The most important bottleneck in artificial intelligence infrastructure right now isn’t the processor. It’s the memory.
That distinction matters enormously for Nvidia, the $3.4 trillion colossus whose data center GPUs have become the backbone of AI training and inference worldwide. And it’s the reason one Wall Street analyst just handed the stock a fresh vote of confidence — not because of anything Nvidia designed in-house, but because of a component it buys from someone else.
Vijay Rakesh, an analyst at Mizuho Securities, reiterated his Outperform rating on Nvidia shares this week and maintained a $180 price target, arguing that surging demand for high-bandwidth memory, or HBM, will be a powerful tailwind through 2026 and beyond. As Barron’s reported, Rakesh sees the total addressable market for HBM ballooning from roughly $25 billion this year to $45 billion by 2026 — an 80% increase that reflects how ravenous AI workloads have become for memory bandwidth.
The logic is straightforward. Every next-generation GPU Nvidia ships requires more HBM. The Blackwell architecture, Nvidia’s current flagship, already uses significantly more memory per chip than its predecessor, the Hopper series. And the forthcoming Rubin platform, expected to arrive in late 2026, will push those requirements higher still. More memory per GPU means more revenue per system — not just for the memory makers, but for Nvidia itself, which captures the value of the total package it sells to hyperscale cloud providers.
This isn’t a peripheral story. It’s central to the investment thesis.
HBM is a specialized type of DRAM that stacks multiple memory dies vertically, connecting them with through-silicon vias to deliver far greater bandwidth than conventional memory modules. The technology has existed for years, but AI’s explosive growth has turned it from a niche product into one of the most supply-constrained components in the semiconductor industry. SK Hynix, Samsung, and Micron are the three primary suppliers, and all three have been racing to expand production capacity. SK Hynix has held a dominant position, particularly with its HBM3E product, which is qualified for Nvidia’s Blackwell GPUs. Samsung has been playing catch-up after quality issues delayed its qualification process, though recent reports suggest it has made progress. Micron, meanwhile, has been aggressively investing in its own HBM capacity and has secured supply agreements with Nvidia.
Rakesh’s analysis, as detailed by Barron’s, highlights that HBM demand isn’t just growing — it’s accelerating faster than many investors appreciate. The transition from HBM3E to the next generation, HBM4, is expected to coincide with Nvidia’s Rubin architecture. HBM4 will offer even greater bandwidth and capacity per stack, which means the dollar content of memory in each GPU system will continue to climb.
For Nvidia, this creates a virtuous cycle. Higher memory requirements raise the average selling price of its GPU systems. They also deepen Nvidia’s competitive moat, because integrating HBM at scale requires extensive co-engineering between the GPU designer and the memory suppliers. That kind of tight collaboration isn’t easily replicated by competitors like AMD or the growing roster of custom AI chip designers at companies like Google, Amazon, and Microsoft.
But there’s a risk embedded in this story, too.
If HBM supply can’t keep pace with demand, Nvidia’s own shipments could be constrained. This has already happened. During the initial Blackwell ramp, memory availability was one of several factors that created supply tightness. Nvidia CEO Jensen Huang has acknowledged on earnings calls that demand for the company’s products far outstrips supply, and memory is one of the key gating factors. The company has been working to diversify its HBM sourcing — qualifying multiple suppliers for each generation — but the reality is that only three companies on earth can manufacture this product at scale.
The memory angle also intersects with geopolitics. SK Hynix and Samsung are both South Korean companies. Micron is American but manufactures globally. U.S. export controls on advanced semiconductors to China have already reshaped trade flows, and any escalation in trade tensions — particularly involving South Korea — could ripple through the HBM supply chain in unpredictable ways. The recent tariff turbulence between the U.S. and its trading partners has kept this concern alive, even as the semiconductor industry has so far avoided the worst-case scenarios.
Nvidia shares have had a remarkable run. The stock is up more than 200% over the past year, driven by relentless demand for AI infrastructure from the likes of Microsoft, Meta, Google, Amazon, and Oracle. The company reported fiscal fourth-quarter revenue of $39.3 billion in February, a 78% year-over-year increase, with data center revenue alone hitting $35.6 billion. Gross margins have remained extraordinarily high, hovering near 73%, in part because Nvidia’s pricing power allows it to capture the value of the entire GPU-plus-memory system it delivers.
Rakesh’s $180 price target implies roughly 30% upside from recent levels around $138. That’s an aggressive call, but it’s grounded in the math of HBM scaling. If the memory TAM really does reach $45 billion by 2026, and Nvidia continues to dominate the market for AI training and inference hardware, the revenue trajectory supports a significantly higher stock price. The question is whether competition, supply constraints, or a slowdown in AI capital spending could derail that trajectory.
On the competition front, AMD has been making incremental gains with its Instinct MI300X accelerator, which also uses HBM. But AMD’s market share in AI accelerators remains in the single digits, and Nvidia’s software platform — CUDA, along with its extensive library of AI frameworks and tools — continues to create enormous switching costs for developers and enterprises. Custom silicon from the hyperscalers is a longer-term threat, but most industry observers believe it will complement rather than replace merchant GPUs from Nvidia for at least the next several years.
The AI capital expenditure cycle shows no signs of cooling. Microsoft alone has signaled plans to spend more than $80 billion on AI-related infrastructure in its current fiscal year. Meta has guided to $60-65 billion. Google and Amazon are each spending in the $50 billion-plus range. Much of that money flows directly to Nvidia. And every dollar spent on a GPU system includes a growing proportion allocated to HBM.
So the memory story is really an amplifier for the broader Nvidia thesis. It’s not just that Nvidia sells more GPUs. It’s that each GPU it sells contains more high-value memory, which raises ASPs, which drives revenue growth even if unit volumes were to plateau — which they aren’t.
There’s a subtlety here that deserves attention. The HBM market’s rapid expansion is also creating opportunities and pressures for the memory makers themselves. SK Hynix has seen its stock surge as investors recognize the profit potential of HBM, which commands significantly higher margins than standard DRAM. Samsung, after stumbling on HBM3E qualification, has been investing heavily to regain ground, and its semiconductor division’s profitability hinges in part on catching up. Micron has positioned HBM as a key growth driver and has been rewarded by investors for its supply agreements with Nvidia.
The interdependence between Nvidia and its memory suppliers is unusually tight for the semiconductor industry. Nvidia doesn’t just buy commodity parts. It works closely with HBM vendors on packaging, thermal management, and electrical specifications that are tailored to each GPU architecture. This co-design process takes months and creates a web of mutual dependency. Nvidia needs the memory makers to hit their production targets. The memory makers need Nvidia’s qualification to access the most profitable segment of the DRAM market.
What Rakesh’s analysis ultimately underscores is that Nvidia’s dominance in AI hardware isn’t just a function of chip design prowess. It’s a systems-level advantage. The company’s ability to integrate cutting-edge GPUs with the highest-performance memory available, package them together with custom networking (via its Mellanox acquisition) and software, and deliver complete solutions to hyperscale customers — that’s what makes it extraordinarily difficult to displace.
Memory is the unsung hero of this story. Or, perhaps more accurately, the unsung profit driver. As HBM content per GPU continues to rise with each architectural generation, the financial impact compounds. Investors focused solely on chip counts or GPU shipment volumes are missing a significant part of the picture.
Nvidia reports its next quarterly results in late May. The market will be watching closely for commentary on Blackwell production ramp, supply chain conditions, and — yes — memory. If Rakesh is right about the HBM trajectory, the current stock price may still be underestimating what’s coming.
WebProNews is an iEntry Publication