
Nvidia's Vera Rubin Holds HBM at 288 GB, Upending Memory Demand Math
Vera Rubin holds HBM at 288 GB per GPU while rack bandwidth jumps to 1,400 TB/s, shifting HBM growth for Micron, Samsung and SK hynix from capacity to shipments and pricing.
- By
- Grace Kim
- Filed
- Channel
- Memory & Storage
- Read
- 4 min read
Vera Rubin keeps high-bandwidth memory capacity flat at approximately 288 GB per GPU while more than doubling memory bandwidth compared with Blackwell Ultra. The change breaks the pattern that ran from Hopper through Blackwell Ultra, when each Nvidia generation carried sharply more memory. For Micron, Samsung and SK hynix, the three principal HBM suppliers, revenue growth now depends more on system shipments, memory technology and manufacturing economics than on gigabytes per accelerator.
What changed in the GPU cadence?
Nvidia's roadmap shows how fast memory requirements once grew. HBM per GPU rose from 80 GB in the H100 (2022) to 288 GB in the Blackwell Ultra B300 (2025), a 3.6-fold increase, according to Nvidia product specifications and The Information Network analysis. The B200-to-B300 step alone raised capacity 50%, from 192 GB to 288 GB.
Vera Rubin, introduced in 2026, stops that progression. It retains 288 GB per GPU while moving from HBM3e to HBM4.
- A100 (2020): 80 GB, HBM2e
- H100 (2022): 80 GB, HBM3
- B200 Blackwell (2024): 192 GB, HBM3e
- B300 Blackwell Ultra (2025): 288 GB, HBM3e
- Vera Rubin (2026): 288 GB, HBM4
- Rubin Ultra (2027, projected): ~1 TB, HBM4e
HBM4 raises value through a wider interface and more demanding manufacturing, not through more installed bits. This creates a gap between HBM bit growth and HBM revenue growth, since revenue can rise on pricing and richer product mix while capacity holds steady.
Why do rack-level numbers matter more?
Nvidia increasingly sells integrated rack-scale systems, so per-rack memory defines demand. The GB200 NVL72 carries roughly 13.4 TB of HBM across 72 GPUs; the GB300 NVL72 raises that to 20.7 TB, a 54.7% increase. Vera Rubin NVL72 holds at 20.7 TB while aggregate bandwidth climbs from 576 TB/s to 1,400 TB/s — about 2.4 times the bandwidth with zero capacity growth, per Nvidia system specifications and The Information Network calculations.
A hypothetical deployment of 10,000 GB300 NVL72 racks requires roughly 207,000 TB of HBM. The same number of Vera Rubin racks needs about the same capacity. Pushing deployment to 15,000 Rubin racks lifts total HBM to roughly 310,500 TB, a 50% increase that comes entirely from more systems.
Three variables now move independently: installed capacity per rack, bandwidth and system shipment volume. Deployment volume can sustain bit-demand growth even when per-GPU capacity plateaus, while slower rack uptake would cut bit demand regardless of GPU performance gains.
Is Nvidia raising memory productivity?
Yes. Nvidia said in a recent technical discussion that Vera Rubin NVL72 can deliver up to 30 times the agentic AI throughput per megawatt of GB300 NVL72 in a specified workload, despite unchanged HBM capacity. The figure is workload-specific. It still shows that architecture and software now stretch each installed gigabyte further, which favors Nvidia's platform economics but prevents memory makers from treating GPU performance growth as automatic bit demand.
What does HBM4 mean for Micron, Samsung and SK hynix?
HBM4 can support higher average selling prices through greater bandwidth, advanced process requirements and tight integration with accelerators. Higher prices do not guarantee higher profit. Yields, manufacturing costs and capital expenditure determine the outcome; a supplier with strong HBM4 yields captures margin, while one with production trouble faces higher costs despite favorable demand.
Micron has invested heavily in HBM technology, advanced packaging and DRAM capacity, and its opportunity hinges on qualification for successive GPU generations. SK hynix holds a strong position through manufacturing experience and customer relationships. Samsung pursues the same market through advanced DRAM and packaging investment, with yields and high-volume qualifications as the gating factors.
There is also a manufacturing constraint: HBM consumes substantial wafer and packaging resources, so expanding HBM output reduces capacity for conventional DRAM. That helps explain why conventional DRAM markets stay tight even amid rising capital spending, though new fab capacity and better yields could create excess supply if demand slows.
The investment risk is confusing performance growth with memory demand. Suppliers that expand capacity on aggressive HBM consumption assumptions could face weaker utilization or pricing pressure, and an individual vendor's share gains can mask slower industry bit growth. Rubin Ultra's projected ~1 TB per GPU in 2027 could resume capacity growth, but investors should treat that roadmap figure separately from the current Rubin architecture rather than assume every annual transition repeats the Hopper-to-Blackwell pattern.
Source: 247wallst.com
More from Grace Kim
Show full bio
Market editor covering people, launches and day-to-day practice in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.
1,130 articles
Related articles
nvidias-nvhbm-sidesteps-jedec-in-the-hbm4-stack
Nvidia’s NVHBM Sidesteps JEDEC in the HBM4 Stack
sk-hynix-wins-second-straight-tsmc-partner-of-the-year-award-1eab6919
SK hynix Wins Second Straight TSMC Partner of the Year Award
samsung-and-tsmc-reportedly-team-up-on-nvidia-s-nvhbm-memory-37b86323
Samsung and TSMC Reportedly Team Up on Nvidia's NVHBM Memory
tsmc-weighs-3d-dram-entry-after-report-of-nvidia-request
TSMC Weighs 3D DRAM Entry After Report of NVIDIA Request



