu Tan said the ratio of CPUs to GPUs inside the data centre had tightened from roughly one-to-eight to one-to-four, and might yet converge on parity or better. His finance chief supplied the granularity: training runs at seven or eight GPUs per CPU, inference at three or four, agentic workloads at approximately one. At a conference last week Mr Tan went further, relaying what frontier-model customers now tell him — “four CPU to one GPU”. Not parity. Inversion.

Figure 1: Parity, and Past It

Then the punchline, which is that supply rather than demand is now the binding constraint. Server CPU prices are up as much as twenty per cent since March. Intel is diverting production from consumer chips. Three new entrants — Nvidia’s Vera, ARM’s first own-brand processor after thirty-five years of pure licensing, and the hyperscalers’ proliferating custom silicon — have crowded in at the moment of greatest scarcity, which is itself a useful signal.

Whatever was structurally undervalued about CPUs eighteen months ago has been recognised. The question is what else the new architecture is buying.

What the agents are buying

Rather a lot, on inspection.

Begin with memory, which has had an even more startling year than compute. SK Hynix earned a first-quarter operating profit of 37.6 trillion won — some twenty-five billion dollars — at a seventy-two per cent margin that reportedly exceeds Nvidia’s own. Samsung’s chip division made 53.7 trillion won, roughly ninety-four per cent of the entire group’s profit. DRAM contract prices rose more than eighty per cent in a single quarter. High-bandwidth memory is sold out across all three suppliers through 2026.

Then packaging, where TSMC reckons demand for the CoWoS that sits between a wafer and a working accelerator runs about three times available capacity. Then the foundry beneath all of it: SK’s chairman warns the wafer shortage may persist to 2030, since adding capacity takes four to five years and the shortfall exceeds a fifth. Then networking, then the identity and security apparatus an autonomous agent requires before any enterprise will let it act.

If the cloud is bolted to the ground, the agent is hard-wired to the data centre — and the data centre is applied semiconductor engineering at half a dozen layers.

The pilot problem

There is a less flattering version of this, in which the agents themselves disappoint. Forrester finds some eighty-eight per cent of agent pilots fail to reach production; Gartner projects two-fifths of agentic projects abandoned by 2027. Meta paid two billion dollars for Manus in December, only for Beijing to order the deal unwound in April — a reminder that the application layer carries political risk as well as technical.

Figure 2: Where the Consensus Stopped

Here the consensus has it half right. The bull case for agents at the application layer is oversold. The bull case at the infrastructure layer is under-priced. The reason is mechanical. Even when a pilot fails, the compute has been billed, the memory allocated, the wafer packaged, the rack built and powered. Forrester’s figure describes a failure of governance and scoping inside IT departments; it does not describe a failure of demand for silicon. JPMorgan runs some four hundred and fifty use cases against a $17.5bn technology budget. Whether all four hundred and fifty justify themselves is, charitably, an open question. The chips were bought regardless, and will be replaced by faster ones in three years.

Concluding thoughts

The positioning we favour is a stack rather than a bet. CPU compute remains a position to hold and no longer one to chase; the easy re-rating is in the price. Memory offers the best combination of structural demand and constrained supply. Packaging and foundry share those characteristics, with the added advantage of a near-monopoly supplier. Networking and security sit at an earlier stage of recognition — generally where we prefer to be entering rather than exiting.

Through our QuadLogic lens this is Reflation in microcosm: structural demand, constrained supply, persistent pricing power, and capital expenditure committed for years rather than quarters. Such cycles reward diversified, stack-level exposure over concentrated single names, particularly once the easy double has been recorded.

The consensus has correctly grasped that the agent revolution needs far more compute than anyone assumed. It has stopped one layer above where the opportunity now sits. The picks-and-shovels trade is well known; the rest of the stack is not wafer thin.

This document is a market commentary intended for professional advisers and institutional investors. It does not constitute investment advice, an offer to buy or sell any security, or a recommendation. Views expressed are those of the author at the time of writing and are subject to change without notice. Past performance is not a reliable indicator of future returns; the value of investments may fall as well as rise. Investors should consult their financial adviser before acting on any view contained herein.