TL;DR

• DRAM is down 30% from June highs, SK -38%, Samsung -32%, Kioxia -40% while MU is down 23% and Sandisk -26%. The sell of driven by positioning, macro (US-Iran War) and AI capex digestion fears. But in my opinion, fundamentals haven’t changed one bit

• Fundamentals have strengthened. Micron’s Q4 2026 report showed record $41B revenue, 85% gross margins, HBM sold out through 2026 and CEO saying memory tightness persisting “beyond calendar 2027”. SK Hynix's CEO calling 2027 "the worst year in the industry's history from the supply perspective". Samsung says the 2027 supply demand gap will be wider than 2026's. The memory cartel have spoken. Supply remains tight at least for another 18 months.

• The demand runway is multi year and multi vector. Hyperscaler capex is guided to $750B in 2026 (80% Yo

  Y) with 2027 forecasts clustering above $1 trillion; new fab capacity cannot add meaningful bits before late 2027/2028; HBM’s rising wafer “trade ratio” cannibalises commodity DRAM and new vectors (sovereign AI, humanoid robotics, HBM4/Rubin, AI PCs/phones, etc) extend the cycle.

• The main risks are the historical cyclicality of memory, Chinese supply (CXMT), and a hyperscaler capex cut.

So Why Is Memory Down 30%?

The trade got too crowded. Oil spiked, SK Hynix listed on NY, and the market got nervous about AI Capex. Notice what is not on that list? Not one of them is about a memory company selling less memory, or selling it cheaper, or losing customers? The stock broke, but the business did not.

1. Positioning: Morgan Stanley called memory one of the most crowded positions in the market . A stock can only go up if there’s a marginal buyer left. When positioning is crowded, there isn’t one. Buying power becomes exhausted and the stock stops rising on good news, because eveyone who would’ve bought on that news already bought. Of course except the Koreans running to grab another loan from the bank. This is more dangerous on the way down. When it starts falling, eveyone is reaching for the exit at the same time. Leveraged positions unwind, and a chain reaction occurs.

   Next MS also flagged “peak rate of change”. Market’s aren’t pricing on price levels. They price the second derivatives. “Are prices accelerating” is the golden question. DRAM contract prices rose 90% QoQ in 1Q2026. But in 3Q26, it’s only rising at 15% QoQ. Deceleration leads to selloff.

   Lastly DRAM EPS revision breadth around 89%, near historic highs. This just means of all analysts covering DRAM names, 89% of their recent EPS revisions were upward. Here’s where the problem comes in. They can’t keep raising estimates forever. The moment breadth falls from 89% to 70% (for example), the narrative flips and multiple compresses even when EPS is still climbing. To put simply: if everyone analyst has already told their clients to buy, there’s no one left to convert.

2. Macro: US Iran strikes on July 12 + blockade = escalating tensions. Oil prices spiked. Markets derisking. Nothing to do with DRAM.

3. SK Hynix sell the news: Record $26.5B Nasdaq ADR listing. Then Korea Investment cut Q2 operating profit 8% below consensus and Seoul shares fell 15% in a single session.

4. Capex Digestion Fears: Meta announced it would launch a cloud business to sell excess AI computing capacity. Stock rose 8%, but AI hardware got destroyed.

Where Do I Stand?

Read more

TSMC, the company that fabricates silicon for Apple, Nvidia, and AMD, signed a 10 year partnership with Amkor Technology to build out advanced semiconductor packaging in Arizona. Not a memorandum of understanding. Not a “we’re exploring possible opportunities some time in the future” press release. A decade long procurement agreement.

The stock then jumped 18.8% in a single session.

Amkor Technology (AMKR) is the company that takes a raw silicon wafer and turns it into a finished, tested chip that actually works inside your phone, your car, or the AI server answering your questions right now. Without packaging and test, a fabricated chip is just a fragile piece of silicon sitting on a wafer. Someone has to connect it, protect it, wire it up, test it, and ship it.

That’s the sector AMKR plays in.

Apple’s COO called them “a critical partner.” Nvidia’s EVP of Operations endorsed their packaging capabilities on the record. AMD is ramping its next generation server CPU packaging on Amkor’s proprietary technology platform. And TSMC just committed to buying their services for the next ten years.

To me, this derisks most of the bear case. We’re not looking at a speculative bet on a company that “might” win AI contracts someday. The contracts are already signed. The factory is already under construction.

The question isn’t whether Amkor is relevant to the AI supply chain. It’s whether the market has fully priced in what’s coming. Management is guiding for revenue to nearly double and EPS to more than triple by 2030. The $7 billion Arizona campus will be the first high volume advanced packaging facility on American soil. And the technology shift from monolithic chips to multi-die chiplet architectures is making advanced packaging the single most important bottleneck in semiconductor performance.

So here’s why I think there is still upside left in this…

Read more

I'm a little worried about those quoting that SMCI is criminally undervalued at 0.8x P/S ratio. The problem I see with SMCI long term is hyperscaler vertical integration.

So here the question. What happens when hyperscalers like Google, Amazon, Microsoft stop buying off the shelf AI servers and start designing their own?

I mean this isn’t a few years away… it’s happening now. They’re already doing it with custom silicon (TPUs, Trainium), and the next logical step is owning the full rack design too.

So here’s how I’ll structure this article.

1. The Thesis — What Bulls Are Betting On

2. The Threat — Vertical Integration Is Already Here

3. Does This Kill DELL Too? Since I hold DELL…

4. Verdict…

   ---

The Thesis — What Bulls Are Betting On

SMCI is the fastest gun in AI servers. Every time NVIDIA drops a new GPU platform, SMCI is first to ship it in a fully integrated rack — weeks ahead of Dell or the ODMs. That speed to market is their moat. Pair that with their lead in direct liquid cooling, which isn’t optional anymore when you’re packing 100kW+ racks into a data centre, and you’ve got a company that’s been in exactly the right place at exactly the right time.

But what happens when the hyperscaler capex wave slows down? That’s where the second act matters. SMCI just signed a $20 billion deal with Saudi Arabia’s DataVolt to build sovereign AI campuses — a customer base that doesn’t have the engineering capability to design their own servers. Enterprise AI adoption is still early innings too. Every corporate IT department upgrading to on-prem AI infrastructure is buying off-the-shelf, not custom.

And then there’s the price. The stock went from $10 to $120 to $30. The market has already priced in a lot of pain. When expectations are this low, the bar for an upside surprise gets a lot easier to clear.

So what happens when your biggest customers decide they don’t need you anymore?

This is what I’m concerned about. That’s the real risk with SMCI. Google has TPU v7. Amazon has Trainium 3. Microsoft has Maia 100. Custom silicon isn’t experimental anymore. It’s deployed at scale and reduces total cost of ownership by 40-65% versus off the shelf NVIDIA GPUs. So if you’ve already spent billions designing your own chip to avoid the NVIDIA tax.

So why keep paying a middleman to build the server around it?

Read more

There are two ways to get light into a CPO module. You either put the laser on the same package as the chip, or you keep it separate and feed the light in from outside. That choice sounds like a minor engineering detail. But it’s not. It might be the most consequential architecture decision in the entire AI infrastructure buildout.

Integrated laser CPO bonds or attaches the laser directly onto the same package as the switch ASIC or GPU. The logic is simple: fewer components, shorter optical path, tighter integration. Intel has historically pursued this approach through heterogeneous integration, fusing III-V materials (like indium phosphide) directly onto silicon wafers.

ELS-based CPO keeps the laser physically separate in its own thermally controlled housing and feeds light into the photonic chip via fiber. Ayar Labs, Broadcom’s CPO switches, and Nvidia’s Spectrum-X Photonics platform all use external laser sources.

So who’s winning? According to a January 2026 CPO market report covering Nvidia and Broadcom ecosystem strategies: “Most leading implementations have converged on external laser source architectures that keep temperature-sensitive lasers separate from heat-generating ASICs, improving reliability and enabling redundancy.”

_{Source: Synopsys}

That’s not a niche opinion. That’s the direction the industry is moving. Here’s why:

Heat kills lasers. This is the fundamental problem with integrated approaches. Lasers are the most temperature-sensitive component in the optical chain. ASICs and switch chips pump out hundreds of watts of heat. Putting both on the same package is like bolting a precision instrument onto a furnace. Even small thermal fluctuations degrade laser performance and shorten lifespan. ELS eliminates this by keeping the laser in a controlled environment, completely isolated from the ASIC’s heat.
That’s the difference between a laser rated for 10+ years and one that degrades unpredictably.

Yield economics favour separation. Silicon photonics chips and InP lasers have completely different manufacturing processes with completely different defect rates. Integrate them onto the same package, and a defect in either one kills the whole unit. You’re throwing away a working ASIC because of a bad laser, or throwing away a good laser because of a bad chip. With ELS, you test and qualify each laser independently before it ever ships. Only known-good lasers reach the customer. At scale, that yield advantage compounds into massive cost savings.

Serviceability. If an integrated laser fails in the field, the entire package is dead. That means scrapping an expensive switch ASIC because a $50 laser component died. With ELS, you swap the laser source without touching the rest of the system. For hyperscalers running millions of optical links, that serviceability difference translates directly into lower total cost of ownership.

Material optimization. InP and silicon want fundamentally different fabrication conditions. Temperature profiles, chemical processes, and crystal structures are incompatible. Forcing them together on one substrate means compromising on both. ELS lets each material be manufactured in its ideal process with zero trade-offs.

Wavelength flexibility. External lasers can be tuned or swapped for different wavelengths without redesigning the photonic IC. As data rates scale and multiplexing demands increase (from 8 to 16 to 32 wavelengths), that modularity matters.

The industry isn’t just leaning toward ELS by default. It’s actively standardizing around it. The CW-WDM MSA (Continuous-Wave Wavelength Division Multiplexing Multi-Source Agreement) is an industry consortium specifically defining standards for external multi-wavelength laser sources for CPO and silicon photonics. The founding promoter members include Arista Networks, Ayar Labs, Intel, Lumentum, MACOM, imec, and Sivers Photonics. When the companies building CPO platforms, the companies making the photonic chips, and the companies supplying the lasers all sit down together to standardize external laser specs, that tells you where the architecture is headed.

Sivers isn’t just positioned for this shift. They helped define it. Their CW-WDM compliant DFB laser arrays are already qualified in Ayar Labs’ SuperNova light source module, already being built into Jabil’s 1.6T transceivers, and already being integrated into GlobalFoundries’ AI chip platform. While the industry debates which CPO architecture will win, Sivers is shipping product into the one that’s actually being adopted.

Every AI data center on the planet is about to hit a wall.

This tiny Swedish company makes the fix.

I bought this stock at 11 SEK and it’s done an (almost) 10x since. But I think we’re still early, so here’s what I’m looking to cover:

1. What Sivers does and why it matters

2. The physical bottleneck forcing a $100B+ industry shift

3. Total addressable market

4. Why their tech is nearly impossible to replicate

5. Catalysts ahead, and why I’m still not taking profits

TL;DR

1. Small cap that makes InP lasers and beamforming chips

2. Copper can’t handle next gen bandwidth… wastes energy… switch to optical

3. Builds ELS for silicon photonics and CPO

4. Industry is converging on ELS. NVDA, AVGO, Ayar Labs all use external lasers

5. TAM is massive!

6. 700M opportunity pipeline

7. Partnerships with Ayar Labs (NVDA backed), Jail, Global Foundries, Win Semi

8. Chips Act, Nasdaq listing, Upcoming volume ramp up

1. So what do they do?

SIVE is a Swedish small-cap semiconductor company with two businesses.

First is Sivers Photonics (indium phosphide lasers).

Second is Sivers Wireless (mmWave RF/beamforming chips). They design and manufacture the tiny laser chips that sit inside optical transceivers, LiDAR sensors, and satellite communication modules.

What problem are they solving? As AI scales, data centers are hitting a physical wall. Copper interconnects can’t handle the bandwidth without overheating and drawing unsustainable power. The industry is being forced to shift from electrical to optical (light-based) interconnects.

SIVE’s InP DFB laser sources are the product that enables that shift.

But here’s where they separate from the pack. Most optical transceiver makers integrate the laser directly into the module. Sivers doesn’t. They specialize in External Laser Sources (ELS) for silicon photonics platforms.

But why does that matter? The next-gen architecture, Co-Packaged Optics (CPO), physically separates the laser from the photonic chip. CPO puts the optics directly onto the switch or GPU package itself, which means it needs a standalone external laser feeding light in. That’s exactly what Sivers builds. While competitors are optimized for today’s pluggable form factors, Sivers is already positioned for the architecture that comes after.

2. The bottleneck forcing a $100B+ industry shift

All the GPUs being shipped by Nvidia, AMD, and Broadcom need to talk to each other. Right now, they do that over copper. And copper is breaking.

TrendForce put it in their March 2026 research: as AI clusters scale beyond single racks (think Nvidia’s NVL576, composed of eight NVL72 systems), copper-based interconnects will no longer meet the required performance and bandwidth demands. Electrical loss on 200Gb/s channels using pluggable modules hits roughly 22 decibels, pushing per-port power to 30W. CPO cuts that to 4 decibels and 9W per port.

Ok in plain english: At next-gen speeds, copper loses so much signal strength over even short distances that data centers burn 3x more power per port just to compensate. Optical eliminates that problem. Power is supply constraint. If optical solves that problem — hyperscalers will not bargain prices (high margin).

Read more

TL:DR PENG calls itself an “AI Factory” company, but the AI segment shrank 42% last quarter while the boring memory business (63% growth) is carrying it. Two businesses: it turns raw memory chips into finished modules, and it builds AI data centers as a contractor. Real seat in the AI buildout, but it’s riding cyclical waves. Trades ~21x forward earnings.

Verdict: buying, starting a position here. Not the perfect entry, but I’m not timing the last 20%.

I’m going to walk you through it the way I had to researched it myself. If you stick with me to the end, I’ll show you the valuation, and don’t worry, I wouldn’t make you read all this if I didn’t think the numbers actually worked.

Quick framing: PENG calls itself an “AI Factory” company. The market is buying the AI story. But here’s the twist that made me look closer, the AI segment actually shrank 42% last quarter. The plain, boring memory business is the thing keeping the lights on.

That contradiction is the whole article. Let’s unpack it.

Part 1: the memory business (the boring half)

This is the part that can get confusing, so I’m going to use an analogy that finally made it click for me.

Think of memory like baking bread. Companies like Micron, Samsung and SK Hynix grow the flour. They manufacture the raw memory chips in factories that cost $15-20 billion each. PENG doesn’t do that. PENG buys the flour and bakes the bread: they take raw chips and turn them into finished, usable memory “modules” which is the actual sticks that go into a server.

So PENG isn’t competing with Micron. PENG is Micron’s customer. They make chips usable.

Why does that niche even survive? Because PENG specializes in the orders the giants can’t be bothered with custom specs, smaller volumes, weird form factors, products that need a 10-year supply guarantee for industrial and defense customers. Micron wants to make ten million identical parts and move on. PENG handles the fiddly stuff.

It’s a real moat, but a narrow one. It works because PENG stays small enough to sit beneath a giant’s attention. Worth remembering when we get to valuation.

This memory segment is ~50% of revenue, and last quarter it grew 63%. Hold that number.

Part 2: the “AI Factory” (the exciting half)

The second business is the one everyone’s actually excited about. Easiest way to think of it: PENG is a general contractor for AI data centers.

NVIDIA makes the appliances, the GPUs. But you can’t just buy a thousand GPUs and plug them into the wall. Someone has to design the building, run the power, handle the cooling (these things run hot enough to need liquid cooling), wire everything into one working machine, and keep it running. That’s PENG. They hand you a finished AI supercomputer.

And they’ve proven they can do it at the highest level. PENG built the “Spectra” system that was just accepted at Sandia National Labs, a US national security facility. That’s a serious credential.

But here’s the honest bit I had to sit with. PENG builds the box. They don’t own the GPU, they don’t own the chip design — the high-margin parts of the stack. They’re the plumber on a very expensive house. Good business, but know which part you’re buying.

And this segment is lumpy. It’s project-based so a few big contracts make or break a quarter. That 42% drop? That’s not the business breaking. It’s just chunky by nature. It can snap back just as hard.

Part 3: so why am I actually interested?

Because the long-term setup is real. AI is shifting from training models to running them constantly, that’s called inference, and inference is extremely memory-hungry. More inference means more demand for exactly what PENG sells.

The clearest signal: management just doubled their full-year revenue growth guidance, from 6% to 12%. Companies do not do that casually. That’s them telling you demand is showing up.

Part 4: the risks (this is important)

Here’s what genuinely worries me, and you should weigh it before anything else.

PENG is surfing two cyclical waves at the same time. First, the AI buildout which is enormous right now, but a lot of it is one-time construction. When most companies have built their AI infrastructure, demand steps down to just replacement. Second, memory pricing has its own brutal up-and-down cycle, totally separate from AI. Right now PENG is riding the up side of both.

Both waves break eventually. I don’t know when and nobody does. For now the buildout is still accelerating, so the cliff isn’t a 2026 problem. But it’s the thing that could break this thesis, so I’m not pretending it away.

However that “demand steps down” worry isn’t a step down to zero. Server memory in a data center gets refreshed on a roughly 3-5 year cycle. Not because the chips die, DRAM is pretty durable, but because after a few years it’s just cheaper to rip out the old hardware and replace it than to keep running it. Newer memory is denser, faster, and uses less power, so the running-cost math eventually tips toward “replace.” Data centers basically regenerate themselves on a rolling cycle.

And AI hardware specifically might refresh faster than normal because AI workloads run the equipment flat out, 24/7, hot. Hard use, faster wear, faster replacement.

So even in a world where nobody builds a single new AI data center, there’s a real, recurring replacement demand underneath. The bear case isn’t “boom then zero.” It’s “boom, then a step down to a recurring replacement cycle.” That floor is what makes me comfortable owning this through a cycle.

Part 5: valuation

This is the important bit. Doesn’t matter how good a company is if it’s already priced in.

On management’s own guidance, PENG trades around 21x forward earnings. That’s not cheap, but it’s not crazy either, it’s a “priced for execution to keep working” multiple.

EV/Revenue around 1.9x, which looks low but a hardware business running ~31% gross margins is supposed to trade at a low revenue multiple. Cheap looking isn’t the same as cheap.

So here’s my honest verdict: I’m buying, and I’m starting a position here.

Not because this is the perfect entry. The stock has already run hard, it’s near 52-week highs, and it’s trading above the average analyst price target. The multiple could absolutely compress toward 18x and I’d be sitting on a paper loss for a while.

But I’m also not trying to time the last 20%. That’s a principle for me, chasing a slightly better entry is how I miss things I actually believe in. And I believe in the long-term thesis: PENG has a real, differentiated seat in the AI buildout, on both the memory side and the integration side. I’d rather own it and be early than wait for a perfect price that may never come.

I’m starting small and I’ll add if Advanced Computing reaccelerates. That’s the plan. But that is my plan. Always DYOR. This is not a recommendation for you to buy, but a journal of my entry.

Key charts

Here are some of my best returns to date.

My approach over the years has been simple. Follow the infrastructure.

It started with NVIDIA at $28 . The AI buildout was picking up steam and I believed the market was still underestimating just how much compute infrastructure would be needed. NVIDIA was the obvious first bet, the company building the GPUs that every AI model runs on. But once that clicked, the next question was: who benefits from NVIDIA’s success? Who’s actually manufacturing these chips?

That led me to TSMC at $113, the foundry behind nearly every advanced chip on the planet.

Then Broadcom at $173, which was quietly building the custom AI accelerators and networking silicon that hyperscalers needed to connect it all together. Not the mention the CEO was a fellow Malaysian…

Then I found AMD.

I first entered in the $120s. Then it kept falling. The market consensus was that AMD couldn’t survive in NVIDIA’s shadow, that the GPU war was already over. I disagreed. I spent weeks going through earnings transcripts, listening to hours of analyst calls, reading deep dives on their data center roadmap, studying management’s capital allocation track record. The more I dug, the more I liked what I saw.

When AMD hit $85, I went all in. Doubled down, tripled down. While the market was pricing in stagnation, I was loading up. My thesis was straightforward: multiple compression on a company with underpriced growth, and a competitive threat that the market was wildly exaggerating. The AI compute market wasn’t winner take all. It was big enough for both NVIDIA and AMD to thrive. That conviction came from research, not gut feel.

AMD became my highest conviction position across all of these names in terms of upside potential.

Google was a similar setup. Everyone was panicking about ChatGPT killing Search and the DOJ antitrust case hanging over Chrome. The stock was unloved. I went deep, studied their cloud business, their AI infrastructure investments, their advertising moat, and built a thesis that the market was overreacting. Bought hard at $180.

Sometimes the best opportunities aren’t obscure names no one has heard of. They’re sitting right in front of you, in companies the market has temporarily given up on.

DroneShield was a different kind of trade. Tactical, not long-term. In November 2025, the CEO sold his entire stake, roughly A$70 million worth of shares. The stock cratered from over A$6 to under A$2. The selloff was aggressive and emotional. I bought at A$2.20, rode the recovery, and exited a month later for +80%. Not every opportunity requires a five year hold period. Sometimes the edge is simply recognising when fear has overshot reality.

Now here’s where my head is at.

I’m still following the AI trade, but I’m moving upstream. The obvious names have already re rated. The question now is: where is the money flowing next? Where are the bottlenecks?

Right now I’m deep in optical infrastructure, AI compute hardware, and neoclouds. The infrastructure behind the infrastructure. Already sitting on some massive gains in names like SIVE and DELL , with several others heating up.

That’s how I think about investing. Research deep, build conviction, and act when the market gives you an opening. If that resonates with you, I write about all of it here.

Subscribe now

Entry: $13.90

Most people hear “Nokia” and think of indestructible flip phones from 2005. That business was sold to Microsoft over a decade ago. The Nokia that exists today is a €20 billion networking infrastructure company that quietly became one of the most interesting plays on the AI buildout.

And the market is only starting to figure that out.

The stock has nearly tripled in 12 months. JP Morgan just raised their price target to $21 and flagged an undisclosed Google win in data center switching. NVIDIA invested $1 billion at $6.01 a share. 9 out of 10 of the world’s top hyperscalers already use Nokia’s optical networks. AI and cloud revenue was up 49% last quarter.

Valuation — Price Targets

There’s no point reading further about how great this company is if that growth is already priced in. So let’s get into the valuation.

NOK’s entire growth story is their AI and Cloud revenue. This is up 49% YoY. 1B in orders in a single quarter. Optical and IP revenue is up 20% YoY, and management just doubled their growth guidance to 18-20% for the full year, up from 8-10% just five months ago. AI and Cloud TAM CAGR assumption for 2025 to 2028 is 27%, up from 16% five months prior. Network Infrastructure overall FY2026 is guided at 12-14%, up from 6-8%.

Nokia’s telco business is 73% of revenue and shrinking, while AI and Cloud is 8% and nearly doubling. The entire re-rating bet is that the mix shift continues fast enough that the growing 8% eventually dominates the narrative.

If the market prices Nokia as a telecom vendor, Ericsson trades at 13 to 15x forward. Nokia at 31x is expensive on that basis. But if the market starts applying an AI infrastructure multiple, Ciena trades at 55x and Arista at 45x, then Nokia at 31x is actually cheap. The thesis is about which direction the multiple goes.

Read more