Nvidia, Alphabet, AMD Secure 85% of Next Year's HBM Supply
Three tech giants have preemptively secured 85% of 2027 high-bandwidth memory production, creating a structural supply constraint that favors established players over new entrants in the AI infrastructure market.
Nvidia, Alphabet, and Advanced Micro Devices have collectively locked up 85% of the projected high-bandwidth memory production for 2027, according to estimates cited by GN auto stocks/technology. This pre-emptive procurement creates a significant barrier to entry for competitors, as the remaining available HBM capacity is insufficient to support the rapid expansion of new AI accelerator deployments. The concentration of supply among these three firms reflects a strategic shift from competing on chip design alone to securing physical manufacturing access.
The scarcity is driven by physical constraints in the semiconductor supply chain. Manufacturing HBM requires three times the wafer capacity of standard DRAM, and the necessary extreme ultraviolet lithography machines are in limited supply globally. Additionally, new cleanroom facilities require a construction lead time of approximately two years. These factors mean that memory manufacturers cannot rapidly scale output to meet the surging demand driven by the transition from model training to inference workloads.
Nvidia Dominates Secure Capacity
Nvidia holds the largest share of secured HBM supply, capturing an estimated 37% of 2027 capacity. The company reinforced this position by signing a $500 billion multiyear partnership with SK Hynix in June. This agreement ensures a steady flow of memory components for Nvidia’s GPUs and CPUs, which are increasingly dependent on high bandwidth to manage data throughput. Securing CoWoS advanced packaging capacity alongside HBM further entrenches Nvidia’s advantage by controlling the final assembly stage of its accelerators.
Alphabet Leverages TPU Ecosystem
Alphabet has secured 36% of the 2027 HBM supply to support its Tensor Processing Unit roadmap. The company will distribute this memory between custom chips co-designed with Broadcom and MediaTek. In July, Broadcom finalized a $200 billion long-term supply deal with Samsung through 2030, explicitly naming Anthropic as its largest customer for these chips. This structure allows Alphabet to monetize its TPU business through high-margin sales to external partners while reducing internal cloud computing costs, achieving a one-year payback period on infrastructure investment.
AMD Faces Tighter Constraints
AMD has locked up approximately 12% of next year’s HBM production, the third-largest share among the major buyers. The company signed a long-term memory agreement with Samsung in March to secure future supply. However, AMD’s new MI455X GPU requires 50% more HBM than its previous-generation MI355X and Nvidia’s Rubin GPU. Management has indicated that the company is currently feeling the pressure of the supply crunch, as demand for its Helios server platform has outpaced the available memory allocation, potentially limiting near-term revenue growth compared to its competitors.
