Two developments point to the same conclusion: the bottleneck in AI chips is expanding from compute capacity into memory, and TSMC is moving closer to becoming a memory supplier.
What this is
At the 2025 VLSI Symposium, TSMC presented a 1T1C memory cell based on oxide semiconductor: each cell consists of one transistor and one capacitor, with the memory array embedded directly in the chip’s back-end interconnect layer. The test chip operates at 0.75 volts, retains data for more than 128 milliseconds, achieves random read/write latency below 8 nanoseconds, and supports endurance of more than 10¹⁴ cycles.
Its intended use is processor cache: compared with SRAM—the small, high-speed cache inside processors—it offers higher density while targeting lower power consumption and latency; compared with standalone DRAM, which is typically fabricated as a separate chip, it can shorten data-transfer distances.
The second development is more direct: according to The Information, TSMC is evaluating the production of 3D DRAM in response to Nvidia’s request. The technology increases density by vertically stacking memory cells. There is currently no mass-production timetable, and TSMC emphasizes that logic-chip foundry services remain its core business.
Industry view
What matters is that TSMC is not merely adding a product line; it is addressing one of the tightest links in the AI chip supply chain. Nvidia and other companies continue to expand their compute investments, but they also need memory for growing training and inference workloads. If storage supply cannot keep pace, additional processors will struggle to translate into usable compute capacity.
The risks are equally clear. Oxide transistors, embedded memory, and 3D DRAM all introduce new manufacturing-yield, reliability, and cost challenges. TSMC has no large-scale DRAM business today, so entering the market means competing with the established supply capabilities of Samsung, SK Hynix, and Micron. Without a clear mass-production schedule and customer commitments, this looks more like a supply-chain contingency than a competitive business already in place.
Impact on regular people
For enterprise IT: When purchasing AI servers in the future, memory capacity, bandwidth, and energy consumption may matter as much as GPU count. Companies must evaluate compute performance not only by graphics-card model, but also by the efficiency of the entire data-transfer system.
For individual careers: The experience of running models locally and using AI tools at work may improve through faster response times and lower power consumption, but meaningful cost reductions will still depend on advanced packaging and memory yields.
For the consumer market: Continued hardware competition among AI phones, computers, and cloud services will keep pushing configurations and prices higher. If memory supply increases, however, shortages and price pressures for those devices may ease over time.