What did The Information report about TSMC and Nvidia?
On 6 November 2026, The Information reported that TSMC had begun preparing plans to make 3D DRAM. The report said the work followed a request from Nvidia, which it described as TSMC's largest customer. It cited two people familiar with the initiative.
According to the report, TSMC was weighing two paths. It could build dedicated memory fabs, or it could buy an existing plant and retool it. The second option would mirror Micron's move in March. Micron agreed to buy a Powerchip Semiconductor Manufacturing Corp. facility in Taiwan for $1.8 billion and convert it for high-end memory.
Buying a plant would shorten the timeline. Building a new fab takes years, while converting an existing site can bring production forward. TechPowerUp noted that difference when it covered the report. The report did not identify a site, a budget or a customer commitment.
The Information said the effort was overseen by Hsu Kuo-chin, a TSMC vice president responsible for advanced packaging. He previously ran Micron's Taiwanese operations. Tom's Hardware and TechPowerUp both reported that detail. Neither TSMC nor Nvidia confirmed the plan when the story appeared.
The claim mattered because TSMC has built its business on logic chips. It makes chips for customers including Apple and Nvidia, but it does not sell DRAM. A move into memory would put it in competition with Samsung, SK hynix and Micron.
The report also landed during a global memory shortage. Prices have risen for data-center parts, PCs and phones, and buyers have been looking for new sources of supply.
What did TSMC say about the 3D DRAM report?
TSMC pushed back publicly. A company spokesperson told Tom's Hardware that TSMC focuses on its core business, which is dedicated IC foundry work.
"Please note that TSMC focuses on our core business at all times, which is dedicated IC foundry," the spokesperson said. "Our mission is to be the trusted technology and capacity provider of the global logic IC industry for years to come."
Tom's Hardware described that statement as essentially a denial. That is a fair reading. TSMC presented itself as a pure-play logic foundry that intends to stay one.
The wording is still notable for what it does not say. The statement does not mention 3D DRAM, Nvidia, Hsu Kuo-chin or the reported planning work. It restates the company's core mission rather than rebutting specific details.
The distinction matters. TSMC's statement rules out a change of corporate identity, not a research project. The report described planning work, not a signed order or a funded fab. Both claims can be true: a company can run early planning while insisting that its core business has not changed.
Tom's Hardware also noted that TSMC lacks the expertise and production capacity to make DRAM, let alone 3D DRAM. That is an assessment, not a company statement.
The reports did not include a public comment from Nvidia. TSMC also did not say whether any early-stage memory research exists inside the company.
Why would Nvidia ask its chipmaker to make memory?
AI accelerators are limited by memory as much as by compute. Each GPU needs high-bandwidth memory stacked beside it, and supply has been tight all year. Nvidia has been redesigning how that memory connects to its chips.
In August 2026, Nvidia announced NVHBM, a custom HBM design that moves the memory controller into the HBM base die. According to Nvidia, the design offers up to 30% more memory bandwidth than standard HBM4E. It also cuts HBM power by 15% and frees up to 25% more area on the XPU compute die.
That change makes the base die more valuable, and it pulls TSMC deeper into memory. SK hynix already uses TSMC's 12nm process for its HBM4 base die. TrendForce reports that TSMC's 3nm process is being prioritized for HBM4E and future custom HBM.
Micron has said its custom NVHBM and HBM4E products will use base dies from an external foundry. Samsung is weighing base dies from either its own foundry or TSMC, according to TrendForce. Both moves reduce the share of memory manufacturing that stays inside a single company.
Custom HBM is already a shared effort. The Elec reported that a custom I/O physical layer for NVHBM cuts the package area used by that interface by as much as 67%. ZDNet reported that Amazon's Annapurna Labs plans to use the design in Trainium 4.
If TSMC already makes the logic that controls memory, the reported request to make the memory itself is a logical next step. It would give Nvidia a supplier inside its own design and packaging chain. It would also reduce reliance on the three companies that control HBM supply.
Hsu's role reinforces that logic. Advanced packaging is where TSMC already joins logic and memory into one system. CoWoS is one of the technologies he oversees, according to TechPowerUp.
What is 3D DRAM and why is it hard to build?
Conventional DRAM stores bits in a single layer of cells. Manufacturers shrink those cells with each generation, but scaling is getting harder and more expensive. 3D DRAM takes a different route. It stacks multiple layers of memory cells vertically.
The idea resembles 3D NAND, which changed flash storage by stacking layers. For DRAM, vertical stacking could increase the number of bits produced per wafer. It requires new manufacturing steps, including wafer bonding, and tight control of heat and defects.
All three major memory makers are working on the technology. Micron has said it will keep scaling planar DRAM for several more nodes before 3D DRAM arrives. Its 1-delta node is due for volume production in the second half of 2027, with a bit-density gain in the mid-20% range. Micron places 3D DRAM between its 1-zeta and 1-eta nodes.
Micron's technology chief has explained why the industry is not rushing. Scott DeBoer, the company's president and chief technology and products officer, said 4F² is not on the roadmap right now after more than 20 years of evaluation, according to TrendForce. He argued that bit density depends primarily on capacitor pitch rather than cell architecture alone. He also said 4F² adds manufacturing complexity and cost. Micron will extend its 6F² design instead.
Further out, Micron's 1-epsilon node will use wafer bonding. Its later 1-zeta and 1-eta nodes may use High-NA EUV and new materials. Those steps show how much engineering remains before vertical DRAM becomes routine. 3D DRAM is not a shortcut around the shortage. It is a bet on the next decade of scaling.
Tom's Hardware noted that Micron's plan includes at least four more planar process technologies before 3D DRAM. That is one reason the outlet argued 3D DRAM cannot ease today's shortage for at least five years.
TSMC has world-class logic process technology, but DRAM fabs use different tools, materials and yield models. Making a few working cells in a lab is not the same as producing billions of good dies. The Investing.com report noted that TSMC holds patents around vertical DRAM. It also noted the execution risk.
How tight is the memory market?
The shortage is visible in prices and revenue. TrendForce estimates that DRAM industry revenue reached $154.73 billion in the second quarter of 2026. That was up 59.5% from the previous quarter.
Samsung held 39.4% of the market, with $60.98 billion in revenue. SK hynix held 24.9% and Micron 23.3%. The three companies account for the vast majority of DRAM supply.
Price growth is slowing but remains steep. TrendForce expects DRAM contract prices to rise about 20% quarter on quarter in the fourth quarter of 2026. That follows increases of roughly 80% in the second quarter and 40% in the third.
The pressure is not confined to contract prices. TrendForce expects the blended average selling price of HBM to rise 121% year on year in 2027, as HBM4 and HBM4E take a larger share of the mix. It also expects memory supply to stay tight through 2027.
China's CXMT has shown a 4F² design that could cut cell area by about 33% compared with 6F². TrendForce cautions that mass production is not expected before 2028, and that yields, leakage and data retention remain challenges. Apple and Tesla have raised the possibility of buying from Chinese suppliers, according to TrendForce, though volumes and pricing are uncertain.
The cost pressure is reaching device makers. The Investing.com report noted that Apple raised prices across its hardware lineup earlier this year after memory procurement costs rose. Nvidia and Apple have both looked for alternative supply, according to that report.
HBM is also consuming capacity that would otherwise go to standard DRAM. TrendForce notes that HBM and conventional DRAM compete for the same limited advanced-process and wafer capacity. That competition is one reason prices are rising across the memory market.
AI servers are the main driver. Memory makers are prioritizing high-bandwidth and high-capacity products, which leaves less capacity for PCs and phones. TrendForce, citing Commercial Times, reported that GPU makers are considering 8-Hi HBM instead of 12-Hi HBM on Rubin Ultra. The aim would be to ship more GPUs from the same DRAM die supply.
Under that scenario, HBM capacity per GPU would fall by about 33%. The same supply could support 50% more GPUs. The trade-off shows how far buyers will go to secure volume.
Why did TSMC's shares jump 7.7%?
Investors reacted quickly to the report. TSMC's US-listed ADRs rose 7.7% in early trading on Friday, 6 November, according to Investing.com. At the time of that report, the ADRs were up 7.23% and Nvidia's shares were up 3.91%.
The move reflected a simple calculation. A memory business would add a new revenue stream and deepen TSMC's role in AI hardware. It would also signal that the company sees the AI build-out lasting long enough to justify new capital spending.
TSMC's core business is already large. The company reported $46.72 billion in third-quarter 2026 revenue, a 66.4% gross margin and a 58.7% operating margin. It guided fourth-quarter revenue to $52.3-54.3 billion.
TSMC is not starting from zero. It agreed in 2024 to make base logic dies for SK hynix's HBM products. That work gives it a place in the memory supply chain without the risk of owning a DRAM fab.
Memory is also a capital-intensive business with sharp downturns. Samsung, SK hynix and Micron have all lived through price collapses. TSMC's investors are used to logic margins, not memory cycles. That helps explain why the company's public answer stressed its core business.
A memory unit would start small against those numbers. It would also put TSMC against competitors that have spent decades learning DRAM manufacturing. The share move was a bet on optionality, not on near-term earnings.
The denial complicates that bet. Investors who bought on the report now hold a stock whose new story has been disputed by the company. That does not make the report false, but it raises the bar for confirmation.
What to watch next
The first test is whether TSMC addresses memory at its next earnings call. Watch the capital-spending line, the advanced packaging roadmap and any expansion of Hsu's organization. Those are observable signals. A single report, even one from a well-sourced outlet, is not.
The second test is Nvidia's NVHBM rollout. The design is central to Nvidia's memory strategy, and its base dies depend on foundry partners including TSMC. If NVHBM expands, TSMC's memory role grows even without a DRAM fab.
The third test is the memory makers' roadmaps. Micron's 1-delta node is scheduled for the second half of 2027, and its 3D DRAM work sits further out. Samsung and SK hynix have their own 3D DRAM programs. Any acceleration would weaken the case for TSMC to enter.
Memory prices are the fourth test. TrendForce expects the pace of increase to slow in the fourth quarter, but not to reverse. Buyers should not expect 3D DRAM to bring relief this decade. The technology is still in development even at companies that have made DRAM for decades.
The next Nvidia platform is another checkpoint. NVHBM is designed for future accelerators, and Amazon's Annapurna Labs is a potential customer for Trainium 4. If more customers adopt custom HBM, TSMC's base-die business grows regardless of whether it ever makes DRAM.
For now, the confirmed facts are modest. A major outlet reported a plan, a key supplier denied it, and the market briefly priced it in. The gap between those three events is the story.