TSMC Is Weighing 3D DRAM Production at NVIDIA's Request

Taiwan Semiconductor Manufacturing is evaluating a push into advanced memory production, including vertically stacked 3D DRAM for NVIDIA, according to a report by The Information citing two sources. The company recently began planning production of the chips in response to a request from NVIDIA, its largest customer. If TSMC follows through, the world's dominant logic-chip maker would enter a market that Samsung, SK hynix and Micron have long controlled — and where an AI-driven shortage has pushed memory prices steadily higher.

The report lands at an awkward moment for the company's communications team. TSM shares trade at $385.30, down 18.4% over the past month but still up 27.77% year to date, according to market data cited by 24/7 Wall St. And on its October earnings call, TSMC pushed back on exactly this kind of story.

What does NVIDIA actually need?

NVIDIA's memory bill explains the request. Its supply obligations rose to $279.00 billion, largely tied to memory procurement for Vera Rubin production. The company also signed a multiyear technology partnership with SK hynix for next-generation memory. Data Center revenue reached $89.02 billion last quarter. NVIDIA CEO Jensen Huang described the demand background in direct terms:

"The AI infrastructure buildout is at full steam. Vera Rubin, now in full production, was built to power exactly this moment."

Conventional DRAM places cells in a single layer. 3D DRAM stacks them vertically to raise capacity. Samsung, SK hynix and Micron are all developing 3D DRAM themselves, and none of them can currently satisfy AI demand at the scale NVIDIA wants. That gap is what reportedly drew NVIDIA to its biggest manufacturing partner.

The stakes are simple. Memory has become a bottleneck for the entire AI hardware industry, and whoever controls 3D DRAM output controls the pace of future accelerator shipments.

Is TSMC denying it?

Sort of — and the denial is where the story gets interesting.

TSMC already packages AI accelerators with high-bandwidth memory through its CoWoS process, which gives it a natural fit for deeper memory work. On the October call, Chairman and CEO C.C. Wei said:

"CoWoS is still the number-one choice for our customers."

That call included no memory announcement. HBM surfaced only in a Citi analyst question about supply bottlenecks. Posts on X relaying TSMC's response said the company rejected reports that it is exploring 3D DRAM production at NVIDIA's request, and reiterated its commitment to logic chip manufacturing.

Still, the boundary between logic and memory is blurring. At TSMC's own symposium in April, SK hynix titled its presentation "Overcoming Technological Limits Requires the Integration of Memory and Logic." A public denial plus quiet planning is not unusual in this industry. The two-source report from The Information suggests something is moving behind the official line.

Why is memory so tempting?

Micron's margins reveal the prize.

Micron Technology (NASDAQ:MU), the only U.S.-listed pure-play memory maker, posted fiscal 2026 revenue of $133.188 billion. Fiscal Q4 non-GAAP gross margin hit 87.0%. Compare that to TSMC's 66.4% gross margin in Q3. Memory, at least right now, is more profitable per dollar of revenue than the most advanced logic manufacturing on Earth.

Micron is already shipping HBM4, with HBM4E volume production expected in calendar 2027. CEO Sanjay Mehrotra framed the company's position plainly:

"In the AI era, memory has become a strategic asset for our customers, and we are investing in our global manufacturing footprint to support their growing demand."

A serious TSMC entry would challenge that pricing power. The incumbents keep one edge, though: years of DRAM process know-how. Memory manufacturing is not a business a logic leader can simply switch on. TSMC would be starting from a deep deficit in production experience against rivals that have spent decades tuning DRAM fabs.

Does TSMC have the money?

Easily. The constraint is capacity, not cash.

TSMC posted Q3 revenue of $46.72 billion, with EPS of $4.60 beating the $4.46 consensus. Q4 guidance calls for $52.3 billion to $54.3 billion in revenue. 2026 capex is expected at $60 billion to $64 billion. Wei signaled the company's ambitions on the October call:

"We have no intention of leaving food on the table for others."

But Wei also said "Demand is still far above our capability to supply today," and management expects overseas-fab margin dilution to extend to 3%–4%. That is the real obstacle. Diverting fabs to memory would compete directly with the N2 ramp and A14, TSMC's next-generation process node scheduled for volume production in 2028.

Every wafer allocated to DRAM is a wafer not allocated to the leading-edge logic that NVIDIA, Apple and others are already fighting over. TSMC is supply-constrained at record margins. Adding a lower-experience product line at the margin is a genuine strategic question, not a formality.

What happens next?

Investors should look for a formal TSMC statement or a capex reallocation, per 24/7 Wall St.'s read on the situation. Until either arrives, 3D DRAM remains an option in planning rather than a committed program — and Micron and SK hynix keep their lead in the most profitable corner of the AI supply chain.

The underlying signal is larger than one product decision. NVIDIA asking its own foundry to build memory would mark a structural shift in how AI hardware gets made: the industry's biggest customer pressuring the industry's biggest manufacturer to break a decades-old division of labor. Whether TSMC says yes, and when, will shape memory pricing and accelerator supply well into 2028 and beyond.