$AMD
We added AMD as a new active Top Pick on 10-15-2025 because the AI boom is expanding the playing field: the insatiable need for AI compute is encouraging industry-wide diversification in hardware.
A transformation may be upon us.
We wrote that “this is our speculation and it could be wrong.”
But something very big changed in Q1 2026, for the better.
AMD is no longer only a speculation on MI450 and Helios.
Q1 showed that the AI buildout is also pulling EPYC CPUs into the center of the infrastructure stack, making the thesis larger than we originally introduced.
And why…?
Because, in the most important part of the server CPU market, high performance AI infrastructure, AMD is increasingly the leader, not a side player as it is in GPUs.
Earnings review
AMD’s Q1 2026 was another clean quarter, but more importantly, it changed the shape of the thesis.
Some numbers fist:
The reported quarter was strong across the board. Revenue reached $10.3 billion, up 37.8% year over year and ahead of the $9.9 billion estimate.
Data Center revenue reached $5.8 billion, up 57.2% year over year and ahead of the $5.6 billion estimate.
Gross margin, operating income, operating margin, and EPS all beat expectations, so this was not a one line revenue beat. The strength flowed through the income statement.
OK, but that’s not really the story.
This is the story:
As we wrote above, the data center story now has two engines.
Instinct GPUs continue to ramp, but EPYC server CPUs are becoming much more important than we originally modeled.
This is real: Server CPU revenue grew more than 50% year over year, and management said cloud and enterprise server CPU sales each grew more than 50%.
That is not background noise.
It means the AI buildout is creating demand not only for accelerators, but also for the CPUs that orchestrate, feed, manage, and parallelize those workloads.
This is where AMD is not a side player. In the most important part of the server CPU market, high performance AI infrastructure, AMD is increasingly the leader.
This is the key change versus last quarter.
Last quarter, AMD was still primarily a hopeful MI400 series story.
If MI450 worked, AMD could earn a meaningful seat at the AI table.
If it did not, the opportunity could compress quickly as Nvidia and custom ASICs absorbed the remaining demand.
That is still true for the GPU side, but it is no longer the entire AMD story.
The AI boom is now expanding the CPU market itself.
Management made a striking revision here.
At its Analyst Day, AMD framed the server CPU market as growing roughly 18% annually to about $60 billion by 2030.
Remember those numbers.
Now, based on customer demand around inferencing, agentic AI, orchestration, data movement, and head node requirements, AMD expects the server CPU market to grow more than 35% annually and reach more than $120 billion by 2030.
That is a massive change in the addressable market and it is directly tied to AI.
This feels like one of those accidental boom moments we are seeing across the AI infrastructure landscape.
It’s a little bit of luck or accidental boom, but still tied to excellent technology and a long-term commitment prior to this to research and development.
It just got a super charged tailwind.
The driver is simple enough.
AI does not just require GPUs.
It requires full systems.
As inference scales and agentic AI begins to create more parallel tasks, the CPU to GPU ratio may shift higher.
AMD even suggested that what used to look like a host node relationship, such as one CPU node for several accelerators, could move closer to one to one in some deployments, and potentially beyond that if agentic workloads proliferate.
That is a real thesis expansion.
Still, let’s not get slap happy, the MI400 series remains the fulcrum for the speculative side and speculations don’t always turn out well.
A failed MI400 series will not be alleviated by the CPU side success. Don’t get overly comfortable with that idea.
But, AMD said customer engagement around MI450 and Helios is strengthening, lead customer forecasts are exceeding initial plans, and the pipeline of large scale deployments is growing.
The company has begun sampling MI450 series GPUs to lead customers and remains on track to ramp Helios production shipments in the second half of 2026.
The cadence is still first half MI350 and CPU strength, then MI450 beginning in Q3 and ramping more meaningfully into Q4 and 2027.
This Meta announcement matters because it broadens validation beyond OpenAI:
AMD said Meta plans to deploy up to 6 gigawatts of AMD Instinct GPUs across multiple product generations, including a custom accelerator based on MI450.
Together with OpenAI, this puts AMD inside the planning cycles of two of the largest AI infrastructure builders in the world.
That does not guarantee success, but it does make the MI450 cycle more credible than it was even three months ago.
The software side also moved forward. ROCm remains the obvious gap versus Nvidia’s CUDA ecosystem, but AMD continues to show progress with broader day zero model support, vLLM integration, stronger MLPerf results for MI355X, and faster development cadence aided by agent based coding workflows.
This matters because the hardware win is not enough and this is crucial to the speculation side.
If customers cannot deploy workloads quickly and reliably, MI450 does not become a platform. It remains a chip. And that means AMD’ stock will lose – likely a lot.
The memory angle also widens the CPU side of the thesis.
This is another, “oh, great, didn’t see that coming, but yes, yes, yes.”
HBM and broader memory supply are tight, and costs are rising across the system, but in data center AI that does not simply translate into weaker demand.
It makes system efficiency more important.
When HBM, DRAM, packaging, and power are constrained, customers need to keep every accelerator and every byte of memory utilized as efficiently as possible.
That increases demand for high performance CPUs because CPUs handle orchestration, scheduling, data movement, preprocessing, retrieval, agent coordination, and the head node work that keeps GPUs fed and productive.
In other words, memory tightness does not directly cause CPU demand by itself, but it raises the value of CPUs inside the AI infrastructure stack, and that strengthens the EPYC side of the AMD thesis.
This is where AMD’s CPU position becomes even more valuable.
In a world where HBM, DRAM, power, and packaging are all constraints, the winning vendors are not just the ones with a good accelerator.
They are the ones that can help customers architect full compute systems around scarce resources.
AMD now has a credible argument across GPUs, CPUs, networking, and rack scale systems. That is a larger argument than we originally underwrote.
Turning back to the numbers, guidance reinforced the momentum.
AMD guided Q2 revenue to $11.2 billion at the midpoint, up 45.7% year over year and ahead of the $10.5 billion estimate.
Gross margin guidance of 56% also came ahead of expectations.
Management expects Q2 data center revenue to grow double digits sequentially, with both server and data center AI growing double digits.
So, the conclusion changes, but it does not become reckless.
The speculative part is still MI450 and Helios. Those products must gain real traction at hyperscale or the GPU opportunity remains constrained by Nvidia and custom ASICs.
But the investment side is now stronger because the CPU business has become an AI infrastructure beneficiary.
AMD is no longer only a bet that MI450 works.
It is now a bet that AI infrastructure needs more of everything AMD is good at.
In short, Q1 made the AMD thesis larger.
Not safer in the sense that MI450 risk disappears, because it absolutely does not.
But larger because the AI infrastructure buildout appears to be pulling forward demand for both accelerators and CPUs, and AMD is one of the few companies with credible products on both sides of that equation.
...1
