Updated todayDay 16in the edition since 24 Oct 202622 sources
Nvidia Rubin memory cuts and HBM shortage
The gist
Latest: TSMC Denies Nvidia-Pushed 3D DRAM Plans, Stays Logic Foundry
The HBM shortage is directly shaping the specs of Nvidia's most advanced chips and constraining supply to buyers at every scale, from Apple's gross margin to Chinese cloud firms receiving a fraction of their GPU orders. TSMC's denial of the 3D DRAM report removes a near-term relief valve, and Micron's construction freeze signals the shortage may extend well past 2027.
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
Check your email.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.
The full picture
TSMC denied a report that Nvidia asked it to draft plans to manufacture 3D DRAM, reaffirming it will remain a pure logic foundry. TSMC already earns memory-related revenue by fabricating HBM base dies and using SoIC hybrid bonding, and noted that even Micron plans several more planar DRAM nodes before adopting 3D DRAM, meaning any entry into that market could not ease shortages for at least five years. Separately, Nvidia reportedly wants Samsung, SK Hynix, and Micron to build its custom NVHBM base dies on TSMC's 3nm process, an arrangement that would generate memory revenue for TSMC without direct DRAM exposure.
The shipping Rubin GPU delivers HBM4 bandwidth of 19.2 TB/s and NVLink bandwidth of 3.0 TB/s, down from preliminary figures of 22 TB/s and 3.6 TB/s; compute remains at 50 PFLOPS NVFP4. Nvidia has not explained the reduction, noting the earlier figures were labeled preliminary. AMD's MI455X lists 23.3 TB/s and 432 GB of HBM4 against Rubin's 288 GB.
For the follow-on Rubin Ultra, TrendForce reports Nvidia originally designed around 12-Hi HBM4e but began evaluating lower-spec alternatives in early Q3 2026, driven by a projected 2027 DRAM shortage and uncertainty around HBM4e validation timelines. The four configurations under consideration are 8-Hi HBM4e, 12-Hi HBM4e, 12-Hi HBM4, and 8-Hi HBM4. SemiAnalysis reports the mainstream Rubin Ultra SKU will use 8-Hi HBM4 at 192 GB, down from Rubin's 288 GB, while peak theoretical FLOPs remain unchanged. TrendForce notes the final configuration depends partly on wafer allocation decisions by memory suppliers. Nvidia has not publicly confirmed either finding.
Supply constraints span the full AI hardware stack. Micron's full 2026 HBM4 output is sold out, with customers receiving only 60-70% of ordered volumes. Micron paused construction of its $3-4.5 billion Fab 11E facility in Taoyuan after the Taoyuan union demanded 15% of operating profit in cash; the plant was expected to come online in 2028, and Taiwan accounts for roughly 60% of Micron's DRAM production. TSMC's CoWoS capacity is fully allocated through at least mid-2027. Top-end GPU lead times stand at 30-45 weeks; hard drive lead times are at 50 weeks; ABF substrate lead times are 48-56 weeks. Rubin's production ramp is bottlenecked by substrates and capacitors running at more than three times their balanced lead times. Hyperscalers including Microsoft, Amazon, Google, and Meta are competing for the same constrained components, leaving smaller players at a disadvantage.
How it developed7 November 2026added 8 Nov
TSMC denied it was drafting 3D DRAM plans at Nvidia's request, closing off a reported bid to add TSMC as a fourth memory supplier.
It said DRAM entry could not ease shortages for at least five years. Nvidia is reportedly asking Samsung, SK Hynix, and Micron to build its NVHBM base dies on TSMC's 3nm process instead, giving TSMC memory revenue without direct DRAM exposure. Chinese AI firms ordered about 6 million RTX Pro 5500 GPUs totaling $70 billion; Nvidia told partners it can produce at most 3 million cards in 2027.
Micron paused construction of Fab 11E, a $3-4.5 billion HBM and DRAM facility in Taoyuan, Taiwan, after a union demanded 15% of operating profit in cash, removing a planned source of 2028 capacity.
TrendForce reported top-end GPU lead times jumped to 45 weeks as Nvidia's Rubin transition begins, bottlenecked by substrates and capacitors running at more than three times balanced lead times. Hard drive lead times reached 50 weeks, ABF substrate lead times 48-56 weeks, and server CPU shortages are also emerging.
Nvidia asked TSMC to manufacture 3D DRAM chips to add a fourth memory supplier beyond Samsung, SK Hynix, and Micron; TSMC is drafting plans but set no timeline.
TrendForce reports Nvidia halved SOCAMM capacity in Vera Rubin Superchip modules after finding LPDDR5X supply will remain constrained through 2027, extending spec cuts from HBM into CPU-side memory. South Korea's SSD exports reached $5.64 billion in October, up 584% year-over-year, as Samsung and SK Hynix attribute the surge to agentic AI expanding KV cache demand.
Nvidia's shipping Rubin GPU delivers HBM4 bandwidth of 19.2 TB/s and NVLink of 3.0 TB/s, below the previously announced 22 TB/s and 3.6 TB/s, Nvidia noting earlier figures were labeled preliminary.
Micron's 2026 HBM4 output is sold out at 60-70% of ordered volumes, Samsung is seeking $4 per gigabit for 2027 HBM4, triple its HBM3E price, and TSMC is reportedly raising CoWoS prices up to 20% with capacity booked through mid-2027. Apple CEO John Ternus called memory pricing 'unprecedented' and forecast holiday gross margin at roughly 45%, down from 48%.