Independent Semi Research: https://xitter.twitterapi.xyz/t.co/Qu61K9tElF Substack: https://xitter.twitterapi.xyz/t.co/60ru2W5jNl Podcast: https://xitter.twitterapi.xyz/t.co/tfzu1g15GY NFA. DYODD
Joined November 2020
- Tweets5.9K
- Following1.3K
- Followers26.5K
- Likes7.9K
Pinned Tweet
I have a somewhat different view of where independent semiconductor research is heading.
The problem institutional investors have isn't a shortage of research.
It's the opposite.
I've spoken with a number of hedge funds and institutional investors, and a recurring complaint is simply: too many reports.
Their inboxes are overflowing with research they would like to read but realistically never will. AI is only going to increase that volume.
So with SemiExponent, I'm building something deliberately different.
Less publishing. More interaction.
The institutional product is centered around direct access to semiconductor expertise: discussing technology, challenging assumptions, arguing through competing interpretations, and red-teaming an investment thesis when the underlying question is technical.
In other words, not another research feed.
A technical sparring partner.
That model is intentionally high-touch, which also means SemiExponent will work with a relatively small number of institutional clients.
I've been discussing the concept with several investment firms and am now beginning to launch it.
If this sounds useful for your team, you can reach me through SemiExponent.
Just leave a note with your email, I will get back to you.
semiexponent.com/contact
Another microLED startup -- Staera -- with some interesting per-lane rates (4-8 Gbps), and they want to go scale in too.
- 4 to 8 Gbps per link, with more than 600 links per millimeter of chip edge
- Designed for under 0.2 pJ/bit, against 5 pJ/bit for co-packaged optics
- More than 100x the performance of electrical links across the millimeter-to-meter range
- Reaches beyond 50 mm, against 2 mm for UCIe-A die-to-die links
- Lightwire-based HBM designed to exceed 24 TB/s, against 2.8 TB/s per HBM4 stack today
Datacenter permit restrictions mean that our priorities are going to shift.
Efficiency over all else.
For every GW deployed, we need to assess what the effective token output is.
All optimizations will go towards squeezing max compute out of every watt.
Who are beneficiaries?
1) 800VDC power -- allows densification and efficiency upgrades
2) Custom inference accelerators -- can reduce tokens/watt dramatically
More explanations in the [FREE] post on Substack:
open.substack.com/pub/viksne…
As a professional writer, I battle this every day.
I still prefer to hand write articles. It brings clarity to thinking and the article itself.
But as a business, it gets difficult to compete with others using AI.
When a subscription gets measured by quantity over quality, or you start to see successful Substacks use AI and people don’t seem to care, you start to question if hand crafting articles is worth it.
An investor I spoke to also admitted that all their subscriptions are read by AI anyway.
What then is the point of hand crafting posts of no human would read it?
Writing with AI has led to a confusing time, especially as a Substack writer.
I’m also leaning towards outlets that are not anonymous, and more human — like @semidoped podcast. You can’t fake real humans talking to each other on video.
Sadly, that transcript is fed into AI as well.
Ultimately, the choice of how much AI one chooses to live with is a personal choice. There is really no wrong answer, just preferences. Just like some people read physical books, while others read ebooks or listen to audiobooks.
35% stock drop in a single day makes it Intel's worst day on the stock market ever.
And we needed to talk about it on @semidoped
.@austinsemis and I chat about what the hell happened, what intel could have done about it, and the lessons we learned from it.
Top points:
- server CPUs were supposed to save the day, they didnt
- Substrate shortages: inevitability or planning disaster
- Their fab lit on fire, and what to do about it
and a whole lot more.
🎙️ NEW EPISODE: Intel's Earnings Disaster: Bad Luck, Bad Planning, or Both?
Intel had a strong quarter. Its outlook told a different story.
@vikramskr and @austinsemis unpack Intel’s Q3 2026 results and the gap between demand for its chips and its ability to deliver them. They cover the AI memory squeeze hitting PCs, the substrate and chiplet bottlenecks limiting server CPUs, and the fallout from the Fab 28 accident. Then they debate what Intel could have done differently and whether its long-term foundry story still holds.
Presented by @CrusoeAI: Fine-tune and serve open models without managing a single GPU. When we sat down with Crusoe’s product team, Austin asked: what will you do with my data? Their answer: They don’t touch or see it. Your training data is encrypted so even Crusoe’s own admins can’t see it, and the fine-tuned model is yours. Deploy on Crusoe in one click, or download and run it anywhere. Upload a dataset, pick an open model, and pay per token rather than per GPU hour. Run a few versions, compare them, and download the one that works best. Accounts are free, and get $5 in credit when you sign up. Try Serverless Fine-Tuning: semidoped.com/crusoe-x
This episode is supported by @meter. You have the GPUs and the power. Now who builds the network, and who do you call when it breaks? Meter designs, deploys, and runs frontend, storage, and management networks on its own switches and its own network OS. One call reaches the team that designed the hardware and wrote the software. semidoped.com/meter-x
Chapters:
0:00 Intel's 35% Stock Drop
1:22 The Pre-Earnings Bull Case
4:03 Reason 1: Weak PC Market
12:25 The Flawed Q2 Pivot Plan
17:17 The IDM Model's Failure
20:23 Reason 2: Substrate Bottleneck
23:18 Chiplet Design's Fragility
25:18 Reason 3: The Fab 28 Explosion
40:46 The GPU Intelligence Gap
45:51 The Foundry Thesis Question
50:22 Packaging as a Hope Story
51:19 Four Lessons from the Meltdown
Get more of Austin and Vik daily, free!
Sign up: daily.semidoped.com/
Connect with Vik and Austin:
Vik's Paid Substack: viksnewsletter.com
Austin's Paid Substack: chipstrat.com
@austinsemis @vikramskr
Yeah $10. Now we speak in Trillions!
TIL: The First Contract in Silicon Valley Was Ten Dollar Bills
open.substack.com/pub/semido…
6
3,595
Just got off a recording with a guest for the @semidoped podcast.
They said that I have a unique way of getting the best out of a guest on the show, while being at ease and locked in throughout the conversation.
I really appreciate this feedback. I've been working hard at becoming a better interviewer
3
27
4,598
We always appreciate our sponsors on @semidoped ... thank you for believing in us.
Our presenting sponsor is @CrusoeAI ... ya'll prob already know what they do.. if not you should check em out.
Our supporting sponsor is @meter -- they build networking from the ground up. check 'em out too!
We will do our best to make great content around semiconductors and markets!
We have a new supporting sponsor for Semi Doped!!
@meter!
Meter builds networks from the ground up. Great folks to get your GPUs hooked up into clusters.
Check 'em out!
It might be market armageddon, but we can still talk about technology -- because what else is there to talk about 🤣
In this @semidoped episode, I talk to the CTO of d-matrix, Sudeep Bhoja, and we go through the evolution of @dMatrix_AI ... starting from their analog in-memory compute chips, to their future product: Lightning.
We also discuss the implications of working with NVIDIA to provide inference solutions, the impact of software, and what the future holds for inference.
Check it out.
🎙️ NEW EPISODE: 3D Stacked DRAM Is Where Inference Hardware Ends Up: d-Matrix's Sudeep Bhoja
Sudeep Bhoja, co-founder and CTO of d-Matrix, joins @vikramskr to walk through the company's path from an early, non-consensus focus on memory-bound inference to a full rack of 3D-stacked DRAM. He covers the pivot from analog to digital in-memory compute, how the Corsair card fits into disaggregated systems, and how the next-gen Raptor chip plugs into Nvidia's ecosystem.
- Corsair delivers 150 TB/s of on-chip SRAM bandwidth, over 8x the 18 TB/s of HBM on an Nvidia Blackwell GPU
- Raptor stacks a logic die on a DRAM die for 32 GB, 16x the prior SRAM capacity, at ~100 TB/s
- 0.37 pJ/bit vs. d-Matrix's estimate of ~5 pJ/bit for HBM4 plus a GPU
- A full rack of 144 Raptor packages, 2.3 TB of high-bandwidth memory, built to drop into Nvidia's VR rack
Presented by @CrusoeAI: Fine-tune and serve open models without managing a single GPU.
When we sat down with Crusoe’s product team, Austin asked: what will you do with my data? Their answer: They don’t touch or see it. Your training data is encrypted so even Crusoe’s own admins can’t see it, and the fine-tuned model is yours. Deploy on Crusoe in one click, or download and run it anywhere. Upload a dataset, pick an open model, and pay per token rather than per GPU hour. Run a few versions, compare them, and download the one that works best. Accounts are free, and get $5 in credit when you sign up. Try Serverless Fine-Tuning: semidoped.com/crusoe-x
Chapters:
0:00 Intro: The Cost of a Token
3:28 The 2019 Bet on Inference
7:43 Nighthawk: Analog In-Memory Compute
12:08 Why Analog Failed
19:02 Corsair: Digital In-Memory Compute
26:04 System Strategy: Speculative Decoding
37:05 Raptor: 3D Stacked DRAM
39:40 The 'Free Lunch' of 3D Stacking
45:00 Trade-off: Logic-on-Top vs DRAM-on-Top
54:14 Go-to-Market: The Nvidia VR Rack
1:03:17 The Future: Lightning & 4-High Stacks
1:12:45 The End State of Inference
Get more of Austin and Vik daily, free!
Sign up: daily.semidoped.com/
Connect with Vik and Austin:
Vik's Paid Substack: viksnewsletter.com
Austin's Paid Substack: chipstrat.com
@austinsemis @vikramskr
English version of the post. 👇
Palomino is engineering microLEDs specifically for datacom, not adapting them from displays.
The emission angle of palomino microLED is narrow and helps with better coupling into fiber.
otcmarkets.com/stock/PALX/ne…
Korean media report that U.S. startup Palomino recently received its first purchase order from a strategic customer, a global hyperscaler, for GaN microLED-based optical interconnect chips for AI infrastructure.
First commercial use of microLED chips for optical interconnects?
etnews.com/20261105000095
1
20
4,182
Been going deep into protection strategies, and came across this terrifying slide.
See stick figure human connected to the output of a 4.6MW solid state transformer?
If the parasitic capacitance at the node exceeds a few microFarad, then the SST discharges through human.
The slide mildly puts it as "risk of fibrillation"
I put the risk as "stick figure human explodes"
Intel’s stock death.
A murder mystery and post mortem.
Featuring star detective @Srasgon
open.substack.com/pub/viksne…
This is kinda big news in my world.
SST-based 800V DC deployments are the holy grail of datacenter power.
China has early deployments.
Source:
prnewswire.com/news-releases…
6
1,873
Scheduled my first expert call on the @FundaAI network.
My conversation will be around the impact of HBM bandwidth/capacity on accelerator performance.
Let's see what I learn. I will publish the transcript.
Finally a $QCOM inference chip is spotted in the wild.
Waiting for benchmarks.
Details of MRVL + LITE
$MRVL to acquire $LITE in an all-stock transaction.
- Implied enterprise value of $115 billion for Lumentum.
- Lumentum stockholders will receive 4.5 shares of Marvell common stock for each share of Lumentum common stock.
- This represents ~$1,132 per share of Lumentum common stock and a premium of approximately 23.9% based on each company’s price as of October 28, 2026.
- Marvell stockholders will own ~67% of the combined company. Lumentum stockholders will own ~33%.
- Matt Murphy to remain as CEO of the combined company.
- Transaction expected to close in H1 2028, subject to shareholder and regulatory approvals.
Marvell and Lumentum to combine.
Here is a post thinking through the consequences:
open.substack.com/pub/viksne…
