Korea

• Korea’s president told a cabinet meeting today there were signs AI had been used in successful attacks on at least seven banks, including Shinhan and Woori. Details of the AI tools used or the scale of the breaches have not been disclosed. Reuters Korea Herald

• Several cybersecurity vendors at or close to the daily limit up including:

  • AhnLab (053800) +15%

  • Genians (263860) +20% - Archetype Capital has been researching this company

  • RaonSecure (042510) +30%

Japan

• Cybersecurity also the theme of the day here, though the spate of cyber incidents in the last week don’t appear to be attributed to AI.

• Broadband Security, S&J, SecureVeil, Cyber Security Cloud and Caulis all closed up more than 20%, several of them limit-up; ZenmuTech, Terilogy, Segue, Global Security Expert and FFRI followed.

OAI apologies to Australian Select Committee for breaches of Australian government websites (The Herald)

Linked to Australian newspaper to sample some Aussie outrage. It is more pointed than international coverage but still diplomatic:

Australia’s Deputy Prime Minister Richard Marles has said OpenAI co-founder and CEO Sam Altman didn’t mention the breach of the Medicare system when the two met in early September. But Kwon told the inquiry Altman didn’t know about the violation at the time, although it was known elsewhere within the company.

A growing number of Australians are also calling for tougher rules concerning AI data centres and AI copyright protection as Prime Minister Anthony Albanese’s government works on new laws to govern the sector. Both OpenAI and Claude maker Anthropic are awaiting clearance for large data centres planned by developers in Australia where they have agreed to be the main buyer of computing power.

With thanks to the Karen Leung Foundation, we’ll be attending Sohn Tokyo on 16 October and look forward to covering that event. Please drop me a line with you are in Tokyo on 15 or 16 October.

Subscribe now

Read more

Thai readers weighed in, noting that the rain was less severe than in 2011. However, as of this weekend, the Bangkok Post is reporting that flood impact is more severe than expected, that flooding has persisted, and that more rain is coming. Even if a facility is not flooded, flooding can significantly disrupt operations.

A WDC short ended up working but for dramatically different reasons. A Nikkei reporter in Manila was covering the first shipment of HDDs from a facility in Manila, writing:

Toshiba will double its production capacity for hard disk drives (HDDs) for artificial intelligence (AI) data centers by fiscal year 2027. The company will invest approximately 60 billion yen to expand its Philippine factory.

The reporter then expressly says the comments from Toshiba appear to be intended to encourage the factory floor:

At the [first shipment ceremony], Vice President Koji Iketani… encouraged the factory floor by saying, "Expanding our nearline HDD production capacity is an important milestone toward achieving our goals. Achieving a 30% global market share is not a pipe dream."

The official release says nothing about market share ambitions, only an aim to double production on a storage capacity basis:

TIP will continue to invest in production expansion, aiming to nearly double its annual production capacity by FY2027 compared with FY2025 on a storage capacity basis.

Bernstein noted that if Toshiba does double exabyte production, assuming WDC and STX grow in accordance with their own long term plans, Toshiba share of exabytes would grow from 11.2% to 16.8%, not 30%. The firm also observed that Toshiba market share fell from 17% to 11% today after it announced 30% market share ambitions in 2022.

We roundup HDD takes from Nikkei, Citrini and others below, and cap off with a short note on ABF and Testing.

I have coffee with some Thai and Filipino business leaders in an hour. If there is a material update it will be in the chat.

Subscribe now

Read more

AT&S has been a recurring topic on this newsletter since last November and has benefitted from substrates being called out by multiple technology leaders. Substrates has arguably overtaken memory as a new favourite bottleneck among investors.

Last week AT&S announced the expansion of its partnership with Marvell, which echoes Matt Murphy calling out substrate capacity at the Citi Conference and saying that he had been visiting suppliers all week:

Adding AT&S to its “Dynamic AI” basket, Citrini cites increased manufacturing complexity, meaningful capacity expansion being years away, and CPU demand growing 25% CAGR for next four years driving ABF substrate demand.

In the CNBC video out over the weekend, AT&S touts co-development with big AI players and is cautiously optimistic about changing their perception from T2 to T1 supplier.

On that note:

Read more

In case you have not noticed, Funda has opened its chat to all subscribers, paid and free, so I would encourage you to check it out / ask questions before they close some threads for paid subscribers only.

I also opened this newsletter’s chat to everyone this week. This is a summary of the discussion - probably 8/10 on the real time calls made this week.

BE: Project Jupiter Force Majuere Notice (BBG)

I used to work on some of the most difficult energy projects in a major industrial’s portfolio. Covid was a flurry of Force Majuere notices, which we dealt with for many years after the event.

Force Majuere mechanisms in contracts typically require you to issue a notice within a certain amount of time of learning about a relevant event / trigger, otherwise you would not be able to rely on the Force Majuere mechanism at all.

Oracle’s statement below is true, and it is remarkable that they would say it out loud, because it compromises their position if they have to litigate this later. Oracle should be behaving as if there is a genuine uncontrollable event affecting the project instead of saying everything is ok.

META: Investors digest the Funda survey

Having spent last week focused on optics and CIOE, where we spent time explaining the implications for NPO and Semtech’s position in China, I had some time to discuss CPUs this week.

It became clear from discussions over the weekend that there was going to be a continued bid for CPUs, which I flagged in the chat on Monday.

This bid really started from the release of Astra and not, as many on social media victory lapping would have you believe, from Muse. This is because strong computer use capabilities bodes well for CPU demand.

I’ll pause here to note that AT&S (which we’ve been flagging since 13 August) has actually outperformed Intel and AMD in the last 30 days, and that substrates benefit from CPU demand.

Since Funda’s Muse note was published I think investor thinking has developed along these lines:

Read more

As we’ve pointed out a few times before, Malaysian semiconductor stocks have been remarkably resilient, and many did not pull back in July at all. It appears that capital is longer duration and semis / DCs/ manufacturing probably has little competition with other sectors as a destination for investment.

ViTrox is a notable example:

Earlier this year, we also flagged a small cleanroom contractor that has doubled since July, which speaks to the amount of investment happening in Malaysia.

By popular request, we’re presenting a bull/bear case for a company that is working with US photonics players like Lumentum and Coherent. If you’ve been following photonics, you’ll understand that demand is not in question - this company will get as many orders as it can reliably deliver, and so its ability to multibag arguably turns only on execution.

I think I should also caveat this by saying that I have not personally met the company or checked the numbers shared with me below, but I did think it was worth publishing because of the identity of the people who shared the idea - there is pretty serious attention on this stock.

Subscribe now

Read more

Kioxia said to consider raising $10 billion in US (BBG)

The Tokyo-based memory-storage maker is seeking to gain more liquidity in the US with the listing after it repurchased billions of dollars’ worth of shares in Japan, the people said.

Selling ADRs may also allow Kioxia to join a semiconductor-focused stock index, they said.

Citi: Continual learning in AI will reshape long-term memory demand inspite of memory de-spec and safety concerns

The firm sees a deepening DRAM shortage in 2027–28, with supply-demand ratios of -8.7% and -9.7%. It forecasts global DRAM demand rising 30% and 35% year-over-year, driven by continual learning that lifts server DDR5 and HBM demand at the same time, while supply grows only 19% and 22%. Personal and physical AI emerging from the second half of 2028 should intensify undersupply through 2028.

For NAND, Citi projects demand growth of 29% and 33% outpacing supply growth of 21% and 25%, as producers add DRAM and HBM capacity at NAND’s expense. Continual learning and Nvidia’s KV-cache offloading trend are expected to lift high-density eSSD demand by 53% and 41%, leaving NAND S/D ratios at -6.1% / -5.5% for 27E and 28E, tighter than the -0.8% Citi models for 2026.

Citi projects blended NAND ASP to grow +34%/ +15% in 3Q/4Q26E.

Citi’s memory top picks are Samsung Electronics, SK Hynix, Micron, Sandisk and Kioxia; its SPE and materials favorites are Montage, Applied Materials, Lam Research, TES, Eugene Tech and TechWing.

Trendforce responds to SA’s Long Live the Short King Piece

SemiAnalysis argues that 4hi is for data-center inference. Decode is bandwidth-bound more than capacity-bound; rack-scale systems already have huge aggregate HBM; frontier-lab ASIC teams are designing around 4hi because it is the lowest cost/token at current HBM prices.

TrendForce argues 4hi is for edge, not the DC. Real inbound interest is autonomous driving and robotics. Mainstream data-center platforms are “relatively low” probability. Near-term volumes from edge are tiny.

While SA argues that 4hi is in part motivated by the constraint on wafers, TrendForce argues that 4hi makes less sense because the scarce input is the foundry logic base die + packaging:

At equivalent wafer capacity, 4hi yields a greater number of stacks, and every stack requires its own base die built on a foundry logic process. Since HBM5 base dies will be manufactured on more advanced process nodes, the constraint shifts to base die and packaging, where capacity is already tight and would prove insufficient. As a result, 4hi HBM is unlikely to provide substantial relief to overall capacity constraints

TrendForce assesses that because advanced logic node capacity is tightly allocated and shifting the same base die capacity to 4hi results in a relatively low HBM output value, HBM suppliers will prioritize allocating resources to mainstream specifications and related customized solutions for data center computing platforms. In the near term, 4hi is likely to remain limited to samples and small-volume custom projects.

MS: Attractive entry points in select upstream bottlenecks

Out last night. MS says AI materials group is up 91% LTM but has corrected 36% since 52-week high. Not all the opportunity is priced in - scope for valuation gap to narrow with more richly valued PCB / CCL peers.

MS prefers high-end glass fabric #1, HVLP4+ copper foil #2, optical fiber #3. Tickers below, as well as an update on Ibiden.

Funda will be stepping up its coverage of upstream companies, so keep a lookout for that.

Read more

Funda will make the argument in a forthcoming note that AI development should not be slowed uniformly because release pace, training pace, and sector readiness differ across industries.

I believe Funda is already read by KOLs like Gavin and Elon, but if you agree with the views in the note, please do your part to amplify it by (can’t believe I’m saying this) liking and subscribing and reposting the content.

In the marketplace of ideas, you can do your part to “vote” for views you agree with.

The Funda platform has summarised the various “camps” that have formed over the weekend. This is free for everyone to read and repost.

Some responses from investors and folks working in AI in the chat this week:

Join our chat - we think its one of the higher signal Substack chats around. We have regular contributions from professional investors and engineers, who have weighed in on names such as Kokusai Electric, Nanya Plastics, Kingboard, Victory Giant, Komico and Yageo.

Get 25% off forever

Separately, our work this week reinforced the value of getting close to the supply chain in Asia.

I think the strength of a company’s partners (relative to a competitor’s) does explain recent surprising market share gains, such as Semtech in TIAs.

We also had comments from Marvell attributing their success to the strength and history of relationships in the supply chain:

If you just popped out of nowhere and you said, hey, I got a data center part now, and I think I need like billions of dollars of revenue like next year, I think people are going to look at you like, where were you 5 years ago? I mean you can’t get some of the substrate capacity until the end of the decade if you haven’t started. How are you going to get it if you didn’t start 5 years ago?

[O]nce we acquired Avera and then Inphi and Innovium, we got these very complex reticle-sized products that required deep relationships in the supply chain in terms of substrates and back-end capacity that wasn’t that fungible.

[Our COO] and I basically went off and struck a bunch of deals, long-term deals with all of our supply chain, not just TSMC, our sole-source vendor on the fab side, but also on the substrate side, back-end capacity, et cetera. And that worked really well. I mean that's basically what's gotten us from, call it, $4.3 billion run rate when we bought Inphi.

…I just got back from an entire week in Asia visiting my supply chain. New vendors, existing suppliers and just kind of reaffirming where we're going. And so that has really helped that combination of credibility, trust, compelling road map and then capital you put in…

…suppliers are a lot like you guys, okay? At the end of the day, they want to figure out who's the management, where is this company actually going? What end markets are they in? Are they credible? Do they have a good plan?

We celebrated our one year anniversary this week. There will be a longer reflection on how we got here, but to put things in perspective, this newsletter went from a post in September 2025 about a Singaporean diesel generator company that was an analogue of Power Solutions (a big retail favourite on X) to a position where we have some exciting things lined up for the next 12 months:

• As Matt Murphy says, the supply chain have views on where the likes of Marvell are going, and I think we’re making progress on getting closer to them so we can get really differentiated content for you.

• We’re going to push a little harder on opportunities in Singapore - there are definitely doors we can try to open here in relation to names like Soitec and our favourite Singapore pick. I’m excited about what we can achieve here - stay tuned.

  Subscribe now

  ---

I think the discussion on multiplayer and multimodal AI, and how the consumer experience should change, was lost in the noise of conference roundups about the usual semi names and the Muse launch:

Read more

"One of the areas where we have made the most progress and gained the most recognition with Samsung Electronics is our joint production and ​research on the next-generation chips we are developing," Harrison Kim, General Manager of OpenAI Korea, said at ​a press conference in Seoul on Wednesday, without elaborating.

The report does not say if this relates to Jalapeno.

Samsung: Taylor fab reportedly fully booked before 2nm production begins (Digitimes)

Kioxia: Priority is keeping prices at high levels, not substantially higher price; dismisses Hynix tie-up (BBG)

Kioxia CEO’s recent comments driving home the point that the trade is about the duration of the up-cycle now:

Ota, who took Kioxia’s helm in April, said he has instructed his sales teams not to push for substantially higher prices from data center operators, out of fear of damaging investment appetite in the AI sector. Ota didn’t rule out price hikes in future, but said Kioxia’s priority for now was on keeping chip prices at their current high levels.

CXMT / YMTC: Enough DUV?

Think this is alluding to a debate about whether YMTC has enough DUV tools and its impact on Kioxia and Sandisk.

Asia Fiber and Cable: Verizon and Corning Fiber deal (RT)

CIOE also kicking off in Shenzen, which Funda is attending.

Our newsletter readers have been following some companies that benefit from fab construction and utilisation in Korea. We flag three below:

• Rigaku - the importance of X-Ray inspection was recently called out by TSMC. Rigaku has been telling investors that demand from its DRAM customers is strong, exceeding Rigaku’s internal targets.

• X-Ray has many applications outside of HBM too - Rigaku recently launched a product aimed at batteries, semiconductor devices and MLCCs.

Read more

Some interesting comments from the Citi and GS tech conferences overnight coinciding with the launch of Muse by Meta.

Meta: Launches Muse, a personal AI agent (PR; Sources)

Mark Zuckerberg’s ambition for Muse is to “give every person in the world a very capable personal agent that can understand their goals and can just work on their behalf 24/7.”

Zuckerberg’s ambition is to build something he can hand to anyone in his family, regardless of their technical ability: “I don’t think there are going to be billions of people who are going to buy a Mac Studio and configure it and run it at their home”.

Muse launches with up to 100 million tokens a week for free, plus a virtual machine to work on your behalf.

Sandisk @ Citi TMT

On the NAND demand forecast they put out for the Analyst Day - “every time we’ve revised it recently, we’ve gone up.”

The two most sensitive variables are:

Read more

Memory / Storage: Price of Fable 5.1 cache reads dropped 75% (Ant)

Anthropic estimates ~25% lower costs for typical workloads and up to ~45% for highly agentic, context-heavy work.

Two more launches this week according to this account:

eSSD: Top five eSSD revenues double q/q (Trendforce)

TrendForce expects enterprise SSD demand to remain elevated in 3Q26 thanks to the growing adoption of generative AI agent services, steady data center infrastructure deployments by CSPs, and large-scale shipments of NVIDIA GB-series AI server racks. However, the firm continues to have concerns about “rising output from Chinese suppliers and the emergence of China’s domestic cloud market”.

Hynix / Kioxia: SK Chairman floats JV (SED)

SK Group Chairman Chey Tae-won said he is considering building a new memory chip plant in Japan, and described joint production with Kioxia, the Japanese NAND flash maker in which SK hynix (000660) holds a stake, as “one option.” He also suggested the group could lay out a concrete investment plan within this year.

Chey made the remarks in an interview with Japan’s Asahi Shimbun on the 2nd. Kioxia is a company with many strengths, he said, adding that the two sides could cooperate on joint production, research and development, and supply chain sharing. At the same time, he said SK would end its investment in Kioxia if no further partnership could be built.

OAI: Astra uses a new reasoning approach called “recurrent depth.” (Information)

OAI’s chief scientist appears to corroborate the story, saying chain of thought monitoring is “fragile and unfortunately trending in a negative direction, for reasons not contingent on architecture changes”. This pushes back on the safety concerns raised in the Information report but not the architecture change.

Discussion on X about whether looped transformer architectures is bearish for HBM:

Subscribe now

Read more

The author of our original Kingboard post was gifted the opportunity to size up his position by 40%, and is now up 15%:

That sort of dislocation is quite encouraging in my view, as it suggests hunting season is back on. His note on earnings is below.

Subscribe now

Read more

Still, there have been bright spots upstream and in optics in Japan.

ABF Substrate

We flagged two weeks ago that ABF Substrate was called out by TSMC at OCP Asia.

Meritz Securities in Korea out with a note on Semco today:

At the recent OCP APAC Summit, TSMC noted that the ABF substrate shortage — driven by larger CoWoS package area and expanding AI demand — is emerging as a key bottleneck in advanced packaging. As CoWoS-based 3DFabric moves toward ever-larger package areas and greater interconnect complexity, the yield impact of small defects and material variation grows accordingly. End customers are also broadening multi-sourcing of ABF substrates to secure supply stability, but differences in substrate characteristics between suppliers have, if anything, increased the burden of yield management and process optimisation on CoWoS. Ultimately, simply diversifying the supplier base will not resolve the ABF shortage, so we expect the value — and the supply-chain bargaining power and pricing leverage — of suppliers capable of stably delivering high-specification ABF substrates to strengthen further.

The die/package area of next-generation AI accelerators — including NVIDIA’s Feynman, slated for launch in 2028 — is expected to expand roughly 3–5x versus 2026. On a simple calculation, this means that even if AI system demand comes in at only half of current expectations, ABF substrate demand on an area basis could still grow 1.5–2x versus 2026.

So why did ABF names trade down after OCP Asia and rebound today?

Subscribe now

Read more

I’ll be in New York from 7 to 14 September to meet the Funda team there and some of you. Please reach out if you’d like to meet.

We have been more cautious about the market in recent weeks, but as we note below, there are significant bright spots:

• Nand pricing is rebounding.

• The Koreans are returning capital.

• Alibaba made numerous helpful comments on ROI and AI investment payback.

Finally, Funda is working on a study of enterprise adoption of AI. This is going to be important as we head into the tail end of this year.

NAND

BofA reporting that NAND wafer prices “rebounded sharply” in early August.

Hearing this being attributed to Vera Rubin mass production ramping up.

Trendforce post on X on the same new has a bearish tilt, saying “overall transaction momentum [in consumer] continues to wither”:

The spot market exhibited a visible vertical consolidation recently still attributed to the overall sluggish consumer demand, and overall transaction momentum continues to wither without sizable improvement to the prospect of the end sector, despite some of the suppliers willing to compromise on prices. Spot prices of 512Gb TLC wafers rose by 0.39% this week (Aug.17), arriving at US$21.208.

An earlier post from 30 July describes Agentic AI as a force that could “significantly increase demand for high-speed SSDs, helping absorb incremental supply” and bring it into “equilibrium”.

Trendforce does not appear to treat AI driven demand as the base case.

Hynix: MS says on track to complete buyback in 37 days

Hynix registered for 650k shrs buyback (12% ADV) today compared to 650k (100% filled) yesterday or 2.7% of the total done i.e. on track to complete the full size 24m shares in 37 days (by Oct 15) at this pace. Separately, US ADR +4.4% (premium 33% or flat DoD).

Samsung: Board meeting to approve shareholder return program expected today (Nikkei)

South Korean semiconductor giant Samsung Electronics is expected to unveil a shareholder return program worth a total of 100 trillion won ($72 billion), Nikkei has learned.

According to sources, Samsung is expected to announce its plan following a board meeting on Friday. The program would mark the largest-ever shareholder return in South Korea.

Alibaba: Comments on ROI and Payback

Some helpful comments from Alibaba for the AI trade:

Read more

Dr He confirming what we’ve flagged for several months now - memory and ABF are the two most severe shortages in semis:

One of the immediate challenges I see right now, and this is true for the whole industry. In the last couple of years, and for the next few years, the whole industry is facing tremendous shortage, not only on memory, but also on something called ABF substrate. Your sourcing team is knowingly taking risk of multiple sourcing because of the capacity shortage.

Ibiden is a key beneficiary. AT&S is a relatively undervalued player who took more space in Kulim last week to support a growing workforce. AT&S is one of our favourite ideas.

Singapore raises GDP growth forecast to 4.5 to 5.5% (PR)

Ministry of Trade and Industry:

[A]cceleration in global AI-related capital expenditure should boost growth in the electronics and precision engineering clusters of the manufacturing sector. Notably, demand for AI-related semiconductors (e.g., networking and memory chips) from the data centre end-market is expected to remain strong amidst the rapid rollout of agentic AI across industries, while capacity expansions by semiconductor firms will raise demand for semiconductor equipment

In the same week:

• AEM, AEHR’s bigger cousin, reported net profit up 876% y/y. AEM has been getting more orders from AMD and Intel in past 2 months, and is likely struggling to fulfil AMD’s demand.

• Metasurface reported its order book grew 2.6x.

YMTC third in shipments but fifth in shipment value (Counterpoint; Nikkei)

As we reposted in early August, only 12% of 73 investors surveyed by MS thought YMTC was the most important factor for Kioxia going forward.

That is not to say YMTC is not important, its just important to have a measured take on their impact on the eSSD market.

Counterpoint:

The global NAND market in Q2 2026 was defined by the migration of AI workloads from training to inference, which pushed enterprise SSDs to 48% of total bits shipped and left consumer supply short. The resulting supply tightness pushed industry revenue to record levels, rising fivefold from Q2 2025.

Nikkei:

YMTC is a semiconductor manufacturer supported by the Chinese government, and is expanding its sales of consumer electronics such as smartphones and personal computers. The company’s shipments increased by 22% compared to the same period last year.

However, in terms of shipment value share for the April-June quarter, YMTC ranked 5th, falling behind Kioxia. This is because Kioxia and others sell advanced products for artificial intelligence (AI) servers, which have high unit prices and high profitability. YMTC has not been able to fully penetrate the server market and is content with 5th place in terms of shipment value.

Counterpoint:

The Chinese vendor plans to shift its mix further towards eSSDs in the second half of the year to cement its third-place position globally, supported by growing avenues for capital support. The same dynamic now frames the industry: with servers set to take most NAND bits, profitability through 2027 will be decided less by who ships the most bits, but who ships the right mix.

JPM 9 August memory note:

On the NAND side, our research indicates YMTC’s bit/wafer is largely on par with the leading memory makers indicating limited technology gap. We expect YMTC’s capacity/bit share to account for ~16% of the global supply by 2028E. This puts the risk of oversupply more in NAND vs. DRAM, on a relative basis. We expect the leading NAND makers to focus on manfuacturing high-end NAND (SLD based NAND or HBF) to widen the gap. Given the existence of US regulations, we believe YMTC’s NAND will likely be for local uses. Recent news stating that YMTC is considering to manufacture DRAM (link) is a positive for the overall NAND S-D, in our view.

What local uses? Apart from consumer electronics, China has its own DC buildout. Given that this newsletter started with coverage of the beneficiaries of the DC boom in Johor, it is striking to now see Mongolia dwarf that market.

Inner Mongolia: 3x the DC pipeline of Johor

UBS out with a note this week reporting that Ulanqab, Mongolia has received 12.5GW capacity commitment from data center operators and internet/AI companies as of Jun 2026, implying 10x+ capacity expansion from its c.1.2GW live capacity as of 2025.

This is three times the total upcoming data center capacity in Johor, Malaysia.

UBS sees GW-scale projects in Ulanqab underway from 3 of the top 4 data center operators in China, namely GDS, VNET and Chindata.

UBS also cites projects from Envision, ZDATA, Centrin, and self-built campuses Kuaishou and UCloud, ByteDance, Z.AI and other leading internet/AI companies

Apple: Chinese suppliers might help with supply but not necessarily pricing

On its latest earnings call, throwing cold water on the idea that CXMT and YMTC will price too aggressively.

In terms of the sources of supply, primarily the DRAM market has 3 suppliers. And obviously, if there were more suppliers, that would be good, and it would help us on the supply side and perhaps the pricing side. It’s unclear on the pricing side, but it could help on the supply side. And so we’re evaluating all options.

Temasek denies new interest in Samsung and Hynix, already has position (BT)

Business Times:

Temasek said in response to queries from Bloomberg News that it did not seek advice from the Korean government on the timing of investments in either SK Hynix or Samsung, and that it first invested in both companies more than two years ago.

Nikkei:

In the near term, Temasek noted “compelling opportunities” in AI, identifying five focus areas that it plans to dive deeper into: energy and data centers, semiconductors, cloud services, foundation models, and related software and applications.

Temasek has already taken bets in U.S. powerhouses that are at the forefront of the tech trend, such as ChatGPT creator OpenAI and Anthropic, which pioneered the Claude platform, in addition to semiconductor player Nvidia.

In South Korea, Temasek has positions in memory chip giants SK Hynix and Samsung Electronics, both of which have caused the country’s stock market to surge this year but have also been drivers of volatility as investor sentiments made short-term swings.

Intel: LBT podcast

'I used to [not] invest in memory because it’s kind of commodity business, right? But now [it has] become different. There’s a lot of new technology come out. So we are kind of looking… one of my pet projects is looking [at] some of the memory new architecture.

And I think you just saw the news — I hire my good friend Seok-Hee Lee. He used to run SK Hynix, right? So you kind of know something that I’m thinking about. We’re not ready to unfold it.'

Discounted exposure to SiTime

Given the (short-lived) success of the Kingboard Holdings idea, here is another hold co idea for value investors.

SiTime is the leader in micro-electro-mechanical system (MEMS) oscillators which have been replacing quartz oscillators in a variety of applications such as computing, consumer, networking, communications, automotive and industrial systems.

MEMS win on durability, shock resistance, and rapid configurability, while quartz crystals generally lead in pure phase noise, low jitter, and ultra-low power consumption. It’s a relatively small market which has been growing rapidly thanks to AI and datacenters, with the serviceable available market (SAM) quadrupling between 2019 and 2027, a CAGR of 27%.

Read more

In July, Sandisk and Kioxia drew down dramatically, more than six months earlier than when NAND prices are widely thought to peak (~H2 2027). This was largely attributed to some investors being forced to sell significant holdings, as well as the “second derivative”, the rate of change of NAND price increase, slowing down.

Since then, Jukan received a bit of heat for this tweet which was viewed 3.6 million times. Those angry about Jukan’s views fail to appreciate that being bullish about memory market fundamentals does necessarily mean being bullish about the price action in the short or medium term.

This is why we can, without contradiction, be bullish about s/d dynamics while also have reservations about pressure on the stock because of a turnover in the shareholder base and because it will take some time for people to persuaded that “this time is different”.

In this note, we summarise one of Dr Xiaoyu Ma’s presentations at the Future of Memory and Storage Convention 2026.

Dr Xiaoyu Ma is a Senior Software Engineer at Deepmind. He created the concept of high bandwidth flash with David Patterson in 2023, and has published numerous important papers.

Dr Ma explains the demand vectors driving the LLM inference / memory crisis. We also flag some honest questions about NAND arising from his arguments.

Subscribe now

First, Dr Ma’s explanation of the basic flow of LLM inference.

There are two stages, prefill and decode. Prefill is the stage where you process all the input tokens (the basic units of text or data) all at once.

This can be hundreds, thousands, or even up to millions of tokens all at once. Prefill is inherently parallel, so it’s compute-bound and runs great on TPUs and GPUs.

The second decode stage, where we generate tokens step by step, one token at a time, is problematic. This sequential step is called autoregressive, and that's the fundamental reason why the decode phase is memory-hungry.

There’s a data structure called KV cache that connects the two stages. The size of the KV cache is proportional to the input sequence length plus the output sequence length. While the auto-regressive decode stage creates a memory bandwidth challenge, the KV cache creates both a bandwidth and capacity challenge.

These challenges are made a lot of more difficult by the six trends below.

First, Mixture of Experts (MoE). Transformers began eight years ago as dense architectures; the industry has since shifted to sparse MoE, moving from low tens of experts to hundreds or even thousands in today’s state-of-the-art models.

The consequence is a blow-up in the size of model weights.

Editor’s note: Model weights currently sit in HBM, so this point is not relevant to SSDs. However, if HBF is successfully commercialised there is a good argument for storing model weights on HBF. HBF could tighten supply for NAND in the same way that HBM has tightened supply for DRAM. We’ll park this argument for now until we have better visibility on HBF - keep a lookout for the Sandisk investor day on 13 August.

Second, multi-modality.

Models have moved from text generation to audio and video generation, and per-token storage scales accordingly: a text token is a couple of bytes, an audio token roughly a thousand bytes (~1,000x text), and video tokens hundreds of bytes up to megabytes — another hundred-to-thousand-fold increase. More bytes per token means a proportionally larger KV cache for the same sequence length.

Editors note: This is indeed a well understood bull case for NAND demand. We are also in the very early innings of this playing out. See tweet below.

Third, reasoning.

Read more

It was quieter than usual here last week - my travel and gearing up for the Future of Memory and Storage conference coincided with a very unusual week in the markets. I think the further drawdown was anticipated, as evidenced by the numerous comments in the chat about how we needed to go lower, though the manner and time of the slide ending was not.

Looking ahead, there are many hot topics that will benefit from a discussion with the big memory players and other memory analysts next week.

We will also get to meet some smaller names like Macronix and Penguin Solutions.

There will obviously be some notes circulated next week but I am also interested in how the Funda platform will draw on everything we learn at these conferences going forward, whether you specifically query it about the conference or not.

SK Hynix / Samsung Electronics: -7/8%

Nvidia has been making Rubin Ultra less powerful than promised - when first shown off in 2025, it was supposed to have a big 4-chip package with 1TB of memory.

That's been cut back step by step, and following two weeks of rumours, it appears to ends up with just 192GB — actually less memory than the standard Rubin it is meant to replace, though the memory runs slightly faster, and 576 GPUs can be linked together instead of 72.

I have been hearing debate about whether this is demand destruction due to price increases or supply constraints, and whether HBM despec causes DRAM price to dip because more DRAM wafers are freed up.

It appears to have at least spooked the market. SemiAnalysis also has a note out highlighting that spend is being reallocated from HBM to interconnect, to connect the 576 GPUs.

Kioxia: MS survey of 73 investors

Interesting results:

Read more

These comments were apparently made a few times over the course of the last two weeks. The note summarising the comments is reproduced below:

Shawn Kim bearish take on memory

NAND

-Module makers - inventory 13 weeks, covid peak was 15 weeks

“Demand evaporated”

- Rate of change of price hikes slowing to 5% qoq in Q4 - less of a need to pull forward, opportunity cost of pulling in demand much lower

- CSPs say they have healthy inventory - one of the CSP Tencent said covered for 90% - NAND spot price dropping last 2m - Think demand and supply converging

-CXMT supply ramping up quickly

- DRAM - if NAND goes down, sell DRAM also

- HBM - capping at +40% (say based on Trend force)

- Joe Moore checks why different? - Joe checks with US CSPs, US too bullish on AI

Funda argues that Mr Kim draws from information in the Chinese market, but it is the US NAND market that is driving pricing.

Mr Kim’s view also runs counter to the sentiment at a small industry event I attended yesterday, where I spoke with folks in South East Asia who deeply involved in standing up data centers as quickly as they can:

Read more

We’re committed to presenting a balanced view, even if it comes at the expense of clicks. As we entered the second half of the year, my conversations with experts and investors surfaced questions that were indicative of a more cautious posture. In particular, there was a growing concern that frontier labs may not be as good a business as advertised, coupled with uncertainty about the implications for the AI buildout.

We then saw more model competition in the form of Grok 4.5, a surge of investor interest in hyperscaler capex plans, followed by concerns about memory costs being too high.

Funda quickly addressed the Meta Cloud concerns, but it seemed to dominate the conversations with clients for an entire week.

Funda’s view on Grok catching up is that it is bullish for compute, and the firm continues to argue that fundamentals are unchanged.

I have had the benefit on sitting in on some calls and reviewing transcripts of others, and as recently as this week we have an expert reiterating memory tightness out to 2028, coupled with an expectation that any retreat in NAND pricing thereafter would be modest. We’re looking forward to making this repository of information more powerful through the new search function. In theory, you should be able to ask any question of the Funda platform and get a view informed by Funda’s proprietary data.

TSMC: Committing 100b to Arizona and raising capex guidance

TSMC - the most prudent and well informed market participant - raised capex guidance.

The reaction to relatively good earnings from TSMC and ASML suggests, and this is my personal view, that we need to go lower and that the next checkpoint will be just before Google’s earnings next week.

When asked to break down capex, CC Wei called out testers:

We try very hard to make sure that our CapEx number is correct, but with the flexibility between the front-end and the back-end. Sometimes we have a bottleneck. So we put more money to buy the bottleneck tools.

For long term, I mean, that's the back end, is about 10 to 20%… actually, I'm very honest to tell you that as time goes by, some of the customers' product need more tester. You cannot believe that. I mean, so the tester in shortage, so we have to put more CapEx in the tester or in the packaging or in other areas.

…

the most important reason [for capex guidance increase] is because of the demand continued to increase, and we feel the pressure from the customer to drive TSMC, not drive actually, to cooperate with TSMC for the capacity increase. That's one of the major reasons. The second reason is inflation. Now we buy the tools with inflation price, okay?

Xi promises AI for all (BBG)

China continues to paint itself as the mirror image of the United States, capitalizing on the recent Fable export controls.

The US and China appear to be running two different AI races, and the uncomfortable question this raises is whether the US should continue inflating its training costs to train next generation models when good enough models can evidently be trained under the constraints faced by the Chinese.

Google Gemini Launch Delayed (BBG)

Some dirty laundry being aired by 10 current and former employees in this article.

…encouraging leadership of every department to move in the same direction is like trying to boil an ocean, one ex-employee said. When mandates shift or efforts end up duplicated in multiple departments, it gets even more difficult to maintain a cohesive strategy, current and former employees said. It’s also a challenge for any one offering to get the resources it would need to succeed, and to gain traction in the market, they said.

Montage Technology positive pre (PR)

In the first half of 2026, the Company expects to achieve a substantial increase in operating results, primarily benefit from the AI industry trend and strong market demand. On the one hand, as the penetration of DDR5 products rose with continuous iterations of their sub-generation, shipments of the Company’s DDR5 RCD chips have increased significantly, with the portion of shipments of Gen3 RCD chips and Gen4 RCD chips further rising. On the other hand, revenue from new interconnect products including MRCD/MDB, PCIe Retimer, CKD and CXLMXC chips has grown significantly. In the first half of 2026, revenue from interconnect chips of the Company reached approximately RMB3,111 million, representing an increase of around 26.4% period-to-period, in particular the revenue from interconnect chips in the second quarter reached approximately RMB1,694 million, representing an increase of around 28.2% period-to-period and of around 19.5% quarter-on-quarter.

Highlights from the chat this week

Tsmc/ASML great earnings, terrible reaction. All of the data is showing that this factor unwind is historical about -30% (varies by broker) in 17 trading days. Typically, this drawdown is -22% over 33 days with the worst being a -40 in 2021.

I think what caught ppl off guard was the lack of a fundamental reason for the quick drawdown

Lack of long only buying exacerbating the unwind:

Read more

Despite those [Meta Cloud] efforts, Meta is still hungry for even more computing power, the spokesperson said. It is still moving forward with plans for expensive new data centers and recently inked major computing deals with CoreWeave Inc., Alphabet Inc.’s Google and Oracle Corp., among others.

In fact, just days before the original Bloomberg article on Meta Cloud, Meta leadership has been making under-the-radar comments about the need for capacity and perceived ROI.

Read more

In news that appears positive for sentiment on Anthropic and OAI:

Chinese authorities have held meetings with top tech firms over the past month ​about potentially restricting overseas access to China’s most advanced AI models, including those yet to be released, three people familiar …

Read more