AMD Taps GlobalFoundries for MI500 Photonics as It Advances CPO Design

Why It Matters

CPO could give AMD a performance edge in data‑center AI workloads while diversifying its manufacturing base, strengthening its competitive stance against rivals.

Key Takeaways

• AMD partners with GlobalFoundries for photonic ICs in MI500 CPO design
• ASE will handle multi‑chip‑module packaging, enabling tighter optical links
• CPO promises higher bandwidth and lower power for data‑center AI workloads
• AMD eyes Anthropic orders for MI450 accelerators amid tight AI chip supply
• Meta contract covers ~6 GW of Instinct accelerator capacity across generations

Pulse Analysis

Co‑packaged optics represents a paradigm shift in how AI accelerators move data. By integrating photonic interconnects directly onto a multi‑chip module, AMD can dramatically increase the data‑transfer rates between compute cores and memory while slashing the energy per bit. GlobalFoundries’ expertise in silicon photonics and ASE’s advanced packaging capabilities provide the necessary manufacturing depth, positioning AMD to deliver a product that rivals the bandwidth of traditional copper‑based solutions without the associated thermal penalties.

The AI market’s appetite for ever‑larger models has strained the supply of high‑performance chips, prompting customers to seek multiple sources. AMD’s pursuit of Anthropic as a new client illustrates this diversification trend, as startups look beyond the dominant Nvidia supply chain. Existing contracts with OpenAI and Meta, which together account for several gigawatts of compute, underscore AMD’s growing foothold in hyperscale data centers. The MI450 and upcoming MI500 CPO variants are designed to meet the latency‑sensitive demands of generative AI, offering a compelling alternative for firms balancing cost, power, and performance.

Strategically, AMD’s investment in CPO could reshape the competitive dynamics of the AI accelerator arena. Enhanced bandwidth and efficiency translate into higher throughput for training and inference, potentially narrowing the performance gap with Nvidia’s latest offerings. Moreover, the partnership with GlobalFoundries reduces reliance on a single foundry, mitigating geopolitical and capacity risks. As data‑center operators prioritize energy‑efficient scaling, AMD’s packaging innovations may become a decisive factor in winning future AI workloads, reinforcing its position as a serious challenger in the high‑performance computing market.