Category: Semiconductor

Nvidia’s CEO Says Memory and Neocloud Demand Are Fine. Here Is What Investors Should Weigh

Nvidia neocloud demand is “off the charts,” according to Jensen Huang, who spoke with CNBC at GTC Berlin on Thursday. He also took on fears of a memory glut. His message on both was that demand is strong, though investors should remember he runs the company selling into both markets.

Memory: no glut in sight

Huang said a surplus of memory supply is unlikely by 2029, despite a sell off in memory stocks over the past week. SK Hynix, Samsung and Micron took a hit after a South Korean opinion column warned of a “red ocean” of excess supply, meaning tougher competition and slimmer profits. The selloff was also driven by China’s CXMT announcing a new design architecture that could lead to greater efficiencies. Nvidia’s Jensen Huang says he sees no sign of a 2029 memory glut: ‘You can always use more’ +2

Huang’s answer was simple: memory is something you can always use more of. He said HBM will stay in use for many years and that Nvidia plans to keep using large amounts of LPDDR, the memory type found in phones, in data centers. He said that LPDDR approach delivers performance at half the energy per unit of work of any alternative. Jensen Huang pushed back on 2029 memory glut fears amid stock selloff +2

Asked about CXMT specifically, he said Nvidia is open minded and noted that nearly the whole world works with Chinese companies somewhere in the supply chain. He also said U.S. regulations have accelerated China’s efforts to build domestic chips.

The takeaway for investors: a glut call is really a supply call. Huang is a big buyer vouching for demand, but the bear case rests on new capacity, including from Chinese producers. Memory has always been a cyclical business, and one CEO’s forecast three years out does not change that.

Nvidia neocloud demand: money is the bottleneck

On the neocloud side, Huang said the demand for neoclouds is off the charts and their capabilities are extraordinary, but the working capital to fund that growth is missing. In his words, the demand for capital is much greater than its availability, and he expects that to get solved. He framed it as a scarcity of capital rather than weak demand, because so many AI sectors are competing for the same money. Nvidia’s Jensen Huang Wants AI Chips To Be A ‘Financeable Asset,’ Says Neocloud Demand Is ‘Off The Charts’ +2

He also pointed to a second constraint. Power is a limiting factor, especially in Europe where electricity costs more, and he argued scarce power should go to companies like Nebius and Mistral if they are to scale.

The market was not reassured. Ahead of Thursday’s open, Nebius fell 3.6%, CoreWeave slipped 4%, and Iren and Applied Digital each dropped nearly 4%. CoreWeave shares are down 50% over six months, even as Nvidia invested another $2 billion in the company in January.

Huang’s bigger pitch is about financing itself. He said Nvidia is trying to establish its compute as a financeable asset, because its residual value is durable enough to become an asset class. If lenders buy that, the capital gap narrows. If they do not, the gap stays.

The other side of the story

Critics see a circular loop. The Information reported that Nvidia is increasingly acting as lender and backstop for the neoclouds that buy its chips, that its AI Compute Partnership is being restructured after established neoclouds such as Nebius pushed back, and that its credit default swap spreads have roughly doubled in recent months, though its credit rating remains near the top.

One widely shared critique compares Nvidia to GE Capital rather than Enron, arguing that its customers are driven by fear of missing out and that its growth requires perpetual acceleration. That is an opinion piece, not a finding, but it names the risk. Huang himself said in April that these neoclouds would not exist without Nvidia’s support.

What I am watching

First, whether lenders accept GPUs as collateral worth financing, since that is the stated fix for the capital gap. Second, credit signals around Nvidia and the neoclouds, such as spreads and bond pricing, which often move before equities do. Third, how fast new memory supply actually arrives compared with what Huang expects.

Huang can be right about demand and investors can still lose money if the financing breaks first. Strong demand does not guarantee that every company serving it survives its debt.

October 22, 2026
Memory Chip Makers Are Locking In Customers for Years. Here Is What the Deals Actually Say
Memory has a famous habit. Prices soar, makers build too much, prices crash, and everyone suffers. A table of memory chip long term agreements from the biggest memory makers suggests the industry is trying to break that pattern.

One note on the data first. Micron’s figures are as of the third quarter of calendar 2026, while the others are as of the second quarter.

Real money is changing hands

Micron reports US$32 billion in deposits and commitments, up from US$22 billion at the second quarter, plus about US$150 billion in remaining contract value from its signed agreements. An unnamed US NAND maker reports US$59.8 billion in remaining contract value at quarter end, rising to US$91.1 billion once contracts signed after quarter end are included. It also cites US$16.8 billion in financial guarantees, but those are mostly held by or through third party institutions, so they are not the same as cash in the bank. Samsung says it has received about a quarter of its prepayments so far, with more expected. SK hynix says deposits are part of its contracts but gives no amounts.

Price floors, and sometimes ceilings

Samsung has set a minimum price floor for mainstream products, designed to reduce future investment risk. The US NAND maker uses a mix of fixed and variable pricing, where the variable part has both a floor and a ceiling, and it says gross margin stays around 80% even at trough pricing. Micron said in the second quarter that it had a guaranteed gross margin above any prior cycle peak. SK hynix describes a pricing band to cope with volatility, but says pricing will not be uniform across customers.

The ceilings matter. Customers are getting protection too, so part of the upside is capped.

How long, and how binding

Micron’s agreements run through 2031, with five years for data center and consumer customers and three years for automotive. Samsung’s are five years but renegotiated annually on a rolling basis. SK hynix averages about five years, and the US NAND maker offers up to five years with a weighted average above four.

How binding they are also differs. Micron explicitly uses take or pay with binding volume commitments. Samsung and Kioxia describe binding contract terms for both sides. SK hynix talks about long term volume commitments plus deposits. The US NAND maker sets supply commitments each year by quarter. So “locked in” means different things at different companies.

Coverage is still partial

In the second quarter, Micron’s agreements covered about 20% of DRAM shipments and about a third of NAND shipments. It expects about 35% of revenue to be covered by 2030, with a goal of 50%. Samsung puts coverage at 60 to 70% of capacity, with agreements finalized with its top five data center customers and five more in final stages. The US NAND maker expects about half of bit volume covered in its fiscal 2027 and two thirds in 2028, and Kioxia targets 50% of shipments by calendar 2028. The spot market still matters.

Where the demand comes from

Most of it is data center and AI related. SK hynix lists server DRAM, HBM and enterprise SSD, Samsung is HBM focused, and Kioxia and the US NAND maker name hyperscale and AI infrastructure customers. Micron also includes consumer and automotive.

What the table does not tell you

It says nothing about whether customers can afford to honor these deals if demand weakens, or how firmly each clause would hold up. It is a snapshot from company data, not a guarantee.

My takeaway

These agreements give memory makers more visibility and some downside protection than they had in past cycles. But the coverage is partial, the terms vary, and some of the upside is capped. I would call it a smoother ride, not the end of the cycle.

Want more on the companies behind these agreements? Here are my earlier posts:
- Micron Technology: Wall Street Keeps Raising the Bar takes a closer look at Micron, the company with the largest contract value in the table above.
- SK Hynix Is Sold Out. Robots May Be Next. covers SK hynix and its high bandwidth memory business.
And for the demand side, the AI infrastructure buildout that these contracts are meant to serve:
- NBIS Stock Analysis: Is Nebius Worth Buying in 2026?
- $IREN: Energy Meets AI Infrastructure
Disclosure: The author holds a long position in $MU and $SKHY stock at the time of writing.
October 6, 2026
Micron Technology: Wall Street Keeps Raising the Bar

Micron Technology (NASDAQ: MU)

Micron closed at $977.50 on September 17, 2026. That’s up 5.5% on the day. In fact, the stock has been on a wild run this year, and the analyst notes keep piling up in its favor.

Analysts Are Overwhelmingly Bullish

Since the last earnings report on June 24th, most major firms have either reiterated or raised their price targets on $MU. Out of the 27 ratings tracked here, only Goldman Sachs is sitting on a Hold. Meanwhile, everyone else has a Buy.

The targets are spread pretty wide. For instance, Melius Research is the most bullish at $2,200. Similarly, Cantor Fitzgerald, Barclays, and Susquehanna are all sitting at $2,000. A lot of firms bumped their numbers up fast over the summer. As an example, Barclays went from $1,175 to $2,000. Likewise, Phillip Securities jumped from $530 all the way to $1,870. Even Melius itself moved from $1,100 to $2,200 in just a few months.

The Real Story Is Memory

The bigger story behind all this is memory. To begin with, Nvidia confirmed it’s raising AI server prices by more than 15%, and the reason is simple. Server DRAM roughly doubled in price during Q1 2026. In addition, some reports show an 80 to 90% jump across DRAM, NAND, and HBM in that same stretch.

Micron sits right in the middle of this. Along with Samsung and SK Hynix, it’s one of only three companies that actually make the high bandwidth memory these AI chips need. As a result, that gives it real pricing power. Furthermore, analysts expect this shortage to stick around too. Gartner sees it lasting into 2027, while Deloitte thinks AI server DRAM prices could roughly quadruple this year.

Demand Keeps Outpacing Supply

There’s a demand side to this too. For example, Nebius, an AI cloud provider, has been benefiting from rising GPU rental rates, with H100 pricing up about 20% year to date. Overall, it’s another sign that compute demand is outpacing supply across the board, which only adds more pressure on the memory side where Micron plays.

The Takeaway

Put it all together and the setup looks straightforward. Heavy AI demand is pushing up GPU and server prices. Consequently, that’s tightening memory supply. Ultimately, that’s putting Micron in a strong position heading into the rest of the year.

Source: Tipranks

Disclosure: The author holds a long position in $MU stock at the time of writing

September 17, 2026
SK Hynix Is Sold Out. Robots May Be Next.
SK Hynix Inc. – ADR (NASDAQ: SKHY)

The world’s top AI memory supplier is riding an unprecedented upcycle and already eyeing what comes after data centers.

SK Hynix has spent the better part of three decades as a cyclical also ran to Samsung a memory maker whose fortunes rose and fell with the brutal boom bust rhythm of the DRAM market. That reputation no longer fits. Since Nvidia’s GPUs began shipping with SK Hynix’s high-bandwidth memory (HBM) stacked directly alongside the compute die, the company has become one of the more consequential suppliers in the entire AI hardware stack not a chipmaker riding the AI boom so much as one of the physical constraints on how fast that boom can grow.

The company controls a majority of the global HBM market and, by most industry counts, supplies more than 70% of Nvidia’s initial orders for HBM4, the next-generation standard now ramping toward volume production. Samsung is chasing hard, and Micron has carved out a smaller but growing share of the same market but for the moment, SK Hynix is setting the pace.

From training runs to permanent memory

For the first few years of the generative-AI boom, the bottleneck was compute raw GPU throughput for training ever larger models. That story is shifting. As AI systems move from pure training toward persistent inference, agents, and retrieval heavy workloads, the industry increasingly needs somewhere to keep data close to the processor, not just crunch it once and discard it. Every additional data center built to serve that shift needs more memory, not just more compute, and the major chipmakers Nvidia, AMD, Micron, Samsung, and SK Hynix among them have been consistent on one point: the memory shortage now underway is a structural feature of this cycle, not a temporary supply hiccup, and it’s expected to persist for the next several years.

That raises the obvious question hanging over every AI infrastructure stock: what happens when the current wave of data center construction matures and spending normalizes? SK Hynix’s own answer, echoed by much of the industry, is that data centers are only the first leg. The next is physical a coming wave of AI-enabled robotics that would extend memory and compute demand well beyond racks in a warehouse and into cars, factories, and eventually homes. Whether that materializes on the timeline bulls expect is very much an open question, but it’s the thesis increasingly used to justify capital spending plans that would have looked reckless in any previous memory cycle.

Beyond HBM: the next memory tier

SK Hynix’s HBM lead is well established. Less understood is what the company is building next. Alongside SanDisk, it has been co-developing a technology called High-Bandwidth Flash, or HBF and it’s worth being precise about what that actually is, because it’s often described loosely. HBF does not stack HBM chips on top of each other. It stacks layers of NAND flash memory using the same through silicon via packaging techniques HBM pioneered, creating a new tier that sits between fast but limited HBM and large but slow SSD storage.

The point of HBF is capacity, not raw speed: a single stack can hold many times more data than HBM at a fraction of the cost, which matters enormously for AI inference, where a model’s cached context can balloon into the hundreds of gigabytes. SK Hynix and SanDisk released the first open HBF specification in mid-2026 under the Open Compute Project, with Google and AI-chip startup Tenstorrent joining the effort a sign hyperscalers see real value in a cheaper, denser memory tier for inference workloads. Samsung appears to be pursuing a related but distinct approach of its own, stacking memory differently rather than adopting the HBF standard outright, while Micron has not yet detailed a competing product in this specific category.

“The HBM era is ending the HBF era is coming,” one SK Hynix executive told an industry conference in late 2025, previewing what the company hopes becomes the next multi-year growth leg once HBM itself matures.

Who’s actually buying this

Geographically, SK Hynix’s customer base is heavily concentrated a fact that cuts both ways. The United States, driven almost entirely by hyperscaler and AI-accelerator demand, accounts for roughly two-thirds to seventy percent of total revenue depending on the quarter, with China a distant but resurgent second on the back of mobile and NAND demand. That concentration means SK Hynix’s results are unusually sensitive to a handful of customers Nvidia chief among them, alongside at least one large, unnamed hyperscaler that has separately grown into a top-tier customer in its own right and to any shift in U.S.–China trade policy or export controls.

What management is promising

SK Hynix’s own guidance has become the thing to watch each quarter, if only because the company keeps having to raise it. Capital expenditure for 2026 has been pushed up repeatedly through the year into the low to mid $30 billion range by the most recent count, with some estimates, once packaging and infrastructure commitments are included, pushing past $35 billion. That’s a record for the company and among the largest capex programs in the memory industry’s history.
- HBM and technology leadership: Management continues to guide to more than 50% global HBM share, with over 70% of Nvidia’s initial HBM4 allocation locked in.
- Shipments and migration: Double digit sequential DRAM bit growth guided for the second half of 2026, alongside a push to get 321 layer NAND to roughly half of total NAND capacity by year end.
- Fab scaling: Plans to expand 1c nanometer DRAM output toward 170,000–200,000 wafers per month by early 2027, and a new advanced packaging facility under construction in West Lafayette, Indiana, at a cost north of $4 billion.
- Contract backlog: Multi year supply agreements now in place with roughly ten major hyperscale customers reportedly including Nvidia, AWS, and Google several of which include upfront cash deposits that lock in pricing years in advance.
What the market is paying for it

Wall Street’s enthusiasm shows up most clearly in earnings-per-share growth, which the sell side expects to keep compounding even as the revenue base gets much larger. On a per common share basis (SK Hynix’s Nasdaq listed ADRs represent one-tenth of a common share, so divide by ten for the per ADR equivalent), consensus estimates have EPS climbing sharply through 2026 before growth normalizes into 2027 and 2028.

Target prices: three ways of looking at the same stock

Because so much of the bull case rests on how long the current pricing power lasts, target-price models for SK Hynix tend to spread out unusually wide. A simple earnings multiple framework 2026 consensus EPS multiplied by a range of forward P/E multiples illustrates just how much the multiple assumption matters more than the earnings number itself:

The bear case assumes the market eventually re-applies the discount it has historically given Korean memory stocks and cyclical semiconductor names generally. The bull case assumes SK Hynix keeps re-rating toward the multiples U.S. peers like Micron have commanded during past AI-driven upcycles. Roll the same framework forward to 2027 EPS and the spread is similar in shape, just shifted higher anywhere from roughly $147 in a trough-multiple scenario to $237 if SK Hynix holds its 2026 multiple into the following year.

Every analyst currently covering SK Hynix’s Nasdaq ADR rates it a buy, with targets clustered between $200 and $320 a genuinely unusual amount of one directional consensus for a memory stock, and arguably a risk in itself. When an entire sell side community agrees this strongly, the disagreement that matters has usually moved into how far and how long the cycle runs, not whether the stock is a buy today.

Who owns it

Ownership is one area where the original draft’s figures didn’t hold up well against SK Hynix’s own disclosures, so it’s worth restating cleanly. SK Square the SK Group holding entity spun out of SK Telecom in 2021 remains the controlling shareholder with roughly 20% of shares outstanding. South Korea’s National Pension Service holds around 7–8%, and foreign institutional investors collectively led by BlackRock, Vanguard, and Singapore’s GIC own more than half the company, on the order of 54–56%. The remainder is split between domestic institutions and retail shareholders. Insider and employee ownership, by contrast, is minimal, consistent with a company of this scale.

Sources: TradingView, TipRanks, Seeking Alpha, TIKR

Disclosure: The author holds a long position in $SKHY stock at the time of writing
September 10, 2026