The Dark Horse of the AI Supercycle: Why CPUs Are Making a Historic Comeback

Everyone's been staring at Nvidia. Meanwhile, the CPU; the chip most people wrote off as yesterday's news, just had its best quarter in history, and the analysts who missed it are now scrambling to catch up. 

On May 22, 2026, AMD CEO Lisa Su flew to Taipei and told the world point-blank: demand for CPU chips is running hotter than anyone expected, and she's asking her manufacturing partners to ramp production right now. That's not a routine corporate statement. That's a CEO signalling that supply can't keep up with what the market is pulling.

Let's break down why the humble CPU is quietly becoming the most important chip in the AI story, what AMD's numbers are screaming, and why this could be the most underfollowed supercycle in tech right now.

The GPU Era Isn't Over, But It's No Longer The Only Story

For the last three years, the AI trade was simple: GPUs accelerate AI model training, Nvidia makes the best GPUs, buy Nvidia. That logic worked brilliantly. But AI has moved on. The industry is no longer just training massive models in data centres. It's running AI everywhere; in search engines, coding tools, customer support, enterprise software, and increasingly, in AI agents.

Think of it this way: training an AI model is like building a new Taj Mahal. You need heavy machinery, cranes, and enormous resources concentrated at one place for a fixed period, that’s GPU. But once it's built, running it, answering millions of visitors every day, needs a completely different kind of infrastructure. That's where the CPU comes in.

GPUs are the cranes. CPUs are the plumbing, the elevators, the security system; the stuff that keeps the building functional 24 hours a day.

How Agentic AI Is Changing Compute Demand

This is the part most coverage misses. The real driver behind the CPU comeback isn't just "AI inference." It's Agentic AI: systems that don't just respond to a question but autonomously execute multi-step tasks, spawn sub-agents, retrieve data, and loop back to check results. Think of ChatGPT going from answering "summarise this" to actually doing your research, writing the report, sending it for review, and filing it.

UBS analyst Timothy Arcuri ran detailed bottom-up modelling on what agentic AI means for chip demand, and the number is startling. Today's AI workloads use roughly 8 to 12 CPU cores per GPU. In agentic AI deployments, that ratio could jump to 80 to 120 CPU cores per GPU, a whopping 10x multiplier. Each agent, and every sub-agent it spawns, requires 1-4 dedicated CPU cores. A single complex task can generate 10 to 100 sub-agents running simultaneously.

In agentic inference, as much as 70-80% of the actual compute workload shifts to CPUs, not GPUs. The GPU still handles the heavy model computation. But the CPU orchestrates, retrieves context, manages memory, calls tools, and keeps everything synchronised in real time. Without a capable CPU, the GPU is essentially sitting idle, waiting for instructions.

The market has barely priced this in.

AMD’s Record Quarter Signals A Structural Shift

AMD recently posted the most important CPU quarter in its history.

Source: AMD Q1 2026 Earnings Report (SEC Filing), Mercury Research

For context: AMD's server CPU revenue share was just 2% in 2017. It now sits at 46.2%. That's a structural takeover. AMD's data centre revenue also crossed Intel's for the first time ever in Q1 2026, with AMD at $5.8 billion against Intel's $5.1 billion.

EPYC processors are so in-demand that AMD is currently selling out every single unit it makes, driven particularly by agentic AI deployments.

AMD CEO Lisa Su In Taipei: Read Between The Lines

On May 22, AMD CEO Lisa Su said in Taipei that the company was asking its partners to ramp up production given strong demand for AI. AMD is a major customer of TSMC, the world's largest contract chipmaker.

This matters more than the headline suggests. When a CEO publicly tells manufacturing partners to push harder, it's an admission that the demand they're seeing has moved beyond their existing plans. It's not a routine ask. Supply chains take months to respond. Su wouldn't be making this call unless the order books were running well ahead of internal forecasts.

AMD has also announced that its next-generation EPYC processor, codenamed "Venice," is now ramping production on TSMC's advanced 2nm process technology; the first high-performance computing product in the industry to enter production at this node. A follow-up chip codenamed "Verano" is already planned to build on this foundation with advanced memory innovations including LPDDR, targeting cloud and AI workloads where power efficiency is becoming a constraint. AMD has also committed over $10 billion in investments across Taiwan's AI ecosystem to scale manufacturing.

The CPU Market Opportunity: $30B to $170B by 2030

UBS estimates the server CPU TAM could grow roughly fivefold by 2030, rising from a $30 billion baseline in 2025 to approximately $170 billion.

Source: UBS Global Research, May 2026

UBS projects CPU revenues could reach $41 billion for AMD and $39 billion for Intel by 2030, while high-end AI CPUs are expected to command prices ranging from $3,000 to $4,000 per unit.

AMD itself raised its own internal server CPU TAM forecast from $60 billion to $120 billion by 2030 and even that looks conservative against what UBS is projecting.

After the server CPU market grew 21% year-over-year in 2025, UBS does not expect a slowdown in 2026, as hyperscaler capital expenditure is projected to rise nearly 81% year-over-year.

Beyond GPUs: The Full AI Hardware Stack

Just as CPUs remained eclipsed under GPUs, there’s more nuance in what’s needed to run AI as a whole. The broader AI hardware story is wider than most investors realise. As Saxo Bank's chief investment strategist Charu Chanana noted in May 2026, AI doesn't run on any single chip. The full stack looks like this:

Source: Saxo Bank Equities Research, May 2026

The GPU gets the attention. The rest of this table gets the revenue.

Our View For Investors: Is CPU The Next AI Supercycle?

Here's the honest take: the GPU supercycle was relatively easy to identify because it was loud. Nvidia's revenue went from $27 billion to over $130 billion in two years. Everyone could see it.

The CPU supercycle is quieter, but potentially more durable. The reason is simple: GPUs face substitution risk (custom silicon, ASICs, TPUs). CPUs don't. Every GPU rack still needs CPUs to function. Every agentic AI deployment multiplies CPU demand. And unlike GPU training, which is front-loaded and episodic, CPU demand from inference and agents is recurring and growing every quarter.

AMD's journey from 2% server share in 2017 to nearly half the market today is one of the most remarkable turnarounds in semiconductor history. And it's not done yet. If UBS's $170 billion TAM projection is even half right, we're still in the early innings.

The risks are real: Intel isn't giving up, ARM is taking share fast, and hardware cycles can be lumpy. Valuations already reflect some AI optimism. But if agentic AI scales the way it currently looks like it will, the CPU may well become the defining chip of the next phase of the AI buildout and not just a supporting actor.