AMD launches Instinct MI350P PCIe AI accelerator for data centers

Advanced Micro Devices has unveiled the new AMD Instinct MI350P, an AI accelerator in a standard PCIe form factor designed to go head-to-head with NVIDIA’s H200. According to Tweakers, AMD is targeting data centers that want to run AI models without immediately switching to complex liquid cooling or custom server architectures.

The new MI350P is built on the same generation as the more powerful MI350X, but delivers roughly half the performance. That gap is strategic: AMD aims to reach a broader market, including cloud providers, enterprise data centers, and companies determined to keep using air‑cooled racks.

The PCIe version makes the MI350P easier to deploy in existing servers, slashing infrastructure costs for companies expanding AI capacity without full data center overhauls.

Many modern AI systems now rely on specialized SXM or OAM modules with extreme power budgets and liquid cooling. They deliver peak performance but add operational complexity. AMD positions the MI350P as a more accessible alternative for organizations running inference, smaller training jobs, or hybrid AI workloads.

Air cooling is a big part of that pitch. Data centers worldwide face rising energy prices and limits on power and cooling capacity. A PCIe card that fits into existing racks makes large‑scale rollouts simpler—and cheaper.

AMD has long competed with Nvidia in AI hardware, but the fight is shifting from raw performance to real‑world deployability. Nvidia currently dominates with the H200 and Blackwell platforms, helped by CUDA and a powerful ecosystem.

AMD is trying to narrow that lead with open software, lower costs, and more flexible deployment options. The MI350P is a case in point: instead of chasing only maximum training throughput, AMD is prioritizing practical scalability within existing infrastructure.

That’s appealing for companies that need AI horsepower but don’t have hyperscaler budgets for full rack replacements or liquid‑cooled clusters.

The AI hardware market is tilting toward inference—the actual execution of trained models—which often calls for different hardware than training.

Training clusters crave peak compute; inference is more about efficiency, density, and operating costs. PCIe cards like the MI350P can therefore shine for AI services in enterprise environments, edge data centers, and cloud platforms.

AMD is tapping into a broader trend: not only are the largest models growing, but demand is also rising for affordable, production‑ready generative AI inside existing IT estates.

The MI350P launch shows AMD widening its AI strategy. The company isn’t just chasing top‑end performance; it’s also going after segments where easier deployment and lower infrastructure costs matter.

That’s timely as more companies push AI features without billion‑dollar bets on new data center designs. As a result, the AMD‑Nvidia contest is becoming less about benchmarks—and more about complete AI infrastructure.