2026.10.07 | Substack | MRVL | Marvell 2026 Investor Day Recap

原标题：Marvell 2026 Investor Day: Irrational Recap

#asic #networking

Substack · 2026-10-07

Irrational Analysis is heavily invested in the semiconductor industry.
Positions will change over time and are regularly updated.
Opinions are authors own and do not represent past, present, and/or future employers.
All content published on this newsletter is based on public information and independent research conducted since 2011.
This newsletter is not financial advice and readers should always do their own research before investing in any security.
Feel free to contact me via email at: irrational\_analysis@proton.me

Hello today I am covering Marvell’s investor day for an unusual reason.

I sold my Ciena position to buy more AXTI yesterday trying to lever up more on optics and somehow Matt Murphy sent Ciena to the moon…

… while only mildly pumping his on stock.

This must be some kind of cosmic justice for all the anti-Marvell trash talk I have put out over the last ~18 months.

I have been consistently an ultra Broadcom bull and moderate Marvell bear. The engineering delta is quite wide. And yet, everyone seems to be using Broadcom as a funding short and Marvell has gone to the moon. I swear this ticker is like meth for the pod monkeys.

What do you have for us, Matt?

What happened to Celestial? I like them but they seem to have disappeared.

Errr… Credo, Lumentum, Ciena, Coherent, Astera Labs, all have higher connectivity ratio?

Who is this slide for, the tourist long-onlys?

Very charitable way of describing Teralynx switches.

Credit to Marvell, their dense SRAM is excellent. No clue how it works but the data they showed a year ago at Hot Chips was A+ quality.

Hot Chips 2025: Irrational Recap

September 13, 2025

Irrational Analysis is heavily invested in the semiconductor industry.

Read full story

In terms of NAND controllers… they have like no market share right? Not sure on this.

I see 5G/telco is dead. Love to see it. Warms my heart.

Buddy this bubble probably gona pop before 2030.

Frankly, when semiconductor companies start to give far out targets like this, it indicates bubble lol. Years from now I need to come back and re-use this slide to show that there were signs.

Half this slide deck is the same bar charts showing CAGR CAGR CAGR.

Next investor day, they should try banning that acronym and see if they can spend more time on substance.

Well you guys certainly need a lot of new grads given the epic attrition this year.

If this is just a wrapper for external LLMs to interact with EDA tools I am going to be very disappointed. Please tell me you at least finetuned an open-source model on your internal data?

Someone ask Innolight about what they think of the Aria 1.6T DSP that had to get re-spun and where a Marvell VP got fired over that temporary mess.

I have a particularly mean comment I will keep to myself for now.

Their die-to-die PHY (bidirectional) is very impressive. Although I don’t think it will function to direct-drive optics in bidi mode.

Irrational Analysis (Public Companies Only) SerDes Rankings (Oct 2026):

There is a limit. It’s called chromatic dispersion.

TFLN is a meme material pitched by PhDs who have a distinct lack of understanding on how the real world works.

You all know what I think about uLED.

Plasmonic I honestly don’t know enough about. Glad Matt went out and collected all of the optical modulators. Maybe they can get a design win now.

Yes when I think SiPho ring modulators, I think “process agnostic”. Sigh.

To my knowledge, Marvell has no hybrid integration InP program. Not sure what they are talking about. Maybe the intern who prompted AI to make this slide forgot to proofread.

Their die-to-die is very good. Wish they spent more time on it.

Are these… rectangular optical lenses?

Uhhhhhh. HMMMMMMMMMMM

Semtech is better.

Yes we know about the out-of-spec concat FEC you implement for Google. Have a cookie.

Again, ask around on what Innolight thinks of Ara.

ENHANCE.

Not great linearity. It’s ok.

This is kickass. Excellent constellations.

Pre-FEC BER of 3e-6 with post-FEC BER of 1e-11.

Very good! Just need a little more improvement. I think target for post-FEC is 2e-13.

This is genuinely impressive. Congratulations to everyone who worked on this.

uVCSEL can go 10 meter easy. Either it does not work at all or it can hit 10 meter.

Please tell me you guys are not serious about TFLN.

See this is a bad constellation. The 1.6T version so much better and yet post-FEC BER is slightly worse. I wonder if the 1.6T coherent ZR/ZR+ sample has CDR bug. Bursty errors from the phase interpolator?

The point of GeSi EAM tech was to put the modulators underneath the silicon and save massive shoreline. WTF are you guys doing, pivoting this tech to NPO/CPO?

Can someone please check in on Preet? Has he been locked away in Matt Murphy’s basement?

Yes, one modulator but 4 drivers and 8 TIAs. Makes perfect sense.

OCI MSA is literally defined at O-band. Your EAM is C-band only.

Are you saying Photonic Fabric (proprietary) or OCI MSA?

This memory expander product uses HBM as a cache for the DRAM. Not sure on the economics of this or if Marvell can even get HBM allocation for this product when most/all of their allocation probably needs to goto XPU.

Cool product though. Curious on the latency. If it is under 100 ns that would be pretty good.