Technical Red-Teaming for Pro Investors: semiexponent.com Substack: viksnewsletter.com Podcast: semidoped.com NFA. DYODD
- Tweets 5,589
- Following 1,236
- Followers 25,916
- Likes 7,393
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
11
8
198
96,416
Spoke 1:1 with the most excellent @FundaAI today, and I will be doing expert interviews on their platform.
Although most people prefer to keep transcripts to institutional clients, they are letting me post to my Substack, and it will become part of their expert network database!
5
45
8,617
The force majeure notice from Oracle over the Project Jupiter datacenter build out in New Mexico is a hard case of NIMBY-ism.
People don't want giant datacenters in their home towns.
While Silicon Valley was going gaga over pacing the frontier, @austinsemis makes this sharp observation from the corn fields of Iowa a week ago.
He points out that local governments pushing back is what will pace the frontier. Not the 'safety concerns' of Silicon Valley tech leaders (who continued to push models out anyway).
Just the kind of insight we provide for free on @semidoped
🫡
Who really paces frontier AI?
It's local governments pushing back on new data center construction. That may be the ultimate governor on AI's scaling curve.
5
1
24
5,339
To clients I've spoken to, I maintain that 3D DRAM integration in HBC is a technically sound approach.
- It expands capacity while maintaining SRAM like performance
- The pj/bit consumed while moving bits makes it a good approach for power constrained edge devices
In this quick episode, @austinsemis sits down with Durga Malladi from QCOM for a quick chat on @semidoped to find out more.
Check it out.
🎙️ NEW EPISODE: Qualcomm's HBC vs HBM, Dragonfly AI 250, and Winning on TCO
Austin sits down with Qualcomm's Durga Malladi live in Maui to talk HBC (High Bandwidth Compute). HBC is Qualcomm's differentiated bet to win data center inference workloads.
HBM shuttles data across wide buses to a separate accelerator. HBC puts compute directly on the stacked DRAM's logic die.
- HBM is fast but power-hungry; HBC drops latency AND power per bit
- 18x effective bandwidth on the AI 250 at the SAME 768 GB and SAME 160 kW
- A trillion-parameter FP4 model on a single card
- Foundry + memory-vendor relationships de-risk multi-megawatt supply at scale
Chapters:
0:00 Introduction
2:49 The Memory Wall Problem
8:05 HBC Physical Architecture
9:33 HBC vs. Custom HBM
11:50 Silicon Proof Points Coming
12:08 Managing Thermals
13:34 Dragonfly AI 250
14:48 Trillion-Parameter Model on One Card
16:56 Tokenomics and TCO
18:26 Why TCO is Key
19:26 Manufacturing at Scale
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 @Qualcomm @Snapdragon
6
2
48
18,302
The original CPUs for agentic AI thesis was published on my Substack on Feb 17, 2026.
viksnewsletter.com/p/the-cpu…
𝗦𝗜𝗚𝗡𝗔𝗟 𝗙𝗥𝗢𝗠 𝗡𝗢𝗜𝗦𝗘
@vikramskr @austinsemis @semidoped
You deserve way more credit than calling out the Agentic #AI inference driven CPU boom just a month ago.
On April 10th I put out a post that CPUs were the new GPU and we were going to see unprecedented demand. That $INTC was going to be a big beneficiary.
April 15th @vikramskr published a graphic showing a detailed schematic of Agentic orchestration workload. This confirmed my thesis.
At that time @Intel was trading about $64.
On 23rd April I published my first ever $INTC earnings forecast. Significantly above the street.
On April 24th $INTC reported earnings and the stock jumped 23% overnight as Xeon CPU demand kept accelerating. My forecast beat the street by miles.
We are now in an age where a single researcher with deep expertise can generate a signal many experienced analysts on the street may not find.
The question is do you know what is the signal and what is noise?
Find the question - you have the answer.
2
35
5,921
Good post on 3D Dram
In the past few months Qualcomm, Samsung, Cerebras and d-Matrix have all teased the same thing. DRAM stacked with compute.
Here's the case for why the trend makes sense. Three reasons.
1/ Every tensor engine gets its own path to memory instead of sipping from a shared straw
2/ A path below 0.1 pJ/bit, which frees the power budget for compute and networking
3/ Unlike HBM, 3D-DRAM access can be made deterministic, like SRAM
Full Article: chiplog.io/p/why-stacked-3d-…
4
7
74
12,201
Since I didn't actually to go to ECOC (epic fail, I should have), I had to live vicariously through announcements. There were still quite a few interesting ones related to optical scale up.
Here are my thoughts on a few topics:
🕊️Teradyne Iris 100 for microLEDs
🕊️ams OSRAM thin-film VCSEL
🔒 Lumentum DWDM ELSFP
🔒 Lumentum + Qualcomm 1060nm VCSEL
🔒 Coherent Photonlink
viksnewsletter.com/p/ecoc-20…
3
2
26
4,744
The AI/agentification of market research has implications to buy-side too.
The easiest thing for a buy-side firm to do, is to build their own AI-driven research platform. It is getting easier to do everyday.
Many independent research firms today provide AI dashboards, which is definitely useful. @FundaAI has a good one for example. @DiligenceStack has the Atlas platform, which is also good.
But when buy-side firms are already spending millions on research subscriptions, their own datasets are massive, and likely to produce far more insights than any one firm's AI research platform.
If the resistance to this agentic world is cultural because investors would rather believe in their spidey-senses instead, they will lose out to the more AI-forward of those who would actually adopt technology.
What this also means that differentiated research in the future will come from people who have domain expertise. The ability to process data, numbers, and figures differently from general sell-side research provides an advantage going forward.
I believe in this approach for my own firm, SemiExponent, and there are others on Substack also branching out into research with differentiated domain expertise. @damnang2 is a prime example.
The world of research is changing, and the question is who chooses to adopt it, and how quickly.
This experience will only spread, in my view.
The historical best practice to get smart on a name was read the filings, read the transcripts, and read a big stack of sell-side research, then model out the company. This hadn't really changed much in decades (outside of innovations in alternative data & expert network transcripts). Having a comprehensive offering of sell-side research was important at the institutional level.
We have reached threshold on numerous fronts where a public market investor can achieve similar or deeper comprehension on a name with AI, and doesn't necessarily need that same comprehensive sell-side offering. That random sell-side report that went deep on a certain aspect of the business or industry can now be created with an AI agent sitting on the right data pipeline.
One example I've shown in the past on DKNG...in the past, if I'm trying to get smarter & sharper on a deep dive on state level taxation, the right sell-side note dropping at the right time is supremely helpful. Now, I can run that analysis when I need it at the push of a button. The cohort of investors who build off of sell-side models will, very soon, be at push-button AI capabilities (and more may move modeling off Excel into JSON). The moat of the sell-side is melting.
And I believe the sell-side has, collectively, overplayed their hand in being adversarial to the agentic path. My view is they will eventually fold, but not before many clients learn to build around the commercial friction and, maybe, eventually come to the same conclusion as Just Another Pod Guy.
Like most things the top decile sell-side analysts will be fine, decades of investor trust and relationships will continue to monetize.
But what happens to the 16th best analyst on a name? It think it's obvious the industry just needs fewer voices on a name, so how do you pivot?
Corporate access has enduring value (if you don't believe it, be a fly on the wall when there is one seat at a key meeting and 5 pods wanting that same seat...).
But I think it's deeper than that...how does the sell-side drive differentiated client insight? Not just regurgitate publicly available information (never much value, and now zero value). Cleveland Research to me is the working mental model of deep embedding of their analysts into the operational flow of industries driving a regular & valuable flow of investible insights.
4
3
63
20,134
We'd like a small victory lap for calling out the CPU boom a month ago on @semidoped ...
This was when @bot dropped, and the demand for CPUs when these personal agents run on VMs was crystal clear.
It seems like the markets woke up only when Muse took off.
Semi Doped and its free daily newsletter (daily.semidoped.com) is all the free alpha you will ever need.
And, this comes from @austinsemis and I, who have high-priced newsletters on Substack 🤣 -- that has deeper content.
Get on the Semi Doped train: podcast, newsletter, and X account.
$0 spent.
🎙️ NEW EPISODE: Grok Bot and How CPUs are used in Agentic AI
The rise of user-friendly agentic AI platforms will create a massive new demand category for dedicated, high-core-count “agentic CPUs” to execute tasks in parallel, fundamentally reshaping the server CPU market beyond just feeding GPUs.
- The GPU is the "genius", the host CPU is the "assistant"
- Agentic tasks create spillover work that swamps the host CPU
- This creates a need for dedicated, high-core-count agent CPUs
- The "Mac Mini craze" was for security and isolation, not just local GPUs. This can be done with cloud VMs
- The key metric is cost-per-core, or "cost per employee"
Chapters:
0:00 Introducing Grok bot
3:45 Grockbot's Cloud VM Architecture
6:28 The 'Mac Mini Craze' Explained
14:38 Three CPU Deployment Models
16:02 GPU as Genius, CPU as Assistant
21:36 The Limits of the Host CPU
25:20 The 'Office Building' Analogy
29:09 Cost-Per-Core is the Metric
31:20 Intel's P-rack and E-rack
37:32 The Orchestration Bottleneck
40:19 Where Grok bot's VM Lives
44:22 The Unanswered Question
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
7
8
104
18,942
Its very strange that ams OSRAM's press release changed after I copy pasted it for the first time.
They specifically removed the 1,024 emitter array detail.
Breakthrough Thin-Film microVCSEL Platform for AI Scale-Up Networks
ams OSRAM continues to expand its photonics portfolio through the development of a next-generation 850 nm thin-film VCSEL platform specifically designed for AI scaling-up networks.
The platform combines several differentiated technologies in a single architecture, including flexible Thin-Film VCSELs, dense 25 µm pitch arrays, advanced optics with pass-through 3D sensing, silicon TSV integration and compatibility with standard multimode fiber infrastructure. This enables highly parallel optical links with exceptional packaging density and power operation.
The company has successfully demonstrated multiple microVCSEL integration on CMOS driver substrates and demonstrated over-the-fiber 32 Gb/s NRZ operation. The first chip arrays comprise approximately 1024 emitters each, offering high-speed optical data transmission.
This innovation achieves substantial improvements in performance per area and energy efficiency, enabling data center and AI cluster scaling while minimizing power consumption.
5
2
41
8,061
Breakthrough Thin-Film microVCSEL Platform for AI Scale-Up Networks
ams OSRAM continues to expand its photonics portfolio through the development of a next-generation 850 nm thin-film VCSEL platform specifically designed for AI scaling-up networks.
The platform combines several differentiated technologies in a single architecture, including flexible Thin-Film VCSELs, dense 25 µm pitch arrays, advanced optics with pass-through 3D sensing, silicon TSV integration and compatibility with standard multimode fiber infrastructure. This enables highly parallel optical links with exceptional packaging density and power operation.
The company has successfully demonstrated multiple microVCSEL integration on CMOS driver substrates and demonstrated over-the-fiber 32 Gb/s NRZ operation. The first chip arrays comprise approximately 1024 emitters each, offering high-speed optical data transmission.
This innovation achieves substantial improvements in performance per area and energy efficiency, enabling data center and AI cluster scaling while minimizing power consumption.
6
7
68
53,626
ams OSRAM's 25um dense pitch is impressive, and is what Avicena mentioned that their next-gen microLED pitch is going to be.
Each lane runs at 32 Gbps, and has 1,024 emitters each. That's 32 Gbps x 1024 / 8 = 4 Tbps communication. The press release does not mention reach or BER.
If the pre-FEC BER is at least 1e-9, which it better be for scale-up applications, and reach is 5m, that would put microVCSELs way ahead of what is possible with microLEDs today.
You see, its not that people have an issue putting 1,024 emitters and running a wide-but-slow approach. The "too many cables" argument was always weak.
The real limitation of microLEDs is per lane speeds. If you can get 32 Gbps out of microVCSELs, then why deal with microLEDs which are limited to at most 8-10 Gbps.
It's not that microLEDs do not work; they do provide a good solution for some lane rates. microVCSELs per lane speeds give it more runway.
There has been some mention of reliability testing in the press release too, which is good. Now I wonder about the temperature sensitivity.
Breakthrough Thin-Film microVCSEL Platform for AI Scale-Up Networks
ams OSRAM continues to expand its photonics portfolio through the development of a next-generation 850 nm thin-film VCSEL platform specifically designed for AI scaling-up networks.
The platform combines several differentiated technologies in a single architecture, including flexible Thin-Film VCSELs, dense 25 µm pitch arrays, advanced optics with pass-through 3D sensing, silicon TSV integration and compatibility with standard multimode fiber infrastructure. This enables highly parallel optical links with exceptional packaging density and power operation.
The company has successfully demonstrated multiple microVCSEL integration on CMOS driver substrates and demonstrated over-the-fiber 32 Gb/s NRZ operation. The first chip arrays comprise approximately 1024 emitters each, offering high-speed optical data transmission.
This innovation achieves substantial improvements in performance per area and energy efficiency, enabling data center and AI cluster scaling while minimizing power consumption.
1
41
6,592
Teradyne Introduces Iris 100: Production-Proven Test System for MicroLED Devices
Teradyne Iris 100 was engineered for high-volume precision manufacturing with its spectrometer and high-resolution camera that measure the spectral response of the array, and per-pixel luminance and uniformity. Advanced test parallelization and image-processing algorithms enable characterization of a full array containing millions of microLEDs with high throughput.
"MicroLED is shifting from lab to volume production for both AR microdisplays and, increasingly, optical interconnects, which are becoming foundational to how AI data centers scale beyond copper," said Shannon Poulin, president of the Semiconductor Test division at Teradyne.
---
Large throughput testing is something I did not cover on my post (below) on microLEDs interconnects.
$TER 's optical test solution, Iris 100, for microLEDs is just as important as whether the tech works in the lab or not.
Each optical transceiver has hundreds of microLEDs which all need to be tested known-good before integration on silicon.
Although this platform is useful for AR displays, it signals growing interest in microLEDs for signal interconnects.
viksnewsletter.com/p/a-micro…
5
1
36
4,552
Realized that my post on microLEDs did not have this nice Astra made animation of the different bit sequences we tested on the optical link.
When looking at victim/aggressor analysis on a link for crosstalk analysis, you should use different sequences on the victim/aggressor.
This way all of them do not move up and down at the same time. You want the aggressor going up, while the victim is going down, and some random combinations of the above when analyzing eye diagrams.
The article has a lot more nuance on optical link testing on MicroLEDs, so you can evaluate for yourself whether you think the link works or not.
Long term, microLEDs are still newer than microVCSELs as a base technology for communication. The eventual adoption of any technology depends on a lot more than just 'if the tech works or not.'
Give it a look:
open.substack.com/pub/viksne…
MicroLEDs for interconnects are becoming increasingly interesting as a technology for AI datacenters.
I went to Avicena's office and spent 3 hours talking to their entire team, from the CEO to senior leadership to engineering working on the tech.
I asked in-depth questions and I took lots of pictures. The latest report combines all my observations in a single post.
Important ideas to take away:
- MicroLEDs are more of a copper replacement for scale up/in. Don't compare to laser optics. They are not the competition.
- Slow lane rates simplify a lot of things; faster isnt always better.
- Crosstalk isn't as big an issue as everyone makes it out to be.
- Fiber bundle does not make the cable fatter. It's a pretty normal sized cable.
- Gearboxing is a requirement for "wide-but-slow" ; more lanes require deskew, slow lanes simplify circuits.
- It is a killer technology for scale-in; the rumblings of which we are just beginning to see.
What I still don't buy: that microLEDs have 30m reach. I have not seen a demo yet. This messaging from companies working on microLEDs draws comparison to optics -- which is missing the point.
There is room for a good interconnect in the sub-10m range.
The post below addresses lots of the nuance that is missing from the dialog on microLEDs for interconnects.
Read it here:
viksnewsletter.com/p/a-micro…
2
2
53
8,144