1/
AMD isn’t Nvidia. It doesn’t need to be.
What matters: they’re finally building a full-stack AI/data center offering that’s real — not theoretical.
Execution is lining up. Revenue is following.
A few key reasons why:
Discussion
2/
The upcoming MI350 (CDNA 4) GPU series is the big unlock.
Mid-2025 launch. 35× perf over MI300X.
No need to redesign racks — it drops into current MI300 systems.
Major deployment w/ Oracle already underway.
That’s not roadmap noise. It’s movement.
3/
But the present is moving too.
EPYC Turin (Zen 5) is in-market, shipping, and scaling.
Used across 150+ server platforms.
30+ new cloud instances live from AWS, GCP, Oracle.
This isn’t theoretical traction — it’s booked compute.
4/
MI300X and MI325X are shipping now.
They’re handling real-world inference, including Meta’s LLaMA 405B.
MI325X builds on MI300X — better memory, better economics.
Sets the stage for MI350 without disruption.
5/
AI PCs are noise in most headlines, but Ryzen AI is actually selling.
+50% QoQ notebook sell-through
+80% YoY commercial design wins
HP, Lenovo, Asus — all in.
Does it matter long-term? TBD. But it’s margin-positive now.
6/
The missing piece? Software.
ROCm isn’t CUDA. But it’s catching up. Fast.
Bi-weekly updates. 2M+ Hugging Face models supported.
LLaMA 4, Gemma, DeepSeek — all Day-0.
This used to be AMD’s moat problem. Less so now.
7/
Acquisition of ZT Systems plugs the final gap:
Full rack-scale systems — CPUs + GPUs + networking.
Think Nvidia DGX competitor — but modular.
ZT gives AMD the vertical muscle to sell full AI infrastructure to hyperscalers.
8/
What matters:
→ Turin is already delivering
→ MI325X is ramping
→ MI350 is sampling
→ ROCm is working
→ ZT is the enabler
This is no longer “wait for next gen.”
They’re in market. And in motion.
9/
Still early. Nvidia’s CUDA + Blackwell + market capture is real.
But AMD isn’t just playing catch-up.
They’re building something that fits the post-monopoly stack.