TSMC Weighs 3D DRAM for Nvidia, Then Insists It Will Stay a Logic Foundry

Taiwan Semiconductor Manufacturing Co has long described itself in narrow terms: a pure-play logic foundry that turns other companies' chip designs into working silicon and leaves the branding, the selling and the customer relationships to them. The model rests on a simple bargain. Customers bring designs, TSMC brings process technology and capacity, and no one competes with anyone else's product. It is the reason Nvidia, TSMC's largest customer, can hand its most important blueprints to a company it does not own.

That bargain came under scrutiny on November 6, 2026. The Information reported on November 6 that TSMC had begun charting plans to make 3D DRAM, an advanced type of memory, at the request of Nvidia. Reuters reported on November 6 that the chipmaker was considering expanding into advanced memory chip production, relaying the account and citing two people with direct knowledge of the matter. DRAM stores data temporarily in devices, and 3D DRAM stacks memory cells vertically rather than in a single horizontal layer, which raises capacity without demanding more floor space.

The timing is not accidental. A severe, AI-driven shortage of memory chips has pushed prices higher and sent chip designers scrambling to look for additional supply. TrendForce said AI demand and extra orders from U.S. cloud providers have left memory makers with extremely low inventory and strong bargaining power. Memory has moved from background component to strategic lever.

A day later, TSMC tried to close the story down. Tom's Hardware reported on November 7 that the company had pushed back on the report, saying it is a pure-play logic foundry and plans to remain one. The sequence, a reported expansion followed by a public denial, is the heart of the matter. In the chip industry, what a company explores and what it says are two different things.

Key Facts

The original report was specific about who is behind the effort. According to The Information, the initiative is helmed by Hsu Kuo-chin, a TSMC vice president who oversees advanced packaging and who previously ran Micron's Taiwanese operations. The report described two routes: building new factories for advanced memory chips, or acquiring an existing plant and converting it, an approach similar to Micron's move this year to buy a fab and convert it for DRAM output. No production start date, fab site, capacity figure or investment total was disclosed. TSMC and Nvidia did not immediately respond to Reuters' requests for comment.

Investors reacted before the company did. TSMC's US-listed ADRs rose more than 6% in early trading on November 6, a move that suggests the market saw a memory entry as a growth story rather than a distraction. The Next Web reported on November 6 that the move would give the world's largest chipmaker an even bigger share of the AI hardware supply chain, putting it up against Samsung Electronics, SK Hynix and Micron. Bloomberg reported on November 6, citing The Information, that the planning was already under way.

TSMC's response was blunt. "Please note that TSMC focuses on our core business at all times, which is dedicated IC foundry," a company spokesperson told Tom's Hardware. "Our mission is to be the trusted technology and capacity provider of the global logic IC industry for years to come." That statement does not deny that planning work exists. It denies that the company's identity has changed.

The competitive field TSMC would enter is unusually concentrated and unusually flush. SK Hynix plans to spend $720 billion on new plants. Micron expects the memory shortage to last into 2028. Both have signed long-term supply agreements with major customers: Samsung has deals with 10 major customers and is negotiating with another 10, while SK Hynix has signed agreements with 14 customers and posted a record 83 trillion won operating profit. Memory is no longer a sleepy cyclical backwater; it is where the AI boom's scarcest input is made.

The price data explains the urgency. Capital Times reported on November 3 that TrendForce expects conventional DRAM contract prices to rise 10% to 15% in the fourth quarter from the third, and NAND flash contract prices to rise 15% to 20%. On November 2, the DDR4 16Gb spot price averaged $82.719, while the DDR5 16Gb chip was unchanged at $61.50. TrendForce attributed the squeeze to AI demand and extra orders from U.S. cloud providers, which have left memory makers with extremely low inventory and strong bargaining power.

Analysis

The contradiction between the report and the denial is less strange than it looks. Large manufacturers routinely study markets they have no intention of entering, and they routinely deny intentions until a decision is final. TSMC's public position is not really about 3D DRAM. It is about the neutrality that makes the foundry model work. The moment TSMC sells memory under its own name, it stops being a neutral partner for chip designers and becomes a competitor to the memory makers whose HBM base dies it also helps build.

What this really means is that TSMC is hedging against a world in which memory, not logic, becomes the binding constraint on AI systems. The technical barriers are real. Building advanced DRAM requires memory intellectual property, process recipes and yields that TSMC does not have today, and the report's option of buying a small specialty memory producer would not solve the hardest part: small makers do not have leading-edge commodity DRAM nodes. Tom's Hardware noted that TSMC would still need to invest billions of dollars and years of research and development even after an acquisition.

There is also a cycle to worry about. Memory shortages reliably trigger capacity expansions, and capacity expansions reliably produce oversupply. A foundry that enters DRAM at the top of a boom risks arriving just as prices roll over. TSMC's existing ties complicate matters further: SK Hynix partnered with the foundry in 2024 to manufacture logic base dies for next-generation HBM, and competing directly with a partner is a different calculation from competing with a stranger.

The bigger picture here is that the wall between logic and memory is thinning. In October, Nvidia reportedly asked Samsung, SK Hynix and Micron to build the base layer of its custom HBM memory on TSMC's 3nm process. That request already pulled TSMC into the memory stack, not as a DRAM maker but as the foundry for the logic portion of high-bandwidth memory. Moving from base dies to the memory array itself is a much larger step, but it is a step in the same direction. My judgement is that the planning is real, the denial is about positioning rather than facts, and any actual production is years away.

Why It Matters

For the AI hardware market, the question is who captures the value when memory is scarce. TrendForce's forecast of a 10% to 15% quarterly rise in DRAM contract prices is not a rounding error for server builders. Samsung executive vice president Kim Jae-june put the third-quarter rise in DRAM average selling prices at around 20% from the second quarter and NAND at a high-teens percentage, while DRAM shipments grew only by a low- to mid-single-digit percentage. Suppliers are earning more by shipping barely more, which is the definition of pricing power.

The costs are already reaching consumers and adjacent industries. Apple raised Mac and iPad prices in June citing memory costs. Elon Musk is planning a $55 billion chip plant to cut his reliance on Micron. Micron's expectation that the shortage lasts into 2028 tells buyers to plan for years, not quarters. If TSMC does enter memory, it would add capacity and competition at the exact point where the market is most brittle, though only after a long and expensive build.

For Taiwan, the stakes are strategic as well as commercial. Taiwan's manufacturing PMI fell 0.9 point to 62.5 in October, still a strong reading but a reminder that the island's industrial cycle is sensitive to semiconductor demand. Global chip sales reached $159.7 billion in August, up 144.3% from a year earlier, according to the Semiconductor Industry Association, a figure that captures how violently the AI buildout has reshaped the industry. A TSMC memory business would deepen Taiwan's centrality in that boom, and also concentrate more of it in one company.

Next Up

The next signals will be mundane and concrete: capital expenditure guidance on TSMC's next earnings call, hiring inside Hsu Kuo-chin's advanced packaging organisation, any small acquisition of a specialty memory producer, and language in exchange filings that goes beyond the current blanket denial. Watch also whether Nvidia says anything publicly, since the request at the centre of the report came from the customer that TSMC can least afford to alienate.

Tom's Hardware's assessment that advanced memory technology remains years from market viability is the right frame for expectations. The story is not that TSMC will ship 3D DRAM next quarter. It is that the world's most important logic foundry is at least studying whether the pure-play model still fits a market where memory has become the scarce good.