Nvidia’s Annual GPU Upgrades Are Changing the HBM Investment Outlook
This post may contain links from our sponsors and affiliates, and Flywheel Publishing may receive compensation for actions taken through them.
- Nvidia’s annual GPU upgrade cycle is no longer producing consistent increases in high-bandwidth memory capacity per GPU.
- Vera Rubin maintains approximately 288 GB of HBM per GPU while delivering more than twice the memory bandwidth of Blackwell Ultra.
- Micron and other memory manufacturers can benefit from HBM4’s higher performance and pricing, but future revenue growth will depend increasingly on system shipments, memory technology, and manufacturing economics.
Nvidia’s Annual GPU Cadence Changes the Memory Equation
The rapid expansion of artificial intelligence has made high-bandwidth memory (HBM) one of the semiconductor industry’s most important growth markets. As AI accelerators become more powerful, memory manufacturers have invested heavily in additional capacity, advanced packaging, and new generations of HBM to satisfy the requirements of increasingly sophisticated computing systems.
However, the relationship between GPU performance and memory demand is changing. Nvidia has accelerated its product development schedule toward annual GPU introductions, creating the expectation that each successive generation will require substantially more HBM. That assumption was supported by the transition from Hopper to Blackwell and subsequently to Blackwell Ultra, when memory capacity per GPU increased sharply.
The introduction of Vera Rubin changes that relationship. Nvidia (NVDA | NVDA Price Prediction) is maintaining approximately the same HBM capacity per GPU as Blackwell Ultra while substantially increasing memory bandwidth and processing performance. This development has important implications for Micron Technology (MU), Samsung Electronics (SSNLF), and SK hynix (SKHY), the three principal suppliers of advanced HBM.
The central investment question is whether future HBM revenue growth will be driven primarily by increasing memory capacity per accelerator or by higher-value memory technology, more systems deployed, and improved manufacturing economics. The distinction matters because these factors can produce very different revenue and profitability outcomes for memory suppliers.
Nvidia’s GPU Cadence and HBM Capacity
Nvidia’s GPU roadmap illustrates how dramatically memory requirements increased during the transition from earlier AI accelerators to Blackwell Ultra. It also reveals a significant change in the latest generation: increasing computational performance no longer necessarily requires increasing the amount of installed HBM.
According to Table 1, HBM capacity per GPU increased from 80 GB in the H100 to 288 GB in Blackwell Ultra B300, representing a 3.6-fold increase. The transition from B200 to B300 alone raised capacity from 192 GB to 288 GB, an increase of 50%. However, Vera Rubin retains 288 GB per GPU while transitioning from HBM3e to HBM4. Consequently, Nvidia’s annual upgrade cycle is producing a substantial increase in memory performance without a corresponding increase in memory capacity per GPU.
| Table 1. Nvidia GPU Generations: HBM Capacity and Memory Technology, 2020–2027 | |||
| GPU generation | Introduction | HBM per GPU | HBM technology |
| 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 projected | HBM4e projected |
| Source: Nvidia product specifications and roadmap disclosures; The Information Network analysis. | |||
The implications extend beyond a single product transition. Between H100 and B300, memory suppliers benefited from both increasing GPU shipments and a dramatic expansion in the amount of memory installed in each accelerator. That combination multiplied the demand for advanced DRAM capacity.
The transition to Rubin is different. HBM4 provides substantially higher bandwidth through a more advanced memory architecture, including a wider interface and more demanding manufacturing requirements. These improvements increase the value and complexity of the memory supplied, even when the number of installed gigabytes remains unchanged.
For memory manufacturers, this creates a distinction between growth in HBM bits shipped and growth in HBM revenue. Revenue can increase through higher selling prices and a richer product mix without a comparable increase in total memory capacity. Conversely, higher bandwidth alone does not guarantee higher revenue if pricing premiums decline or production costs increase.
Rubin Ultra could eventually reverse the capacity plateau if Nvidia delivers approximately 1 TB of HBM per GPU. However, investors should distinguish that projected future increase from the current Rubin architecture rather than assuming every annual GPU transition follows the same pattern.
Rack-Level HBM Demand Reveals a More Important Trend
Individual GPU specifications provide only part of the picture because Nvidia increasingly sells integrated rack-scale computing systems. These systems combine dozens of GPUs, CPUs, networking components, and memory into a single computing platform. The transition from individual accelerators to integrated racks makes system-level memory requirements especially important for evaluating future HBM demand.
According to Table 2, the GB200 NVL72 contains approximately 13.4 TB of HBM across 72 GPUs, increasing to 20.7 TB in the GB300 NVL72, a 54.7% expansion in installed memory capacity. However, Vera Rubin NVL72 maintains approximately 20.7 TB while increasing aggregate HBM bandwidth from 576 TB/s to 1,400 TB/s. This represents approximately 2.4 times the memory bandwidth without any increase in installed HBM capacity per rack.
| Table 2. Nvidia Rack-Level HBM Capacity and Bandwidth Comparison | |||
| 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 |
| Source: Nvidia system specifications; The Information Network calculations. | |||
This comparison is critical because it separates three variables that are frequently combined in discussions of AI memory demand: installed HBM capacity, memory bandwidth, and the number of systems deployed. Installed capacity measures how many memory bits are required for each rack. Bandwidth measures how quickly those bits can be transferred to the GPUs. System shipments determine how many racks are purchased and therefore how much total memory the industry must manufacture.
The three variables are related, but they are not interchangeable. For example, consider a hypothetical deployment of 10,000 GB300 NVL72 racks. At 20.7 TB per rack, the installation would require approximately 207,000 TB of HBM capacity. Replacing those racks with 10,000 Vera Rubin NVL72 systems would require approximately the same aggregate HBM capacity, despite delivering substantially greater memory bandwidth and computing performance.
If deployment increased to 15,000 Rubin racks, total installed HBM would rise to approximately 310,500 TB, an increase of 50%. That growth would come from additional systems rather than greater memory capacity per system.
The distinction becomes increasingly important as Nvidia expands the scale of its AI infrastructure deployments. Memory manufacturers can experience substantial demand growth even during periods when HBM capacity per GPU remains unchanged, provided the number of systems shipped continues increasing. Conversely, slower rack deployment could reduce memory bit-demand growth even if GPU performance continues advancing rapidly.
Nvidia Is Increasing the Productivity of Installed Memory
The changing relationship between HBM capacity and GPU performance reflects a broader shift in Nvidia’s system architecture. In a recent technical discussion of memory optimization, Nvidia explained how hardware and software improvements can increase the amount of useful computation performed with 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, even though installed HBM capacity remains approximately unchanged. That figure is workload-specific and should not be interpreted as a universal 30-fold performance improvement. Nevertheless, it illustrates the importance of memory efficiency, system architecture, and software optimization in determining the economics of AI computing.
For hyperscalers, the relevant measure is increasingly the cost of completing an AI workload rather than simply the number of GPUs or gigabytes of memory installed. Higher memory bandwidth can reduce performance bottlenecks, while improved software can make more efficient use of available memory. Together, these improvements can allow more AI computations to be performed without proportionately increasing installed HBM capacity.
This development is favorable for Nvidia because it improves the economic value of its computing platforms. For memory manufacturers, however, it means that GPU performance growth cannot automatically be translated into equivalent growth in HBM bit demand.
What This Means for Micron, Samsung, and SK hynix
The transition from HBM3e to HBM4 presents significant opportunities for memory suppliers, even if installed capacity per accelerator remains unchanged. Micron has invested heavily in HBM manufacturing technology, advanced packaging, and DRAM capacity to expand its participation in AI infrastructure. Its investment opportunity increasingly depends on qualification for successive GPU generations, manufacturing yields, and the profitability of higher-value memory products.
The shift toward HBM4 can support higher average selling prices because the technology requires greater bandwidth, more advanced manufacturing processes, and demanding integration with AI accelerators. However, higher selling prices do not necessarily produce proportionately higher profits. Manufacturing costs, yields, and capital expenditures must also be considered. A supplier that achieves strong HBM4 yields could capture meaningful margin benefits, while one experiencing production difficulties could face higher costs despite favorable demand.
SK hynix has established a strong position in advanced HBM and benefits from its manufacturing experience and relationships with leading AI accelerator customers. Its challenge is to maintain that competitive advantage as HBM4 increases technical complexity and competitors expand their capabilities.
Samsung is pursuing the same opportunity through advanced DRAM manufacturing and HBM packaging investments. Its ability to improve yields and secure high-volume customer qualifications will influence its participation in the next generation of AI memory demand. These companies face a common economic reality: increasing the value of each gigabyte sold can be just as important as increasing the number of gigabytes shipped.
There is also a broader manufacturing constraint. HBM production consumes substantial wafer-processing and packaging resources, and increasing HBM output can reduce the capacity available for conventional DRAM products. Consequently, investments in HBM manufacturing do not necessarily translate into proportionate increases in conventional DRAM supply. This helps explain why conventional DRAM markets can remain constrained even as memory manufacturers expand capital expenditures. Over time, however, new fabrication capacity and improved yields could still create excess supply if demand growth slows.
The Investment Risk: Confusing Performance Growth With Memory Demand
Investors have become accustomed to treating each new generation of AI accelerators as an automatic source of additional memory demand. That assumption was reasonable during periods when memory capacity per GPU increased sharply. It becomes less reliable when system designers achieve substantial performance improvements through bandwidth, architecture, and software optimization.
The risk is particularly important when evaluating memory manufacturers’ capital expenditure plans. HBM4 requires advanced manufacturing capabilities, and suppliers must invest ahead of demand to qualify their products and secure customer orders. These investments can be highly profitable when production volumes and selling prices meet expectations. However, if memory manufacturers expand capacity based on aggressive assumptions about future HBM consumption, they could eventually encounter weaker utilization or pricing pressure.
The industry therefore faces two simultaneous developments. Advanced AI systems continue requiring more sophisticated memory technology, supporting demand for premium products. At the same time, improvements in memory efficiency can moderate the growth in installed capacity required for each system. Neither development eliminates the other.
Investors should also distinguish between an individual supplier gaining market share and the overall HBM market expanding. Micron, Samsung, or SK hynix could report substantial revenue growth through customer wins and improved product mix even if total industry HBM bit demand grows more slowly than previously expected. This distinction is especially relevant when evaluating companies whose share prices already reflect expectations of sustained AI memory shortages.
Investor Takeaway
Nvidia’s annual GPU upgrade cadence is changing the economics of high-bandwidth memory demand. The transition from H100 to Blackwell Ultra produced a 3.6-fold increase in HBM capacity per GPU, but the subsequent move to Vera Rubin maintains approximately the same installed memory capacity while substantially increasing bandwidth and computational performance.
For Nvidia investors, this represents an important architectural advantage. The company is improving system performance and memory productivity without requiring proportional increases in installed HBM capacity. That can strengthen the economics of AI infrastructure deployments and help customers obtain greater computational output from their capital investments.
For Micron, Samsung, and SK hynix investors, the implications are more complex. HBM4 offers opportunities for higher selling prices, advanced product differentiation, and improved margins. However, future growth will depend on production yields, customer qualifications, rack deployment volumes, and the balance between manufacturing capacity and market demand.
The most important conclusion is that the AI memory cycle has changed, but the fundamental risks of semiconductor supply and demand have not disappeared. Nvidia’s annual GPU cadence will continue driving demand for more advanced memory technology, yet investors should no longer assume that every increase in GPU performance produces a corresponding increase in installed HBM capacity. The companies that benefit most will be those capable of manufacturing increasingly sophisticated memory products at attractive margins while maintaining disciplined capacity expansion.
Contact [email protected] for any questions or corrections.








