Hello world, it’s Monday, November 2nd.
Big news today is Marvell + Lumentum merger deal.
CXMT is posting 613% year-on-year revenue growth in Q3 and swinging to profit while also validating its 4F² DRAM architecture. Elsewhere, Sam Altman flew to Seoul to lock in HBM supply for OpenAI, Nvidia is set to become Naver’s third-largest shareholder, Kioxia posted a near 15-fold profit surge but weak guidance spooked investors, and Samsung Electro-Mechanics is flagging a high-capacity MLCC shortage peak sometime in the first half of 2027.
Let’s get into it. — Austin & Vik
Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player!
Presented by Crusoe
Crusoe builds the full stack for AI, from power and data centers to the cloud. It can also help you tune a small model on your own data for just a few bucks.
For repeatable work like reports, an open model trained on your terminology and formats can match a frontier model, and it’s cheaper and faster to run.
Crusoe’s Serverless Fine-Tuning handles the training: upload your dataset, pick an open model like Llama, DeepSeek, or GLM, and pay per million tokens used, not per GPU hour.
About $2 buys roughly 5M training tokens on smaller models.
Deploy on Crusoe Cloud in one click or download the weights and deploy on your own.
Marvell to buy Lumentum for $115B in stock
Marvell will acquire Lumentum at 4.500 Marvell shares per Lumentum share. That values Lumentum at about $1,132 per share, a 23.9% premium to the October 28 close. Lumentum holders will own about 33% of the combined company.
Marvell gets Lumentum’s InP lasers, transceivers, optical circuit switches and fabs to pair with its DSPs, switching and custom silicon. It expects more than $1B in cost synergies and EPS accretion in year one. Close is targeted for H1 2028.
Marvell fell more than 13% on October 29 when the talks leaked. Nvidia invested $2B in Lumentum in March.
Vik: This is really big news in the AI connectivity world. Austin Lyons and I discussed this whole thing with experts from Cignal AI on Semi Doped podcast. See the full episode below. I wrote a Substack post on it as well.
CXMT Posts Record Quarter and Validates 4F² DRAM Architecture
Chang Xin Memory Technologies reported Q3 revenue up 613% year-on-year, swinging the company to profit and pushing its nine-month revenue gain above 700%, according to Global Times. CXMT simultaneously announced that its sixth-generation DRAM technology platform, built on a 4F² cell structure, has completed engineering sample validation, with first samples expected to ship before year-end. The 4F² architecture departs from the 6F² layout used by Samsung and SK hynix, compressing the cell footprint and enabling higher density per wafer. Digitimes reports that DDR5 momentum is building inside CXMT’s 2027 mass-production roadmap, as the company pursues a distinct process path rather than simply matching rivals on output volume.
Vik: 4F² bit cell in DRAM will increase DRAM scale and is the theoretical density achievable in DRAM memory layout. It’s like the holy grail of DRAM cell structure but is very difficult to manufacture which is why it has been elusive for so long. Looks like CXMT has something working. Interested to check it out.
OpenAI’s Altman Flies to Seoul to Lock In HBM and AI Memory Supply
Sam Altman traveled to South Korea this week for back-to-back meetings aimed squarely at securing the memory chips OpenAI needs to keep scaling its models. He sat down with roughly 50 Samsung presidents to discuss HBM and AI chip collaboration, a gathering unusual enough in size that Altman’s parting gift, a Codex keyboard for each executive, drew its own press coverage. SK Group Chairman Chey Tae-won also held separate talks with Altman, according to Korean financial media, covering AI infrastructure and memory supply. Bloomberg reported the visits as OpenAI moving deeper into semiconductor procurement as a chip shortage tightens, with HBM availability now a direct constraint on compute scaling.
Vik: Sam Altman said on the Cheeky Pint podcast that people are wasting RAM by leaving it unused in their PCs. This has led to an explosion of memes on the internet. It’s too funny. I want to enjoy some memes after a grim week in the markets last week.
Here’s how to trap RAM Altman.
Nvidia to Take Equity Stake in Naver, Surpassing Founder’s Position
Nvidia is set to become Naver’s third-largest shareholder through a capital increase, overtaking founder and chairman Lee Hae-jin, who drops to fourth place. According to a regulatory filing with South Korea’s Financial Supervisory Service dated October 30, Naver revised the payment date for the transaction to next month. The arrangement makes Nvidia not merely a hardware supplier to Naver but an equity partner in one of Asia’s largest AI platform companies, a structure that puts a GPU vendor directly inside the cap table of a sovereign-scale internet operator. Lee Hae-jin, who built Naver from a search portal into a technology conglomerate, cedes his ranking to the chipmaker as a result.
Austin: Nvidia’s equity is the small piece here. Brookfield is arranging the other $9B for the GPUs and the Sejong data center, so Naver doesn’t carry all that capital up front. Same story as the Nscale credit guarantee talk last week... access to capital is the hard part for buyers, and Nvidia keeps helping solve it.
Kioxia Posts Record Profit but Weak Guidance Sends Shares Tumbling
Kioxia’s net profit is surging roughly 15-fold to $8.7 billion this quarter, driven by AI data-center storage demand that lifted its operating margin to a record 80.4%. The Japanese flash-memory maker also announced a 90-yen dividend payout and an Iwate plant merger, moves that under rosier conditions might have impressed. They didn’t. Forward guidance came in below analyst estimates, and Tokyo shares fell between 11% and 16% in the sessions following the report, with peer SanDisk absorbing a roughly 12% decline in sympathy. The market’s read is that NAND price momentum, spectacular as it was, is leveling off, and a crowded supplier field will make the next leg harder to climb.
Vik: NAND demand tied to AI has always been difficult to size in my opinion. DeepSeek for example uses NAND at its core. Storage demands are more vague.
Samsung Electro-Mechanics CEO Dates MLCC Shortage Peak to First Half 2027
Samsung Electro-Mechanics has signed an 880 billion won MLCC supply deal for AI servers, pushing its cumulative AI server component contracts well past 5 trillion won. The company’s CEO has told investors the shortage in high-capacity MLCCs will peak in the first half of 2027, putting a specific date on what the broader industry had been treating as a vague bottleneck concern. That shortage call arrives as order intake continues to swell: the 880 billion won contract alone translates to roughly $659 million, and total AI server commitments have now cleared the equivalent of $3.7 billion. Capacity constraints in MLCC production are now consequential enough to affect AI server build schedules at scale, with supply unable to keep pace through at least mid-2027.
Sector Watch
Compute & Edge Silicon
MediaTek beats Q3 guidance with 18% sales growth and ships its first AI cloud chip, with the CEO targeting 20% of the global AI chip market. (Telecompaper)
Nvidia updates Rubin GPU specifications as first production hardware begins shipping to customers. (ServeTheHome)
Foundry & Packaging
Ibiden plans expanded AI packaging substrate capacity for Intel customers starting from 2029, adding to its ongoing substrate investment cycle. (digitimes)
Shinko Electric advances plans for a dedicated AI packaging hub at its Mobara facility to meet rising advanced substrate demand. (Shinko Electric)
Power & Energy Infrastructure
Schneider Electric invests MX$3.6 billion in a new Queretaro manufacturing facility and signs a collaboration with NUS on tropical data center innovation. (PR Newswire)
Lite-On announces $320 million additional investment into two Vietnam subsidiaries to expand AI power-supply manufacturing capacity ahead of 2027 CSP demand. (digitimes)
Data Centers
Nebius leases 50 MW of AI data center capacity from AIB under a 12-year agreement to expand its GPU cloud footprint. (ET Datacenters)
Oracle plans to deploy an Arm-based AGI CPU on OCI in 2027 and shifts to per-task AI pricing as annual CapEx approaches $95 billion. (Engineering.com)
Microsoft rolls out its Azure Maia 200 AI accelerator across data centers in the USA, Europe and Australia. (Technology Record)
Optics & Networking
Nokia runs AI-RAN trials with Marvell Technology at VMO2, demonstrating the architecture requires no Nvidia silicon. (Light Reading)
Siemens EDA extends its Nvidia collaboration to deliver automated 3D IC electrothermal design workflows for advanced packaging. (EE Journal)
Policy & Trade
Anthropic plans to sell $3.5 billion of IPO shares to Japanese investors, with its S-1 disclosing compute costs have fallen from 251% of revenue to 48% over recent quarters. (Reuters)
Tesla and Samsung jointly oppose proposed US tariffs on Vietnam as a Commerce Department probe deadline approaches. (Reuters)
AI Compute Finance & Hyperscalers
OpenAI CEO Sam Altman signs energy and investment deals in the UAE with ADNOC and G42, extending OpenAI’s sovereign compute partnership network into the Gulf. (Bloomberg)
Meta has its S&P rating outlook cut to negative as 2027 AI CapEx projections top $225 billion, raising cash-flow concerns. (PPC Land)
Amazon taps an $80 billion supplier loan facility to help fund its AI infrastructure buildout as Q3 CapEx surges beyond investor expectations. (Bloomberg.com)
Memory & Storage
Hana Micron secures a $37.5 million loan from Korea’s National Growth Fund to expand memory packaging capacity and close the gap with Taiwanese rivals. (Seoul Economic Daily)
Macronix expands its mature flash portfolio targeting AI edge and inference applications as high-bandwidth memory demand reshapes the broader market.
Phison wins two Micron supplier awards for the third consecutive year, reinforcing its position as a key NAND controller and SSD solution partner. (Phison)
Tell us how to make Semi Doped better! Give us your feedback! ❤️