AMD Ryzen AI Max PRO 400 targets 300B local LLMs
Key insights
- AMD's Ryzen AI Max+ 495 PRO supports 192GB unified memory, enabling 300B-parameter LLM inference locally for the first time on x86.
- HP and Lenovo are the launch OEM partners for AMD's 'Agent Computers' lineup, with systems shipping in Q3 2026.
- The new series raises AMD's memory ceiling from 128GB on the Max 300 to 192GB, targeting workloads previously requiring multi-node cloud setups.
Why this matters
Local inference at 300B-parameter scale means enterprises can run frontier-class models on-premises, removing the compliance, latency, and cost constraints that currently push sensitive AI workloads into managed cloud services. The 192GB unified memory target places AMD in direct competition with Nvidia's GB10 and Apple's M-series Ultra chips, and HP and Lenovo's distribution reach gives AMD a credible commercial path that GPU-first competitors lack in the enterprise workstation channel. For founders and AI infrastructure teams building for regulated or air-gapped environments, this accelerates the timeline for private deployments that previously required custom server builds or multi-GPU rack configurations.
Summary
AMD's Ryzen AI Max PRO 400 Series, announced at Computex on May 31, is the first x86 chip AMD claims can run 300B-parameter LLMs locally at 4-bit quantization without cloud offload.
The flagship Ryzen AI Max+ 495 PRO tops out at 192GB unified memory and 160GB VRAM, up from the 128GB ceiling on the current Max 300 series. HP and Lenovo are the launch OEM partners, with systems shipping in Q3 2026 under AMD's 'Agent Computers' positioning.
Essentially: (AMD, HP, Lenovo) are moving data-center-scale model capacity to a single workstation.
- 192GB unified memory is the unlock that allows full 300B-parameter models to run without multi-node infrastructure.
- AMD's 'Agent Computers' framing targets enterprise buyers who need autonomous AI task execution without cloud dependency.
- Q3 2026 puts AMD in direct competition with Nvidia's GB10-based Project DIGITS for the local frontier-inference market.
Workstation hardware is now closing the gap with cloud inference at frontier model scales.
Potential risks and opportunities
Risks
- If AMD's 300B-parameter performance claims do not hold up against independent benchmarks at launch, HP and Lenovo face credibility exposure on 'Agent Computers' marketing that enterprise procurement teams acted on during Q3 2026 buying cycles.
- Nvidia could accelerate GB10 Project DIGITS availability or cut pricing to undercut AMD's OEM channel before Q3 2026 systems ship, leveraging existing CUDA ecosystem lock-in with enterprise LLM inference stacks.
- AMD's 160GB VRAM figure relies on unified memory architecture; workloads optimized for discrete GPU VRAM may underperform relative to marketing claims, creating post-sale support burden for HP and Lenovo enterprise accounts.
Opportunities
- LLM inference optimization vendors targeting CPU and APU architectures (Ollama, LM Studio, llama.cpp maintainers) gain a large-memory target platform that expands their enterprise user base ahead of Q3 2026.
- Enterprise AI software vendors building on-premises deployment options (Glean, Writer, Cohere) can now credibly market 300B-parameter private deployments without data center hardware, expanding their addressable market into mid-market regulated industries.
- AMD's OEM partnerships with HP and Lenovo give systems integrators (CDW, SHI, Presidio) a new high-margin AI workstation SKU to position with federal and financial services customers who have strict data sovereignty requirements.
What we don't know yet
- Pricing for HP and Lenovo systems running the Ryzen AI Max+ 495 PRO has not been disclosed ahead of Q3 2026 availability.
- Whether AMD's 300B-parameter claim applies to specific model architectures such as Llama 3 405B or Mistral Large, or is a theoretical VRAM capacity argument, has not been clarified.
- Inference throughput benchmarks at 4-bit quantization for 300B models have not been released, making direct performance comparisons with Nvidia GB10 or Apple M3 Ultra impossible at this stage.
Originally reported by amd.com
Read the original article →Original headline: AMD Ryzen AI Max PRO 400 Debuts at Computex — First x86 Chip for 300B-Parameter Local LLMs, 192GB Unified Memory, Ships Q3 2026