Situation assessment: Nvidia is taking control of the memory supply chain.
Nvidia has told its memory suppliers to use TSMC’s manufacturing process for next-generation memory chips, according to The Information. That makes the world’s most valuable AI chip company the one deciding how its memory gets built, and not only who builds it. The Information’s headline is short on details, but the direction is clear. Nvidia wants one foundry standard across its whole AI stack.
What stands out here is the leverage. Memory makers have long chosen their own processes. Now their biggest customer is telling them where to go.
🎯 What’s Happening
Modern AI accelerators run on high-bandwidth memory, or HBM. HBM is a stack of DRAM chips sitting on a “base die,” a logic chip at the bottom that handles communication between the memory and the GPU.
In earlier HBM generations, memory makers built that base die themselves on their own older processes. With HBM4 and the generations after it, that changes. The base die is turning into a real logic chip that needs advanced manufacturing, and the most advanced logic manufacturing in the world belongs to TSMC.
The Information’s report suggests Nvidia is turning that industry trend into a requirement.
📋 Key Points
- Who’s affected. Nvidia’s three HBM suppliers are SK Hynix, Samsung and Micron. SK Hynix already works with TSMC on HBM4 base dies. Micron has also moved toward TSMC for future base dies. Samsung is the awkward one, because it runs its own foundry and has been pitching its own process as the full package.
- Why TSMC. TSMC already makes Nvidia’s GPUs and handles the advanced packaging (CoWoS) that joins the GPU and the memory. Putting the base die in the same house means fewer integration problems and tighter performance tuning.
- Why now. Memory bandwidth is the real choke point for AI workloads. Big models are often limited more by how fast data moves than by raw compute. Whoever controls the memory interface controls a big share of system performance.
- Custom memory is coming. Logic base dies let chip designers build custom features right into the memory stack. Nvidia wants those designs made on a process it knows inside and out.
⚠️ Threat Assessment
For Samsung, this is a direct hit. Its pitch has been that one company can make the memory, the logic and the packaging. If Nvidia sends the base die to TSMC, that pitch loses a lot of its force with the industry’s largest buyer.
For TSMC, it’s another layer of dominance. The company already sits at the center of AI chip production. Now it’s moving into memory too.
For the wider market, concentration risk goes up again. More of the AI hardware stack now runs through one company, mostly in one place: Taiwan. Any disruption at TSMC, whether it’s capacity limits, geopolitics or a natural disaster, now hits GPUs, packaging and memory all at once.
🧭 What This Means for Practitioners
Most AI teams don’t buy HBM directly. But they feel its effects in GPU prices, availability and performance.
- Supply constraints may tighten. TSMC’s advanced capacity is already booked solid. Adding memory base dies to its workload could squeeze timelines for everyone.
- Performance gains should keep coming. Tighter integration between GPU and memory usually means better bandwidth per watt, which matters most for inference at scale.
- Watch the competitors. AMD, Google and the other custom chip builders also depend on HBM. If Nvidia sets the standard for base dies, rivals may have to follow it or fight for whatever supplier capacity is left.
- Samsung’s response matters. If Samsung goes along with it, TSMC’s position gets even stronger. If it pushes back, it risks losing share in the most lucrative memory market there is.
🔭 Outlook
This fits a pattern. Nvidia keeps moving past the GPU and shaping everything around it: networking, systems, software and now memory manufacturing. It doesn’t make memory, but it increasingly decides how memory gets made.
Expect memory makers to go quiet publicly while they negotiate behind the scenes. Expect TSMC to talk up its logic-for-memory capacity. And expect HBM to stop looking like a commodity part and start looking like a custom component built to Nvidia’s spec.
More details are available in the original report from The Information.