AMD Bets on Inference and Agentic AI as Nvidia's Training Moat Holds
Analysis comparing Nvidia and AMD for long-term AI investment, outlining Nvidia's dominance in training and complete infrastructure plays and AMD's growing advantages in inference and agentic AI workloads.
Key Takeaways
- AMD holds two $100 billion GPU inference deals, targeting a market expected to eventually exceed the training segment.
- AMD acquired MEXT, whose AI-driven software moves cold data from DRAM to flash memory, reducing costs without impacting performance.
- The GPU-to-CPU ratio shifts from 8:1 for training to 1:1 for agentic AI, and AMD sees a $120 billion addressable market for its data center CPUs.
- Nvidia trades at under 16x forward P/E for fiscal 2028 estimates; AMD trades at 39.5x one-year forward P/E with a market cap under $900 billion.
- Nvidia's CUDA software platform, seeded into universities and research labs over years, remains its primary competitive moat for AI model training.
Why It Matters
The AI compute market is splitting into two segments — training, where Nvidia's CUDA ecosystem remains entrenched, and inference, where AMD's memory advantages and lower cost structure are gaining traction. AMD's MEXT acquisition and ZT Systems integration give it an end-to-end inference server play at a time when inference is projected to grow larger than training. The agentic AI shift, which demands a 1:1 GPU-to-CPU ratio versus 8:1 for training, plays directly to AMD's data center CPU leadership. Watch whether AMD's Helios rack-scale systems, expected in H2 2026, deliver the software maturity and multi-node scaling needed to convert design wins into sustained revenue.
Additional Context
The two $100 billion inference deals referenced in the article align with AMD's recent marquee agreements. In February 2026, Meta and AMD signed a multi-year partnership covering up to six gigawatts of AMD Instinct GPUs, with the first one-gigawatt deployment expected in H2 2026; The Wall Street Journal reported the deal is worth more than $100 billion and includes warrants for up to 10% of AMD shares (per AI Wire Media, February 2026). OpenAI signed a structurally similar 6GW agreement with AMD in October 2025, also reportedly including a 10% equity component. Both deals center on inference workloads rather than training, marking a strategic divergence from the Nvidia-centric training paradigm. AMD's MEXT acquisition, announced June 15, 2026, addresses a worsening constraint: DRAM now accounts for nearly 60% of server cost, up from roughly 50% in 2023 (per EE Times, June 2026). Gartner has forecast a 130% increase in combined DRAM and SSD prices by end of 2026, with memory prices rising nearly 4x since Q3 2025 (per Network World, June 2026). MEXT claims its AI-driven tiering can expand effective memory capacity 2-4x by predictively moving cold data to flash and restoring it before it is needed (per The Register, June 2026). Market share data underscores the scale of AMD's challenge. Nvidia holds approximately 80% of the AI accelerator market by revenue with $193.7 billion in FY2026 data center sales, versus AMD's estimated 5-7% share generating roughly $7-8 billion in Instinct revenue (per Silicon Analysts, April 2026). Custom silicon from Broadcom, which hit $20 billion-plus in AI ASIC revenue in FY2025, represents a larger and faster-growing competitive threat to Nvidia than AMD does. Meanwhile, AMD's Helios rack-scale architecture with MI450 GPUs is planned for H2 2026 (per AMD, March 2026), and SemiAnalysis InferenceX benchmarks show AMD's MI355X delivering comparable or better cost-per-token than Nvidia's GB300 NVL72 at high concurrency in FP8 configurations (per AMD technical article, March 2026).
Read full article at theglobeandmail.com
Get this in your inbox → Subscribe
Enjoy our coverage?
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source