Nvidia shipping a flagship GPU with memory bandwidth below its own announced figures reflects direct product impact from the HBM supply crunch. Supply constraints projected to persist through 2027 mean the pressure on AI accelerator specs and HBM allocation is not a near-term blip.
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
Nvidia's shipping Rubin GPU delivers HBM4 bandwidth of 19.2 TB/s and NVLink bandwidth of 3.0 TB/s, reduced from previously announced preliminary figures of 22 TB/s and 3.6 TB/s respectively. Compute performance at 50 PFLOPS of NVFP4 inference is unchanged. Nvidia has not explained the change, noting the earlier figures were labeled preliminary. AMD's MI455X lists 23.3 TB/s bandwidth and 432 GB of HBM4 against Rubin's 288 GB.
The spec reduction sits within a broader pattern of HBM supply pressure. TrendForce reports Nvidia began evaluating lower HBM configurations for its next-generation Rubin Ultra platform in early Q3 2026, weighing 8-Hi HBM4e, 12-Hi HBM4, and 8-Hi HBM4 alternatives beyond the original 12-Hi HBM4e design. TrendForce attributes this to a projected overall DRAM shortage in 2027 that will limit wafer capacity memory suppliers can allocate to HBM, and to uncertainty around HBM4e validation schedules and production ramp. TrendForce states Nvidia's primary objective for Rubin Ultra is to increase I/O speed, while expanding GPU shipment volume is a secondary priority. Several cloud service providers are similarly evaluating lower HBM capacities for their next-generation in-house AI ASICs, per TrendForce.
Micron's full 2026 HBM4 production is sold out, with customers receiving only 60-70% of ordered volumes. Elevated HBM average selling prices benefit SK Hynix and Samsung but represent a margin headwind for Nvidia and hyperscale customers.
Broader DRAM contract prices rose roughly 80% in Q2 2026 and approximately 40% in Q3, with Q4 projected to add around 20% more. Bernstein argues the apparent deceleration is misleading: with roughly 40% of supply locked into long-term contracts, a 30% spot move appears as only about 18% in reported aggregate prices. Bernstein also argues a 15-20% quarterly price jump would still rank in the top 5-10% of the past 15 years. Customers are paying large upfront deposits to lock in supply, with Sandisk holding $16.5 billion and Micron approximately $32 billion against contracted revenue. New mature DRAM supply remains limited, underpinning elevated prices.
How it developed1 November 2026added 2 Nov
Nvidia's shipping Rubin GPU, confirmed November 1, carries HBM4 bandwidth of 19.2 TB/s and NVLink at 3.0 TB/s, both down from preliminary figures of 22 TB/s and 3.6 TB/s, while compute holds at 50 PFLOPS NVFP4.
AMD's MI455X lists 23.3 TB/s and 432 GB against Rubin's 288 GB. Bloomberg reported Sam Altman visited roughly 50 Samsung executives in South Korea, with the talks building on an October 2025 deal targeting 900,000 DRAM wafers per month from Samsung and SK Hynix for OpenAI's Stargate project.
Samsung EVP Jaejune Kim said during the company's Q2 2026 earnings call that DRAM and NAND shortages will worsen in 2027 before easing, remaining tight through 2028.
Samsung disclosed it has locked 60 to 70% of its memory capacity into multi-year AI and data center contracts, constraining supply for consumer and GPU markets. The chip division posted operating profit roughly 19 times higher year over year, with memory accounting for nearly all of that gain.
SK Hynix reported quarterly revenue of 104.6 trillion won at an 80% operating margin, driven by AI memory chip demand, placing a second major manufacturer's earnings alongside Samsung's Q2 2026 disclosures that it has locked 60 to 70% of capacity into multi-year AI and data center contracts and expects tightness to persist through 2028.
First in the edition · 28 October 2026
26 October 2026
DRAM contract prices reported up roughly 80% in Q2 and 40% in Q3; Bernstein argues the apparent deceleration is misleading due to long-term contract effects.
24 October 2026
Micron publishes technology roadmap through 2030, including vertical HBM by 2028 and AI-class NAND flash
22 October 2026
Report notes switching to 8-layer HBM cuts per-GPU capacity ~33% but allows same supply to cover ~50% more GPUs; TrendForce expects 8-layer to cost 10-20% more per gigabit in 2027
1 October 2026added 27 Oct
Micron CEO Sanjay Mehrotra told investors demand will exceed memory supply through at least 2028, with 2027 prices already set above 2026 levels.
Silicon Analysts reported Nvidia is testing at least three downgraded HBM configurations for Rubin Ultra, from 16-stack HBM4E to 12-stack and 8-stack variants. Micron entered high-volume production of a 36 GB 12-layer HBM4 stack, standard HBM4, not the customized HBM4E the original design required. Lam Research raised its advanced packaging outlook from greater than 40% to greater than 70%.
Micron reported to have entered high-volume production of 36 GB 12-layer HBM4 stack for Vera Rubin at speeds above 11 Gb/s per pin and over 2.8 TB/s bandwidth
Studio Global AI corroborates Nvidia's four-configuration HBM evaluation for Rubin Ultra amid HBM4 shortages expected through 2027
6 August 2026added 31 Oct
TrendForce reported Nvidia began reviewing lower-spec HBM for its Rubin Ultra in early Q3 2026, after 12-layer HBM4e served as the design baseline through H1, citing supply constraints and uncertainty over whether that configuration can reach 14-16 Gbps or will be capped at 11-12 Gbps.
Nvidia also halved SOCAMM memory on its Vera Rubin Superchip after TrendForce projected LPDDR5X constraints through 2027, while Silicon Analysts reported Micron's entire 2026 HBM4 output is sold out with customers receiving 60-70% of ordered volumes.
TrendForce reports Nvidia began evaluating lower HBM configurations for Rubin Ultra in early Q3 2026, citing projected 2027 DRAM shortage and HBM4e validation uncertainty.