Emergency memo on the Reuters report today and the chaos that followed.

There is a lot of nuance and complexity here.

Will also be snipping some of the work by my friends at .

FUNDA

Research|Optics: FCC Curbs on Chinese Transceivers, US Vendors and Globalized Leaders Better Placed

According to Reuters, the U.S. government is drafting restrictions on new models of data-center optical transceivers from China, potentially to be implemented by the FCC through its equipment authorization and Covered List frameworks. However, no formal rule text has been released. It remains unclear whether a “Chinese optical transceiver” would be defined by place of manufacture, customs country of origin, place of incorporation, ownership and control, or the location and composition of its R&D and supply chain. The proposal could also be revised or shelved…

Read more

5 hours ago · 9 likes

Contents:

1. Reuters Report Re-Cap

2. Overview Featuring Funda

3. Practical Transceiver Engineering

4. Stocks, Trading, Market Vibes, and You

[1] Reuters Report Re-Cap

https://www.reuters.com/world/trump-administration-drafting-ban-chinese-data-center-devices-sources-say-2026-08-04/

This would allegedly only effect new transceiver models. Everyone has all the flavors of 1.6T qualified so not a huge deal.

This “stealing data” line is total bullshit. Transceiver is just a PHY that converts electrical to optical and back. All data is encrypted at the host endpoint. Transceivers cannot install malware unless the datacenter operator is completely incompetent. Arbitrary code execution via CMIS would need next-level opsec failure.

This same line of nonsense argument used again Huawei remote-radio-heads (RRH) in 5G networks. Just admit you want to ban Chinese stuff for economic reasons. That’s fine. I hate these obviously bullshit security excuses.

Divyansh Kaushik of Beacon Global Strategies is a complete fucking moron who understands nothing about how transceivers or networks function.

You know what does pose a risk? Brain-dead fake experts who spout verifiably false nonsense and drive policy.

Tell me, if data is encrypted and decrypted at host endpoints, how does the transceiver sniff data? How can a transceiver install malware over protected CMIS registers which link to a BMC bus which is itself another layer of security? Idiot do you even know what CMIS is?

https://www.oiforum.com/wp-content/uploads/OIF-CMIS-05.4.pdf

CMIS is the framework in which a transceiver interacts with a host system.

Essentially there is a register map of allowable commands between the host and the MCU inside the transceiver. In order to present a security risk, the transceiver must:

1. Somehow store a payload that can exploit an arbitrary code execution vulnerability on the host system.

2. Execute this payload via a very limited register space as the host is running normal state machine operations.

3. Do all of this after defeating the signed key authentication system as describe by OIF-CMIS-05.4 section 9.9.2

There is a signed security system via public+private keys.

[2] Overview Featuring Funda

Funda did great work and I want to add here and there. Few points of disagreement.

Innolight, Eopto, Accelink, and several others have created “more than a shell” companies that are USA registered but have all their footprint in southeast Asia. So Malaysia, Thailand, Vietnam, and such.

All IP (product designs, patents, know-how) is licensed to the “more than a shell”. Nearly all major production processes are done outside of mainland China. From active alignments of the TOSA/ROSA to PCB assembly, to module assembly, to burn-in test. All of the administrative stuff too.

It’s easy to get Chinese transceivers that are technically/legally “not Chinese”.

I am more cautious. China probably will retaliate and that will result in a nice 15-20% single-day drop in .

Long-term, the InP wafers will make it out one way or another. More on this later.

De-coupling is the point. I’m not sure why Funda frames this in a tone as if China cares about preventing acceleration of USA de-coupling.

[3] Practical Transceiver Engineering

The key difference between China and the west in optical transceiver competence is active alignment.

Aligning optical elements (fiber, lenses, isolator, DFB, EML, SiPho PIC, photodiodes) is a nightmare. This is where your yield goes to die.

Low-single-digit micron accuracy is common. Some designs require sub-micron accuracy.

THE CHINESE ARE INCREDIBLY GOOD AT THIS. ORDER OF MAGNITUDE BETTER THAN THE WEST. THEY MAKE THEIR OWN CUSTOM MACHINES WHICH ARE ACTUALLY QUITE GOOD. WAY HIGHER YEILD, COUPLING EFFECIENCY, AND THROUGHPUT.

Companies like Lumentum and AAOI try to claim they have automated lines.

Everyone claims to have custom tooling for precise, high-TPUT active alignment.

Only the Chinese actually have it. TFC, Inno, Accelink, Eopto wipe the floor with the west.

And don’t get me started on Fabrinet.

Fabrinet is the worlds best non-Chinese optical contract manufacturer because everyone else is unusably dogshit.

Dont want to use the Chinese but also want to have hope of a viable product? Fabrinet!

The key differentiation the Chinese have in transceivers is much better active alignment. This shit is very difficult. Lot of bespoke know-how that only they have figured out.

Later in the stocks section, I need to discuss EML vs SiPho manufacturing pivot. At a high-level, SiPHo transceiver has 50% reduced InP content versus and EML based transceiver while having significantly higher active alignment content. It’s all the same shit. Everyone has a MZI-based PIC for 1.6T. It is trivial to pivot manufacturing from EML based to SiPho based. The issue is active alignment capacity/TPUT, not the design itself.

[4] Stocks, Trading, Market Vibes, and You

This news has let do a lot of chaos and it’s important to understand that even if you are right, other large market participants are wrong and more powerful. In which case you lose money while being right.

Let’s start with the most important stonk, .

Long term, this ticker is going to the moon. Easily a 2-4x within the next 12-18 months. Funda and myself are bullish for the same reasons. Problem is, idiots exist and it takes one bad news piece from China to tank the stock 20% in a single day.

I have done ok trading this ticker. Sold out today because risk too high. Busy playing around with Sandisk and AAOI.

Seeing AXTI go down and IQE go up off this news is hilarious and tells you how much tourism there is in this sector. Idiots, where does IQE buy the InP wafers so they can epi them?

Tower went down, then went back up but weakly. Again this is stupidity.

1. The “more than a shell” companies will ship plenty of transceivers. Who gives a shit if Innolight logo is not on the finished part. Same PIC design ordered from Tower and shipped to Terahop.

2. Existing transceiver models will (probably) be allowed to continue selling and everyone has already qualed all the models so this becomes issue at 3.2T which is far away.

3. Lumentum, Coherent, and AAOI all use Tower for SiPho PIC so production shift means Tower still win.

4. Tower has unique IP via Openlight PDK for integrated InP on SiPho which is very valuable in a world where active alignment efficiency becomes a bottleneck.

Whatever happens, this is good for Lumentum, Coherent, and AAOI.

AAOI is interesting as their capacity expansion for transceivers is massive. Fun degen play that has better short-term risk/reward than AXTI in my opinion.

Fabrinet wins from these trends. They may be incompetent but they are not Chinese and their western competition even more stupid.

I don’t know why Ciena and Nokia moved from this. They play in more long haul scenarios that don’t get effected by todays news.

China cannot block InP wafer exports and USA cannot cut off transceiver imports. There is some mutually-assured destruction here. De-coupling will take 2-3 years and there will be some spicy news release and volatility along the way.

My sense is the Trump admin is trying to push hyperscalers harder. For what it’s worth, the hyperscalers have already been trying to AGRESSIVLY design-out Chinese tranceivers. Honestly I am not sure why Trump admin chose to do this. Industry was already doing the “desired” thing at a pretty brisk timeline. If you know you know.

Do you recognize this man?

Many of you do.

For those who do not, his name is Bill Hwang and he is a hero. A legendary degenerate gambler.

I want to start with his story before discussing a new hero and legend of finance, Leopold Aschenbrenner. Leopold was already famous but what happened the last two weeks cemented his place as the greatest degenerate, brain-damaged gambler of all time. You need to be aware of his situation.

Contents:

1. Bill Hwang

2. Leopold Aschenbrenner

   1. Social/industry background

   2. What happened and where we are today.

   3. A 2-week chronological re-telling, featuring 30+ contacts.

3. State of The Market

4. Irrational Analysis Consolidated Perf+Holdings+Records

5. STONKS

   1. Trading Account Names

   2. Long-Only Account Names

[1] Bill Hwang

The above 16-minute video is excellent summary of this magnificent man.

A two sentence summary of his life up to the event we care about is as follows.

Bill Hwang was a highly experienced hedge fund guy who was trained by one of the best in the industry. After committing some minor crimes, he closed is hedge fund and started a family office.

For those who are not aware, family office is a term for institutional investor entities that manage the money of a single private entity. Some of these family offices speak with me from time to time.

There are far fewer rules for family offices, both in how they choose to operate and SEC reporting requirements. Some family offices are simply one guy investing his families $100M of wealth like a normal retail investor, but with a prime brokerage account. Other family offices are staffed with professional hedge-fund people who make the decisions on behalf of the rich person/family. Strategies vary in terms of if using leverage, long/short, diversification requirements, risk rules and so on.

Bill Hwang’s family office was called Archegos Caprial Management. It’s important to understand some unique attributes.

We have a family office that…

• Is run by a highly-experienced hedge fund guy playing with his own money.

• Run by someone literally convicted of financial crimes in the past.

• Has a PM with a risk-seeking personality that somehow (yes really) is strongly linked to his devotion to evangelical Christianity.

It get’s funnier but first I need to explain what prime brokerages are.

I personally have normal brokerage accounts with E*TRADE, Charles Schwab, Robinhood, and Interactive Brokers (lite). My trading account is with a sub-prime broker called Webull (this is a joke I love u Webull). These are all normal accounts.

Prime Brokerage accounts are with large banks such as Goldman Sachs, Bank of America, Morgan Stanely, and so on. These accounts are different than what a normal retail investor can get and typically come with.

• Much higher leverage limits.

• Bespoke products such as total return swaps, hedging instruments, private placements, convertible bonds, and so on.

• Advanced order execution and tools.

• High fees.

Bill Hwang’s trading strategy was quite simple. He had a list of stocks he liked and would lever up to the tits.

Mostly legacy media and Chinese tech stocks.

But he did not buy the shares himself. He used a bespoke instrument called a total return swap (TRS). TRS is essentially a contract where the prime brokerage client gets the economic exposure (gains/losses) of an instrument (single stock, ETF, whatever) without actually owning it. The prime broker (bank) owns the securities, not the client.

If this sounds like a way to lever up without disclosing what you are levered on, congrats you guessed right.

Now… what if someone went to like 7 different banks, got total return swaps with massive leverage on the same small handful of correlated stocks, lied to the banks about the concentration of his holdings, and the banks where fucking retarded and never bothered to talk to each other?

All right enough back story, let’s look at the shitshow and admire the chaos that occurred back in 2021.

On March 23rd, 2021, ViacomCBS announces a $3B stock offering. Management knew their stock was over-valued and wanted to take advantage by diluting shareholders and getting some cash money in dah bank.

As expected, the stock price tanked. Discovery stock also tanked because these are two highly-correlated businesses and the same psychotic degen is levered up on both lmao.

Then things are flat for about 24 hours because all the prime brokers are trying to figure out what is going on and Archegos staff are obviously in damage control mode.

The zoomed out chart is more useful.

The initial selling was Archegos trying to save themselves. After a brief period, the banks removed heads from asses and started liquidating Archegos. When a large block (chunk) of shares of a stock are sold by a bank to a private buyer all at once, that is called a block trade.

#Forshadowing

There are some truly glorious CNBC clips of Faber and Cramer talking about these Arcehgos-related block trades and the prime broker who ate the most shit, Credit Suisse.

“Credit Suisse has been the center of it. It was originally a $4.7B loss, last week you may recall after the close I was reporting on those Discovery [block] trades… we didn’t think… we thought they were done. They weren’t done. And one of the key questions was are they [Credit Suisse] losing even more money[?] and the answer is yea they are. Almost another $700M.”

The face and body language of Faber is so good. Seriously click video its in the first 30 seconds.

Bill Hwang lost it all, and (more importantly) made a bunch of banks loses a hilarious amount of money too. Years later on November 20th, 2024, Bill Hwang was convicted of a variety of financial crimes.

Without commenting on legal matters, I want to say I think Bill Hwang is a hero. A true degenerate gambler.

Frankly the key lesson of this saga is prime brokers should do more due-diligence on clients.

If I go and try to get a home-equity loan on the same house from 10 banks with the intent of YOLO-ing all the money into Lumentum stock on 4x margin, I would be immediately caught. The credit-reporting system would find out immediately.

So why the hell does the prime brokerage system not have something similar? A single phone call from one bank to any other of the ~6 banks could have prevented all this.

Bank 1: “Hey do you have Archegos as a client too?”

Bank 2: “Yes”

Bank 1: “Does he have a strangely concentrated book using total return swaps?”

Bank 2: “Shit.”

Apparently, many of the banks had no clue Bill Hwang had accounts with others.

Bill, I don’t believe in God but God bless you. ✝️

[2] Leopold Aschenbrenner

Leopold is an AGI cultist.

But is not a brain-dead AGI Cultist like Dwarkesh.

Leopold is reasonably smart and several of the people he hired are also smart and (more importantly) experienced.

I have really struggled with the structure/ordering of this section. Had to delete and re-write several times.

[2.a] Social/industry background

The finance world has a variety of social cliques. Several of these cliques do not like each other. The most obvious divide is between pods and non-pods.

“Pod” refers to the multimanager strategy that major firms like Citadel, Point72, Balyasny, and Millennium.

The way pod shops work is they aggressively manage risk. If you work there and your book (portfolio) goes down more than 5% in a single day, you are literally fired and escorted out the building. This core premise of multi-strategy investing has many consequences. “Pod-brain” and “Pod-monkey” is slang attacking this group of people.

There is this concept called factors. To explain, I need to use a few toy examples.

Let’s say you are bullish AMD GPUs for whatever reason. You want to buy $100M worth of AMD stock but don’t have $100M. Instead you only have $10M. What to do? You could buy $20M worth of AMD and short $10M of an index like SOXX or SMH. But $20M is in fact a smaller number than $100M. We are professionals here and demand leverage.

So what if we buy $100M worth of AMD, sell short $100M of Nvidia, and keep the $10M of cash we started with in something safe like treasury bonds?

That would be much more allowable by the risk department.

But the problem is, AMD is a CPU-heavy business while Nvidia is just starting to sell CPUs. AMD sells FPGAs while Nvidia has none. Nvidia has a strong networking group that sells switches while AMD has zero networking switch department. Long AMD, short Nvidia, is not truly a factor neutral pair trade. We need to go deeper to lever up more.

Suppose you really like Coca-Cola stock (KO) for whatever reason. Maybe you think their brand distribution and growth is better. Maybe you like a new flavor of Coke. Maybe you think Dr. Pepper is about to die and people who like Dr. Pepper will migrate to Coke over Pepsi (PEP), leading to asymmetric market-share gains.

You still have only $10M in starting capital but dammit Coke is better than Pepsi and you want to make a $1B bet on this idea.

Buy $1B worth of KO.

Sell short $1B worth of PEP.

Pay your prime broker $10M in fees up-front and pray.

At a high-level, allowable leverage scales with three portfolio attributes.

Assume $10M in starting capital.

1. Net exposure.

   1. Long $10M, short $10M, net 0%

   2. Long $20M, short $10M, net 100%

   3. Long $20M, short $0M, net 200%

2. Gross exposure.

   1. Long $10M, short $10M, gross 200%

   2. Long $20M, short $10M, gross 300%

   3. Long $20M, short $0M, gross 200%

3. What are the things you own.

   1. Borring stable shit like carbonated sugar water?

   2. Or volatile levered-up, questionable financial platinum dogshit?

Most of the internal social/politics within finance world revolves around if you are happy with your employers’ rules and what you think of other institutions rules.

(making up some directional numbers)

Pods are levered up 5-10x but always net 0% with extremely strict factor-based risk rules.

Regular hedge fund maybe does 3x leverage, net +/-15%.

A conservative family office might go for 1.5x leverage, net 50%.

Leopold apparently (25+ sources) went 4-5x leverage at an unknown net. But I do know his shorts WERE NOT ANYWHERE CLOSE TO FACTOR NEUTRAL.

Situational Awareness LP (henceforth abbreviated to SALP) was long AI winners, short AI losers.

So say long semiconductors like Intel and SK Hynix and short traditional SaaS like Adobe.

I don’t have solid information on SALP net exposure, only gross. But I don’t think that net number matters because I am reasonably confident a large portion of SALP shorts were anti-correlated with their longs, so these shorts provided little risk-related benefits and were simply more leverage on the same factor (AI).

In short, SALP did not behave like a normal hedge fund. They had the leverage of a pod with the risk-management of a psychotic family office. Perhaps this is not far from the truth given that SALP is basically the AGI cultist family office lol.

Citrini had a great tweet on how the LPs of SALP were true believers and they would get more finding in time and don’t count Leopold out which… aged poorly over less than 24 hours.

I’m sorry not trying to be mean but the pun potential of this fiasco is too good.

Anyway, just FYI… 80% of the 30+ people who chatted with me over the last two weeks hate SALP. 20% like SALP.

Here is a very generic conversation from… two days ago. 6PM Wed July 29.

This person is part of the 20%. He was of the mindset that SALP is fine an no margin call.

My view at the time was they definitely got margin called but probably only down 20-30% and will live.

Four hours later, the FT article came out and I became aware that the situation was far worse than I thought.

[2.b] What happened and where we are today.

In the hedge fund world, there is a tradition where if one fund gets in really big trouble, Ken Griffen of Citadel (king of the pod monkeys) will come and buy your book at a steep discount and thus give you a dignified death with some hope of coming back someday.

It’s basically hedge-fund seppuku.

Sepuku is a traditional Japanese samurai ritual where in order to maintain honor, the fallen samurai kills himself with a short sword by stabbing is abdomen and dragging the sword to make sure all the internal organs are sliced and he really has no chance to live. Then a friend of the samurai “assists” the seppuku by chopping his head off.

In this analogy, the prime broker (Goldman Saches and others) represent the short sword. The margin calls become mortal wounds to Leopolds squishy vital organs. And in their hour of greatest need, facing a slow, painful, and humiliating death, Ken Griffen came to give them an honorable one instead.

Here is a snip from the SALP fund letter.

, the large cap semiconductor index is up 50.1% YTD and my psychotic trading account (typically 150-190% gross and net (I dont short other than for fun)) is up 56.8%.

So SALP still outperforming me YTD. But the thing is, I now have over-paid taxes with all these short term losses. It is unclear what the tax implications SALP have to deal with due to castration via Citadel are.

Let’s talk about how much liquid assets they have. My sense is not much and there is a very specific reason for this suspicion.

https://www.wsj.com/finance/investing/situational-awareness-down-67-in-july-in-ai-stock-rout-cd19901f

The AGI cultist family office was thinking about selling some Anthropic shares (blasphemy! Future Claude ASI will look down upon this decision) and changed their mind. The only reason you would want to sell some private stuff to get $3.B in cash is to keep more public stocks. This shows you there was some internal conflict between SALP employees who wanted to have more capital to re-build the public portfolio and AGI cultist LPs who offload all their critical thinking to Claude, because they had none to begin with.

[2.c] A 2-week chronological re-telling, featuring 30+ contacts.

I want to try some creative writing for fun.

Our story begins last week, around Wednesday July 22nd. There are (weak) rumors that SALP is in trouble. Maybe even asking LPs for more capital to meet margin calls. Not that bad. Market looks sick, as if someone big is vomiting blood and taking the rest of us with them.

Over the weekend, there was speculation that if Leopold is in trouble, Kioxia would shit itself violently on Monday July 27.

Kioxia indeed violently shits itself.

The entire hedge fund community knew that SALP had massive Kioxia and SK Hynix positions. It is unclear if SALP owned shares or total return swaps, or some other leveraged thing. Only the SALP and the prime brokers who vaporized SALP know what glorious bespoke financial instruments were needed to create this entertainment value.

I know Leopold is German but would love to see one of those DNA tests. He might be ethnically 10% Korean or something.

The next day is Tuesday July 28th and both Bloom Energy and SK Hynix are up for earnings. Once again, these are two names everyone knew SALP was massively long.

Bloom has a blowout earnings. Best earnings call ever. Hyper omega bullish. And yet, the pop got faded after hours as SH Hynix started to violently convulse from a bad miss.

Made like $40K day-trading Bloom correctly. Usually I am an idiot and hold into the down-cycle but the moment I saw weakness in SK Hynix, I knew SALP was gona start selling more Bloom.

The next day (Wed July 29) it becomes increasingly obvious that SALP is in deep shit. During the day, there are (bullshit) claims of how everything is fine and they are unwinding in a orderly way. By the evening, the chatter gets meaningfully worse. And then the FT article comes at like 9 PM. It is at that moment I realize that SALP is probably dead.

12 hours later, Faber from CNBC comes forth with better info.

https://www.cnbc.com/2026/07/30/leopold-aschenbrenners-hedge-fund-is-facing-steep-ai-losses.html

WSJ corroborates a few hours later, only disputing the “all public positions” part.

The funny thing is, you could see the seppuku in real-time as Faber published.

Wednesday was by far the most fun part of this saga.

And so here we are. Thursday and Friday get skipped because it’s mostly AGI cultist cope.

But let it be known that Leopold was right and a good stock picker, just a little too delusional.

Looking forward to SALP resurrection.

[3] State of The Market

The market was frankly confused last week. Yes, drawdowns happen but this was bubble-pop level of violence and we definitely have a lot more bubble to go.

Early this week, people in the know (me included) knew that someone big (SALP probably) was getting vaporized. Thursday was violent green day because everyone found out the coast is clear and SALP self-immolation had finished. Friday was also up but there was a fade due to Iran war fears. I personally de-levered a bit after hours.

[4] Irrational Analysis Consolidated Perf+Holdings+Records

CSV of trading records YTD here.

Ia Trading July 2026

55KB ∙ XLSX file

Download

Download

[5] STONKS

Presently trying to re-build my trading account. Things are very fluid.

[5.a] Trading Account Names

Lumentum, Intel, Tower remain ultra-core positions.

Bloom and Semtech are core positions. Unsure on what sizing I want to do but going to try and keep this unless something crazy (again…) happens.

Everything else I am clueless on sizing so dont take seriously.

AXTI is a giga-long. if you look at the chart, it started to collapse a month or two earlier than everything else. This is because China banned total return swaps and a bunch of Chinese and Hong-Kong institutions became mini-Lepolds.

AEHR very interesting high beta. Their most recent earnings and guide were a smash hit. I think it still has legs but this ticker is really violent.

Like both Ciena and Nokia but same reasons (long haul optics) but can’t decide which one better or sizing.

AAOI has lot of torque (USA based transceivers) but same problem as Aehr. Violent shit.

And finally we have my one share of Cerebras for the activist campaign. Learned some interesting things about Cerebras and they deserve better coverage in a dedicated post. Also it’s 3 AM and im hungry so wana finish this and go cook.

[5.b] Long-Only Account Names

Free-cashflow from my dayjob goes into the long-only accounts and I am still happy with everything except Samsung. Wana buy a lot more.

Leopold preferred SK Hynix but I like Samsung more for DRAM. Really betting on synergy of internal foundry + design team to co-design HBM base die better than SK Hynix.

Welcome to a super interesting earnings call. Qualcomm Q3 2026 is full of fascinating details.

Yes, I have puts. Hoping to survive IV crush but whatever. Tiny position anyway.

IDK who is buying after hours. Clearly they missed the important info.

Taking screenshots of the relevant part of above post from September 2025 because people are lazy and don’t click links.

=================

=================

=================

=================

Ok now that you have the necessary background information from, let’s go over the 7/29/2029 (today) earnings call and slides.

Slides first.

1. I don’t think China smartphone bottomed this quarter. Strip out seasonality and you will see smartphone (industry wide) continue to implode.

2. Yes, Qualcomm has several long-term contracts and cannot raise prices as fast as they want. However, a particular comment by Akash gives away the bigger issue. More on this later.

3. This is bullshit. Selling modems to Apple is highest margin QCT business. Much more on this soon.

This is a very very bad comment. I am surprised Akash actually said this. Credit for being honest.

I knew down-spec was gona propagate as memory ate more of the BOM. The value of Qualcomm is being diluted. This is a structural shift. When smartphone OEMs realize selling down-spec of this years premium chip or last years premium chip, they won’t go back.

If you can save money on a component and consumers don’t notice, why not keep doing that?

What if the memory issue continues for another two years or more? Smartphone chips have gotten very good. I still use my Galaxy S23 Ultra (QCOM powered) and its great. Need to replace the battery at some point but other than that performance and features and functionality all good. No need to upgrade.

This is the key info.

Apple is accelerating the design-out of Qualcomm modes ahead of the April 1st 2027 licensing deal expiration.

My sense is Verizon green-lit all next-gen iPhones in 2027 to not have mmWave. This was the last real gate to a global Apple modem. Likely because Verizon has finished deploying their C-band purchased in March 2021.

Apple is preparing to re-start their holy war against QTL, as the memory crisis obliterates Android.

In the slides, there was the bullshit excuse of “capacity constraints”. Who the fuck put such a lame ass excuse into official documents? Was it you Brett Simpson?

Stacy called this excuse out and Akash immediately folded.

Do any of you know how painful FCC (and foreign equivalent) wireless certification is? Apple and Qualcomm did not on a whim at the last minute decide this! Apple must have been working on (almost) full QCOM design-out back in March/April and told them at the last minute to make it hurt as much as possible.

Nice so a small revenue from new business severely harms QCT gross margins. Nice diversification guys. Maybe Qualcomm deserves the valuation of a GUC or Alchip. Certainly they deserve a lower valuation that Mediatek, who has a real datacenter business with real scale and real accretive gross margins.

What the fuck you had this huge datacenter investor day event, showed all these fake racks and have no fucking silicon to even test?

You don’t even have a basic prototype test vehicle? It’s just RTL and Virtuoso simulations?

You clowns. 🤡

Change the brand name from Dragonfly to Vaporbullshit.

I think handset is going to get worse in 2027. Possibly even 2028.

The key takeaway from this call is that Apple is preparing for war.

• Silently started FCC qualification of a lineup of iPhones that give QCOM low-single-digit share (foldable only?) versus the prior expected 20% share around April/March.

• Waited until the last minute to inform Qualcomm to maximize pain.

• Banked modems in the meantime to have inventory once the war starts.

Someone please talk to Verizon and ask if their C-band deployment is complete.

Welcome to a very stupid follow-up to an even stupider idea.

For technical background, see this post.

Idiot-proof summary:

• SOA = optical amplifier = a DFB/laser without gratings

• Above point is over-simplified. Designing an SOA has some unique complexities. Pre-amp section, tapering, weird optical beam shape…

• Only a very small number of LiDAR shitcos know how to make high-power SOA.

InP world is apocalyptic. Shortage is so bad. People really don’t understand. Last week I met a new person (whos name I promptly forgot) and he gave a nice intuitive analogy as to why InP situation is worse than DRAM.

“Only 25% of the worlds DRAM goes to AI. Can just take more from consumer. 80% of InP already going to AI.”

Ignore the absolute percentages he said. The core argument is true. InP has ZERO use-case for consumer markets. DRAM… plenty. We can always just (metaphorically) shoot the smartphone industry and annex their DRAM allocation. Can’t do that with InP.

Suppose you have a design objective to make an ELSFP module for NPO/CPO that has the following core attributes.

• Single wavelength (1310nm)

• 22 dBm power coupled to each output fiber.

• 8 output fibers.

Here is my mock-up of the normal way (8x discrete DFB) and the LiDAR shitco way (SOA strat).

It’s not completely stupid.

Let’s do the math to figure out:

• BOM + assembly cost

• InP intensity

• gross margins

I will post screenshots of the model and a full XLSX for spreadsheet enthusiasts to play with.

As you can see, a 650 mW SOA has better InP area efficiency (more photons per unit area of InP) but this basic model at a device level misses two critical attributes.

1. You still need a CW laser to seed the discrete SOA.

2. Coupling losses from many lenses and isolators hurt a lot. Especially coupling losses after SOA output to accommodate beam shape issues.

Now on to product-level (ELSFP) cost modeling.

Normal way gets 51% gross margin while the LiDAR shitco way gets only 17%.

Given that LiDAR is fucking dead, I think they would be delighted to have massive revenue even if the margins are ass.

The math works. Kinda hilarious to be honest.

Here is spreadsheet. Have fun.

Irrational Analysis Inp Model Lidar Shitco Edition

1.24MB ∙ XLSX file

Download

Download

So now that we know this stupid idea can work, how to play it?

As I have said many times, only LiDAR losers know how to make high-power SOA. The usual suspects (Lumentum, Coherent, Broadcom, Sumitomo, Furukawa, AAOI) do not know how to make this niche device because they have been too busy making money in real markets.

The vast majority of LiDAR shitcos (both public and private) are fucking liars. I don’t want this to end up being a situation where people ping me “what about this one”.

There are only two that have viable, high-quality SOA appropriate for NPO/CPO market. One is privately held IDM (designs and manufactures) and I’m not going to tell you their name because reasons.

The other is who manufactures with Seivers as the InP fab partner and whose public claims (which I have independently verified) were used to build the COGS and InP model.

Thank you to the anon person who pointed out the public info that strongly indicates Aeva uses Seivers.

Seivers is one of my mute works on X/Twitter because I find their fans (particularly the Serenety simpleton) to be quite annoying.

Getting tired of shooting down their 8/16 wavelength monolithic array and supposed link to Ayar/AMD. Idiots, Lumentum cut everyone except Nvidia off. Ayar did not remove Lumentum from their website. Lumentum removed Ayar.

Go fucking ask Matt Sysak what is going on.

You can also ask literally anyone who has actually used the Seivers DWDM laser or Ayar Labs Supernova what they think. Everyone in industry knows what is going on and you clowns keep making the opposite conclusion from reality.

Frankly, I can still hate Seivers DWDM product and be mildly bullish the company because of Aeva SOA. The yield on the DWDM monolithic array is horrific. Seivers will may way more money (in revenue $ and gross margin $) off Aeva than they ever did off Ayar Labs. LOL.

But frankly Aeva more interesting buy than Seivers and more interesting that both of them.

Have fun degenerates. I only own Lumentum, AXTI, and AAOI.

Not bothering with this crap but dogshit is fun to write about.

This post is directly inspired by ’s excellent deep-dive on InP DFB laser physics.

Vik's Newsletter

Lasers for CPO/NPO: Part 1 – The InP DFB Laser

There is a lot of FUD in the market about ultra high power (UHP) lasers for CPO and who the leading provider is in this segment. Everybody I’ve asked about UHP lasers unequivocally says Lumentum is in the lead, but few are able to discuss the engineering challenges in making UHP lasers, where the Lumentum advantage lies (if it even exists), and what the real alligator-filled moat is that keeps competitors out…

Read more

20 days ago · 40 likes · Vikram Sekar

Go sub to him to get the next part of the series as soon as it arrives.

Broadly, I agree with 90% of the content and just have a few moderate differences in opinion.

First I want to re-cap his work before pivoting to semiconductor optical amplifiers (SOA), a LiDAR and long-haul niche device that may have new use-cases in ultra high-power CPO laser market.

And yes, I’m fine.

Sometimes I like to zoom out to remind myself I am not completely incompetent.

Contents:

1. Vik’s Newsletter DFB Recap

2. “Normal” Optical Amplifiers: EDFA and PDFA

3. Semiconductor Optical Amplifier (SOA) Basics

4. Four-Wave Mixing (FWM): Blessing and Curse

5. SOA for CPO/NPO Market: Featuring Aeva, Active Alignment Hell, and Back-reflection

[1] Vik’s Newsletter DFB Recap

My goal is to describe things from a practical perspective. Physics blah blah blah. Let’s just focus (pun intended) on how thing work and why we would choose to use thing.

DFB lasers have two parts.

Active region converts electricity to light.

Gratings (tiny mirrors) bounce the light back and forth, recursively amplifying it until the light blasts past the last mirror out of the laser causing it to go pew pew pew.

Mode-hopping is not just undesirable. It is death. Most transceiver link flaps are laser mode hops. People go to great lengths to design test lasers and find “mode up free” regions in the LIV curve.

Temperature drift effects wavelength much more strongly than drive current. As a rule of thumb, 0.1 nm/kelvin drift is common. As Vikram mentioned before, the grating design/spacing determines the laser wavelength, but there is manufacturing variation Thus in a real system, heavy temperature tuning is necessary. Imagine a ELSFP module with 8 wavelengths. Each DFB is designed to operate at 40C but may be operating at any temperature within +/- 10C at up to two decimal point accuracy. 39.85, 45.11, blah blah blah.

Everything Vikram wrote is correct but he is missing two critical topics.

Laser reliability (does thing burn itself out after running 6 months) depends on charge density within the cavity. A shorter cavity at the same output power means higher charge density and serious reliability concerns.

Mode-hop risk (how likely the laser is to flicker) scales with cavity length.

This is absolutely critical to understand. Lumentum and Broadcom can achieve acceptable linewidth for CPO/NPO applications at around the same cavity length. Not only does this save them money (more chips per InP wafer), it also drastically reduces mode-hop risk. All competitors need around 60% longer cavity length which makes the lasers at far higher risk of mode-hop/instability.

[2] “Normal” Optical Amplifiers: EDFA and PDFA

Suppose you have an optical signal and need amplification. What is the default solution?

If you are in C-band, this is called EDFA.

If in O-band, this is called PDFA.

They are usually depicted as a spool in block diagrams.

This is because EDFA/PDFA are indeed long-fibers that have been doped with fancy materials.

Broadly, the performance of these amplifiers is excellent. Lots of gain and very low noise figure.

The problem is cost and physical footprint.

[3] Semiconductor Optical Amplifier (SOA) Basics

SOA is basically a DFB laser without any gratings.

This type of amplifier is more noisy than a PDFA/EDFA but is tiny and dirt cheap (relatively). It also has some properties that are either a blessing or a curse depending on what you are trying to accomplish. More on that later.

First let’s go over the basics on how these SOA things work and behave.

ASE = amplified spontaneous emission = crap SOA outputs if you turn it on but dont have any input optical power.

ASE goes down if you seed with something at the input.

Traditional applications of SOA (long-haul comms, LiDAR) involved seeding a very weak signal (say -25 dBm to -5 dBm) such that the small signal gain of 20-30 dB boosts the output to around 13 dBm at most, typically lower.

Think using a demux in a long-haul DWDM system to pick out one wavelength and amplify it.

Note that only one wavelength of light enters the SOA in the vast majority of traditional applications. This is important.

[4] Four-Wave Mixing (FWM): Blessing and Curse

All amplifiers have intrinsic (common sense) impairments.

1. Amplify input noise.

2. Add additional noise. (typically from driver)

3. Add non-linearity.

Optics is different because fuck you. There are these “non-linear” impairments. I have no idea what non-linear means but let’s roll with it.

One of the more notorious non-linear impairments is four-wave mixing (FWM). This process happens in basically every medium (glass fiber, SiPho PIC, SOA, EDFA, whatever) but is much stronger in certain situations.

FWM shows un in optical fiber…. after a few kilometers.

FWM happens in SOAs in a meaningful severity very quickly.

FWM scales quadratically/cubicle with input optical power and cavity length. It also scales based on how close your wavele3ngths are together. Tighter spacing is exponentially higher FWM efficiency.

So what am I talking about? Let’s look at an example where two wavelengths enter and SOA.

Intuitively, you would think the output of the SOA would be those two wavelengths amplified with some broadband noise assed.

WRONG.

Because of physics reasons, some of the photonics teleport to create two extra junk spikes on the sides. The spacing of these junk spikes depends on the spacing of the input signals. Also, these spikes have conjugate phase, which means the frequency domain spectra is flipped. Unless you are intentionally trying to use this information to back out chromatic dispersion, this is bad. Conjugate phase noise spikes is not helping the situation.

FWM is either a blessing or a curse. You either want this to happen in order to accomplish useful tasks such as wavelength conversion.

Or you want to avoid this shit like the plague.

CW laser amplitude noise (RIN) is defined as an integration from 10 MHz to Nyquist frequency. The lower 10 MHz cutoff is due to photodiode and TIA responsivity. Any noise that slow in theory does not matter.

[5] SOA for CPO/NPO Market: Featuring Aeva, Active Alignment Hell, and Back-reflection

So far 5 investors have pinged me about , a LiDAR shitco, because sell-side picked up on this SOA thing. Let’s back up a bit.

Traditional long-haul datacom systems only need 20 to maybe 100 mW output power.

LiDAR needs much more power so all the high-power (400 mW and above) SOAs on the market are from degen shitcos desperately trying to pivot. Aeva is one such pile of flaming dogshit.

The Aeva SOA is quite good. Rated to about 650 mW output power with excellent properties (WPE, low ASE).

In theory… someone could take weak 20-100 mW discrete DFB lasers and just put an Aeva SOA (or similar) behind it to amplify up to 400 mW. Boom, CPO/NPO viable laser using crap low-power lasers.

There is nuance. Calm down finance friends. It’s really funny how excited finance people get over detritus that has a chance to re-rate from a total business pivot.

There already is a strategy called MOPA (main oscillator, power amplifier). Take one monolithic InP chip. Print a low-power (100 mW) DFB then add a short waveguide then add a SOA.

Furukawa has a MOPA targeted for CPO/NPO. Coherent is pivoting to MOPA after botching their UHP DFB.

So why is this MOPA thing not super common? Why are the two dominant UHP laser players (Broadcom, Lumentum) not using this strategy?

First, it costs more. A lot more InP area.

But the real reason is HORRIFIC MODE HOPING DUE TO THE CAVITY BETWEEN THE DFB AND THE SOA.

DFB lasers are very very fucking sensitive to optical back-reflection.

Any light that gets reflected backward goes into the gratings and destabilizes the shit out of the laser.

This is why there is always a isolator between the DFB and whatever it is shining light into.

Optical isolators are (well try to be) one-way gates that let light go forward but not backwards. You can’t have an isolator on a monolithic InP MOPA chip because of manufacturing limitations. The isolators are made of materials not InP. Terbium whatever.

https://www.coherent.com/news/glossary/faraday-rotators-and-isolators

Fun fact, makes around 60% of the worlds isolators I think.

Anyway. MOPA bad because mode hop issues that are kinda unfixable. All you can do is mitigate and pray this shit passes reliability testing.

Using a discrete SOA in conjunction with a DFB solves this issue. So why has no one done this? The answer is simple. Cost.

You see that? Four lenses. The isolators are missing from the above diagram too. An isolator between the DFB and SOA input is a must. Isolator on the output you can skip in theory but it’s tricky.

Each element in an optical path (DFB, lenses, isolator, SOA, fiber, PIC, whatever) generally requires active alignment, an expensive process that requires specialized equipment (FiconTec or custom) to align with low single-digit micron accuracy and less than 2 degrees angle accuracy.

This shit is expensive and it gets out of control real quick. Just with some napkin math, a 8-wavelength ELFSP module using shitty lasers and Aeva (or equivalent LiDAR shitco SOA) requires…

• 8x 20-100 mW DFB

• 8x 400-650 mW class SOA

• 32-40 lenses

• 8-16 isolators

• 56+ active alignments

On the surface, this seems completely retarded. But… the CPO/NPO market is so starved that any functional supply will sell. If we call Broadcom and Lumentum module-level (ELSFP) gross margins at 50-65%, I can see a half-dead LiDAR shitco partnering with TFC or Fabrinet to sell the same thing with worse power efficiency at 5-20% gross margin.

If anyone knows which InP fab (please let it not be CPFC) Aeva uses or their SOA cavity length, please let me know. It is trivial to model COGS and yield of this stuff.

Quick note on an easy topic as I spend the day watching world cup quarterfinals at a bar.

So many stupid takes on CPO and NPO recently. CoWoS counters have graduated to COUPE counters. Idiots still think Coherent’s UHP laser is viable. And Ashkan’s legendary career choice is being mis-interpreted.

Saalam Ashkan! Mobarak!

Contents:

1. CPO vs NPO: For Dummy Supply-Chain Degenerates

2. Re-affirming Lumentum Superior UHP Laser

3. VCSEL CPO: The Ashkan Arc

4. Jihad Against Jim Anderson

[1] CPO vs NPO: For Dummy Supply-Chain Degenerates

A lot of idiots who understand nothing on the engineering of advanced optics (CPO, NPO) are spouting bullshit while blindly re-gurgitation Taiwan rumor mill.

To be clear, CPO and NPO are conceptually very similar. NPO has worse power efficiency and worse channel density. But the underlying tech is very similar.

TSMC seems to have botched their high-density 2D grating couplers. That issue, alongside delayed SiN PDK development has given a temporary advantage to Tower Semi. Nvidia moved their program from TSMC to Tower and everyone who actually follows this space properly knew 6+ months ago.

The downside of a Tower-based NPO solution is twofold.

1. Lower channel density. (need more wavelengths)

2. Worse power efficiency

#1 is because of longer electrical channel with more reflections (bump capacitance is killer) requiring strong electrical SerDes. More EQ and more complex driver.

#2 is because of the previously mentioned items + exponential laser power and noise requirements. To compensate for fewer channels, modulation speed has to go up. This means SNR requirements go up and laser burden also goes up.

There exists a sweet-spot for datarate of a CPO system between 32G and 64G NRZ. Personally I prefer 64G NRZ but there are arguments for any datarate in that range.

Too slow and your heater power explodes. Too fast and the laser power + SerDes power explodes. These are not linear penalties.

Nvidia has decided to temporarily move away from TSMC COUPE. This is because TSMC has dragged their feet on SiN and (more importantly) botched 2-D grating couplers.

Plan-A was slow-and-wide (~50-64G NRZ) 8-wavelength DWDM on TSMC COUPE CPO .

Plan-B is 200/400G PAM4 on 16-wavenlgth DWDM on Tower SiPho NPO.

Everyone who follows this space properly knew this 6 months ago. The entities who blindly peddle Taiwan leaks with zero understanding of the underlying technology have been spreading panic derived from their own incompetence.

At higher modulation rate, the noise requirements (RIN, linewidth) of the laser go way up.

Laser linewidth murders the extinction ratio of ring modulators.

This means longer cavity length, bigger InP chip, and new (higher price) SKU for Lumentum. Absolutely fucking nobody other than Lumentum and Broadcom can meet the laser specs needed for Nvidia NPO. Retards, content goes way up with Nvidia moving from CPO to NPO. Reality is the opposite of what you pod monkeys think.

You can see I trimmed Tower for risk management reasons but not a single share of Lumentum was sold this year. I dare every institutional investor reading this to short Lumentum. Free market will decide who is right, me (engineering-driven investment analysis) or the fools who parrot Taiwan leaks using ClaudeChatGPT.

[2] Re-affirming Lumentum Superior UHP Laser

The data is public. Only Lumentum’s data mind you. Everyone else is a coward who wont publish proper information outside of NDA.

Lasers can hit a wide range of power levels depending on the temperature. For example, Lumentum’s UHP (ultra-high-power) laser is officially rated for 350mW of output power with a soft cap at 400 mW. You can run it at 450mW but reliability becomes questionable and wall-plug-efficiency (WPE) AKA power conversion efficiency suffers.

The output power and efficiency depends strongly on temperature. Higher temperature means less power and worse efficiency. Most people choose to operate the hot-side of the TEC (thermo-electric-cooler) at 50C and the cold-side at 40C. This provides a good balance between system-level thermal efficiency and laser performance. At OFC 2026, Lumentum’s demo was at 30C cold-side TEC setpoint to make themselves look better. I am telling you 40C is the correct (reasonable/realistic) comparison point.

Lumentum also shows public noise data across output power. Nobody else has done this. There is a reason for this.

I would like every single buy-side and sell-side analyst to print out the above Lumentum slide and request Coherent to share the same data in the same format under the same test conditions.

I know the answers. No need to send me your notes.

Coherent has failed in the UHP laser market. They know they have failed. This is provable given their roadmap.

There are two ways of making a high-power laser.

1. Make big laser.

2. MOPA

MOPA stands for main-oscillator power-amplifier. Essentially it is a smaller 100mW laser feeding into an amplifier all on the same InP chip.

The benefit is the noise of the final output depends on the smaller 100mW laser and is thus much easier to manage.

The drawback is the monolithic InP chip is much larger and there is significant mode-hop (flickering/instability) risk due to the cavity between the DFB and the SOA.

Furukawa already has a MOPA for UHP market. Coherent is working on it as a pivot given how dogshit their existing 400 mW laser is.

Lumentum and Broadcom are running circles around literally everyone else. There is no Chinese competition. I dont give a fuck what the rumors about JY 300mW DFB are. Show me the linewidth. Show me the cavity length. Show me the WPE curve.

Jim Anderson is bluffing. He has nothing until the MOPA is ready next year and even then it might fail due to mode hops.

[3] VCSEL CPO: The Ashkan Arc

Ashkan Seyedi is a very important person. If you worked in optics world, you know who this guy is. He previously was the head of Nvidia optics before Jensen poached Ling Liao from Intel to take over. Ling and Ashkan are both very smart but Ling is sharper IMO.

Within Nvidia, there were two factions. Team ring-modulator (Santa Clara) and team VCSEL CPO (Isreal/Mellanox). Team ring-modulator won last year. Ashkan was a part of team ring. Go look at his academic paper publishing history.

https://scholar.google.com/citations?hl=en&user=TVhiqXwAAAAJ&view_op=list_works&sortby=pubdate

His PhD was on GaAs nanowire photodiodes arrays (for micro-LED systems) and all his recent work is on ring-modulator DWDM. Ashkan is a vocal hater of micro-LED.

So when this Nvidia hot-shot defected to a dying euro-poor LED shitco, every optics tourist immediately assumed he has become a micro-LED supporter.

Fundamentally what you need to understand is micro-LED and VCSEL CPO/NPO are conceptually very similar. The key delta is max datarate.

LEDs are limited to 2-5 GHz bandwidth, with a (very generous) cap around 10 GHz is the LED people come up with something groundbreaking.

VCSELs meanwhile can hit 50-70GHz bandwidth without issue.

Optics people often ignore electrical reality. I don’t give a damn how nice your shitty LEDs are. Having a huge number of lanes and single-digit GHz will result in disastrous EMI and crosstalk. It won’t work without a super-heavy FEC on all the hundreds of fucking lanes. Anyone with a basic system-level understanding realizes this immediately.

VCSEL CPO/NPO is literally what micro-LED wishes it could be. Objectively better in every way except cost.

Yes, micro-LED is cheaper. It also does not and never will work at scale so who the fuck cares.

AMS OSRAM is a dying company. They have both LED and VCSEL but in reality most of their business is LED. That existing business (LED lighting) has been annihilated by the Chinese.

What you have to understand is VCSEL CPO/NPO quality depends on the quality of the packaging, integrated drivers, manufacturing, and yield of the system. The VCSEL quality itself is not that important. Many can make a 32-64G NRZ VCSEL.

Broadcom is the king of VCSEL. Coherent has very good VCSEL. In fact, Coherent’s VCSEL NPO demo at OFC 2026 was excellent.

Lumentum, Furukawa, and many others also have good VCSEL.

My gut feeling is Ashkan saw something in AMS OSRAM historical LED packaging expertise and decided this was his chance to make VCSEL CPO/NPO work at scale. He is very smart. He knows something. I am certain.

They have a packaging/integration ace up their sleeves. Something to take meh internal VCSELS and make a real NPO/CPO solution that can credibly compete with SiPho DWDM rings.

[4] Jihad Against Jim Anderson

I am not short .

It’s just I don’t like their CEO.

Trash talking semis CEOs amuses me.

A particular moment in the last earnings call highlights why I do not like Jim Anderson.

The entire point of SiPho transceiver over EML is to sacrifice modulation performance in favor of saving money by reducing InP content. Historically, every datarate transition (100G to 400G to 800G to 1.6T) resulted in the same pattern. EML first because it has the best performance. Then 6-18 months later SiPho transceivers show up (usually from the Chinese) that are good enough performance but meaningfully cheaper.

This market is different. InP shortage is so fucked, SiPho took majority share much earlier.

THE GROSS MARGINS OF EML BASED AND SIPHO BASED TRANCEIVERS SHOULD NOT BE THE SAME. SOMETHING IS VERY WRONG. EITHER LUMENTUM IS OBLITERATING YOU BY SPIKING EML PRICES CAUSE YOUR 6-IN INP PROCESS IS BROKEN OR YOUR ACTIVE ALIGNMNET YIELD OF SIPHO TRANCEIVERS IS HORRENDUS. OR MAYBE BOTH.

Jim Andersion has been CEO of Coherent for 2 years and he still lacks a basic understanding of the business he is running.

I’m pretty sure sell-side knows more about the business. That’s why the Citi dude asked the question. He probably has followed optics for many years and knows the gross-margin profile of different types of transceivers.

So when the same dude who does not understand transceiver basics claims he has a viable UHP CW laser for CPO/NPO, yea I don’t fucking believe him.

Print out the Lumentum slide showing detailed data on UHP laser and ask Jim Andersion to share the same data in the same format under the same conditions.

Ask him how GR-468 qualification of 200G VCSEL devices is going.

Ask him why EML and SiPho transceivers have the same gross margin.

Ask him what the relative yield of 200G EML on 3-in and 6-in is.

Ask him what the relative yield of 100/250/400mW CW laser on 3-in and 6-in is.

I am obsessed with the Matrix movies. Only the first three. Fourth one is garbage.

Recently re-watched them and one particular scene in the second movie has seeded a splinter in my mind, driving me mad.

The world has changed in the last two years. Micron is at 86% gross margins. AI coding tools (allegedly) make it much easier to port software to alternative (not Nvidia) architectures.

And yet, the world has not changed. Nvidia still has 75% gross margin which is applied to the DRAM too mind you. Nvidia still has vast majority of CoWoS allocation and new supply-chain moats like that deal with Lumentum for lasers. Mr. Leather Jacket is still the king.

Today’s incoherent bullshit writing will be largely abstracted from technical matters for a variety of reasons. Humor my vague philosophical nonsense.

Contents:

1. A tale of two startups. (who almost died)

2. Know Thy Enemy. Structural [Dis]Advantages.

3. Optimism is a disease. Death by default. People problems.

4. Are you solving the right problem?

5. Life lessons from the wrong places.

6. Stealing Scraps from the King

[1] A tale of two startups. (who almost died)

Groq and Cerebras are quite interesting. Those of you who follow hardware startups know that both of these chip companies went through ugly periods. Near death moments.

And yet, Groq got a $20B exit via Nvidia deal and Cerebras managed to IPO at a > $40B valuation on tiny float. We shall see what the valuation is after the insiders finish dumping shares but it will probably be at > $20B.

Thanks Altimeter for letting us know the schedule on which you will be unloading your bags to retail.

Jokes aside, both Cerebras and Groq got excellent exits given the very ugly place they were in just a few years ago.

A funny problem that happens from time to time is when correlated sources tell me the same thing.

For example, early this year three hedge fund contacts told me Lumentum could do $60 EPS in 2028. When the fourth hedgie told me this I asked, “did <retracted> tell you this?” Answer was yes. All four of these people were parroting the same source.

Rumors tend to be correlated and propagate from the same one or two people. I generally try to get 3-4 uncorrelated sources before publishing something.

Uncorrelated generally means one or more of the following is true:

1. I know these people don’t talk to each other.

2. Background variety (Person A is finance, Person B is engineering)

3. Bias variety (Person A hates corp XYZ, Person B likes corp XYZ)

4. Information asymmetry. People tell me different things that only lead to a more interesting conclusion if you know all the info.

Ok enough with the background. Juicy conclusion time.

I am 98% confident that Nvidia made their deal with Groq primarily in response to the OpenAI/Cerebras deal.

Alluded to this in the Cerebras IPO report because at the time there were only two uncorrelated sources. Got a third one and now feel confident enough.

For a variety of reasons, I am absolutely certain these three individuals do not talk to each other.

One contact is finance, the other two are engineering.

One contact has strong bias in favor of Cerebras, one contact hates both Cerebras and Groq, and the third has a general dislike of both but is mostly indifferent. All three are regular readers.

Most importantly, the information nuggets provided by each of the three are completely different. In isolation, each scrap of info means little. But together, they corroborate a far more interesting story…. perfectly.

Mr. Leather Jacket was caught off guard and needed a solution to Cerebras fast inference urgently. Credit to him, Jensen was decisive and managed to get a full product (LPU rack for disagg inference) together in extraordinary time. Most executives would spend months or even years in denial, refusing to accept the competitive threat is real. Jensen’s actions here show why he is the greatest semiconductor CEO of all time.

But he was caught off guard. Keep that in mind.

[2] Know Thy Enemy. Structural [Dis]Advantages.

A mistake I see frequently is startups focusing on other startups as competitors.

Let me be very clear.

The entity that kills you will be a large corporation, not some other startup.

In the case of AI ASIC, there are four comparative threats, in the following order:

1. Nvidia/GPU

2. Google/TPU

3. Amazon/Trainium

4. AMD/GPU

These are the only four entities that matter. Four well-funded megacorps with massive structural advantages.

• Preferential access to HBM, advanced packaging, and other supply-chain items.

• Volume discounts, warrants, financing deals, and other bean counting tricks.

• Much larger teams that can try more ideas.

• Insourced key IP (SerDes, custom standard cells, controller digital IP, custom tooling).

• Experienced commercial teams who know how to scale into mass production.

• Funds to spam tapeouts.

• A single lab (one of many) with equipment worth more than your entire funding round.

• Better relationships with powerful people.

• Bundling opportunities.

If you (AI ASIC startup employee/founder/investor) think competition from anyone other than these four (NVDA, GOOG, AMZN, AMD) matters, you fundamentally do not understand business reality. You are a moron and I shall die upon this hill.

[3] Optimism is a disease. Death by default. People problems.

Recently, I took a daytrip up to SF to catch up with a friend. This friend is much older and wiser than me. He has worked at megacorps and startups. Been an IC, manager, and VP/executive. We share personality traits to an extent.

These days, I get too many emails to respond to, but I make an effort to respond to students. There are many things I wish I knew when I first entered the workforce.

This section is somewhat skewed. Those of you who have been working a while should know most of these things. For everyone else, I hope this helps you in some way.

I have never seen a project fail due to technical or engineering reasons.

This statement may sound strange but it’s true. It only makes sense if you understand my definition of failure.

Every project starts with resources. People, capital, IP, equipment, whatever.

If those resources are used to find the optimum outcome in a reasonable amount of time, that is a success.

If the optimum outcome is “the technology does not work but we found out quickly and pivoted to a new project”, great! That is not failure.

If a competitor out of the blue came up with something innovative that invalidates what you are working on, that is fine, so long as you recognize this quickly, kill the project, and re-assign resources to something that has a chance.

My job function involves talking to (and arguing with) people who are at least an order of magnitude smarter than me within their specialty on a regular basis. Here are some amusing observations I have come to realize over the years.

• None of these people talk to each other or understand what other disciplines/departments are doing. This leads to hilarious problems. Seriously I have so many genuinely funny stories of genius-level people (within their specialty) completely failing to understand basic concepts outside their niche. It’s almost as if most people have no interest in exploring other fields. Maybe that’s why I am good at my job. Too stupid to be good at anything in particular but obsessively curious enough to be mediocre at everything and ask experts the right questions.

• Positive bias is a huge problem. Everyone is intrinsically proud of their own work.

• Nobody properly tests their own work.

• Factions always form within a project. Does not matter if it’s within a megacorp or tiny startup. People naturally congregate and start engaging in dumb tribalistic behavior. It’s human nature. Good management minimizes this phenomenon while bad management accelerates it like gasoline on a dumpster fire.

• Nobody wants to be at fault, but somebody is always at fault. Root causes always have a technical and human endpoint. This does not mean the person(s) to blame for a particular fuck up should be punished. Just that fault assignment always exists. “No one is to blame” is an attitude for invertebrate losers.

So, what kills projects? One word.

Optimism.

Optimism is a disease that must constantly be attacked, especially at a startup.

Think of this as my version of “only the paranoid survive”.

Every project needs a “corporate terrorist”. Someone who does not give a damn, aggressively delivers bad news, challenges optimism, and forces serious discussions about the core viability of the product.

The trick is to put people on the spot in a forceful way without pissing them off (too much). Some controlled hostility is both necessary and beneficial.

Problem is not your fault? Prove it or help me prove it. Don’t deflect to others. I’m talking to you. Heat is on you. The only way to get rid of me is to help me prove or disprove your involvement in the ongoing disaster.

A little targeted internal panic goes a long way.

Projects fail because of people problems. Root causes always have a technical and human endpoint. Attack both. Be harsh but fair.

Startups are a little different in that there are some extra (advisable) psychological best-practices.

1. Working at a startup is a high-risk life choice. Behave accordingly.

2. The default outcome of a startup is death. Every day you should remind yourself of this.

3. You are always behind the competition.

4. The competition has mulligans. You don’t.

5. The competition is large corporations, not other loser startups who don’t have anything. If any of you had something, you would show financials publicly in a SEC S-1 filing and IPO.

6. Your equity is a zero until liquidity. No liquidity, no value.

7. Startups do not have the luxury of time for the usual level of corpo bullshit. Move fast or die.

8. Large profitable corps can afford to engage in the sunk-cost fallacy. Startups cannot and must cut losses aggressively. Wasting money on a dead project threatens the survival of a startup.

[4] Are you solving the right problem?

AI inference is extraordinarily memory bandwidth bound. If you are not solving the memory issue in some way, I’m not sure you are solving the right problem.

If your design uses HBM, you have chosen to accept a major structural disadvantage.

Good luck getting allocation or prices remotely close to what the big four can achieve.

Advanced packaging is another major structural disadvantage for similar reasons, but Intel EMIB ramp is helping to an extent. HBM situation is fucked now and still going to be fucked for at least another three years.

Have you ever been involved in BOM cost analysis at 10K, 100K, vs 1M units volume?

I have. Small players get squeezed. Paying double vs the big guys means you lucky.

So, if you [startup] have designed an AI ASIC that needs HBM, are you solving the right problem?

*I wish my friends at Etched and MatX all the best.

[5] Life lessons from the wrong places.

I take life lessons from video games and movies. Probably not a good thing.

Three in particular stand out.

• Far Cry 3 (Ubisoft)

• The Matrix

• Dishonored (Arkane Studios)

The corporate world is insane. So many people doing the exact thing over and over again expecting shit to change. Perhaps many of them have given up and are running on autopilot.

There are these personality architypes you quickly learn to spot. The empire builders who derive their self-worth from how many people report to them. The status collectors. The brown-nosers. The bullshiter who talks a lot, sounds important, but contributes nothing while slowing everyone else down. The dead weight that was hired so some manager could get promoted. The H1B indentured servants who are too afraid to leave.

The vast majority of large corporations are full of worthless do-nothing projects that exist to feed the egos of parasites wearing suits and justify the titles of worthless middle management. Corporate boards of directors are almost universally incompetent and spineless.

My views on corporate America are similar to Agent Smith’s view of the Matrix.

I hate this place… this zoo… this prison… this reality… whatever you want to call it.

I can’t stand it any longer.

It’s the smell, if there is such a thing. I feel saturated by it.

And finally, we have Dishonored, my all-time favorite video game. The premise is interesting. You play as an assassin in a city overrun with a rat plague. Think bubonic plague but there are swarms of rats roaming the streets eating people alive.

Interestingly, the optimal way to play is to not kill anyone. Every target has a non-lethal option. Have you heard the phrase “there are many fates worse than death”? Yea, that concept is taken to the extreme. Your choices also matter, leading to darker outcomes. Lore reason for the “chaos” (morality) system is more people you kill means more dead bodies which means more rats and more plague.

Dishonored is a special kind of game called “immersive sim”.

https://en.wikipedia.org/wiki/Immersive_sim

By far my favorite genre.

So why am I telling you all this?

There are many fates worse than death. For example, spending your entire life in a corporate zoo, exchanging your sanity and dignity for money.

The point of a startup is to take a risk and escape.

Working at a startup should not be just a job. It should be your life.

With great risk comes great reward.

An escape from insanity. Purpose.

Fuck the (theoretical) money.

[6] Stealing Scraps from the King

Attempting to compete with Nvidia seems like a suicidal endeavor. Mr. Leather Jacket is king, seemingly immortal.

Amusingly of all the people to show the king can bleed, it ended up being the Cerebras clowns.

The thing is, basic math tells you that only stealing a few scraps from Mr. Leather Jacket’s banquet table is enough.

Let’s call Nvidia’s terminal revenue at $500B/year at 75% gross margins.

AI ASIC startup captures 0.5% of the market. —> $2.5B/year

Subtract $500M opex. Assume 50% gross margins on the remaining $2B revenue.

$1B operating profit per year * 40x bubble multiple = $40B market cap

We are on the dumbest timeline.

The golden age of unhinged silicon.

Take your shot or forever wonder what could have been.

The king bleeds, long live the king.

_{Did I ever tell you the definition of insanity?}

Disjoint note on some (mostly) unrelated ideas.

Contents:

1. Applied Materials DRAM and Advanced Packaging Master Class

2. Erock initial analysis.

3. HBF Updated Views

4. VCSEL CPO Updated Views

5. 800V DC Practical Limits

[1] Applied Materials DRAM and Advanced Packaging Master Class

“Master Class” == technical marketing event

It was very well done and I enjoyed it a lot. Fair warning, I am a material science and process technology tourist. However, AMAT appears to be an optics tourist lol. You will see what I mean.

Good overview. Note the very tall capacitor. This is the main delta between DRAM and logic nodes.

EUV intensity is indeed going way up in DRAM. This is one reason I own ASML in the long only book. AMAT is arguing that this trend also benefits them due to stricter etch requirements.

Another argument is that AMAT is bringing SiGe Epi to DRAM world. Sure this seems believable and valuable.

Array and periphery disaggregation in DRAM seems like a stretch. At the very least far away. A subset of NAND world (Sandisk+Kioxia) has started to do this. Unsure if DRAM world will adopt this technology anytime soon.

Skipping vertical cell transistor and 3D DRAM sections as that stuff is far away and speculative.

I would like to know what the hell AMAT is smoking. 0.05 pj/bit for CPO? LOL

SiPho CPO is 2-3 pj/bit ideally. VCSEL CPO is 1 pj/bit.

Fundamentally, if you ignore everything else (laser, PIC heaters, driver) a typical SiPho CPO system will have hybrid bonds and TSV in the datapath. Applied Materials bullshit numbers on same slide imply CPO should be 0.6 pj/bit at a physical interface level. Nonsense numbers.

Ok well thinner DRAM dies means hybrid bonding pushed out. AMAT wins either way I guess but perhaps this is not the message they want to send given their stake in BESI.

Idiots. The laser is 99% of the time NOT IN THE PIC. Do some research next time this is embarrassing.

Waveguides do not go vertical. To make light go up, you use a grading coupler. Somehow the AMAT tourist clowns who made this slide depicted fiber attach in the wrongest way possible. Vertical waveguide into an edge coupler that somehow has z-axis angle tolerance as well.

Ok so eBeam beats optical inspection. What about Rigaku x-ray inspection? Someone go as AMAT IR about this and send me your notes I am curious.

[2] Erock initial analysis.

Taking a first pass at their S-1. If anyone has more info please send it my way via email.

Fast response natural gas generators.

Conveniently left Bloom out of the comparisons.

Dogshit gross margins. Nice.

Sigh the story seems to be recurring service revenues. Yay something I can’t properly think through.

Ok fuck this the S-1 has no useful information. Time to go to patents.

They appear to have one patent. It is for a engine with a tall vertical cooling stack.

HBF Updated Views

I have had a virulent hatred of HBF (high-bandwidth flash) ever since Sandisk presented the idea in their investor day. So before the spinoff from WD.

Several have pushed back and I would like to thank these people for providing useful feedback.

Most of my opposition was due to endurance concerns. I assumed HBF was going to use TLC or QLC. The people who pushed back tell me SLC is going to be used.

Using SLC eliminates most of my concerns regarding HBF viability. The one remaining issue (big problem…) is digital logic integration.

NAND needs a lot of fancy math to perform at it’s best. These algos typically run on a dedicated logic controller. In the case of HBF, this logic needs to be integrated into the base die. Space will be tight but it seems doable. What concerns me is heat. The logic base die of a HBF stack is gona kick out a lot of heat. Likely on the order of 10 watts. Sandisk and SK Hynix (the two companies working on HBF) have a difficult problem to solve and I see no public evidence of progress on this front.

Keeping an eye on this space but for now, softening stance from “I hate HBF” to “it could work but thermal issues need to be solved”.

VCSEL CPO Updated Views

Within Nvidia, there are two camps.

Santa Clara group is pro ring modulator SiPho.

Isreal (Mellanox) is pro VCSEL CPO.

Last year Santa Clara won. The roadmap for next 2-3 years is set in stone. This is true for most of the industry. SiPho based NPO and CPO is ramping across Nvidia, Broadcom, Innolight, Eoptolink, Accelink, and AMD. Microsoft and Amazon both have committed programs. InP CW laser supply is an unmitigated disaster. Three of the entities in the above list are in a panic trying to find supply.

But this move by Ashkan is very interesting for the future. He was a part of the Santa Clara group, team ring modulator. Additionally, if you are plugged into this space, you know that Ashkan’s PhD thesis is on micro-LED tech and he publicly shit on micro-LED at a conference last year in his keynote address. It was hilarious. Avicina had a booth with a (horrific) live demo and Ashkan was just roasting micro-LED on stage as the highest profile speaker.

For now, nothing changes in terms of investment view. InP CW driven SIPho (NPO or CPO, MZI or ring) is unstoppable for the next 2-3 years and there is a crippling panic spreading. Ask around. You will quickly find out who has secured supply and who is shitting themselves.

But… VCSEL CPO just got a huge vote of confidence from someone very important. I am keeping an eye on this.

The challenge is not the VCSELs themselves. It is packaging them in a dense way with high yield. Broadcom is VCSEL king with Coherent in close second. Lumentum and Furukawa also have decent VCSEL tech. For CPO/NPO, 50G NRZ is the sweet spot in my opinion. At that data rate, the quality of the VCSEL itself is not important. What matters is the integration and packaging of VCSEL+Fibers+drivers.

800V DC Practical Limits

Given the continued momentum of Bloom and behind-the-meter (not grid connected) power, I wonder how much reach 800V DC has. Literally how far can the transmission lines between datacenter and Bloom cells be before resistive loss kills you.

800V direct disproportionately benefits GaN and Navitas. Although it concerns me Navitas management keeps shilling the stupid “we have both GaN and SiC and this is a advantage” nonsense.

Navitas SiC is dogshit btw. I did a deep analysis or SiC and GaN across all players.

Ok well 10AWG limit is 30A so lets go lower.

There is no reality in which these lines run at 20C so bias up resistance.

P_loss = 375*375*(pessimistic 0.4 ohm) = 56 KW loss

Assuming 10K feet run.

So… 56/300 = 18.6% transmission loss which is dogshit.

Assuming you want less than 10% transmission loss, max is 500 feet or 150 meter which is not great.

I must be missing something. If anyone has insights on high power transmission systems please send me an email.

Contents:

1. Holdings and Documents

2. Do I look like a guy with a plan?

3. Hot Topics

[1] Holdings and Documents

June account statement not ready yet. Posting March account statement as it was not included in last update.

YTD transactions in trading account available in CSV form.

Webull is my trading account.

Long-only is a collection of ~7 accounts across Interactive Brokers, E*TRADE, Schwab, and Robbinhood. 401K accounts excluded as those are all mutual funds I do not manage.

As a reminder, I benchmark myself against the market-cap weighted semiconductor index, and have a baseline goal to outperform this index by at least 30%.

At the time of writing this delta is 65.7%, an acceptable outcome given the mistakes in June.

Clean Ia April

27.8MB ∙ PDF file

Download

Download

Clean Ia Mar

30.2MB ∙ PDF file

Download

Download

Clean Ia May

29.1MB ∙ PDF file

Download

Download

Ia Trading 2026 Upto Jul1 Options

20.8KB ∙ XLSX file

Download

Download

Ia Trading 2026 Upto Jul1 Stocks

47.1KB ∙ XLSX file

Download

Download

Some YTD stats from the trading account.

[2] Do I look like a guy with a plan?

People keep asking me questions like…

• What is your exit/entry for <stock>?

• What is your price target for <stock>?

• How do you choose sizing?

• Why are some stocks in only your trading account or only in long-only?

The long-only accounts are my “real money”. Personal safety net. My annual spend is ~$80-90K so I think of the long only accounts as “how many years can I be unemployed”.

Overall strategy is to buy and never sell, unless a re-balance is necessary due to concentration issues or a major long-term (3-5 years minimum) thesis change.

Right now, the long-only book is mostly semis/AI because semis is all I know. Over time, my plan is to add some more diversification. Stuff that is not semis/AI (or at least less correlated) and preferably has a decent dividend. You can see the small BHP position that is under construction. Yes, I know BHP correlated with copper and thus correlated with AI buildout. Also considering Verizon.

For now, I am generally happy with the long-only book except for two names, Samsung and Musashi.

Want to double the Samsung position. Musashi want more but have not decided how much. Free cashflow from dayjob will be put into those two names.

As for the trading account, I have basically stopped depositing money in. Kinda pointless to deposit $10K/month into an account that regularly swings +/- $250K in a single day.

Money in the trading account is considered to be not real until it is withdrawn. Believe me, I feel very little other than mild happiness/frustration on this crazy daily swings. This is all a game to me. Numbers on a screen to feed my narcissism.

That said, it has been nice to withdraw a nice chunk of money to fund expansion of the (real) long-only accounts.

Apparently Interactive Brokers has a weekly deposit limit of $100K lol. At least on the pore account tier I use.

I am terrible at poker, but like to use poker concepts for the trading account. There is a pool of chips (overnight buying power) and I make bets of various size. Blocks of $50K worth of shares. Starter positions at 1/4th desired size. modest position at 1/2 size. Gives the opportunity to double down. Generally don’t like to trim and prefer to “fold” (liquidate entire position). Taking chips off the table means transferring funds to a long-only account.

Any position over $300K I self-ban on adding more but if a position goes nuts and the engineering thesis has not changed, I hold. You can pry my Lumentum shares from my cold dead hands, optics tourists. At some point, I will transfer some of the Intel shares to a long-only account but for now they can serve as an anchor position to lever up on degen shit.

But to be honest, there is largely no plan.

I just… do things.

I’m an agent of chaos.

[3] Hot Topics

Have a more in-depth writeup on some of these topics planned for the weekend. For now, some quick low-quality takes.

Applied Materials had an event recently on DRAM and advanced packaging and it’s excellent. Full coverage coming.

On high-bandwidth flash, some people have pushed back against my (virulent hatred) stance. The pushback is interesting and appreciated. Apparently the plan is to use SLC not QLC. Need to think about this more. My militant opposition to HBF was based on a (wrong) assumption that it would use QLC/TLC.

Power semis I am debating if Wolf or Navitas deserves a size up. Thinking through the practical distance limits of 800V DC. Bloom is another vector to play this but has less torque given how much it already ran up.

Visited and we discussed a lot of things. He brought up Erock and this seems interesting. Need to properly read their S-1.

Fermi is a degen casino-level play with a lot of noise. Matt Dratch and a mutual friend are interested in this flaming dogshit and… yea this could be fun. You can see my YOLO options lol.

A friend told me Nvidia was out buying a lot of long-haul networking gear for some mysterious reason three weeks ago. Now sell-side has picked up on this. Something is going on. Benefits Ciena and Nokia. Trying to track this down.

MediaTek TPU share gain narrative is in full swing. Jeff Pu and all of Taiwan are gleefully gloating over the alleged demise of Broadcom. Regardless of what side you are on, you should own Semtech to play this. The size of your Semtech position should scale with how Mediatek-pilled you are.

Optics is out of fashion due to idiot tourist analysts and clown pod monkeys who understand nothing about laser linewidth, RIN, cavity length, InP yield, or mode hops. Go short Lumentum. I dare you. You can see Intel and Tower both got trimmed this quarter but I refused to sell a single share of .

Speaking of optics, AXTI. A lot of the price movement seems to be because Mainland-China financial institutions (hedge funds, family offices) levered up on total-return swaps of AXTI and recently government cracked down on this.

For more information on what a total return swap is, see this video on the legendary Bill Hwang, a hero who deserves a presidential pardon.

Hello and welcome to an exciting first edition of Cerebras earnings call coverage.

Two months ago, I wrote a detailed report for the IPO. Typeset in LaTeX for the special occasion.

Afterwards, one of my hedge fund friends called me, confused. He could not tell if I was bullish or bearish.

I believe the following statements to be true and do not view them as contradictory.

• I have been a fan of Cerebras since 2019.

• Cerebras should have died due to incompetence years ago.

• Cerebras ultra-fast inference is an incredible niche with enormous untapped potential.

• The WSE has several fundamental flaws and management is in partial denial and may be appropriately compared to clowns.

Most of the time, activist investors get involved for lame reasons. Force management to cut costs, more buybacks, “strategic review” (sell yourself for scraps).

Today marks the glorious beginning of the dumbest, most entertaining activist campaign in history.

I am a proud shareholder of .

You could say I am a holder of share.

We shall have such fun along the way.

Contents:

1. Updates on Wrong Parts of IPO Report

2. Earnings Call Transcript

3. Guide for Financial Analysts

[1] Updates on Wrong Parts of IPO Report

After publishing, several people pinged me on several wrong assumptions, particularly in the modeling sections.

Biggest info was that ASP of a WSE-3 rack is closer to $700-850K. A few other tweaks were needed. See the bold comments on what I changed and why.

There is a lot of (pointless) debate on wafer yield vs packaging yield.

I believe the wafer yield is quite high, easily above 90%. When Cerebras management claims near-100% wafer yield because of routing around defects, sure that is not a crazy claim.

But the only way to explain ASP and gross margins is assuming total yield is horrific. I think they get killed on packaging yield.

Maybe 100% of wafers from TSMC are good but 4/5 wafers fucking die in the packaging process? How the fuck do you explain these atrocious numbers?

Second major issue was my assumption on Tok/s/user.

Apparently the openrouter 3K number is bullshit. Updated to 0.7K.

Also updated ASP from $300K to $700K.

Finally, I managed to get a third uncorrelated source on the Cerebras/Groq/Nvidia palace intrigue.

Invertedly.

The person who gave me this nugget info has no idea on the correct extrapolation.

Information asymmetry is a funny thing.

Debating if I want to make this call without a 4th uncorrelated source.

For now, let’s just say what happened with Groq is (90% confident) very linked to Cerebras. Highly correlated events.

[2] Earnings Call Transcript

At what economics??? Everyone with a basic understanding of this space knows KV cache kills your economics.

People are offloading KV cache to fucking NAND boxes and you clowns are storing it on SRAM.

KV cache competes with model weights for the same scarce resource.

YOU NEED TO OFFLOAD KV BUT I/O ISSUES BLOCK STREAMING AS A VIABLE OPTIONS. ADMIT THE PROBLEM AND GO FIX IT. YOU RAISED MONEY FROM IPO GO SPEND IT TO FIX DISASTER BOTTLENECK.

Bro you seriously bragging about yield? Your yield has to be complete dogshit even after all these years. No other way to explain the financial numbers. Why don’t you share the numbers? I am at 100% wafer level yield, 20% packaging yield.

Feldman makes very good points here that I completely agree with.

Imagine if they had enough I/O to offload KV to NAND boxes. Think of how much bigger models and how much better the economics would be if KV did not compete with model weights in SRAM.

This is what is so frustrating. Cerebras really has something and could be so much better. They just need to fix a few (rather serious, but largely solvable) problems.

I agree TAM is infinity. Too bad it takes a comical number of WSE to hold a model because KV CACHE COMPETES WITH MODEL WEIGHTS IN SRAM and WSE-3 DOES NOT SUPPORT FP8 OR LOWER.

The constraint is your horrific packaging yield.

Ok let’s play a game.

Here is the full Excel model for both COGS/GM and inference tokenomics.

Cerebras Irrational Analysis 062326

1.16MB ∙ XLSX file

Download

Download

I would like every sell-side and buy-side analyst to download the above model, make your own changes, and decide for yourself who has a “misunderstanding”.

Joshua Buchalter of TD Cowen asked a direct question and Andrew Feldman refused to answer.

[3] Guide for Financial Analysts

There are only three factors that matter.

[1] WSE-4 Floating-Point Format Support

WSE-3 is from 2024.

WSE-4 should be fairly far-along in the design process. Ideally it needs to support FP4 but if it’s only FP8 maybe that is forgivable. Willing to hear arguments from them that FP8 is appropriate for the architecture.

I refuse to believe they are stupid enough to stay on FP-16. Someone please confirm this.

[2] WSE on/off Wafer I/O

This is the blocker for KV cache offload/streaming. Two un-correlated sources told me Cerebras is working on “face-to-face” I/O. Unfortunately I don’t have more details as both the sources are not technical.

My best guess is for some reason the I/O is limited to being a NoC extension alongside one axis. NoC probably runs at something slow-ish like 1 GHz.

Right now, only two edges are useful for I/O. Or might be only one edge IDK the numbers are so bad lol.

“Face to face” I/O likely means using the copies trapped between rows of reticles. Electrically, this makes sense. Physically (cooling, alignment, packaging) this is a nightmare.

Very curious on their progress and the underlying details.

[3] Horrific Gross-Margins —> Horrific Packaging Yield

They should honestly have an investor day and dedicate half of it to the yield improvement roadmap.

Again, I am modeling 100% wafer yield, 20% packaging yield to get 42% hardware GM.

If you bump the packaging yield up from 20% to 50% gross margins go from 42% to Nvidia-level 72%.

So much potential!

Extra Intro Note:

Thank you.

Never expected to make it this far.

100% of content will always be free. The pledge thing is for a theoretical future donation box. A backup in case my dumbass gets fired and blacklisted from industry.

But honestly, this…. hobby…. is now net positive for my career lol. Life it seems is stranger than fiction and some things going on behind the scenes. Don’t have to worry about being fired and unemployed anymore.

Hello and welcome to a brief note on two completely unrelated topics. Recently read an excellent article by on coherent lite transceivers. (not to be confused with the corporation named Coherent or the other corporation with ticker LITE…)

FundaAI

Deep|Coherent Lite Reshapes AI Optical Interconnects

Key Takeaways…

Read more

2 months ago · 31 likes · 3 comments

Go sub to them. I am a paid sub on my real substack account.

One table bothered me….

EML having less area than a low-power CW laser is obviously wrong. Then there is the area issue. Photonic chips, particularly lasers and InP, tend to be narrow and long. So using area intensity is problematic. Need to use a wafer map calculator.

After spending half an hour running some quick sanity checks, I realized this table is completely wrong lol.

But Funda’s conclusions are mostly correct. Marvell biggest winner. Other investment conclusions also correct but InP intensity not as high as they think.

One core theme of the Funda piece is InP PIC for coherent lite. I can tell you that Tx will not be InP monolithic. Rx, there is nuance. I did some basic calculations on how much InP would be eaten if both Tx and Rx of coherent lite is monolithic InP and quickly realized this is crazy town. InP situation is already an unmitigated disaster. Adding this extra demand on top is like pouring all the stranded crude oil in the strait of Hormuz onto an existing InP dumpster fire.

Funda friends were kind enough to share the PDF of the Google paper (OFC 2026 tutorial) related to this topic.

https://x.com/phasentrpolator also sent me a great paper. Spent a couple hours flipping through textbooks and got enough material to work with.

Contents:

1. Coherent Component Crash Course

2. Google Paper Quick Overview

3. Summary of FundaAI’s Arguments

4. Wrong Parts

5. InP Yield and Area Crash Course

6. Irrational Analysis InP Intensity Model

7. Theoretical Optimal Coherent Lite Design

8. Winners and Losers

9. [Bonus] AMD CPU Opportunity

10. Trading Account Snapshot: 06/13/2026

[1] Coherent Component Crash Course

I don’t want to bore you too much with how Coherent transceivers work. My primary objective is to teach you what the unique (compared to traditional IMDD) components are, why they exist, and how they affect system design.

If you understand the structures, you can easily understand the delta between coherent and coherent lite.

Fundamentally, coherent optics is the same as how wireless communication systems work. Data is encoded on two axis (I and Q), transmitted, and decoded later.

Another way of looking at this is as follows.

How to transmit complex numbers in the real world?

(use a sinusoid + 90 deg phase shifter to mix each channel)

The benefit of this strategy is a lot more bits per symbol. Wireless goes up to 1024QAM these days. Optics seems to stick with 16QAM.

In a previous life I worked on 1024QAM 5G/6G bullshit. Stupid feature. Tx EVM of commercial base stations cannot support unless ultra ideal conditions and even then not at max coderate.

Anyway, we know the benefit, so what is the drawback?

PHASE NOISE MANAGMENT

All components intrinsically have phase noise. Some more than others. Here is what your constellation looks like when phase noise is allowed to run amok.

You get a useless doughnut.

Phase noise problems are very difficult to debug and design around. Often the solution is “just use more expensive components bro”. Alternatively, people use complex (very area and power expensive) DSP (math) to characterize the phase noise of the system and back-out that transfer function in Rx EQ.

Always remember that engineering is all about tradeoffs. Understand these tradeoffs and you understand why people make design choices.

Back to coherent optics, the LO in this case is a laser. The noise performance and polarization of the LO laser (ok every laser in a coherent system) is critical.

Laser linewidth is a measure of phase noise.

Phase noise is super bad in coherent systems, so the linewidth requirements are very strict. Easily < 0.2 MHz.

Managing phase noise, is painful. Trust me on this.

Leveraging symmetry is critical to removing LO noise. If both arms and photodetectors are precisely matched, large portions of the LO laser will be canceled out.

On the flipside if you screw this up everything becomes garbage and is unfixable.

RIN kills you.

Linewidth kills you much harder.

[2] Google Paper Quick Overview

Bottom (C) is a normal coherent transceiver architecture).

Top (A) is coherent lite.

Middle (B) is ugly, unnecessarily complex, and just worse than (A).

The fundamental delta between coherent and (polarization-folding) coherent lite is… polarization so let’s take a quick detour.

Polarization is the angle of light. Sunglasses work by filtering out some angles.

In optics, the nomenclature for polarization is TE and TM.

Broadly speaking, photonic structures (on Silicon, Silicon-Nitride, Indium-Phosphide, …) have one of three states:

1. Only accept TE mode.

2. Only accept TM mode.

3. Does not care about polarization.

For example here is a structure that rotates light from TM to TE.

And here is a structure that separates TE and TM modes.

Broadly speaking, dealing with polarization on SiPho is much more annoying than InP…

… at the cost of area. Remember, tradeoffs.

Now let me annotate the Google coherent lite block diagram to help you understand why it represents beautiful engineering (good tradeoffs).

Notice how nothing in the Tx block diagram involves polarization. This is a good thing!

Massive cost-down enabled by SiPho PIC.

The key to understanding coherent lite is polarization-folding.

Why go to the trouble?

The answer is surprisingly simple.

The transmitter output it going to have random polarization. If you get unlucky, the receiver will get very little light because most of the light is on the wrong polarization/mode.

The entire receiver in coherent lite is designed around a specific polarization (prob TE). We cannot afford to waste light/energy trapped in TM mode.

Very clever trick. Dramatically reduces complexity and cost on Rx.

The DSP needs to do some extra math though. Intuitively, the transfer function of the polarization folding unit is 1 + Z^-1. Dug up the partial response signaling paper out of curiosity…

Yea this is straightforward. Use all energy, force it to be on TE mode, and have the DSP back out a simple transfer function before running anything else.

[3] Summary of FundaAI’s Arguments

1. Coherent lite is going to be huge because of Google.

2. Driven by OCS insertion loss budget.

This is correct although 3 dB loss on MEMS OCS sounds high. Wonder what size of OCS switch induces such loss.

[4] Wrong Parts

Now for the wrong parts.

This is completely wrong. Transmit is not more complex. Line rate for each MZI drops to 168Gbps. SiPho already works at 224Gbps line rate.

168 is in fact a smaller number than 224.

Nothing in the transmit side is polarization sensitive. There is no need to use InP. SiPho will do just fine.

(Tower Semi sipho, not that GloFo dogshit)

A MZI modulator can easily hit 5 mm long. Using InP for coherent lite Tx PIC would obliterate the industry. Also would obliterate cost.

I will come back to the estimates. Entire FundaAI table is wrong as a reminder. I’m not kidding or trying to be mean. All the numbers are wrong.

[5] InP Yield and Area Crash Course

Indium-Phosphide is a satanic nightmare. As a ballpark, yield of 50% is considered normal for discrete devices (CW laser, EML).

Unlike silicon/logic, there are no tuning parameters to harvest parametric issues.

Some of your InP devices will be at the wrong wavelength. Too far away from target to thermally tune via TEC. You fucked.

Some InP devices won’t hit target power at target temperature (typically 40C or 50C). Nothing to be done.

Some InP devices will be flaky, exhibiting erratic behavior at target setpoint. This is called mode hopping. Can’t do anything. YAY INDIUM PHOSPHIDE IS AWESOME.

Also InP is a very brittle material. Lots of dice die as you try to package and wirebond.

[6] Irrational Analysis InP Intensity Model

I have build a model using numbers pulled from my anus. Maybe I know things. Maybe I don’t. Wink wink.

[7] Theoretical Optimal Coherent Lite Design

It is unclear to me what the actual architecture will be but here are my best guesses. The design space is pretty open.

1. Tx

   1. DFB (no ITLA)

   2. Tower Sipho with MZI modulators.

   3. Each modulator signaled at 168Gbps.

   4. Marvell or Semtech Drivers

2. Rx

   1. Monolithic InP PIC (Coherent the corporation or Nokia)

   2. TIA from Marvell or Semtech.

3. DSP

   1. This is Marvell.

   2. 8x 336G PAM6 SerDes receive from host

   3. 16x 168G PAM4 SerDes to PIC

[8] Winners and Losers

Marvell is the biggest winner by far. They gona make so much money on the DSP and have great driver/TIA bundling oppertunities.

Tower wins because more sipho and GloFo is actually incompetent.

Semtech win but maybe not as much as I hope. Marvell might take most of the sockets inside the transceiver.

Macom lose because photodiodes integrate into InP PIC.

Coherent (the corporation) and Nokia win with 6in InP process for monolithic InP PIC. Other InP players won’t want to burn their 4in capacity on this.

[9] [Bonus] AMD CPU Opportunity

Suppose AMD gets more wafers from TSMC. Assume this but don’t think about it, unless you were on the recent Raymond James bus tour.

Thos incremental N3P wafers would all go to datacenter CPU, the hottest and by far highest gross margin product line AMD has.

Been chatting with some contacts and I don’t think any of them understand the money tsunami Dr. Lisa Su is capable of unleashing IF AND ONLY IF SHE GOT MORE N3P WAFERS.

AMD Turin uses N3P for the CPU core chiplets and N6 for the I/O die.

By far the best implementation of chiplets. Their numbers are dirty. Let me show you.

[10] Trading Account Snapshot: 06/13/2026

First, congratulations to Matt Murphy and the non-trivial number of Marvell employees who are subscribed.

Up 60% in one week lol.

Don’t worry about me I was not short.

Gave up trying to short back at $172.

Marvell is by far the most asked about stock from institutional investors. There are at least 10 hedge funds and 3 family offices who constantly ask me about this ticker.

I don’t know what is happening.

But I will write a public note on some best guesses. Please friends lets talk about something more interesting. AMD is interesting.

No, I am not going to buy into Marvell. In truth, AMD CPU story is more interesting and I am debating on sizing up the position. You will see my trading book and long-only book at end of this post. Generally happy with the state of everything but need more Samsung, considering more AMD, and want to buy back into Ciena but too levered and AMD probably takes priority.

Contents:

1. Politics and Conspiracy Theories

2. Dense SRAM

3. DDR4 CXL Re-Use

4. Holdings Update

[1] Politics and Conspiracy Theories

Jensen $2B into Marvell has vastly out-performed his own stock. Why buy-back shares when you can make meme investments in competitors and suppliers?

The motivation of Mr. Leather Jacket is unclear. I have some guesses but zero conviction behind said guesses so take with large grain of salt.

• Force Amazon to use NVLink Fusion instead of Celestial?

• Block AMD from custom high-performance switchs?

• Weaken Broadcom?

• Get access to Celestial IP as a backup plan if ring modulators don’t scale?

• Get access to Marvell dense SRAM IP?

There was quite a bit of friction between Nvidia and Marvell over optical DSP pricing and performance. Nvidia has their own internal DSP which seems pretty good (recent OFC demo) but may or may not be designed into 1.6T transceivers. So much drama around this I gave up. Waste of thinking time.

[2] Dense SRAM

Marvell showed excellent results of their custom dense SRAM IP at Hot Chips last year. In truth, I don’t understand how it works but the data they showed is rock solid and excellent. Only caveat is their comparisons are against (my guess) Synopsys stock IP. Nvidia, AMD, and have better internal SRAM IP but maybe Marvell’s is still a lot better.

They can run at 500 mV (450 if you push it and ignore cold corner) which is kinda incredible. Usually the floor is around 630 mV… maybe 600 mV.

Licensable = Synopsys

Still, gap is so wide this IP is probably better than most if not all internal SRAM IP from Nvidia, AMD, Broadcom, ect.

This data is super legit. Not cheesed. Just want to point out that in this plot the lowest voltage corner is at 630 mV for N3 while the previous data was for N2. If you ignore < 10C as an operating condition (reasonable for datacenter) then they can drop the voltage way more.

[3] DDR4 CXL Re-Use

At the time of the Hot Chips presentation last year, I viewed CXL as a dead technology. With DRAM going to the moon, well now view has to change.

When old CPU servers are de-commissioned, the DDR4 is not useful anymore. Say you consolidate 3 old DDR4 CPU servers into 1 DDR5 CPU server. That DDR4 usually gets fire sold on eBay or scrapped.

With Marvell Structera (CXL card), Hyperscalers can re-use all that DDR4 in a memory pool. Sure, the latency is dogshit but its way better than NAND storage and DRAM prices have gone exponential with a crippling shortage.

Amusingly this Structera product went from worthless garbage to extremely valuable and it seems nobody has noticed.

This one is for DDR5 but they have a DDR4 version as well.

Not kidding, Marvell’s highest gross margin product is probably the DDR4 Structera card and this most profitable chip is gona go from zero to a large revenue number.

In a world where DRAM prices have gone up like crazy and there is a crippling historical shortage, the residual value of old fully-depreciated DDR4 DRAM sticks from old de-commissioned CPU servers is extremely high.

Marvell can probably charge like 85% margins on this Structera chips and get away with it. The value of re-using old DDR4 sticks is so high nobody will complain.

Astera Labs is the other CXL play but the Marvell product is just better. Way better interface IP than the Synopsys stuff Astera licenses. And remember, dense SRAM! You want SRAM as a cache to hide the latency of the CXL controller.

[4] Holdings Update

I am in the process of finishing up a (rare) re-balance of the long-only accounts. Would say 90% of the way there. Want to make Samsung a > 5% position but perhaps it will have to wait for cashflow via dayjob. Maybe cut GEV in half again and send proceeds to IBKR for buying moar Samsung. Trading account generally happy but want to size up AMD… a lot. Still thinking this thru. Also, I have no plan on how to pay the estimated $56K in tax pre-payment due by July 15th. Problem for future me.

Also if Ciena shits itself this week from earnings I will buy back in. If they moon I will also probably FOMO in anyway. Very unhappy with how I played Ciena this year.

Welcome to an updated view (partial capitulation) on memory. Specifically DRAM. Nand/flash is outside the scope of this post. Go away Sandisk bulls.

I have been negative on DRAM (Micron/Samsung/SK) as an investment due to persistent fears of cyclicality.

There are two camps:

1. Team price/book

2. Team price/earnings

Each side thinks the other is completely retarded.

Still consider myself in team price/book but… willing to make an exception and consort with the other side temporarily.

Funny story….

Two years ago at GTC 2024, I was at one of the after-parties waiting to meet Doug for the first time in real life. At the party, Dylan and a few random hedge fund people were chatting. One of the hedgies was talking about how his PM/boss said “if you see a bubble forming, run towards it”.

*This quote is from George Soros. Random hedgie minion def does not work for Soros so he mis-attributed the quote.

Spicy Takes:

1. I believe HBM is a mistake and will be phased out (90% drop in volume from peak) within the next 7-10 years.

2. DRAM stocks prob have another double or triple left in them.

3. At some point in the next 3-10 years, DRAM stocks will drawdown at least 70% from peak.

Contents:

1. Engineering

   1. HBM is a mistake.

   2. [short term] Hybrid-bonded HBM is necessary.

   3. [partial solution] CXL

   4. The optimal (future) approach.

2. Entertaining History

3. Agentic Reality: Featuring Nick/Coatue

4. Pitch: Samsung

[1] Engineering

Let’s start with engineering topics. Some level of background is needed.

Memory is in two categories:

1. DRAM

2. NAND aka Flash aka NAND Flash

Within DRAM, there is DDR, LPDDR, GDDR, and HBM. I view all of them as equivalent and you will see why shortly.

DRAM, NAND, and logic all have separate process technologies. The factory capacity is not fungible.

All DRAM works the same in principal. A transistor controls the charge of a capacitor. Reading the bit drains the capacitor. The capacitor also naturally loses charge over time so it constantly needs to be refreshed. This is why DRAM is referred to as volatile memory, it needs power to store data.

DRAM has only four core attributes:

1. Latency

2. Bandwidth

3. Capacity

4. Power Consumption

Changing the structure of the grid of capacitors and transistors gives you DDR vs LPDDR vs GDDR but it really is all the same at a high level.

HBM is different because there is vertical stacking.

Some of the DRAM die area is occupied by TSVs (through silicon vias), vertical wires.

There are also micro-bumps to connect the dies together.

HBM (high-bandwidth memory) gives you high-bandwidth, but at a high cost. This cost (economic, thermal, power) is going up at an untenable pace.

Don’t believe me? Take a look at the Rubin HBM4 fiasco.

JEDEC is the standards body that governs DRAM. Generally, when a new spec comes out (DDR5, LPDDR5X, HBM4, whatever) there is an initial speed agreed to in spec and over time industry adopts higher speeds over time.

With HBM4, JEDEC spec said 8 Gbps/pin and immediately Nvidia Rubin (first product to ramp) demanded 11 Gbps/pin. All three major HBM vendors (SK Hynix, Samsung, Micron) proceeded to fail to meet this request at various severities. Endless drama from the supply-chain leaker degenerates.

The root cause of this fiasco is simple. Bump capacitance. Will get to that in a later section.

HBM4 rollout is an industry-wide fiasco.

JEDEC are idiots for releasing a spec that is immediately obsolete.

All the consumers of HBM4 look like fools for making last-minute demands that cause chaos and delay their own roadmaps.

All the HBM vendors look stupid for claiming they can meet 11 Gbps or 10 Gbps then failing Nvidia qualification process because none of them know how to use high-speed probe stations.

[1.a] HBM is a mistake.

Memory bandwidth is an I/O density problem which has been solved in the wrong way by the wrong people.

You see, 11 Gbps pin speed is still… shit compared to everything else on the market outside of DRAM-land. 32G short-reach SerDes (UCIe AP, NVLink D2D, Marvell D2D, Broadcom MAX) are ubiquitous. Hell the 64G class versions are either already in prod or close.

DRAM crowd was given a PHY problem and botched it. Now, other companies are making custom solutions to bypass the fail. See this example from Marvell.

Diagram is not to scale and just illustrative but the point is clear. If you don’t use a retarded PHY, you can save a lot of ASIC shoreline and shove more of them in. Orange box gets smaller.

In this case, beachfront width does not improve because the TSVs of the HBM still limit even though the custom Marvell PHY is way better.

There are gains to be had from custom HBM base die. Marvell claims 25% more bandwidth. Great… but people want way more. AI is insanely memory bandwidth bound.

Key Points:

1. HBM PHY is dogshit.

2. Custom HBM base die partially mitigates this issue.

3. The TSVs are the real limit, locking up precious shoreline.

There are a variety of startups trying to solve this memory bandwidth issue. Some go for SRAM-heavy designs. Others use exotic processes. There is one startup that is my favorite and that is Positron. They solved this problem the “correct” way.

Unfortunately under NDA so can only give vague info that Thomas already shared publicly.

• Positron does not use any HBM.

• Only use commodity LPDDR5X.

• Highest memory bandwidth of any AI ASIC.

• Some clever architectural tricks to actually use all that bandwidth effectively.

Also know everything about Etched but under NDA. Architecture is very cool but they still suffer the same HBM limits as everyone else. Hopefully in the next few months I can negotiate proper coverage.

Positron gave me all their slides. Etched showed me the material but still wont send the PDF. MatX wont give me anything so IDK maybe I need to threaten them. Those Pope podcasts are useless and have zero real information.

Positron is the only startup that has properly addressed the memory bottleneck in my opinion. There is a better way that involves optics but in the short term, Positron’s approach is optimal.

[1.b] [short term] Hybrid-bonded HBM is necessary.

The HBM4 fiasco was driven by two primary factors, base die and pump capacitance.

Samsung used their internal SRF4X logic node (perf ~TSMC N6/N7P) and had the easiest time. Enough margin from the transistor performance to fix issues quickly.

SK Hynix chose TSMC N12 and… had like 6 re-spins.

Micron went full retard and tried to stay on internal DRAM process node for base die.

But even with a good base die, there is still the parasitic bump capacitance issue.

Let’s look at some squiggly lines!

This is a TDR trace (time domain reflectometer):

You can use a machine to map out the impedance of a system. Stare at enough of these squiggly lines and it becomes second nature to spot what part of the system made what part of the plot.

Probe a HBM stack (without the base die) and you will get a TDR trace with one spike per layer. The spikes correspond to parasitic bump capacitance.

Only way to (properly) fix this is to move to hybrid bonding.

Everyone keeps viewing hybrid-bonded HBM from a thermal or stack height perspective… which is valid.

My point is speed/performance is the real gate and what will finally push DRAM vendors to suck it up and use hybrid bonding.

Think of capacitors as speed bumps. You can make the base die as amazing as possible but those parasitic capacitors in the stack will kill your signal. Either blow up PHY power or accept lower speed.

Speaking of which… this SK Hynix innovation is more evidence HBM is a mistake.

Why yes, the PHY power is gona blow up as speeds go up. Your solution is to have an overhang and place a thermal conductor pillar.

This is like trying to eat while wrapping your arm behind your head. Technically possible but stupid.

HBM is a mistake. The solution is to disaggregate memory with a real PHY that drives co-packaged optics.

[1.c] [partial solution] CXL

CXL is a protocol built on top of PCIe that allows for memory expansion and pooling. Initially designed for CPUs, it was abandoned. Now it’s coming back.

Great for Marvell and Astera Labs.

But still does not solve the problem properly. Remember, we have a PHY problem and CXL is just a protocol built on top of a sub-optimal PCIe PHY.

[1.d] The optimal (future) approach.

At ISSCC 2026, Nvidia showed a glimpse of the future. The correct approach to I/O beachfront problem, and by extension the memory bandwidth problem.

Clock-forwarded SerDes directly driving optics.

On the ASIC side, you have essentially no latency added to the system as there is no re-timer or need for a heavy (latency inducing) FEC. Shoreline density of clock-forwarded XSR SerDes is very high. Optics enable reach from 1 to 30 meter easily, once you go through the pain and suffering of getting the data on the optical fiber.

Clock forwarding allows for the entire Rx EQ to be gutted. No CDR, FFE, DFE, MLSE, or compound CTLE. Just a simple TIA.

Sub 3 pj/bit, up to 10 meter reach easily. More likely 30 meter should work assuming chromatic dispersion does not mess with the delay compensation circuitry.

Shoreline density is not great and you need to look at the die shot to see why.

Each lane is in an 81 micron square. The Tx/Rx PLLs are the same size as an entire lane AFE. This does not make sense from an electronics perspective. My guess is photonic issues, particularly thermal crosstalk, forced the EIC to blow up in area. You can see the digital backend and JTAG separating Tx and Rx AFE. This was likely intentional to isolate ring modulator self-heating from Tx rings polluting Rx rings. Point is, the end state of this technology at commercialization is (well… should) going to be way better than 0.4 Tbps/mm unidirectional.

Because the system is optical, you can just route with an OCS that also has effectively no latency when compared to a traditional electrical switch.

On the device side, all you need is a chip that converts optical back to electrical (minimal latency) followed by some LPDDR PHYs. The latency will be dominated by the LPDDR PHY and controller. Perhaps clever optimizations will be enough. Perhaps a stack of HBM somewhere in the system used as a cache will be needed.

But this is the future in my opinion. Optically attached, disaggregated LPPDR memory pools. Nvidia is 2-3 years away from implementing this at scale. Broadcom and Marvell 3-5 years.

The benefits are enormous. Way higher bandwidth/mm shoreline density. Way more capacity. Minor latency penalty if done right. And most importantly, the ability to re-use and re-allocate DRAM across chips instead of packaging them forever with the ASIC while also enabling the ASIC to draw more power due to higher thermal headroom.

It get’s even better. You probably have heard of the HBM trade ratio. How many GB of regular DRAM capacity is consumed to produce 1 GB of HBM.

It is like 3:1 or something. The actual number does not matter. Its a big number. Massive supply destruction has HBM mix goes up.

If that reverses, you would think the DRAM vendors would get screwed but I disagree.

Right now, they are making 80% ish gross margins. At that profitability, all the customers know they are getting screwed and are thus unhappy and strongly incentivized to find a way out.

Imagine a world where HBM is replaced by LPDDR, DRAM supply is 3x, gross margins go down from 80% to 60%.

Everyone wins in this scenario.

DRAM makers end up with 2.25X profits (do the math retards) with 3x revenue at modest margin cut while also not pissing off all the customers.

DRAM customers get 3x supply at much more palatable prices to satisfy insatiable agentic AI demand.

DRAM customers also don’t have to worry about advanced packaging yield as much and can scale their logic chips more instead of wasting space on HBM stacks.

[2] Entertaining History

To understand the history of DRAM companies, you must read the transcripts of past earnings calls. This is typically a boring process so to keep people interested, I will roast Sanjay Mehrotra (Micron CEO) for entertainment purposes. There is a particular feature of Micron earnings decks that makes this goal easier to achieve.

We are going to look at every earnings call transcript from June 2017 (first call with Sanjay as CEO) to June 2025, the last earnings call where a negative dip happened before the crazy rally started.

You really should read this section in it’s entirety. It’s important.

June 2017: $30/share

Sanjay did time his join date of March 2017 well. It was at the bottom of the cycle. 2017 cycle was particularly nice for the DRAM players. I remember as I was in college trying to buy DRAM for a gaming PC build. Fucking DDR tripled in price or some shit.

Sept 2017: $39/share

Ah yes, strategic value, not a commodity. Yes Sanjay pump your RSU.

Dec 2017: $43/share

One of the biggest red flags you can see is mentioning cost competitiveness in one sentence then literally mentioning high-value in the next sentence. It is a sign of someone who knows he is selling a commodity but wants to pretend otherwise.

Mar 2018: $57/share

Well buddy the cycle is about to turn so I hope you got some of those high-value solutions to insulate you.

June 2018: $60/share (TOP OF CYCLE)

Ah yes starting buybacks at the top. Great start.

This time is different guys. Stability in the market!

You can see the Nomura analyst shitting on Micron management. Thanks for keeping important negative guidance info at the end of the presentation you little shits.

Guys it’s just a 1-quarter blip. Next quarter prices will be healthy again. This is def not peak cycle and we certainty are not gona crash.

Are your gross margins deteriorating because of price cuts (very bad) or mix shift (ok excuse).

Hint: it was because of price cuts. OVERSUPPLY BAYBEEEE.

Analyst: Is this cycle (2017-2018) different this time?

Sanjay: Yes.

(it was not different that time)

Sept 2018: $45/share

Sounds like youssa in a down-cycle.

Dec 2018: $32/share (CYCLE BOTTOM)

Oh look this time was not different! Youssa eating shit in downcycle and resorting to decisive (desperate) actions.

Datacenter NAND/SSD is going to be great for us.

Also we are going to eat shit in this exact segment for the next year.

O V E R S U P P L Y

June 2019: $42/share

Ah yes you finally said the dirty word.

OVERSUPPLY

We made it past the usual euphemism of “inventory adjustment/digestion”.

The last sentence is a public call for Sanjay’s degen Korean friends to have mercy.

Sept 2019: $46/share

Amusingly this was probably true… and only a few months later COVID lockdowns hit right after inventory over-stocking due to tariff fears.

Dec 2019: $56/share (local* top in upcycle)

*due to pandemic. Under other circumstances upcycle would have continued prob another 12-18 months cleanly.