NVIDIA CEO Jensen Huang has sent a clear message: the primary bottleneck in the AI revolution has shifted from GPUs to memory and advanced packaging.
This means the biggest challenge now isn't just producing the AI 'brain' (the GPU), but also the components that feed it data at lightning speed. We're talking about High-Bandwidth Memory (HBM), which acts like a super-fast short-term memory for the GPU, and advanced packaging like TSMC's CoWoS, the technology that skillfully bundles the GPU and HBM chips together. These two areas are now facing a severe supply crunch.
So, how did we get here? First, the demand for AI is simply voracious. Tech giants are building massive AI models that require enormous amounts of memory, and suppliers can't keep up. Major memory makers like Samsung and SK hynix have already sold out their entire 2026 HBM capacity. Second, the advanced packaging needed to connect these components is also in short supply, with TSMC's capacity being a well-known chokepoint. Third, geopolitics casts a long shadow. Huang’s comment that China's access to advanced chips is "up to China" underscores how U.S.-China trade policies can suddenly shift where this scarce supply goes, creating more uncertainty for everyone.
The effects are already rippling through the market. Companies like Dell, which build AI servers, are in a scramble to secure long-term supply contracts and are passing the rising costs on to their customers. This isn't just talk; the soaring revenues of memory producers like Micron confirm that this scarcity is real and profitable for those who control the supply.
In short, Huang's warning crystallizes the new reality of the AI industry. The pace of the AI buildout for the next decade will likely be dictated not by the availability of GPUs, but by the supply of HBM and advanced packaging. This makes the entire ecosystem more fragile and highly sensitive to both production capacity and political decisions.
- HBM (High-Bandwidth Memory): A type of high-performance computer memory used in conjunction with high-performance graphics accelerators and network devices.
- CoWoS (Chip-on-Wafer-on-Substrate): An advanced chip packaging technology that allows multiple chips to be placed side-by-side on a silicon interposer, enabling high-speed interconnection.
- OEM (Original Equipment Manufacturer): A company that builds products, like servers or computers, using components from other manufacturers.