AMD (NYSE:AMD) or Arm Holdings (NYSE:ARM), Which Is the Better AI Buy Right Now?

The AI trade is moving beyond GPUs

AMD and Arm Holdings are starting to sit at the centre of a new AI narrative.

For the last few years, semiconductor investors have focused on GPUs. That made sense. The first stage of AI was about building and training bigger models, and GPUs were the obvious winner in that capex cycle.

But the market is now shifting from building AI models to running them at scale. Once AI moves into everyday workloads, the focus becomes cost, efficiency and handling huge volumes of tasks reliably.

That is where CPUs become much more important. They sit underneath a large part of the inference layer and help manage the workloads that make AI applications usable in the real world.

This gives Advanced Micro Devices (NASDAQ:AMD) and Arm Holdings (NASDAQ:ARM) a stronger place in the AI discussion. The next leg of AI growth may not just be about who sells the most GPUs. It may also be about who powers the inference layer beneath the applications people use every day.

The CPU is becoming a structural bottleneck

Inference is now becoming a much larger driver of AI compute demand. Across the major hyperscalers, it is estimated to account for around 60% to 70% of demand, up from roughly 40% in 2024.

That changes the way investors should think about the semiconductor cycle. In 2024, many investors worried CPUs were becoming less relevant as GPUs dominated AI infrastructure spending. That view missed how CPUs actually function inside data centres.

The AI infrastructure spending race has also accelerated sharply. Amazon has committed to around US$200 billion of capex, Google is guiding to US$175 billion to US$185 billion, Meta has outlined US$115 billion to US$135 billion, and Microsoft’s annualised run rate points toward nearly US$150 billion.

A growing share of that spend is now moving into inference. That is where Arm is gaining momentum through its newer CPU designs, while AMD continues to benefit from demand for more efficient server CPUs.

Server ratios are moving in favour of CPUs

Intel CFO David Zinsner recently described a major shift in server deployment ratios. Data centres have already moved from roughly 1 CPU chip for every 8 GPU chips to around 1 CPU for every 4 GPUs.

Over time, that ratio could move closer to 1 to 1. If that plays out, the demand uplift for CPUs would be significant.

Agentic AI systems are not just about brute-force GPU computation. They also depend heavily on latency, orchestration and CPU-side processing to manage tasks quickly and efficiently.

That puts AMD and Arm in a stronger position as inference becomes the next major growth leg of AI infrastructure.

Agentic AI changes the CPU equation

As AI workloads become more agentic, the compute profile changes quickly.

The number of tokens generated per prompt can rise by 15x or more, and a large part of that workload becomes CPU-bound. Arm’s management has estimated that CPU core demand per gigawatt of data centre capacity could rise from around 30 million cores to 120 million as agentic AI scales.

That is a 4x increase in CPU cores for the same amount of data centre power.

When this is combined with the US$700 billion plus AI capex wave being pushed through by the major hyperscalers, the setup becomes clear. This is no longer just a GPU story. CPUs are becoming a structural bottleneck in the next stage of AI infrastructure.

AMD looks cheaper on growth

If we look at AMD using Capital IQ and market growth estimates, the setup looks interesting.

From 2026 to 2028, AMD’s average forward EBITDA growth is estimated at around 63%. With demand accelerating and supply still constrained for CPUs, the valuation starts to look more attractive.

Based on the forward EV to EBITDA multiple as of 7 May 2026, the blended EV to EBITDA growth ratio sits at 0.48x. That places AMD inside what we would call the opportunistic buy zone.

The caveat is important. This thesis depends on CPU demand accelerating quickly and AMD continuing to capture share in the data centre market. If that growth fails to arrive, the valuation support weakens fast.

Arm is higher quality, but valuation matters

Arm looks different under the same methodology. Its EBITDA to EV growth ratio sits at 1.36x, which is meaningfully higher.

To be fair, Arm should command a premium. IP licensing is a higher-margin business model, and Arm’s revenue base is more diversified across end markets.

But valuation still matters. Arm remains a high-quality business with strong exposure to the same CPU and inference theme, but the current multiple leaves less room for error.

The investors takeaway for AMD and Arm

On that basis, AMD looks like the better buy candidate today.

It offers cleaner upside if the CPU demand thesis keeps strengthening, and its valuation looks more attractive relative to expected EBITDA growth. Arm remains one of the highest-quality ways to play the CPU architecture shift, but the current valuation makes the risk-reward less compelling.

We would watch Arm closely on any pullback. But right now, if investors are looking for the better AI buy between the two, AMD looks like the more attractive setup. The market is starting to recognise that AI is no longer just a GPU story, and that shift could keep bringing CPUs back into the centre of the semiconductor debate.