Nomad Semi

Why betting on Samsung’s HBM comeback is a long shot 📉 1️⃣ Bonding tech: SK Hynix’s proprietary MR-MUF offers better thermal dissipation and a higher yield rate than the TC-NCF technology used by their peers. 2️⃣ 1a nm problem unresolved: Samsung faces additional issues in its front-end DRAM process since the start of 1a nm. Lower yield rate on the front-end impacts overall HBM yield rate with multiple dies (12-Hi) stacked. They tried to redesign the base die to meet NVIDIA’s HBM3e 12-Hi qualification, but did not make changes to the 1a nm die. Switching production back to DDR5 is also challenging given their 1a nm process is uncompetitive. 3️⃣ 1c nm is not yet mature: For HBM4, the use of 1c nm process for DRAM and 4nm in-house foundry process make it hard to believe Samsung can catch up. Progress in 1c nm has not been smooth and they had to redesign 1c nm process earlier. Samsung’s 1c nm process is not yet mature, while both SK Hynix and Micron are sticking to their mature 1b nm process for HBM4. On the other hand, SK Hynix’s 1c nm has reached mass production readiness. Time to market is important in HBM as the 1st mover typically gets a big volume share through annual volume contracts lock-in. 4️⃣ Hybrid bonding wildcard: Hybrid bonding is Samsung’s best chance to catch up in the back-end packaging race. However, the higher cost of hybrid bonding has been delaying the adoption to 2028. Front-end process remains Samsung’s biggest problem. HBM qualification issue is just a symptom of it. For deep dive on HBM, check the link 👇 nomadsemi.com/p/deep-dive-on… $mu #Samsung #Hynix #HBM

@SaViewer13042 Yes

@MooreMorrisSemi Is Marvell cooked? Looks like the allocation and AWS business is going from them to AIChips post 2025?

Latest CoWoS allocation until 2027. Source: JPM $tsm $nvda $avgo

No surprise that Hynix is the 1st to complete development of HBM4 at 10Gbps

@BenBajarin @TeemuMtt3 Please read up more on TPU

@TeemuMtt3 AMD is the answer if they want to balance NVIDIA. ASICs are to specialized in architecture, require higher lift and shift in software. Third parties are running on GPUs today, and will for the long haul, because they are more general purpose (flexible) for AI workloads.

The idea of Google looking to shop TPUs is actually not a good sign. They still can't get adoption on the CSP portion of GCP. So did they overbuild and are looking to recoup those? Are they prioritizing GPUs and not TPUs internally for more workloads so they have excess inventory? Regardless, if GCP is having trouble getting software devs to lift and shift for TPU, why would Neoclouds have any more success? Lot's of puzzling questions.

@Jukanlosreve W2W HB for 12-Hi HBM. Please pray that you have no defects on any single wafer

What are you talking about? The difference between NAND hybrid bonding and HBM hybrid bonding doesn’t matter at all. The important thing is that YMTC has been researching the 4F² (4F square) structure for a long time, and has been trying to apply NAND hybrid bonding technology to HBM. This is something everyone in the industry already knows.

@godofprompt Gemini

Gemini 2.5 is one of the smartest AIs available right now. But most people don’t know how to use it properly. That’s why we created the "Gemini Mastery Guide" designed to help you unlock its full potential. → Prompt engineering breakdown → 10+ mega-prompts → Google Workspace integrations Like + comment “Gemini” and I’ll DM you the doc. (Must be following)

MS on Rubin Timeline $nvda Rubin is on track: NVIDIA's first Rubin silicon should be out from the TSMC fab this October (first tape-out was in June). Meanwhile, our checks suggest NVIDIA's Rubin engineering sampling (ES) is on track for 4Q25, although the fine-tuning of the Rubin chip design continues. The chips and their system design should be finalized next March, with the chip entering mass production in 2Q26, followed by server racks ramp in 3Q26. So it appears there will be no delay to the Rubin schedule, despite some investor concerns.

Please give Fubon’s analyst credit. He did the channel checks and is reporting data. The alternative is for you to purchase @SemiAnalysis_ Accelerator Model. To me, this reads as glass half-full / half-empty for both $NVDA and $AMD. This isn’t the first delay, and a 3 months slip doesn’t change the long-term thesis for $NVDA. Take a chill pill. The CoWoS figures still imply FY2027 revenue well above Street expectations. You should be questioning the Street instead. It isn’t particularly bullish for $AMD either. We’ve seen higher CoWoS counts before that were later revised lower. Scale-up is still a big hurdle that remains unresolved. We need less narrative and more math

@BlameQuants @BenBajarin As for the AMD bulls, this is also not the 1st time we have seen a higher CoWoS number only for it to be revised down later

Precisely, not sure why he is making such a big fuss out of it. Didn't we see a delay with Blackwell previously? Fubon's CoWoS numbers still give a revenue for Nvidia in 2026 higher than any other sellside. Why not go complain about why the street estimate is too low? Taiwan analysts are closest to the supply chain, not the West

Fubon is wrong, has a bad track record, is an uncredible source, and can't spell NVIDIA right lol.

@JShpigler AMD's CoWoS include CPU

@MooreMorrisSemi I believe you misread, AMD is the fastest CoWoS grower

1) Rubin redesign leads to more limited volume 2) TSMC CoWoS capacity to hit 130k in 2027 3) Blackwell Volume: 750k in Q1 25, 1.2mn in Q2, 1.5mn in Q3, 1.6mn in Q4 4) Broadcom is fastest growing CoWoS customer in 2026 Source: Fubon $tsm $nvda $avgo

@icnewer A broker in Asia

@MooreMorrisSemi Who is CLSA?

CLSA thinks MediaTek has a good chance to win Meta's 2nm Arke project

@SKundojjala Correlation might break down since ASML's EUV lead time is much shorter than in the last few years

In 2021-2023 cycle, a 5-6 quarter lag between ASML EUV revenue peak (& trough) with Lam& AMAT's combined non-China revenue held up well. ASML EUV revenue peaked in 3Q21 and then hit cyclical trough in 1Q22. Lam & AMAT followed, peaking later in 4Q22 and then hitting trough in 3Q23. Ignoring 2023 altogether, 2024-25 EUV peak should give an idea of next peak for Lam & AMAT. EUV is a leading signal for other equipment makers given its 12-18 month lead times.

HSBC upgrades $AMD to TP of $200 as MI350 series pricing will be USD 25k instead of USD 15k $nvda

Morgan Stanley's opinion on MTIA project $avgo $MTK