Dan Ives Says Investors Miss Nvidia’s Korean Memory Engine

Dan Ives has sharpened a bullish argument around Nvidia despite a volatile summer for Korean memory stocks. The Wedbush analyst says investors still underestimate the earnings power created by AI infrastructure demand. Google News readers may recognize the familiar Nvidia thesis, but his supply-chain focus changes the story.

The argument is no longer simply that Nvidia sells the processors behind generative AI. It is that each new accelerator requires a wider system of specialized memory, networking, packaging, power, and data-center investment. Samsung Electronics and SK Hynix sit near the center of that system because they produce high-bandwidth memory, commonly called HBM.

That connection creates the real tension. Nvidia’s data-center business shows enormous demand, while Korean semiconductor shares have experienced sharp reversals after extraordinary gains. Ives treats those selloffs as pauses inside a longer AI investment cycle. Skeptics see something more familiar: a capital-intensive memory boom that can eventually produce excess supply, lower prices, and compressed margins.

This distinction matters beyond semiconductor portfolios. Cloud providers, AI developers, and enterprise buyers depend on the same chain. If memory remains scarce, computing capacity stays expensive and deployment schedules remain constrained. If suppliers expand too aggressively, today’s shortage economics can reverse faster than expected.

What Dan Ives Actually Changed in the Google News Debate

Ives shifted attention from one winning chip designer to the entire system required to turn AI processors into usable computing capacity.

In an August report distributed by Stocktwits and syndicated through AI market coverage, Ives described South Korea as an important center of the AI supply chain. He highlighted Samsung and SK Hynix alongside Nvidia rather than treating memory as a secondary component.

That framing extends the familiar Nvidia investment case. Nvidia designs accelerators, networking products, and software that let customers build large AI systems. However, an accelerator cannot keep its computing units busy unless data reaches them quickly enough.

HBM addresses that bottleneck. Manufacturers stack multiple memory layers and connect them through very short electrical paths. This design delivers far more bandwidth than conventional server memory placed farther from the processor.

The result is a tightly coupled product. Nvidia can raise accelerator performance, but the complete system still needs qualified HBM with enough capacity and acceptable production yields. A shortage in one component can delay the sale or installation of an entire server rack.

That is why Ives’ focus on Korean memory stocks carries more weight than a routine list of bullish picks. It identifies the physical constraint behind rising AI capital spending. The argument links Nvidia’s order book to the factories, packaging lines, and engineering work needed to supply memory.

Nvidia’s own filings support the existence of extraordinary infrastructure demand. The company reported $51.2 billion in data-center revenue for its fiscal third quarter of 2026. That figure increased 66% from the prior year and 25% from the previous quarter.

Data-center compute revenue reached $43 billion, while networking revenue reached $8.2 billion. The latter grew 162% from the prior year, according to Nvidia’s quarterly filing. Those results show that customers are buying systems, not isolated processors.

Ives interprets that spending as an early stage of a broader cycle. His case assumes demand will move outward from accelerators into hyperscale infrastructure, enterprise software, and AI-enabled services. Memory suppliers benefit earlier because their components must be installed before those services can operate.

The distinction is important for readers following Google News headlines about NVDA. Nvidia remains the visible platform company, but its sales pull revenue through several less visible layers. SK Hynix and Samsung turn that pull into a Korean memory demand story.

This does not mean every supplier captures equal value. Nvidia controls the software platform and much of the system architecture. Memory remains a more standardized business, even when advanced HBM requires difficult manufacturing and qualification.

Still, HBM is not interchangeable at the moment a new platform launches. Suppliers must meet demanding targets for bandwidth, heat, power use, reliability, and packaging compatibility. That creates a temporary advantage for vendors that qualify early and manufacture reliably.

The event, then, is not a new Nvidia product announcement. It is an analyst’s attempt to redraw the market map. Instead of asking whether AI demand helps Nvidia, Ives asks which suppliers participate when Nvidia’s platforms keep scaling.

Nvidia Memory Demand Turns Korea Into a Core AI Market

The AI infrastructure trade now pressures investors to evaluate Korea as part of Nvidia’s production system, not as a separate regional market.

South Korea matters because Samsung and SK Hynix hold enormous positions in global memory production. Together, they make about two-thirds of the world’s memory chips, according to reporting on their Korean chip expansion.

That concentration gives the country unusual exposure to both AI demand and memory-cycle risk. Strong accelerator orders can raise HBM demand, improve supplier bargaining power, and redirect factory capacity. A slowdown can transmit through the same chain.

HBM production also affects conventional memory markets. Advanced memory consumes manufacturing resources, engineering attention, and packaging capacity. When suppliers prioritize HBM, fewer resources remain available for ordinary DRAM products.

This reallocation can tighten the broader market even when personal-computer or smartphone demand remains modest. Higher utilization and a better product mix can lift earnings across a supplier’s portfolio. That mechanism helps explain why investors have treated AI memory as more than a niche category.

SK Hynix entered this cycle with an especially strong position. The company’s 2026 outlook cited Counterpoint Research estimates showing a 62% HBM shipment share during the second quarter of 2025. Its revenue share stood at 57% during the following quarter.

The same HBM market outlook said HBM3E should represent roughly two-thirds of 2026 HBM shipments. HBM3E is an advanced generation used with current AI accelerators, including systems based on Nvidia’s Blackwell architecture.

SK Hynix also said it had established a mass-production system for HBM4. That newer generation targets forthcoming accelerator platforms and raises bandwidth further. However, readers should treat the company’s market positioning and production claims as corporate statements until customer shipments confirm them.

Samsung presents a different profile. It has a broader semiconductor operation, large conventional memory exposure, and extensive manufacturing resources. Its challenge is converting that scale into timely qualifications for the most valuable AI products.

Competition between the two Korean companies therefore works inside the larger Nvidia demand story. SK Hynix seeks to defend an early HBM lead. Samsung seeks to use its scale and engineering depth to capture more advanced orders.

Micron adds a third major supplier from the United States. Its presence gives customers another qualified source and reduces dependence on two Korean companies. It also increases the chance that future capacity growth weakens supplier pricing.

For now, the market is rewarding access to scarce memory. South Korea’s policy response shows how seriously the country takes that opportunity. Samsung and SK Hynix announced plans for four additional fabrication plants in the nation’s southwest.

The companies placed their combined long-term investment plan at 800 trillion won. They did not provide completion dates for those particular plants. SK Group Chairman Chey Tae-won also warned that such projects require extensive land, water, power, and skilled workers.

That warning reveals a constraint behind every large capacity announcement. A semiconductor plant takes years to plan, construct, equip, and qualify. Spending commitments do not immediately become usable HBM output.

This lag supports Ives’ bullish view in the near term. Nvidia memory demand can grow faster than supply because customers order systems now, while new factories arrive much later. Existing suppliers can therefore retain bargaining power during the gap.

The same lag creates long-term uncertainty. Projects announced by Samsung, SK Hynix, and Micron can eventually reach the market together. If AI spending slows before those lines fill, a shortage can become overcapacity.

Korea is consequently both the beneficiary and the pressure point. Its memory manufacturers must invest enough to support customers without repeating the industry’s history of excessive expansion. Investors must decide whether HBM changes that history or merely extends the current upcycle.

The Mechanism Linking NVDA Orders to Korean Earnings

Nvidia’s demand reaches Korean earnings through qualification, capacity allocation, packaging, and product mix, not through a simple rise in chip volume.

The first link is architectural. Modern AI models move large amounts of information between processors and memory. Accelerator performance suffers when memory cannot deliver that information fast enough.

HBM places stacked memory close to the processor and connects both components within an advanced package. The arrangement increases data throughput while managing space and power constraints. It also makes manufacturing more complicated.

A supplier must produce working memory dies, stack them successfully, and integrate them with packaging partners. Yield, meaning the share of manufactured units that meet specifications, becomes a decisive financial variable. Poor yields raise costs and restrict shipments.

The second link is customer qualification. Nvidia and other accelerator designers test memory against specific platform requirements. Early qualification can give a supplier a favorable position during the steepest part of a product ramp.

That advantage is not permanent. Samsung, SK Hynix, and Micron continually improve products and manufacturing processes. Customers also prefer multiple suppliers because diversification reduces operational and geopolitical risk.

The third link is capacity allocation. HBM uses more wafer and packaging resources than conventional products with comparable bit capacity. Expanding HBM production can therefore limit supply growth elsewhere in the memory portfolio.

This tradeoff can support prices for server DRAM and other products. It can also expose manufacturers to execution risk if they allocate too much capacity toward one demand forecast. The best product mix today can become the wrong one after customer plans change.

The fourth link is system scale. Nvidia no longer sells only a processor that customers install independently. Its data-center offering includes compute modules, networking, interconnects, software, and complete rack-scale designs.

That wider system increases the number of components tied to each deployment. Nvidia’s 162% annual networking growth in fiscal 2026 illustrates the effect. Customers require fast connections within and between racks, not simply faster individual GPUs.

Memory suppliers benefit from the same shift. Larger models, longer context windows, and inference workloads can require greater memory capacity. Inference is the process of running a trained model to generate answers or predictions.

Training created the first visible wave of accelerator demand. Inference broadens the potential workload because deployed applications serve users continuously. Yet it does not guarantee unlimited HBM consumption.

Model developers are improving efficiency through quantization, smaller specialized models, caching, and better software. Quantization reduces the numerical precision used by a model, lowering its memory needs. These techniques can reduce hardware required for a given task.

Efficiency can also increase total consumption. Lower costs make more applications economically practical, bringing additional workloads online. The final effect depends on whether usage grows faster than the memory saved per task.

This is where Ives’ argument becomes more than a forecast about chip shipments. He expects the economic value of AI adoption to spread across the technology stack. Hardware demand remains central because every software service ultimately needs computing capacity.

Enterprise adoption provides a practical example. A company deploying internal AI search must process documents, store representations, run models, and return answers within an acceptable delay. Teams can organize those inputs through a searchable knowledge base, but the underlying service still consumes infrastructure.

Millions of small interactions can create sustained inference demand. The resulting workload reaches cloud providers, accelerator vendors, network suppliers, and memory manufacturers. That is the demand chain behind the bullish Dan Ives AI stocks thesis.

However, earnings do not move in perfect alignment across that chain. Nvidia can preserve high margins through software integration and platform control. Memory suppliers remain more exposed to manufacturing costs, pricing cycles, and competing capacity.

SK Hynix’s early position gives it leverage, while Samsung’s scale gives it recovery potential. Micron offers geographic diversification and additional competition. Nvidia benefits when all three improve supply because abundant HBM supports more accelerator shipments.

That last point creates a subtle conflict. Memory manufacturers benefit from scarcity and favorable prices. Nvidia benefits from enough supply to ship complete systems at scale. Their demand outlook aligns, but their ideal supply conditions do not.

What the Bullish Memory Thesis Does Not Settle

The evidence supports strong AI demand, but it does not prove that today’s margins, market shares, or stock valuations will persist.

Memory is one of technology’s most cyclical industries. Producers spend heavily during shortages, add capacity with long delays, and sometimes bring that capacity online after demand has weakened. Prices can then fall rapidly.

HBM contains more differentiation than conventional memory. Its design, stacking, thermal behavior, and qualification requirements create meaningful barriers. Those barriers can slow commoditization, but they do not eliminate it.

SK Hynix itself acknowledges the risk. Its 2026 outlook notes concerns about price corrections after 2026 as competition intensifies and production expands. The company argues that technical gaps make a sudden near-term shift unlikely.

That assessment is useful, but it comes from a market leader with an interest in defending its position. Independent investors still need evidence from shipment volume, selling prices, customer concentration, and manufacturing yields.

Samsung’s response is especially important. Successful qualification for more Nvidia platforms would improve supply diversity and intensify competition. It could strengthen the overall AI system while reducing SK Hynix’s individual bargaining power.

Micron creates similar pressure. Three credible suppliers can support faster accelerator growth, but they can also narrow excess returns. The outcome depends on how quickly Nvidia memory demand expands relative to qualified supply.

Market concentration adds another risk. Samsung and SK Hynix have become central to South Korea’s stock market performance. By August, the two companies represented roughly half of the KOSPI index, according to an AI volatility analysis.

That concentration can magnify both rallies and selloffs. A change in global risk appetite, interest rates, or leveraged positioning can overwhelm favorable operating news. Share-price weakness does not automatically indicate weaker AI demand.

SK Hynix’s response illustrates the gap between business conditions and market behavior. The company announced a 40 trillion won share repurchase and cancellation plan after a two-month share decline. The move followed a substantial loss of market value despite continued enthusiasm around AI memory.

Ives views such volatility as a pause rather than a broken thesis. That interpretation remains an analyst judgment, not an established fact. Markets can price future growth too aggressively even when the underlying industry continues expanding.

Customer concentration deserves equal attention. A supplier that depends heavily on Nvidia gains exposure to the leading AI platform. It also becomes vulnerable to design changes, qualification losses, or shifts toward custom accelerators.

Google, Amazon, and other cloud operators develop proprietary chips for selected workloads. Those chips still use advanced memory, so they can broaden HBM demand. However, their designs can change supplier relationships and reduce Nvidia’s share of certain deployments.

Export controls create another uncertainty. Nvidia’s filing says changing trade rules can affect investment decisions, supply operations, costs, and customer purchase timing. Restrictions can close markets or force companies to redesign products.

Power availability also constrains deployment. Customers cannot operate new AI racks without electrical capacity, cooling, networking, and suitable buildings. Chip orders can move ahead of completed data centers, creating timing differences between reported demand and productive use.

Financing matters as well. Large cloud providers can fund long construction programs, but smaller operators depend more heavily on capital markets. Higher borrowing costs or weaker utilization can delay projects.

The demand forecast ultimately rests on customer economics. Enterprises must obtain enough value from AI applications to justify recurring infrastructure spending. Experiments and pilot projects do not necessarily become large production workloads.

Developers are also working to reduce inference costs. Better models, optimized software, and specialized chips can process more work with the same memory. A bullish forecast must account for those improvements instead of assuming fixed hardware intensity.

The counterargument does not require an AI collapse. Demand can remain healthy while supplier earnings disappoint elevated expectations. Slower growth, falling prices, or a shift in market share can weaken the investment thesis without ending the infrastructure cycle.

Google News readers should therefore separate three questions. Is AI computing demand growing? Are Korean suppliers capturing that growth? Are their shares pricing the opportunity conservatively? Each question requires different evidence.

Ives offers a confident answer to the first two. The third depends on market prices, future profits, and each investor’s assumptions. It cannot be settled by a strong quarter or a bullish television interview.

Three Signals That Will Test Dan Ives’ AI Demand Case

The next test will come from platform shipments, HBM4 qualification, and cloud spending rather than another round of optimistic commentary.

The first signal is Nvidia’s data-center growth. Its fiscal third-quarter numbers established a demanding baseline: $51.2 billion in data-center revenue and 66% annual growth. Future reports must show that Blackwell-based systems continue converting orders into recognized revenue.

Networking deserves particular attention. Strong NVLink, InfiniBand, and Ethernet sales suggest customers are building complete clusters. Weak networking growth beside strong processor shipments might indicate a less balanced infrastructure ramp.

Investors should also monitor inventory and purchase commitments. Rising commitments can support a durable demand case, but they increase exposure if customer schedules change. Nvidia’s disclosures can reveal whether supply remains the constraint or demand is becoming less certain.

The second signal is HBM4 qualification and production yield. HBM4 represents the next major transition for advanced memory. Early supplier positions will influence market share as Nvidia’s forthcoming platforms scale.

SK Hynix says its HBM3E leadership is carrying into HBM4. Samsung wants a larger role, while Micron continues developing its own supply. Actual customer qualifications and commercial shipments matter more than sample announcements.

Yield will determine whether qualification produces attractive earnings. A technically successful product can still disappoint financially if too many units fail manufacturing tests. Higher usable output would support system shipments but could eventually ease scarcity.

This signal can strengthen or weaken Ives’ view in different ways. Smooth HBM4 ramps across several vendors would reinforce the broader AI infrastructure thesis. A major qualification loss would weaken the case for the affected supplier without necessarily hurting Nvidia.

The third signal is capital spending from Microsoft, Alphabet, Amazon, Meta, and other large operators. These companies fund much of the accelerator and memory demand moving through the supply chain. Their budgets provide an external check on vendor optimism.

Spending alone is insufficient. Readers should compare capital expenditures with cloud growth, AI service usage, and management comments about capacity. More construction supports the demand chain only when customers can deploy and monetize the equipment.

A coordinated reduction in data-center plans would challenge Ives’ central argument. Continued expansion, backed by higher AI revenue or usage, would strengthen it. Delays caused only by power or construction could preserve demand while shifting its timing.

These signals should be evaluated together. Nvidia can report strong sales while memory suppliers lose pricing power. HBM companies can report excellent margins while cloud customers question future spending. No single metric proves the entire chain.

The larger insight from this Google News story is that the AI market has moved beyond a one-ticker narrative. NVDA remains central, but accelerator performance depends on memory, packaging, networking, electricity, and customer economics.

Dan Ives is betting that demand across this chain remains underestimated. The verified numbers show a large and expanding infrastructure market. They do not remove cyclicality, competition, or the possibility that investors have already priced years of favorable growth.

For developers and enterprise buyers, the immediate question is practical. Do HBM availability and cloud capacity improve enough to lower the cost of deploying useful AI systems? For investors, the question is stricter: does revenue growth exceed the expectations embedded in these companies?

Watch Nvidia’s complete system sales, not only GPU announcements. Watch which memory suppliers qualify for HBM4, then examine their yields. Finally, compare hyperscaler spending with measurable AI usage. Those three checks will reveal whether the Korean memory engine is durable or simply running at the peak of another cycle.