Nvidia's Vera Rubin Holds HBM Flat at 288 GB While Bandwidth Jumps
Nvidia's Vera Rubin keeps 288 GB of HBM per GPU while more than doubling bandwidth versus Blackwell Ultra, upending the assumption that every GPU generation means more memory demand.

Why it matters
- Vera Rubin keeps approximately 288 GB of HBM per GPU while delivering more than twice the memory bandwidth of Blackwell Ultra.
- HBM per GPU grew 3.6x from H100's 80 GB to Blackwell Ultra B300's 288 GB before flat-lining with Rubin in 2026.
- Vera Rubin NVL72 raises aggregate HBM bandwidth from 576 TB/s to 1,400 TB/s with the same 20.7 TB of memory per rack.
- Rubin Ultra, projected for 2027, could return to ~1 TB of HBM4e per GPU.
- Nvidia says Vera Rubin NVL72 can deliver up to 30x the agentic AI throughput per megawatt of GB300 NVL72 in a specified workload.
Nvidia's next flagship GPU will ship with roughly the same amount of high-bandwidth memory as its predecessor — the first break in years of relentless per-chip memory growth. Vera Rubin, due in 2026, carries approximately 288 GB of HBM per GPU, identical to Blackwell Ultra B300, while delivering more than twice the memory bandwidth, according to Nvidia product specifications and roadmap disclosures analyzed by The Information Network.
That single spec change reframes one of the semiconductor industry's biggest investment questions. For three years, memory suppliers rode a simple formula: each new Nvidia accelerator needed dramatically more HBM, so every GPU generation automatically multiplied memory demand. Vera Rubin breaks that link between GPU performance and installed memory capacity — and the implications reach straight into the stock stories of Micron (MU), Samsung Electronics (SSNLF), and SK hynix (SKHY), the three principal suppliers of advanced HBM.
How big is the capacity change?
The numbers from Nvidia's GPU roadmap show how sharply the trend reversed.
- 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
From H100 to Blackwell Ultra, HBM capacity per GPU rose 3.6-fold, from 80 GB to 288 GB. The B200-to-B300 step alone added 50% more memory, from 192 GB to 288 GB. Then Rubin arrives and holds capacity flat while switching from HBM3e to HBM4 — a substantially higher-bandwidth, more complex memory architecture with a wider interface and more demanding manufacturing requirements.
Rubin Ultra could eventually restore the growth curve with its projected ~1 TB per GPU in 2027. But the analysis warns investors against extrapolating that projection backwards and assuming every annual GPU transition follows the old pattern.
What does rack-level demand reveal?
Individual GPU specs tell only part of the story, because Nvidia increasingly sells integrated rack-scale systems that combine dozens of GPUs, CPUs, networking components, and memory into a single platform. The rack numbers make the plateau unmistakable.
| Metric | GB200 NVL72 | GB300 NVL72 | Vera Rubin NVL72 |
|---|---|---|---|
| GPUs per rack | 72 | 72 | 72 |
| HBM generation | HBM3e | HBM3e | HBM4 |
| HBM capacity per rack | 13.4 TB | 20.7 TB | 20.7 TB |
| Capacity growth | Baseline | +54.7% | 0.0% |
| Aggregate HBM bandwidth | 576 TB/s | 576 TB/s | 1,400 TB/s |
The GB300 NVL72 expanded installed memory 54.7% over the GB200. Vera Rubin NVL72 adds zero capacity while lifting aggregate bandwidth from 576 TB/s to 1,400 TB/s — roughly 2.4 times the bandwidth, per Nvidia system specifications and The Information Network calculations.
This separation matters because three variables get routinely conflated in AI memory debates: installed capacity, bandwidth, and system shipments. Capacity measures how many memory bits each rack needs. Bandwidth measures how fast those bits reach the GPUs. Shipments determine how many racks — and therefore how much total memory — the industry must manufacture.
The analysis offers a concrete illustration. A hypothetical deployment of 10,000 GB300 NVL72 racks at 20.7 TB each requires approximately 207,000 TB of HBM. Swapping in 10,000 Vera Rubin NVL72 systems requires roughly the same aggregate capacity, despite far greater bandwidth and compute. If deployment grows to 15,000 Rubin racks, total installed HBM rises to about 310,500 TB — a 50% increase that comes entirely from additional systems, not from more memory per system.
That cuts both ways for suppliers. Memory manufacturers can see strong demand growth even when capacity per GPU is flat, provided rack shipments keep climbing. But slower rack deployment could curb bit-demand growth even while GPU performance races ahead.
Why is Nvidia getting more from the same memory?
The flat capacity line reflects a deliberate architectural shift. In a recent technical discussion of memory optimization, Nvidia explained how hardware and software improvements increase the useful computation drawn from a given quantity of memory. The company reported that Vera Rubin NVL72 can deliver up to 30 times the agentic AI throughput per megawatt of GB300 NVL72 in a specified workload — a workload-specific figure, not a universal 30-fold speedup — while installed HBM stays roughly unchanged.
For hyperscalers, the metric that increasingly matters is the cost of completing an AI workload, not the raw count of GPUs or gigabytes installed, with higher bandwidth clearing bottlenecks and better software extracting more from available memory. This is favorable for Nvidia, which improves the economic value of its platforms. For memory manufacturers, it means GPU performance growth no longer translates automatically into equivalent HBM bit demand.
What does this mean for Micron, Samsung, and SK hynix?
The HBM3e-to-HBM4 transition still creates real opportunities for suppliers, because the new generation supports higher average selling prices through greater bandwidth, advanced manufacturing processes, and demanding integration with AI accelerators. But higher prices do not guarantee proportionate profits — manufacturing costs, yields, and capital expenditures all enter the equation. A supplier with strong HBM4 yields could capture meaningful margin; one struggling with production could face higher costs despite favorable demand.
- Micron has invested heavily in HBM manufacturing technology, advanced packaging, and DRAM capacity. Its opportunity increasingly hinges on qualification for successive GPU generations, yields, and the profitability of higher-value products.
- SK hynix holds a strong position in advanced HBM, backed by manufacturing experience and relationships with leading accelerator customers. Its challenge is defending that lead as HBM4 raises technical complexity.
- Samsung is pursuing the same market through advanced DRAM manufacturing and HBM packaging investments, with yields and high-volume customer qualifications determining its share of the next demand wave.
There is also a manufacturing constraint beneath the surface. HBM production consumes substantial wafer-processing and packaging resources, so expanding HBM output can shrink the capacity available for conventional DRAM. That helps explain why conventional DRAM markets can stay tight even as memory makers expand capital spending — though new fabrication capacity and improved yields could eventually produce excess supply if demand growth slows.
Where is the investment risk?
The core risk, per the analysis, is confusing performance growth with memory demand. Investors have treated each new accelerator generation as an automatic source of additional memory demand — a reasonable assumption when capacity per GPU was climbing sharply, and a less reliable one when system designers extract performance from bandwidth, architecture, and software optimization instead.
The risk bites hardest on capital expenditure plans. HBM4 requires advanced manufacturing capabilities, and suppliers must invest ahead of demand to qualify products and win orders. Those bets pay off when volumes and prices hold. If manufacturers expand capacity on aggressive assumptions about future HBM consumption, they could face weaker utilization or pricing pressure down the line.
Investors should also separate an individual supplier gaining share from the overall market expanding. Micron, Samsung, or SK hynix could report substantial revenue growth from customer wins and richer product mix even if total industry bit demand grows more slowly than expected — a distinction especially relevant for stocks that already price in sustained AI memory shortages.
The takeaway: the AI memory cycle has changed, but the fundamental risks of semiconductor supply and demand have not gone anywhere. The companies that benefit most from here will be those that can manufacture increasingly sophisticated memory at attractive margins while keeping capacity expansion disciplined — and investors can no longer assume every GPU performance jump buys a matching jump in HBM.
Source: 247wallst.com
More from Elena Vasquez
Show full bio
Market editor covering technology and product news in artificial intelligence and machine learning at AI In Context.
744 articles
Related articles
- TSMC Is Weighing 3D DRAM Production at NVIDIA's Request
- DRAM prices forecast to rise 400% as AI redirects memory supply
- Nvidia's Rubin HPC GPU Packs 50,000 CUDA Cores, Ships 2028
- Analysts: Nvidia's Wobbling Chip Dominance Hands Samsung an Opening
- Samsung and SK hynix cash piles hit 329 trillion won on AI boom