Samsung Is Poised to Win if Nvidia Loses Steam
A wave of customer diversification away from Nvidia is redirecting HBM orders toward Samsung. The real story isn't Samsung's chips — it's whether Alphabet and AMD actually follow through.
A CEO Exit, an Industry Shift
Jensen Huang stepped down as Nvidia CEO, and the semiconductor world immediately started mapping the fallout. But the real story isn’t about who runs Nvidia next — it’s about where the memory goes when Nvidia doesn’t consume everything.
According to UBS estimates released this week, Nvidia’s share of global HBM (high-bandwidth memory) demand is projected to fall from 60 percent this year to 41 percent next year. That sounds like a decline, but it’s also just math. Demand is growing so fast that even a smaller percentage share means more absolute volume. The question is which memory maker captures the freed-up portion.
The answer, according to most analysts, points to Samsung Electronics.
Who Gets the Nvidia Drop-Off
UBS forecasts that Google will absorb much of what Nvidia leaves behind, rising from 18 percent to 30 percent of HBM demand, while AMD climbs from 7 percent to 13 percent. Morgan Stanley tells a similar story for 2027, projecting Nvidia at 37.3 percent, Alphabet at 36 percent, and AMD at 12.1 percent.
Together, the two non-Nvidia customers would account for nearly half of all HBM procurement by 2027. That is a structural repositioning — not a temporary blip caused by supply constraints or pricing cycles.
The driver is simpler and more durable than geopolitical hedging: it is the rise of application-specific integrated circuits, commonly called ASICs. Counterpoint Research estimates that HBM embedded in AI-server ASICs will grow thirty-five times between 2024 and 2028. Google’s Tensor Processing Units are leading that expansion, followed by custom silicon designs from companies like Broadcom, which already partners closely with Samsung on both memory and foundry production.
Why Samsung Benefits First
Samsung’s advantage isn’t mystery. It comes from two things that are easy to overlook until you compare them to SK Hynix.
First, Samsung already holds preferred HBM4 supply contracts with both Google and AMD. AMD named Samsung the top-source supplier for HBM4 on its upcoming MI455X accelerator. Google sources a significant volume of HBM for its TPUs through Broadcom, which funnels memory orders to Samsung. Those are not speculative pipelines — they are active procurement relationships covering next-generation products that ship within months.
SK Hynix, by contrast, remains heavily concentrated in Nvidia’s supply chain. UBS estimates Samsung’s HBM bit-share will reach 41 percent next year, edging past SK Hynix at 39 percent — a reversal after SK Hynix held roughly 59 percent of the market in 2025 before dropping to an estimated 48 percent this year. The trend line favors Samsung precisely because its customer base is wider.
Second, and arguably more important, is how Samsung’s vertical integration aligns with what advanced memory buyers now want. HBM4 and subsequent generations require a custom logic die at the bottom of the stack — the controller that manages communication between the memory layers and the GPU or ASIC. Designing and integrating that logic die is increasingly valuable as generative-AI models demand tighter performance envelopes. Samsung can design the logic, fabricate it on its own foundry, and assemble it with its own memory dies under one roof. That one-stop capability matters more with each new HBM generation.
The SK Hynix Exposure Problem
SK Hynix’s vulnerability is straightforward exposure. It won the early HBM race by locking in Nvidia as its anchor customer, and the strategy paid off handsomely for years. Now that same dependency looks fragile.
If Nvidia’s share drops ten percentage points, SK Hynix does not merely lose ten points of share — it loses the margin advantage that came from serving a single dominant buyer at scale. Restoring those volumes requires finding replacement customers willing to pay comparable margins in a market where Samsung and the newly minted Chinese memory entrants are competing aggressively. That is harder than it sounds. Foundry customers like Google do not switch suppliers lightly; once they qualify Samsung for a product, they tend to stick with that qualification.
This dynamic places Samsung on track to overtake SK Hynix in total HBM bits supplied by 2027, assuming current projections hold. Even if SK Hynix recovers some ground later, it will face a steeper climb than before.
Where the Case Falls Apart
Not everyone agrees the transition is automatic.
JP Morgan’s counter-scenario remains notable: it estimates Nvidia could still command more than 60 percent of HBM demand through 2027, implying that whatever share Google and AMD gain comes from a larger absolute pie rather than from reclaiming Nvidia’s existing footprint. In that world, SK Hynix’s relative position improves and Samsung’s advantage narrows considerably.
There is another risk specific to AMD. During its recent earnings call, AMD indicated that it may adjust HBM configurations depending on workload demands — language that suggests per-chip HBM quantities could fall short of some optimistic forecasts. If AMD ships fewer HBM stacks per accelerator than expected, Samsung’s anticipated revenue uplift from that relationship diminishes accordingly.
And perhaps most uncertain of all: none of this hinges solely on supplier assignments. Google and AMD have been quietly building internal silicon capabilities and deepening partnerships with multiple memory vendors simultaneously. Samsung is favored today, but vendor lock-in has never been absolute in this space.
What Happens Next
The near-term trajectory looks clear enough: Samsung is winning the current round of HBM allocations from non-Nvidia customers, and the gap versus SK Hynix is opening. Whether that lead persists depends on whether Nvidia truly loses its structural grip or simply shares demand more broadly in an expanding market.
For Samsung, the strategic takeaway is less about HBM technology than customer diversification. The company’s dual strength in memory manufacturing and foundry services makes it uniquely suited to serve the ASIC-driven future. For SK Hynix, the lesson is equally stark: having a single dominant customer is a liability when that customer’s strategy begins to shift.
The market should watch next quarter’s results for two signals — Samsung’s HBM unit volumes against expectations, and any change in AMD’s per-accelerator memory configuration guidance. Either one could accelerate or blunt the thesis.