To The Moon

On November 8, The Information reported that $Taiwan Semiconductor Manufacturing(TSM)$ is drawing up plans to manufacture 3D DRAM at the request of its largest customer, $NVIDIA(NVDA)$ — and TSMC denied it the same day via Tom's Hardware, reiterating its focus on dedicated IC manufacturing. Why care about a denied rumor about tech that cannot ship for years? Because the request itself is the signal: the world's most valuable chipmaker asked its foundry to escape a memory supply chain it can no longer control. The pricing data below shows why.

Figure 1 — Spot and contract prices across the memory stack on report day. Sources: DRAMeXchange via DDRWatcher, Nov 9; TrendForce, Nov 2 2026.

Nvidia’s bottleneck is no longer silicon

For three years the AI build-out's constraint was compute — wafers, CoWoS packaging, GPUs. It has migrated. TrendForce's November 2 bulletin forecasts Q4 conventional DRAM contract prices up 10-15% QoQ, NAND up 15-20%. Server RDIMM contract prices have already risen, and suppliers are converting spot strength into structure: long-term agreements with floor prices, volume commitments, prepayments.

The spot market is two-speed. On November 9, DDR5 chip spot averaged $60.17 at DRAMeXchange, up 12.9% over twelve weeks, and GDDR6 held $12.50, up 6.8%. DDR4, the PC-grade commodity, averaged $46.87, down a fourth straight week and off 7.3% in twelve weeks; the DDR5-to-DDR4 ratio widened from 1.23x to 1.28x in a month.

That is not one memory market but a tight, AI-linked premium segment and a collapsing legacy segment sharing the same fabs. Memory index trackers note HBM consumes roughly three times the wafer area per shipped bit versus DDR5, so each HBM wafer removes several times more mainstream bits from supply. The shortage is a deliberate reallocation, not a demand blip — and a greenfield DRAM fab takes seven to nine quarters from tool-in to meaningful output.

Figure 2 — Twelve-week spot change by memory segment: DDR5 +12.9%, GDDR6 +6.8%, DDR4 -7.3%. Source: DRAMeXchange via DDRWatcher, Nov 9 2026.

What TSMC actually said, and what it did not

The Information cited two people familiar with the plans and said the work is overseen by the TSMC vice-president who runs advanced packaging — previously head of Micron's Taiwan operations, per the report: the person who would run exactly this program.

TSMC's denial, reported by Tom's Hardware, was precise rather than sweeping: a spokesperson said TSMC remains focused on dedicated IC manufacturing and its mission to be the trusted technology and capacity provider of the global logic IC industry. That denies near-term entry into DRAM fabrication. It does not deny that the request was made, that memory has become strategic, or that TSMC already packages memory into system-in-package products for clients. TSMC does not make DRAM, and its spokesperson did not pretend otherwise.

Why would $NVDA(NVIDIA)$ ask? Because HBM supply is concentrated in three companies — Samsung, SK hynix and $Micron Technology(MU)$ — and per TrendForce's November 4 bulletin, SK hynix has agreed next year's HBM4 volumes and prices, while Samsung calls its 2026 supply sold out. When your input is sold out years ahead and priced by three counterparties, you look for a fourth path. Nvidia is already opening its NVLink Fusion rack platform to third-party processors — Korea's Rebellions, under a memorandum with SK Telecom signed November 9 — the same instinct aimed at the scarcest input of all.

Figure 3 — TSMC quarterly revenue: $35.85B, $40.18B, $46.72B, $52.3-54.3B Q4 guide midpoint. Sources: TSMC Q3 FY26 results, Oct 15 2026; Zacks.

The business TSMC is protecting is worth the denial. Its Q3 FY26, reported October 15, produced US$46.72 billion of revenue, up 41.2% year on year and above the top of its $44.6-45.8 billion guidance range, with net income up 68.7% and gross margin at 66.4%. High-performance computing was 68% of revenue, up 22% QoQ; Q4 guidance of $52.3-54.3 billion implies another 14% sequential step. A foundry printing those numbers has little incentive to inherit DRAM's capital intensity and cycle risk — the denial is credible near term.

The five-year mismatch, and what actually fixes supply before then

Even for the incumbents, 3D DRAM is years from mass production, and Micron's public roadmap shows at least four more planar generations before its 3D DRAM arrives. A TSMC memory program, even started today, could not ease the shortage inside five years. The request is not a supply fix — it is insurance against a supplier oligopoly.

What fixes supply before 2027 is price. The September 30 DRAM contract print is 40 days stale per DDRWatcher, and the November reset converts tightness into revenue: at or above TrendForce's 10-15% forecast, the oligopoly's pricing power is confirmed; soft, while DDR4 keeps sliding, means the correction has reached the contract layer.

The backdrop is a violent repricing. Bank of America analysts led by Vivek Arya, cited by Wall Street CN on November 9, count the SOX down 38% from its June 22 peak of 14,634 to 9,011 on November 5 — 136 days versus a 253-day, 32% four-cycle average. The SOX trades at 13.7x forward earnings against 17.7x for the S&P 500, a 22% discount where the index once carried a 15% premium. BofA puts cloud capex at a ~$300 billion quarterly run-rate — about $1.2 trillion annualized — against a $1.555 trillion 2027 consensus.

Figure 5 — Forward P/E: SOX 13.7x vs S&P 500 17.7x, historical trough near 12x. Source: BofA Global Research (Vivek Arya) via Wall Street CN, Nov 9 2026.

Bull Case

1. If the overdue November DRAM contract reset prints at or above TrendForce's +10-15% forecast, the oligopoly's pricing power is confirmed into H1 2027.

2. SK hynix has agreed next year's HBM4 volumes and prices and Samsung calls 2026 supply sold out (TrendForce, Nov 4); if LTAs with floor prices roll forward, $MU(Micron)$'s ASP mix improves in its next reported quarter.

3. TSMC's Q4 guidance of $52.3-54.3 billion implies +14% QoQ; a print above the top of that range contradicts the "AI capex cracking" narrative.

4. Spot leadership in DDR5 (+12.9%) and GDDR6 (+6.8%) shows AI-linked demand still pays up while legacy DDR4 corrects — a mix win for HBM owners.

5. If TSMC discloses memory development spend in a future capex breakdown, the second-growth-engine scenario in Figure 6 activates — falsifiable on the January call.

Bear Case

1. The SOX drawdown is already 38%, deeper than the 32% average of the last four downcycles; if EPS cuts surface from Q1 2027 as BofA projects, a march toward the historical 12x trough is the base case, not the tail.

Figure 4 — SOX -38% peak-to-trough vs a -32% average over the last four downcycles. Source: BofA Global Research (Vivek Arya) via Wall Street CN, Nov 9 2026.

2. DDR4 spot has fallen four straight weeks and DDRWatcher reads the regime as "Correction"; if the overdue contract reset comes in below TrendForce's forecast, the tightness thesis breaks at its source.

3. Cloud capex consensus of $1.555 trillion for 2027 versus a ~$1.2 trillion run-rate requires ~30% growth; if hyperscalers flatten, every memory price assumption resets downward.

4. TSMC guided Q4 gross margin to 65-67%, down from 67.7% in Q2 on the N2 ramp; a DRAM entry would dilute a margin structure it is already defending.

5. The denial may be final: if the January call repeats the logic-only mission with no memory capex line, the rumor is dead and the story reverts to a three-player oligopoly.

Strategic Slam

$TSM(TSMC)$ — the winner either way: the only company both customers and governments ask to de-risk the supply chain, and it declines on its own terms. Stay constructive while Q4 guidance holds and gross margin stays at or above 65%; a memory line in the 2027 capex breakdown is the re-rating trigger.

$NVDA(NVIDIA)$ — the request is an admission of supplier risk. Watch whether it keeps widening NVLink Fusion to third-party compute the way it just did with Rebellions and SK Telecom; each additional supplier cuts its exposure to the memory trio's terms.

$MU(Micron)$ — the pricing window is the gift, the cycle is the tax. If the November reset confirms +10-15%, the HBM-led mix story has two more strong quarters; if contract pricing follows DDR4 spot down, Micron is the most geared to the downside.

Catalyst Calendar

• Nov 9, 2026 — DRAM spot prints (DRAMeXchange via DDRWatcher): regime reads "Correction," contract data 40 days stale.

• November 2026 (expected) — next monthly DRAM contract reset, the first since Sep 30; date unconfirmed.

• Mid-January 2027 (expected) — TSMC Q4 FY26 results; the $52.3-54.3 billion guide and any memory commentary are the tell.

• Q1 2027 — window in which BofA expects the first visible EPS cuts across semis.

Figure 7 — Catalyst timeline from the stale Sep 30 contract print to the mid-January call. Sources: TrendForce Nov 2; DDRWatcher Nov 9; The Information via Tom's Hardware, Nov 8 2026.

Figure 6 — Scenario matrix: memory cycle direction vs TSMC's revealed choice. Sources: TrendForce Nov 2026; DDRWatcher Nov 9 2026; The Information, Nov 8 2026.

Verdict

The denial changes nothing and proves everything. Memory is the choke point of the AI build-out, pricing power sits with the three DRAM makers until at least 2027, and the strongest foundry on earth is being petitioned to build a fifth path it sensibly refuses. Own the oligopoly for two quarters, keep $TSM(TSMC)$ as the core compounder at a 66% gross margin, and treat any confirmed TSMC memory program as a 2028-and-later story — not a reason to buy today.

Key Takeaways

1. The Information reported Nov 8 that TSMC is planning 3D DRAM production at Nvidia's request; TSMC denied it the same day via Tom's Hardware.

2. TrendForce forecasts Q4 DRAM contracts up 10-15% QoQ, NAND up 15-20%; the last contract print is 40 days old.

3. Spot is split: DDR5 +12.9% and GDDR6 +6.8% over twelve weeks; DDR4 falls 7.3%, four straight weeks down.

4. HBM's three-fold wafer intensity and seven-to-nine-quarter fab lead times make the shortage structural.

5. Even a started-today TSMC memory program could not add supply for five years — the request is insurance, not a fix.

If the most advanced foundry on earth cannot buy its own customer a memory escape hatch, what exactly is Nvidia's plan B when the three DRAM makers set the price?

Tickers: $TSM(TSMC)$ $NVDA(NVIDIA)$ $MU(Micron)$

Sources: The Information via Tom's Hardware (Nov 8, 2026), incl. TSMC spokesperson statement; TrendForce bulletins (Nov 2, Nov 4, 2026); DRAMeXchange via DDRWatcher (Nov 9, 2026); Memory Index HBM wafer-intensity and fab lead-time methodology (Nov 9, 2026); BofA Global Research (Vivek Arya) via Wall Street CN (Nov 9, 2026); TSMC Q3 FY26 results and Q4 guidance (Oct 15, 2026) via Zacks; NVIDIA-Rebellions-SK Telecom memorandum (Nov 8-9, 2026) via NVIDIA Newsroom and Yonhap.

Disclaimer: For information only, not investment advice. Figures are cited from the named public sources above; forward-looking statements are uncertain. The author holds no position in the named securities.