SemiLover

[SemiLover @iammoai1]

💡 Don’t Doubt Bill McDer“Moat”

$NOW $CRM $SAP $ORCL

1. The surge in CPU demand driven by agentic workloads may — as Jensen Huang’s logic implies — signal rising software consumption more than inference itself.

[JabroniCoin.USD @TheBenSchmark]

JPMorgan makes asic point and thinks $AVGO bears will be COT offsides.

There were drops. Now ramps. Pat $580

[SemiVision 🇹🇼 👁️👁️ @semivision_tw]

An LPX rack consists of 256 LPUs, collectively offering 128 GB of SRAM and an aggregate bandwidth of 40 PB/s. The high bandwidth of the LPU primarily comes from massive parallelism across numerous memory banks and on-chip interconnects, enabling extremely low read/write latency. This allows token generation throughput to scale from roughly 100 tokens per second to over 1,500 tokens per second.

https://t.co/LVWvTp3Pfm

[MacroValue @pradeeepk]

Imo $CRM is buying Fin for its inhouse AI model Apex which is custom built for CRM & performs better vs generic LLMs for CRM usecase

Benioff has said they will spend 300mio on Anthropic this year

if Apex can replace Anthropic in Agentforce, the price they are paying actually looks cheap

[Gaetano @crux_capital_]

The Optics sector momentum won't last forever...

There will be winners and losers

Those that outlast the rest and those that get crushed

So my goal is to find the companies that I believe are the most durable that I want to hold in size long term

I did 2 different reports on this.

The one for $COHR is here : https://t.co/LUOe1rMx3L

The one for $GLW is here :
https://t.co/1b1DyRc8Tp

You can search on X for all the free content I put out on these names. There is a lot

But if you want it all compiled in one place, with all my thoughts laid out, youll want to read these articles!

[Max Weinbach @mweinbach]

Specs have dual Snapdragon processors!

Wonder if this will become more common on glasses, dual chip. Makes a ton of sense.

[Sean @sean_________]

Mizuho semi supply chain tracker: no delays 800VDC & strong demand $NVDA CPO switch:
"In the HVDC/800VDC investment case, the deployment schedule should be an average of 12-month lead-time from the initial system (like power rack) shipment by vendors, in our view"
"Lumentum ( $LITE ) highlighted a rising adoption rate of InP-based NPO from its clients and we also expect stronger Google TPU demand to boost $AVGO VCSEL-based NPO sales.
"micro-LED (uLED) based optical solutions that are developed and promoted by other chipmakers, such as Credo ( $CRDO ) in ALC (LED cables) and Avicena (private), to capture future scale-up (XPU to XPU) and scale-in (XPU to HBM) opportunities."

[Jordy Beuving @BeuvingJordy]

Token estimates Goldman Sachs (May 2026)

And $NVDA shares still trading sub index level 😄

[SemiVision 🇹🇼 👁️👁️ @semivision_tw]

The future of Al infrastructure is being built at the intersection of compute, networking, and photonics.
As the industry transitions from traditional pluggable optics to Co-Packaged Optics (CPO), new manufacturing and test methodologies are required to support the next generation of high-bandwidth, energy-efficient optical interconnects.

ficonTEC is proud to be collaborating with NVIDIA to advance scalable manufacturing and test solutions for next-generation CPO and Al-driven optical interconnect infrastructures. By combining NVIDIA's leadership in Al infrastructure and optical system architectures with ficonTEC's 25 years of expertise in high-precision photonics automation, assembly, alignment, and test, we are helping accelerate the industrialization of silicon photonics technologies for hyperscale Al deployments.

ficonTEC’s advanced wafer-level and die-level insertion and test platforms have been developed to address the increasing complexity, throughput, and scalability requirements of CPO manufacturing environments. These developments include some of the industry's earliest integrated automation approaches supporting both double-sided wafer-level and singulated die-level photonic test insertion for CPO applications.

[TBPN @tbpn]

What happens to Chinese AI once American companies stop releasing frontier models internationally?

@gavinsbaker says that China's made a "terrible mistake" by relying on its own chips and distilling American models:

"It feels like the administration has been willing to let them buy H200s or B30s. And I think [the Chinese] have this crazy belief that, 'Oh, our own internal chips are good enough.' They're not."

"And what's making them think that is that Chinese labs are very, very good at distillation...we've all seen those iPhone farms in China. You know they've got the same thing for distillation. They've got 100,000 endpoints distilling these models across every API available."

"They've gotten really, really good at that. But all that goes away if people stop releasing these models at the frontier. And I think Mythos is a sign of things to come there."

[Jukan @jukan05]

Supply security over cost: Google leads CSP charge to diversify InP substrate sourcing

China has recently eased controls on some indium phosphide (InP) substrates, relieving a bottleneck in optical communications capacity for the second half of the year. But supply chain players say the long-term priority is still to expand substrate capacity from non-China sources, with supply security for the AI industry outweighing price.

Representatives from Google and other cloud service providers (CSPs) have personally visited overseas substrate makers to signal that demand for their products is real and substantial.

Given the technological rivalry between the US and China, Beijing began imposing export controls on InP substrates in February 2025, a move the industry has viewed as retaliation for US restrictions on advanced chip exports.

Shortage hampers high-speed transmission
As AI infrastructure drives greater demand for high-speed transmission, the shortage of compound semiconductor substrates caused by China's export controls has become a major bottleneck for the optical communications industry.

Sources familiar with the compound semiconductor supply chain say the InP substrate market has long been dominated by AXT, Japan's Sumitomo, and JX Advanced Metals. AXT and Sumitomo each hold about 40% of the market, together controlling roughly 80% of global InP substrate capacity.

Recently, the industry has raised concerns that the shipment schedule for co-packaged optics (CPO) could be delayed. Meanwhile, industry insiders say that CSPs have already raised their inter-rack transmission requirements from 400G to 800G. Even though in-rack compute nodes still rely mainly on copper wiring for now, CPO scale-out will become the mainstream trend, pushing optical communications demand to peak in 2027 and 2028.

The search for non-China substrates
There has been some good news on China's InP substrate export controls. After a batch of substrates was released in August 2025, the first 2026 batch of InP substrates began shipping at the end of May, which could support second-half operations for Taiwan-based compound semiconductor suppliers including Visual Photonics Epitaxy (VPEC) and Global Communication Semiconductors (GCS).

But industry sources stressed that supply chain security is currently a top concern, one that even overtakes price considerations. Even if China temporarily loosens its substrate export controls and optical communications shipments remain stable in the second half of the year, companies still need to aggressively expand their sources from Europe and Japan over the longer term.

Sources said that non-China supply must eventually exceed 50% of total substrate sourcing to ensure supply chain stability.

However, under China's export controls, large substrate inventories have accumulated inside the country, and domestic makers such as Yunnan Germanium are aggressively expanding capacity, triggering price competition among local substrate suppliers. Overseas makers are also worried that if China fully opens exports in the future, low-priced InP substrates will flood other markets, discouraging local producers from expanding production.

German manufacturer Freiberg Compound Materials originally specialized in gallium arsenide (GaAs) substrates and currently ships only small volumes of InP substrates, with yields still needing improvement. Japanese suppliers have also been relatively conservative about expanding production.

Substrate makers need more certainty before committing to capacity expansion. To that end, Google and other CSP giants have been visiting suppliers in person to make one thing clear: demand for their products is genuinely substantial.

[SemiVision 🇹🇼 👁️👁️ @semivision_tw]

CPO and 2.5D Packaging — The Wiwynn is working with GUC and Ayar Labs, using TSMC’s COUPE 2.5D technology to integrate electrical and optical signal layers. Each GPU directly packages eight optical engines. Large-scale mass production is expected in 2028. To address the high temperature sensitivity of external laser sources in ELS architectures, Wiwynn has introduced liquid cooling to ensure that the temperature difference remains below 3°C, maintaining frequency stability.

[Citrini @citrini]

AI workloads are creating new memory objects that are too large for HBM but too valuable for cold storage. It’s simply to expensive to use DRAM for everything if it’s possible to use NAND for some of it. We’ve gotten three announcements focused on flash from three major players in the past couple weeks:

Nvidia is using flash to store reusable KV cache via CMX. This targets long-context, multi-turn and agentic inference by extending GPU memory with a shared KV-cache tier, allowing flash to become context memory for long-running agents.

AMD is using flash to store some system memory until it is predicted to be hot enough to go to DRAM. They’re buying MEXT as a software layer to accomplish this.

Apple is using flash to store some model weights until needed. Apple stores inactive experts in flash on the device side (AFM 3 Core Advanced, published a week ago).

If you can get the system to know what data is likely to be needed next, flash goes from simple storage to cheaper memory.

[Citrini @citrini]

AMD and AAPL both, in different ways, have announced a focus on flash as a way to reduce the AI memory tax. We wrote about this in our State of the Themes last week.

[리얼맨 @X_Real_Man_]

주식 처음할 때 TQQQ 연평균 수익률로 엑셀 돌렸더니 매달 10만원씩만 넣으면 15년 뒤에 15억인가 되는거 보고 자동 주문 걸어두고 잊은게 1.2억 됨

절세를 위해 팔고 사고 몇번 해서 수익률은 조금 낮게 나왔네...

[Zephyr @zephyr_z9]

@KairosPraxis @austinsemis NVIDIA did a talk recently on using VSCELs for Scale Up
Will be interesting to watch
But if VSCEL is adopted then lots of hopes and dreams are about to get crushed

[Ben Bajarin @BenBajarin]

Right in line with mine and @jaygoldberg's thesis. If true of course. Would be very bullish for QCOM if this happened.

[Ben Bajarin @BenBajarin]

MS on storage. I actually think its much worse than many realize in terms of what agentic storage demand will cause. Detailed in our full report on agentic AI impact on storage.

https://t.co/QQJkCPjLSi

[Jeff Pu @sssjeffpu]

Abstract of Marvell (MRVL) Note

🚄 Custom XPUs, Google attach silicon (multi-billion TAM), AI interconnects
•Google’s CXL/DPU attach content/rate per TPU higher than expected; v10ax bidding still ongoing
•MSFT on track for Maia300 ramp (~300k units in 2027)
•Likelihood of winning one config of AWS Trainium 4?
•Optics demand to be strong on 800G/1.6T pluggables + scale-up optics via Celestial AI
📈 FY27/28/29 EPS to $4.1 / $7.6 / $14. New TP $350

#MRVL

[Marcelo P. Lima @MarceloLima]

$SNDK CEO keeps talking about earning a decent return on investment. Sounds great.

But invested capital at Sandisk is about $10 bn. Earnings est for FY27 are $27 bn. That's an insane ROIC. Why wouldn't Samsung eventually divert wafers to NAND?

[Gaetano @crux_capital_]

If you care about $LITE $COHR $NOK $GLW $CIEN

Then bookmark and read this when you can!