The Re-Rating of the Central Processing Unit
The AI hardware trade started with graphics chips. The central processing unit — cheaper to build, harder to replace, and increasingly the binding constraint in AI deployments — is the trade now underway.
Investment Thesis
The first act of the AI hardware cycle had one trade: own the graphics processing unit (GPU), and that call proved correct — the market has since priced it substantially. The second act belongs to the central processing unit (CPU), and while several names have re-rated meaningfully, we believe the runway ahead is considerably longer than current prices imply.
To understand why, it helps to think about what an AI agent actually does. It does not just generate text — it receives a task, breaks it into steps, and goes off to complete each one: searching the web, reading documents, running calculations, checking databases. The GPU handles the narrow moments of actual thinking. Everything else — fetching information, routing results, keeping track of what has been done — falls to the CPU. In real-world agentic deployments, the GPU is actively working for as little as 10–12% of the time. The CPU is running for the other 88–90%.
As AI moves from simply answering questions to autonomously completing tasks, the demand for CPU capacity per AI system grows continuously. Intel’s management confirmed that the CPU-to-GPU ratio has shifted from roughly 1:8 during the model training era to 1:4 across today’s inference deployments, with agentic workloads pushing the balance toward one-to-one. AMD raised its market size forecast for server CPUs to $120 billion by 2030. NVIDIA — the world’s largest GPU company — disclosed that it expects nearly $20 billion in revenue this year from its own new CPU product alone, entering a $200 billion market it had never previously addressed. These are not analyst projections — they are the companies actually building this infrastructure describing what they see in front of them.
The industry tends to view one CPU per GPU as the natural endpoint. We think that underestimates where this goes. The cost gap between GPUs and CPUs is substantial — GPUs carry a significant manufacturing premium because of the specialised High Bandwidth Memory (HBM) packaging bonded directly to the chip, a component that accounts for a large share of production cost and is not required in CPUs. When an operator’s GPU sits idle for most of the working day waiting for the CPU to finish its coordination work, adding more CPUs is not just logical — it is inexpensive relative to the asset being protected. That arithmetic does not resolve at parity — it keeps moving in the CPU’s favour.
Five Ways to Own the Trade
| Company | Investment Thesis | Latest Developments |
|---|---|---|
| AMDPreferred Pure Play | The only company selling CPU, GPU and networking as a single AI rack platform — with no direct x86 competition until mid-2027. Re-rating depends on customers buying the full stack, not the GPU alone. | Q1 2026: rev $10.3B +38%; DC $5.8B +57% EPS $1.37 vs $1.27E Q2 guide $11.2B Server CPU TAM $120B+ by 2030 |
| IntelPreferred Pure Play | Unable to meet its own demand, and chosen by NVIDIA for its flagship AI system despite NVIDIA having an in-house alternative. Foundry needs an external customer to re-rate fully. | Q1 2026: DCAI $5.1B +22%; op. margin 31% Q2 DCAI up double digits QoQ 18A yields 2 qtrs ahead Xeon 6+ (288 cores) launched Computex June 2026 |
| NVIDIADiversified Beneficiary | NVIDIA built a CPU to stop its own GPUs sitting idle — the most powerful confirmation this bottleneck is real. US export controls removing China from guidance are the main near-term overhang. | Q1 FY2027: rev $82B +85%; DC $75B +92% Q2 guide $91B (excl. China) ~$20B Vera CPU revenue FY2027 Vera Rubin NVL72 ships Q3 2026 |
| Arm HoldingsNew Entrant | Inside virtually every major AI chip being built today. Launching its own chip shifts Arm from royalties to full chip economics — demand doubled within weeks of launch. Competing with NVIDIA and Amazon puts the royalty base at risk. | FY2026: $4.92B +23% DC royalties 2×+ YoY Q1 FY2027 guide $1.26B +20% AGI CPU demand $2B+ (FY2027–28) |
| QualcommNew Entrant | Priced as a fading smartphone business — yet has a confirmed AI chip order from a major cloud company shipping this year, valued at zero by the market. Apple’s modem exit and an unproven data centre business are the two risks. | Q2 FY2026: rev $10.6B; EPS $2.65 Auto $1.3B +38% — run rate above $5B Hyperscaler DC shipment Dec 2026 Q3 guide $9.2–10.0B |
| MediaTekNew Entrant | The world’s largest smartphone chip company — with manufacturing scale, cost structure and Asian cloud relationships most rivals lack. Data centre CPU ambition is early but the industrial logic is compelling. | No DC CPU product yet Dimensity edge AI chips in production Taiwan-listed (2454.TW) ~15× forward P/E AI partnerships with Google, Microsoft, NVIDIA |
The Ratio Shift — How CPU and GPU Balance Is Changing
CPUs cost a fraction of GPUs — partly because they do not require the expensive High Bandwidth Memory (HBM) packaging bonded directly to the chip that makes GPUs so costly to manufacture. That cost gap makes provisioning more CPUs economically compelling. The industry estimates a 1:1 ratio as the endpoint. We think it goes further.
The Case for More Runway
In the era of training large AI models, one CPU could support seven or eight GPUs. That ratio made CPU demand look irrelevant against the GPU boom. It no longer holds. Intel confirmed on its Q1 2026 earnings call that in today’s AI deployments, the ratio has already moved to one CPU per four GPUs — and in agentic deployments (where AI completes tasks autonomously), it is approaching one-to-one. More strikingly, Intel’s own customer data shows that GPUs in these systems are actively working only 20–30% of the time. The GPU, the most expensive component in the rack, sits idle for the majority of every working day. Server CPU prices have risen 10–20% since March 2026 as the market begins to price this in.
Running an AI data centre costs an enormous amount of electricity — and power capacity has become a genuine constraint on how fast the industry can grow. Arm projects that next-generation AI workloads will need four times as many CPUs per megawatt of electricity versus today’s systems. CPUs now represent 44% of total power consumption at this scale, making energy efficiency the decisive battleground. Arm-based chips win that battle, which is why they now power roughly half of all deployments at the world’s largest cloud providers.
Better software normally means you need less hardware, but in AI the opposite is happening. A recent update to Python — the coding language most AI systems are built on — unlocked the ability to run four times as many tasks at once. More tasks simultaneously means more demand on CPUs, not less. Add to that the rise of multi-agent systems, where dozens of AI agents work in parallel on different parts of the same problem, and the multiplier effect becomes significant. Qualcomm’s CEO summarised the trajectory at Computex 2026: the volume of AI processing the world demands every ten seconds will be forty times larger by 2030.
A Bottleneck That Cannot Be Solved by Better Software
When GPU chips became scarce and expensive, engineers responded brilliantly — they found ways to make the same hardware do ten times more work by compressing models and cutting computational corners. This is why a chip shortage that looked like it would last years was partly resolved in months. Investors have rightly come to expect that AI hardware constraints get solved.
Every previous constraint in AI hardware was eventually solved by making software more efficient — doing the same work with less compute. That will not happen here. The CPU’s job in an agentic system is not approximate. A database must return the right result. An API call must complete. A web page must load. You cannot compress or shortcut any of those tasks. The bottleneck that cannot be engineered away is the one worth owning.
ISA (Instruction Set Architecture) — The Architecture That Determines Who Gets Paid
Not every company benefits equally from rising CPU demand. The key is something called the Instruction Set Architecture (ISA) — essentially the language a chip speaks. It determines who can build competing chips, who gets paid when demand rises, and how much they get paid. There are three architectures that matter in this story.
Only Intel and AMD hold x86 licences. Decades of enterprise software certifications and switching costs constitute a moat that technical performance alone cannot displace. When CPU demand rises, Intel and AMD capture the premium directly at the socket level — server CPU prices already up 10–20% since March 2026.
- Moat Type Enterprise software lock-in
- Weakness Less efficient per unit of electricity
- Who Profits Intel & AMD capture it fully
Arm designs chip blueprints and licenses them to others — Apple, NVIDIA and Amazon all build on Arm’s architecture. Its chips use less electricity than x86, which matters enormously when data centres are fighting over power capacity. Arm has recently launched its own chips for sale, shifting from collecting a small royalty on others’ sales toward earning the full selling price — a transition that significantly changes the economics of this business.
- Moat Type Best performance per unit of electricity
- Evolution Transitioning from royalties to chip sales
- Tension Competing with its own paying customers
RISC-V carries no licence fee. The ecosystem is not yet competitive in production data centres, but strategic investment activity is accelerating. NVIDIA has backed SiFive and opened NVLink Fusion to RISC-V CPUs. Qualcomm acquired Ventana Micro Systems in December 2025 — a RISC-V specialist — principally to deepen its Oryon CPU team. These are hedges: if ARM’s own chip ambitions create licensing tensions, RISC-V becomes the alternative of first resort.
- Moat Type No fees, fully customisable
- Weakness Not yet ready for production at scale
- Signal NVDA & QCOM have invested as a hedge
Risks to the Thesis
| What Could Go Wrong | Why It Matters |
|---|---|
| Software gets more efficient | Better scheduling and memory management will reduce CPU demand at the margin, but cannot eliminate the core requirement — every external task an AI agent completes needs a CPU to handle it. AMD and NVIDIA are both building their businesses on the assumption that this bottleneck persists. |
| The big cloud companies build their own | Amazon, Google and Microsoft are all designing their own server chips rather than buying from Intel or AMD. If this deepens, it narrows the addressable market for x86 incumbents — though Arm collects a royalty regardless of who manufactures the chip. |
| AI agents take longer to scale than expected | Most enterprises are still using AI as an assistant rather than an autonomous agent. If that transition takes one to two years longer than expected, the CPU demand inflection is real but deferred. Intel and AMD’s latest earnings suggest it is already underway. |
| Intel’s factory story needs proof | Intel’s chip business is delivering — Xeon 6 Plus launched on 18A in June 2026 with yields ahead of plan. The foundry-for-others business still needs a named external customer. Without one by year-end 2026, the foundry premium in Intel’s valuation compresses. |
| Arm’s customers start building around it | By launching its own chips, Arm now competes with the customers paying its royalties — NVIDIA, Qualcomm and Amazon. Qualcomm’s acquisition of a RISC-V specialist in December 2025 is an early signal that alternatives are being explored. |
| Export restrictions cut deeper | NVIDIA has already excluded China from Q2 2026 guidance entirely. Huawei’s Ascend 950PR entered mass production in March 2026 and is filling the gap domestically. Further tightening could extend the impact across the coverage universe. |
Please refer to our full disclaimer for important disclosures and regulatory information.
.png)
