Speedrun of the things to watch out for this week.

Tomorrow is Monday! How is the market looking like right now?

I am keeping this article short as there has not been any material change this week.

Investors should be focused on:

• Iran Conflict Update

• Earnings Calendar

• Economic Calendar

Iran Conflict Update

The U.S. has resumed strikes on Iran…first major strikes in WEEKS.

Monitor how oil prices will react this week.

In relation to Iran conflict, the U.S. has reached a deal with Venezuela where they would get control of 65 billion barrels of its oil reserves.

This sounds good, because the SPR continues to go lower but there is a discrepancy in the oil grade. This oil is heavier, and will need to be refined in order for it to be used.

Earnings Calendar

Big AI names are still reporting this week. Here are some notable names:

• Dell

• Credo

• Broadcom

• Hewlett Packard Enterprise

• Ciena

On the software/cybersecurity side:

• Palo Alto Networks

• MongoDB

• Zscaler

• Snowflake

• NetApp

Economic Calendar

A good site to bookmark: https://finviz.com/calendar/economic

September 4th:

• Non Farm Payrolls (45K est. vs -23K prior)

• Unemployment Rate (4.2% est vs. 4.1% prior)

Important report to gauge how the Fed will decide interest rates in the coming months.

That is your speedrun for the week.

Disclosure: My portfolio is my personal risk management. This doesn't mean you should follow me. It is to give you an idea of how I am positioned and where my money is. Not financial advice. Do your own research.

Happy Friday! September begins next week!

Read more

Interested in listening to the audio conversation of the X space? Click here.

Participants: PJ - P Equity Research, Arvind Srinivas, Hong Chen, and Nick Doyle from SemiAnalysis. Please leave Nick a follow, he started a X account just for this.

Free to read white paper on HBF done by Arvind. Click here.

Highlights:

On cost

• It’s too early to know how much HBF’s cost premium over HBM will shrink. Current figures are just vendor claims: Sandisk says ~1/8 the cost of HBM per bit; a Hot Chips speaker put it at ~10x the cost of SSDs per GB.

• Some costs (yield, testing, controller amortization) will improve with scale — but structural costs (TSVs, stacking, testing every stack, running NAND in fast pSLC mode) are permanent, the same way HBM will never be as cheap as plain DRAM.

• The real wildcard is endurance — nobody knows the wear spec yet, and if cells wear out faster than expected, vendors will need to ship spare capacity, driving costs up further.

On use case — it’s not an HBM replacement

• HBF is narrow: AI inference only, specifically MoE models at low batch size and long context.

• The often-cited 3TB/s bandwidth is an OCP spec ceiling — the actual taped-out die targets 1.6 TB/s, which is only HBM3E-class and already a generation behind incoming HBM4 (2–3 TB/s).

• It’s best at sequential reads (loading model weights), much weaker at small scattered reads.

• Real fit: on-prem/private enterprise with a handful of GPUs that need capacity (fitting a whole model in one box) rather than hyperscale bandwidth. Hyperscalers will keep pooling HBM instead.

On thermal/reliability — an unresolved problem

• Flash degrades faster under heat, yet Sandisk’s proposed architectures place HBF right next to hot GPUs — something flash was never engineered for.

• Mitigations exist but aren’t proven: a UCIe/chiplet pivot to physically separate HBF from the GPU, and scheduled daily refresh cycles to manage retention at 85°C (essentially a daily rewrite for 6 years) — something hyperscalers won’t love.

• No published thermal spec, no demonstrated solution yet.

On manufacturing

• Sandisk/Kioxia bring deep 3D NAND expertise (BiCS8 already bonds logic under flash — useful groundwork), but they’ve never mass-produced HBM.

• SK Hynix’s involvement matters because HBM-style stacking/testing know-how is the missing half — still unproven at real volume.

Big picture

• Memory is shifting from a commodity market to a tiered, specialized one.

• By 2030, ~60% of DRAM bits are expected to go into HBM/servers (up from ~1/3 in 2020); memory as a share of hyperscaler capex has jumped from ~8% to ~30% this year.

• NAND/Sandisk shortage expected to persist through 2027 into 2028, with HBF as a call option that could meaningfully tighten (or eventually loosen) that dynamic.

• Tema’s DISK ETF is positioned overweight NAND (via Sandisk and Kioxia) on this thesis.

Transcript

Introduction

PJ: Welcome to the space. Today we’re going to talk about High Bandwidth Flash, or HBF. I’m PJ, and we have Arvind, Nick from SemiAnalysis, and Hong from Tema ETFs.

Arvind, do you want to go ahead with the questions?

Arvind: Before that, I think this is the first time we’re having Hong and Nick on. Maybe Hong and Nick can each take a couple of minutes to introduce themselves, and then we’ll take it from there.

Hong: Hey guys, this is Hong. I manage the Tema Memory ETF, ticker DISK. We invest in the memory space across both DRAM and NAND.

In particular, we’re overweight NAND in the ETF, which obviously benefits from this move into HBF that we’re following. I’m excited to hear what Nick has to say about HBF.

Nick: My name is Nick Doyle. My background is semiconductor equity research; I’ve been doing that for the last eight years. I joined SemiAnalysis about a year ago, almost to the day.

I work on the core research team, which essentially sits underneath our different verticals — accelerator, data center, foundry, and so on. The core research team publishes financially focused semiconductor and AI research, often in conjunction with those teams.

Thanks, PJ and Arvind, for having me.

Arvind: Great. Welcome, Nick and Hong. Hopefully we can have a good session over the next 30 to 40 minutes.

The format will be that PJ starts with a couple of questions, I’ll take another couple, and then we’ll finish with closing remarks from Nick and Hong.

Without further ado, PJ.

HBF Cost Structure

PJ: My first question is about the cost structure of High Bandwidth Flash.

We know it uses more expensive NAND — SLC rather than TLC — and it has to be stacked, similar to HBM, which makes it more expensive today.

How much of that cost gap do you think can be closed through better manufacturing and design over time? Do you have a realistic timeline for when that might happen, or is it still too early to tell?

Nick: Thanks. First, I’ll say to PJ, Arvind, and the team that your HBF paper was great. It covered a lot of the bases I expected to talk about. I also really liked the simple “what it solves / doesn’t solve” framing. It made the paper much easier to read, so I’d highly recommend it to anyone interested in HBF.

For your question, the short answer is: no, it’s too early to tell exactly how much the cost gap will close.

Today, the only cost data points we have are claims. SanDisk says HBF is one-eighth the cost of HBM per bit. Then, at Hot Chips this week, we had a speaker put it at roughly 10 times the cost of SSDs per gigabyte. Those two numbers bracket things pretty consistently.

We know yield learning can improve. The 16-high stacking, controller amortization, and testing are all defect-rate and cost problems, and those improve over time. HBM showed us that.

But the structural costs don’t really go away. You still have to drill the TSVs, you still have to stack, and you still have to test every stack. And, as you mentioned, we’re running NAND in a low-density, fast mode — pSLC. Those are permanent premiums.

To simplify it, asking whether HBF’s cost can eventually become similar to standard NAND is a little like asking whether HBM will ever cost the same as DRAM. The answer is essentially no. You’re paying for the advanced stacking and testing required to achieve the bandwidth improvements, and the same applies to HBF.

The wild card is endurance.

Flash wears out as you write to it, and nobody knows the final wear specification yet. If the cells wear out faster than expected, you have to ship spare capacity to compensate, which makes everything more expensive.

One final point: Micron is pushing on-package LPDDR, and NVIDIA is shipping GDDR7 with Rubin CPX. Both are alternatives to DRAM and could compete with HBF.

So real cost discovery probably starts once actual devices begin shipping, potentially next year or in 2028.

PJ: Great. Thank you. That was a good explanation.

HBF’s Target Use Case

PJ: My second question is about bandwidth.

High Bandwidth Flash has worse bandwidth than HBM. So in terms of use cases, it doesn’t seem like it can replace HBM, despite some of the excitement around that possibility.

Is HBF better suited for local or private enterprise workloads?

Nick: I think you guys agree with the SemiAnalysis view here: it’s a very narrow use case.

It’s really for AI inference, specifically mixture-of-experts models at low batch sizes and with long context.

HBM is really a precision instrument, not a hammer.

But I do want to clarify the bandwidth numbers, because there are some very different figures being quoted.

The OCP specification defines grades up to 3 TB/s. But the die that has actually taped out targets 1.6 TB/s. That’s the number we think really matters.

At 1.6 TB/s, that’s HBM3E-class bandwidth — essentially what’s in the installed base today. But HBF is going to arrive over the next year or two, and by then we’ll have HBM4-class memory coming out at roughly 2–3 TB/s per stack.

So the key point is that HBF’s bandwidth is effectively a generation behind when it ships.

It’s also basically read-only, and the quoted bandwidth assumes large sequential access sizes. That’s important.

You can hit those peak speeds when loading model weights, which is exactly what HBF is good at. But small, scattered reads should come in well below that peak speed.

So going back to the use case: HBM wins almost everywhere except low-batch, read-heavy workloads.

Once you get to rack scale, HBM always wins.

Who runs low-batch inference? It’s probably not the hyperscalers. They’re batching aggressively and pooling HBM across enormous scale-up domains.

That pooling comment is almost a perfect signal for what we think AI labs want: 4-high HBM. They want HBM, and the industry potentially moving in that direction is enabled by optical scale-up.

But who actually runs low-batch inference?

Smaller deployments, on-premise deployments, and private enterprises using a handful of GPUs.

Those deployments don’t have the huge scale-up fabric. That’s where the 512 GB HBF stacks become interesting, because HBF can make the entire model fit inside the box.

And with only a few users — not millions — they don’t need HBM’s bandwidth. They need the capacity.

That’s where HBF can fit.

I also think it’s interesting that even SanDisk has essentially said they’re not competing for the same HBM market. They’ve shifted their messaging somewhat toward HBF being a capacity tier for inference.

That’s the sweet spot: local workloads.

But again, LPDDR and GDDR7 are competing for that same area.

PJ: That’s awesome. Thank you for the explanation. Arvind, over to you.

Thermal Challenges

Arvind: Thanks, Nick, for the feedback and for explaining those points.

I want to circle back to the degradation issue you mentioned with flash memory.

I was surprised to see the architecture proposed by SanDisk, where almost all of the proposals had HBF next to the GPU.

Correct me if I’m wrong, but my understanding is that flash memory degrades faster under high temperatures. GPUs today generate a lot of heat.

So how does this work? The thermal environment of putting flash next to a GPU doesn’t sit well with me. How do you see this aspect?

Nick: You’re right. That’s been a concern from the beginning.

We don’t think SanDisk has really shown the complete answer yet. There are mitigations, but this is a problem flash was never engineered around.

Retention and endurance both degrade significantly at high temperatures.

Flash was never designed with the expectation that it would sit next to kilowatt-class GPUs or ASICs. Now that’s becoming a major issue.

We’ve seen two main mitigations.

The first is the interface pivot to UCIe. That gives them the ability to move onto chiplets or use chiplet links to place the HBF farther away from the GPU and away from the HBM.

That helps because you can put the flash somewhere cooler, although you pay for it in terms of the data path.

The second mitigation is scheduled refresh cycles to manage retention at 85°C.

The specification is roughly 24 hours of retention before you lose older data. So you’re essentially rewriting the data every day, for potentially six years.

You schedule these refreshes every 24 hours — effectively a daily maintenance window.

And a hyperscaler is not going to love that.

So those two mitigations exist. They’re not great outcomes, but they’re helpful.

The main point is that there’s still no demonstrated thermal solution or definitive thermal specification.

That’s one of the reasons we were skeptical about HBF last fall.

Arvind: That’s interesting because that was the first thing I thought of when I saw those architectures.

SanDisk, SK Hynix and Manufacturing

Arvind: The second thing that confused me is that SanDisk has SK Hynix as part of its consortium, but SanDisk hasn’t historically done anything like 3D NAND in the HBM-style stacking sense.

Now HBF involves stacking NAND on top of each other.

How do you expect SanDisk to overcome that manufacturing challenge? Is there some additional JV with SK Hynix? How do you see the leap from where they are today to actually building 3D NAND for HBF?

Nick: It makes sense that SK Hynix is involved. Samsung may also be involved eventually. Micron seems to be the least aggressive of the DRAM players here.

There are really two different types of know-how involved.

One is NAND expertise.

SanDisk and Kioxia are arguably the deepest team in the world on NAND. Toshiba invented 3D NAND, and BiCS is their platform.

The current generation, BiCS8, already bonds a separately manufactured fast-logic wafer underneath the flash.

That’s essentially the logic layer that HBF needs to run the HBM-class interface.

I think that’s the big reason the first HBF die taped out at SanDisk. They already had many of these technologies in their manufacturing process.

So the first HBF die tape-out demonstrates that the NAND side is real.

But, as you alluded to, there’s another half to the challenge: HBM-style manufacturing.

We’re talking about 16-high TSV stacking, known-good-stack testing, and more.

It’s not just known-good die anymore. You need the entire stack to be tested on top of the die.

Then you have the base die generation and the custom, analog-heavy controller.

Those are all significant manufacturing challenges that will take time.

SanDisk and Kioxia have never mass-produced HBM — that’s the domain of the DRAM manufacturers.

So SK Hynix being part of the consortium matters because they can theoretically help close that gap given their HBM manufacturing expertise.

But regardless of the manufacturer, I think it’s unverified until we see the actual stacks shipping in volume.

HBF vs. HBM4, ZHBM and THBM

Arvind: That brings me to my final question.

There’s HBF, and recently there’s been a lot of focus in HBM on ZHBM, where memory is stacked on top of GPUs.

Then we have Kim Jong-ho, often described as the father of HBM, talking about his lab working on something called THBM, where they want to stack HBM and put the GPU on top of that.

And then we have HBM4.

There are so many developments happening in memory, both NAND and DRAM. How do you see the entire sector moving forward with so many architectural changes and different solutions?

And then there’s the underlying issue of supply tightness.

How do you see the overall memory segment?

Nick: I think the big thing is that memory has always been a commodity market, and we’re continuing to see these different dynamics.

HBM is driving a whole new margin profile for these companies.

As we invent new solutions for different specific tasks, it moves memory toward being less of a commodity market overall.

That’s great.

From a markets perspective, though, it’s a pretty crowded area, and expectations are really high.

SemiAnalysis is still very bullish — don’t get me wrong. Clients are always asking us for our views on the memory cycle, so I want to make sure that’s clear.

But from a technology perspective, the HBM roadmap is incredibly interesting.

It’s very different from my first years covering memory, when we basically cared about cost per bit and bit crossover. That was the primary focus when we modeled these companies.

Now we have all these new inventions.

HBM4 doubles the interface to 2,048 bits, giving you a major jump in bandwidth.

HBM4E is right behind it, and then we have HBM5.

We also have the other technologies you mentioned, including ZHBM.

Samsung laid out a three-phase evolution of the base die. ZHBM is phase three, where they’re eliminating the interposer and other components, and they’re claiming roughly 70% better power efficiency.

That’s a huge improvement.

We’re continuing to move from having a couple of different memory types to having a tiered memory system, and that has implications throughout the supply chain.

Two big numbers we’ve discussed this year help illustrate the scale of what’s happening.

By 2030, we think 60% of DRAM bits will be going into servers and HBM, up from roughly one-third in 2020.

And memory as a percentage of hyperscaler capex has gone from around 8% to approximately 30% this year.

Those are massive numbers, and that’s why the memory trade has worked so well to date.

There’s a lot to be excited about.

I’d recommend that anyone who really wants to learn about memory attend the major memory conferences. FMS is one that immediately comes to mind.

I’ll close by saying that we view HBF as a great call option on the SanDisk story.

But what really drives the broader theme is the SanDisk NAND shortage.

As I mentioned, we’re pretty bullish.

We’re confident in the shortage continuing through 2027 and into 2028.

We already know enterprise SSDs are sold out for next year.

2028 may be the first credible point where the market starts to loosen.

But any meaningful HBF adoption could change that dynamic significantly, just as HBM did last year.

Closing Remarks

Arvind: I think that’s been a very insightful session into the memory landscape.

Before we close, Hong, would you like to give some closing remarks?

Hong: Absolutely.

If you want exposure to the memory cycle, we have the Tema Memory ETF, ticker DISK.

We invest across both DRAM and NAND, and we’re particularly overweight NAND through our holdings in SanDisk and Kioxia, the two largest pure-play NAND stocks.

We’re bullish on NAND because, as Nick mentioned, the market is likely to remain very tight into 2027 and probably 2028 as well.

AI is generating a huge amount of content, particularly images and videos, which is driving storage demand.

And then you have the potential upside optionality from HBF, which could create a new vector of demand growth.

If you’d like to learn more about the ETF, visit Tema ETFs.

Thank you very much, Nick, and thanks to PJ and Arvind for hosting.

Arvind: All right, that brings us to the end of the session.

Thank you everyone for your time.

Bye, everyone.

This is a cleaned-up, lightly edited transcript of an automatically transcribed audio recording. Filler words, false starts, and transcription artifacts have been removed for readability; technical terms have been corrected where clear from context (HBF = High Bandwidth Flash, HBM = High Bandwidth Memory, SLC/TLC = single/triple-level cell NAND, TSV = through-silicon via). Some residual ambiguity may remain in a few places.

Disclaimer: This is not financial advice. Do your own research.

I will be covering highlights from the company’s earnings call today.

Nvidia reported Q2’27 earnings today, and they beat the top and bottom line estimates set by analysts, with a beat on guidance:

The blip? On gross margin estimate for Q3 - but I will address that later in the article.

[1] Revenue

Overall revenue growth remains robust. Last quarter, they grew revenue at 85% YoY, accelerating to $96.2B in Q2’27 or 102% YoY. Next quarters guide is $108 billion, +/- 2%, which sets YoY growth at ~89%. A period of deceleration is expected after quarters of reacceleration in revenue.

Based on personal forecasting, I have Q3’27 revenue landing at $112-$114 billion, or 16% to 19% QoQ, attributed to higher recognition of sales from recent H200 chip sales to China. Important to note, sales to China are not part of the company’s guidance as noted by management time after time in each quarter.

Per Reuters:

“Small batches of Nvidia’s H200 chips, one of the company’s ‌most powerful AI chips, have been allowed ‌to enter mainland China, the Financial Times reported on ​Tuesday, citing two people with knowledge of the matter.

ByteDance and Tencent have each received about 10,000 H200 processors in recent weeks, while ‌a few other ⁠Chinese technology firms could soon secure similar shipments, the report said.”

Assuming an ASP of ~$30,000 (this is the estimated current pricing) on 20,000 H200 processors sold, that is revenue of $600 million.

Because it is Nvidia, assuming a premium beat on guidance by 4-5% is reasonable and in line with the most recent quarter performance.

For FY27, I forecast almost $420 billion in revenue, jumping to $734 billion in FY28 or +75% YoY.

Coming into earnings, I had FY28 growth around 50%, with sell-side around 40%…we all need to revise our numbers.

That said, I am willing to give another premium to the YoY growth rate that Colette mentioned in the earnings call:

“We are seeing demand acceleration even at our scale. Customers forecasts point to our growth doubling next year. However, as I mentioned earlier, we expect to grow approximately 70% as we are supply constrained.”

Due to supply constraints, they will not be able to achieve 100% growth YoY but I do believe they are good enough to come in higher than their internal forecast of 70%. They are, of course, Nvidia.

In other words, analysts will need to come to their desk tomorrow and revise both revenue and earnings BIG TIME. As we stand, Nvidia is about to do FY29 revenue AND better in FY28 alone.

[2] Margins

I did not expect this but maybe I should’ve…

Q3 margins are expected to come in at 74% +/- 500 bps, vs 75% estimate. But the more interesting commentary on margins came in the earnings call from Colette:

“We expect margins to bottom in Q4 in the 71 to 72% range, before settling at 72 to 73% in fiscal year 28.”

“We are experiencing extreme pricing conditions in memory. The magnitude of the price increase has exceeded our prior expectations and are headed even higher into next year.”

It seems higher component pricing (memory) is eating into the gross margins of Nvidia by 400 bps in Q4, and 200-300 bps in FY28. Overall, margins remain stable in the low to mid 70s range - as most sell-side has estimated. However, I am surprised by the fact this dip is even happening.

It seems they are willing to eat up some of the cost, and truly, the overall demand is not as inelastic as previously assume. They cannot pass along 100% of the memory costs to customers as they may have anticipated.

In other news, this may drown some of the fears in DRAM ASP increase over the coming quarters and next fiscal year.

“The magnitude of the price increase has exceeded our prior expectations and are headed even higher into next year.”

Overall, the gross margin difference should not impact the EPS generation for the company, and total gross profit as much.

Based on my forecasts, 75% GM would generate ~$550.8 billion in gross profit, at 73% it falls to $536.1 billion, and at 72% it falls to $527.8 billion. For the bottom line, after operating expenses and tax, this may be a headwind to EPS by ~$0.50-$0.80/share in FY28.

Overall EPS may be higher or lower due to “other income” as a result of investment gains/losses - which I did not factor as they are prone to variability.

[3] Supply Chain

Nice job Jensen…A+++ on securing the supply chain.

“Our commitments increased from $119 billion last quarter to $279 billion.”

In terms of demand visibility, I’m 100% positive the company has a good forecast on demand and revenue for FY29 which they did not speak on, but wouldn’t be surprised if midway through FY 28, they start talking out the next fiscal year’s revenue. They have spent a lot of money on memory, especially, and other parts of the supply chain to keep potential competitors out of their playing field. When it comes to procurement of supply, they remind me a lot of Apple, but this time it’s from a company with a lot more useful free cash flow on their hands.

I also believe it’s time for sell-side to change their views on TPU shipments from Broadcom and MediaTek.

UBS: ~10.4 million

Morgan Stanley: ~7.4 million

GF Securities: ~8.8 million

If Nvidia is growing 70% YoY on tight supply, how is Google and other ASICs users going to be able to ship the number of chips they’d like? I assume some of this revenue being ahead of estimates is a result of higher ASP, but not all.

The following is an expert call conducted by P Equity Research on August 17, 2026.

Expert Background:

• Working in the mobile phone industry, familiarity with consumer electronics.

• Been working in the Middle East, Africa, and Southeast Asia for over 15 years.

• Knowledge and experience in semiconductors as a researcher.

Questions:

1. Which Chinese AI labs are genuinely ahead right now, DeepSeek, Qwen, Moonshot, Zhipu, MiniMax? and what’s that judgment based on, beyond social media talk?

2. Is the chip/compute constraint actually limiting model quality today, or is that narrative lagging behind what’s really happening inside Chinese labs?

3. Is China’s open-weight strategy a genuine technical edge, a market-share play, or mainly a response to being locked out of certain markets?

4. How far behind Nvidia’s current-gen chips do you think Huawei’s Ascend line actually is — one generation, two, or is the gap closing faster than people assume?

5. Which Chinese company do you think makes the best compute chips right now — Huawei, SMIC-fabbed alternatives, or someone less talked about — and what’s your basis for that ranking?

6. Do you think China’s chip yield/output numbers being cited by sell-side analysts are realistic, or inflated relative to what’s actually shipping and usable at scale?

7. If compute is no longer the binding constraint for Chinese labs, what is? Talent, data, power/energy, something else?

8. Do you expect capex on AI infrastructure in China to rise sharply over the next few years, and if so, who’s spending it? State-backed entities, big tech platforms, or new entrants?

9. Sell-side models assume China’s chip output scales up significantly. Where do you think that forecast is most likely to be wrong, on the upside or downside?

10. Do you have any insight into DeepSeek’s or other Chinese labs’ margins, especially on inference? How are they able to price so cheaply, is it genuine cost efficiency (cheaper compute, better engineering), state subsidy, or are they actually operating at a loss to gain market share?

Read more

Disclaimer: If you are an existing paid sub, these prices do not impact you as you are automatically grandfathered for life unless you decide to unsubscribe or have failure of payment in the future.

Dear subscribers,

This article is a short notice on the pricing update of my substack. See graphic below for a brief overview.

My current pricing is:

• $7/month (or $84/year)

• $42/year

• $50/founding

Updated pricing (starting September 1, 2026):

• $9/month (or $108/year)

• $55/year

• $60/founding

Details on what the subscription offers is mentioned when you decide to subscribe for paid.

The annual plan is ~50% cheaper than monthly of $9 (or $108/year). The founding plan is ~45% cheaper.

Why these prices?

I don’t really like to raise prices, but more work has to be put in to expand offerings for subscribers. I am contacting numerous people to do my own expert calls, find research information to share, and even network with people to try to do independent supply chain checks.

I look at my pricing plan as fairly reasonable right now. With the annual plan, you are paying ~$2.3/month or the cost of one candy bar. Believe it or not, I initially started substack at the minimum required rate ordered by substack.

I intend to cap long term pricing rate below $100/year.

Over the past couple weeks, I have had some subscribers asking me why I don’t charge more, and the reason is really that I want to be affordable to everyone, but also provide more research information to everyone.

However, these prices will not update until 7 days from now or at 12 AM of Tuesday, September 1st.

If you decide to join, you will be automatically grandfathered for life.

Thank you for your support ❤️

Disclaimer: This is the authors opinion. This is not a buy/sell recommendation.

Welcome back to late night thoughts. This is a series I am doing where I dump all the thoughts I have into an article, to look into the “what can go wrong” scenario of the AI trade. For reference, here is my first one from a little while back.

P Equity Research

Late Night Thoughts on AI CapEx

Read more

a month ago · 25 likes · 5 comments · P Equity Research

Intention of these articles are to leave you with an open, questioning mind…and not blindly follow every bullish comment made about AI. There is no precedent in history that tells you being positive 100% of the time goes well.

Table of Contents:

1. Double Ordering

2. Sam Altman’s Podcast

3. Siemens & Murata’s Warning

[1] Double Ordering

“We ran a survey of 50 data center procurement leaders (~40% hyperscalers, ~60% colos / neoclouds / enterprise DCs) making purchase decisions for North American projects... Double ordering seems to be a phenomenon for ~50% or more of the respondent set.”

A lot of equipment for data centers has begun to see a higher % of double ordering. If you are unfamiliar with the concept of double ordering, then let me explain.

In simple terms, double ordering is where a company or person orders more than personal demand anticipates due to fear of scarcity of tight supply which would otherwise cause the order to not be fulfilled. Its a reaction to high demand that may not get met by suppliers.

This isn’t the greatest analogy because lemons are perishable, but imagine you had 100 customers standing in a line for your lemonade. You realize you only had 5 lemons but you need 50 lemons to fulfill demand. You call Walmart, Kroger, Target, Costco and tell each of them to have 50 lemons ready. 4 stores, 50 lemons each…you have asked for 200 lemons but you haven’t paid for them yet. You get lucky and Kroger gives you 30 lemons, Walmart gives you 20 lemons. You got your 50 lemons, now what? Well, you call Target and Costco and pull back your orders. What Target and Costco had was a purchase order at one point, and now that is gone.

That is double ordering. But the supply chain dynamics are even worse in this environment.

All of these companies are expanding capacity significantly to fulfill demand, but the real demand is essentially inflated by double ordering. What happens when the order books get pulled back?

The semiconductor cycle is built on this phenomenon. When new capacity comes online and order books get pulled back, companies see a drop in utilization rate which forces them to cut prices to increase demand.

Although this survey applies to data center equipment only, I believe it is reasonable to believe the same reaction is occuring in other areas of the market, such as wafer fab equipment and passive components.

That also takes me back to my supply chain class in college. Let’s talk about the bullwhip effect.

The AI supply chain is the most complex supply chain in human history. If someone else knows a more complex supply chain, then let me know.

When a customer starts signaling softer demand, it does not end with a simple phone call and everyone goes back to business.

When Google calls Nvidia for chips, Nvidia calls TSMC for fabrication, then TSMC calls ASML for extreme ultraviolet (EUV) lithography machines, who calls Carl Zeiss for ultra-precise optics, and then TSMC calls Shin-Etsu for silicon wafers, Applied Materials for deposition tools, and SK Hynix for memory to package it all together.

All of these companies are reporting high demand, seeing tight supply and it is forcing them to expand capacity to fulfill demand needs and maximize potential revenue, but what if the end customer starts pulling back on orders?

This is how the bullwhip ripples backward through that exact chain:

• Seeing massive orders from the tech giants, Nvidia panics and requests an inflated volume of chips from TSMC.

• Seeing those massive requests, TSMC rushes to secure more capacity, pushing heavy demands back to ASML for lithography machines and raw material suppliers for silicon wafers and memory.

• By the time the signal hits the companies making ultra-precise optics (like Carl Zeiss) or specialty gases at the very base of the supply chain, a slight bump in end-user demand or a temporary scramble for chips looks like an unprecedented shortage. Everyone up the chain builds extra factories, buys more equipment, and piles on inventory.

• Eventually, the hyperscalers get their chips, realize they ordered way too many, and cancel their excess purchase orders. That cancellation ripples backward, leaving TSMC with idle billion dollar packaging lines, equipment makers with canceled tool orders, and raw material suppliers sitting on mountains of unused inventory.

This is how double ordering leads to the bullwhip effect, and a result nobody wants to see.

[2] Sam Altman

Sam Altman did a sit down recently with David Senra where he said:

(Cred firesidealpha on X)

"I thought when we got to GPT-4, which was back in 2023, that very quickly after that there was going to be much more disruption, software businesses up for grabs right away, than it turned out to be."

"I think I was wrong about a few things, but one in terms of the speed: the economy just has so much inertia."

"People keep doing the same things, buying from the same company, wanting to use their tools the same way. I think this is actually a positive in many ways, and it's going to make this big transition go smoother and slower. I'm grateful for it."

"But it means we've all been too ambitious on timelines. Even with this incredible technology, society and the economy will adapt more slowly."

Sam has changed his notion on AI adoption being fast to now it being slow, which actually makes a lot of sense, but we were being fed by him and Dario that AI would AGI very quickly and adoption would go to the moon very fast.

These things take time, as do all technologies. That said, the rate of adoption for AI is already incredible…just not good enough.

The rate of adoption for AI is the only way we can measure when the productivity boom will happen for the U.S. economy as a result of AI. Looking back at the internet in the 90s, much of the adoption worldwide only happened after the dotcom bubble burst.

Many companies went to zero but the internet still stood and has been life changing for all of us and has made our lives easier. I am able to do this article because of the internet. That said, it gives another perspective into whether we are spending too fast and should we start backing up and rethinking our spending approaching?

Maybe hyperscalers are spending with brute force to create magic out of thin air?

[3] Siemens & Murata’s Warning

First, it was Murata.

Now it is Siemens.

“We are actively working to not be wholly dependent on data centers…We understand at some point there could be a bubble and we’re racing to pay back the investments as quickly as possible.” - Executive at Siemens

Two companies supporting critical AI infrastructure have sent a warning that spending may slow down eventually resulting in softening demand - not surprising.

The thing people have to understand now is that law of large numbers is going to play an effect on how much more CapEx can go up from him. The larger the numbers get, the harder it is to grow. CapEx for 2027 is still robust, expected to grow ~40% YoY, but this % is going to be closer to 10-20% by year end 2027.

2025, 2026 have been the years where companies have been expanding capacity, with 2026 being the most aggressive year so far. When the spending curve starts to flatten and more supply comes online, there is a high probability that a lot of companies may start losing their pricing power from before where supply was extremely tight and companies were willing to pay any price, especially for companies selling passive components like Murata.

When pricing power comes down, the lever to see generate a higher EPS growth may start fading away, resulting in an EPS revision for several companies.

The risk is that it happens to one company of significant stature in the market…one of the GPU vendors, or equipment companies…and that could trigger a market wide selloff or even be a peak.

A lot of semiconductor firms say they have a clear sight of demand, until they don’t.

I have already addressed this briefly on my X post.

In conclusion, my article is to leave you all with something to think about and something to monitor as you manage risk in the market. It is important to have a skeptical mind and be ready for the worst case scenario.

It is interesting to explore the other side of the argument time to time, to not go insane in the AI bull thesis 😜

Speedrun of the things to watch out for this week.

Tomorrow is Monday! How is the market looking like right now?

Once again not much to cover this week - keeping it very short.

Investors should be focused on:

• Iran Conflict Update

• Earnings Calendar

• Economic Calendar

Iran Conflict Update

MoU expired last week, and the U.S. is preparing to impose new sanctions on Iran.

Interesting enough, Hormuz traffic has picked up quite a bit over the weekend, surging by 400%. But the problem? Somewhat misleading, somewhat of an illusion.

"Weekly traffic rose from ~40 vessels two weeks ago to nearly 200 last week (103 inbound, 89 outbound), an absolute increase of about 160 ships. Pre-war February levels averaged 130-140 vessels daily (~910-980 per week). Current volume is still only around 20% of that.”

Earnings Calendar

CEO of AI is reporting on Wednesday after the close: Nvidia

There are some other interesting companies reporting that I have my eyes on like:

• Semtech

• Intuit

• Salesforce

• Synopsys

• Crowdstrike

• Marvell

Economic Calendar

A good site to bookmark: https://finviz.com/calendar/economic

August 26th:

• Core PCE

• GDP Growth Rate - 2nd Estimate

• Personal Spending MoM

August 28th:

• Kevin Warsh Speech

Probably the most critical event outside of Nvidia’s earnings on Wednesday. He will be speaking at Jackson Hole and will give commentary on bond yields and inflation.

That is your speedrun for the week.

Disclosure: My portfolio is my personal risk management. This doesn't mean you should follow me. It is to give you an idea of how I am positioned and where my money is. Not financial advice. Do your own research.

Happy Friday!

Read more

The following is an expert call conducted by P Equity Research on August 20, 2026.

Expert background:

• Started at AMD as a GPU design engineer

• Worked at Samsung in DRAM, NAND, displays, and foundry

• Now works at Renesas in automotive electronics

Questions:

Analog:

1. Analog and MCU suppliers, TI, ADI, Microchip, Renesas, are reporting improving bookings even as smartphone SoC demand falls. Is that a genuine industrial/auto recovery, or customers front-loading orders because they’re scared the AI driven memory crunch spills into their segment?

2. Most MCUs rely on embedded flash, which competes for the same fab capacity now prioritized for HBM and DRAM. How much of the “MCU shortage” is really MCU specifically versus a memory story?

3. Are margins for analog going to remain durable for the coming years? And in terms of analog players, do you think the ones with fabs are perhaps the most valuable? Ex. Texas Instruments, Infineon

4. AI servers and GPU/ASIC racks still need a surprising amount of analog content, such as power delivery and voltage regulation, around all that compute. As AI infrastructure buildouts scale, is analog still not being appreciated enough? Is demand greater than people think?

Compute:

5. I have seen conflicting messages about compute shortage, mostly people saying we are short in compute but also a select few saying we have enough chips but the issue is we don’t have the power and DCs needed to plug them all in and serve people’s needs. In your opinion, are we short number of chips needed?

6. Nvidia has scale and negotiating leverage that hyperscaler ASIC teams don’t. Does the HBM shortage actually favor Nvidia’s GPU dominance in the near term, or does it push hyperscalers to double down on in-house ASICs specifically to guarantee memory supply?

7. What are your opinions on AMD’s Helios rack as a competitor to Nvidia? I saw a Futurum report that AMD’s rack costs more than Nvidia’s Vera Rubin, which seems counterintuitive for a company seeking to gain market share.

Memory:

8. Memory supply is so tight, and companies are negotiating for HBM4 supply with memory makers as we speak, but is there really any kind of workaround with memory you see coming in the years or are hyperscalers stuck paying? Some sell-side, like UBS, has hyperscalers spending about 63% of capex in 2027 on memory alone - including NAND.

9. How long do you see supply/demand conditions remaining tight for memory? I think 2028 is the year where some capacity starts coming online, but per estimates, it will still be nowhere close to the amount of memory we need.

10. Is cyclicality in memory over with? Is this business going to be completely different than it used to be? For instance, Dan Ives a few weeks ago was talking about how it is still cyclical and what will happen is the spot prices will be way down after LTAs expire, and when they are renegotiated, that is where memory makers start seeing margins collapse.

• 10a. Do the hyperscalers have the choice to break the LTA whenever they want?

Read more

Every clean energy name in this market is priced for a Congress that spent the last year actively hostile to the sector. It is a situation where the regulatory surprise can only come from one direction. We will be honest, clean energy is a slippery trade and has not been easy the last 5 years.

and I went looking early at what ideas could benefit from the midterm election results. We wanted a clean energy business that does not need Washington to change its mind in order to survive, and that has a second story aided by the data center buildout. Something where you get paid for the recovery and get the AI infrastructure angle thrown in for free.

We found a name with a pristine balance sheet and a product line aimed squarely at data center power.

Below the paywall: the full name, two expert calls including an independent power-electronics engineer who pressure-tested the architecture, the competitive map, and where I think the stock goes in a base and bull case.

Disclaimer: This is copy and pasted from original work of for my paid subscribers to be able to read.

Read more

The following is a Q&A session done by "Parrot” - one of my great subscribers - who was able to contact JX Advanced Metals on my behalf and hold a private Q&A session with the them and ask about InP.

Below is a word for word reply from Parrot - all credit to him ❤️

I have posted it on substack for user convenience to refer to with ease.

Read more

Speedrun of the things to watch out for this week.

Tomorrow is Monday! How is the market looking like right now?

Not much to cover this week - keeping it very short.

Investors should be focused on:

• Iran Conflict Update

• Earnings Calendar

• Economic Calendar

Iran Conflict Update

In 60 days, we made zero progress…the MoU between Iran and the USA is set to expire tomorrow.

Additionally, now Marjorie Taylor Greene is claiming the White House is discussing using nuclear weapons on Iran in strategy meetings, stating “im not speculating.”

First of all, I am not sure how legitimate this is - no one has verified this and it is unlikely the USG would even admit to this being true.

Second, who the hell is telling the former Congresswoman on what’s going on inside the White House?

The Hormuz remains closed still, and no effort has been made on re-opening it. Trump declares they own the Strait but that means nothing…traffic has essentially collapsed by 80-90% since the start of the war…

On the bright side, re-escalation has not happened.

Earnings Calendar

Mostly retail taking over this week:

• Home Depot

• Target

• Lowes

• Walmart

• Ross, etc.

I would briefly look at their earnings results and guidance to see how consumer spending is doing.

But that’s not all, some AI names are reporting as well:

• Fabrinet

• Baidu

• Keysight Technologies

• Alibaba

Economic Calendar

A good site to bookmark: https://finviz.com/calendar/economic

August 19th - FOMC minutes

Must read! To understand where FED members are positioned in terms of interest rates and thoughts on inflation.

That is your speedrun for the week.

The following is an expert call conducted by P Equity Research on August 13, 2026.

Expert background:

• Mechanical engineer working in optical networking

• Knowledge in optics, lasers

• NDA - cannot disclose company they work for

Questions:

1. What’s the single biggest bottleneck in optical networking right now — is it lasers, packaging, testing, or something else entirely?

2. If you had to rank the top 3-5 companies in optical components for AI data centers today, who would make the list and why?

   1. Given your list splits components, lasers, and equipment vendors, which of those three layers do you think captures the most margin/value in the AI data center buildout right now?

3. Which part of the supply chain do you think is most at risk of becoming a chokepoint as AI data center demand keeps rising?

4. A lot of retail investors on X talk up Sivers Semiconductors’ laser tech — in your view, is their CW laser actually competitive, or is the hype ahead of the technology?

5. Who do you see as making the best CW lasers for optical networking / co-packaged optics right now? One user on X that goes by insane_analyst has said Coherent has a worse laser than Lumentum and has issues, while Lumentum has the best. Is there merit to that?

6. There’s been chatter on X — supposedly from someone who spoke with a hedge fund analyst — that China is preparing to flood the market with lasers. Have you heard anything like that, and separately, do you think Chinese manufacturers currently have lasers that are good enough to compete on performance, not just price?

7. How long does it typically take to qualify a new laser or optical component supplier for a major customer — is switching suppliers fast or slow?

8. Some sell-side research suggests hyperscalers are still reluctant on CPO adoption, and that if anyone moves first, it’s likely Meta. Does that match what you’re seeing, and do you think 2028 is realistic for scale-up CPO to hit high-volume production?

   1. If Nvidia is the only player positioned to make CPO work because they control the full stack, does that mean CPO adoption is really a bet on Nvidia’s own roadmap (e.g., specific GPU/switch generations) rather than a broader industry trend, and if so, what would you watch for as the signal that it’s actually happening at scale?

   2. If CPO stays niche/specialized rather than high-volume, what happens to the pluggable optics business, does it just keep growing steadily, or does something else disrupt it before CPO ever could?

Table of Contents:

[1] Optical Bottleneck & Top Vendors

[2] Potential Supply Chain Chokepoint

[3] Sivers Semiconductor, CW Lasers, Qualification

[4] China Laser Dump

[5] Switching Suppliers

[6] CPO

Read more

Disclosure: My portfolio is my personal risk management. This doesn't mean you should follow me. It is to give you an idea of how I am positioned and where my money is. Not financial advice. Do your own research.

Welcome August!

Read more

The following is an expert call conducted by P Equity Research on August 13, 2026.

Expert background:

• Working at Capgemini Group - ranked as one of the world's largest IT consulting, outsourcing, and professional services companies.

• Initially started as a cloud engineer, 5-6 years ago, working with Microsoft Azure to help customers migrate from on-premise infrastructure to the cloud.

• Spending time advising clients on AI strategy and building applications that use both open- and closed-source models.

Questions:

1. What’s the single biggest blocker to adoption you see — trust in outputs, unclear use cases, or workflow friction?

2. Where are clients spending most of their AI budget right now — tools/licensing, custom development, or training and change management?

3. Has client willingness to spend on AI shifted recently — more cautious, or more aggressive than a year ago?

4. Do clients budget AI as a one-time project cost or an ongoing operational spend, and how does that change what you propose?

5. How visible is ROI to clients on their AI spend — do they have clear metrics tracking return, or is it still hard to quantify?

6. When choosing between open-weight and closed/proprietary models, what usually drives the decision — cost, data control, performance, or compliance?

7. Do you expect the open vs. closed model balance to shift in the next year or two, or has it settled into a stable pattern?

8. Do you have any insight into inference margins?

Read more

The following is a transcript of today’s conversation on X, in a space hosted by Yuri Khodjamirian of Tema ETFs, Arvind Srinivas, and P Equity Research.

To check out Arvind’s substack, please click here.

To listen to the audio recording of this space, please click here.

Summary:

Introduction & Participant Backgrounds

• Host (PJ / P Equity): Operates a channel/Spaces focusing heavily on semiconductors and information sharing. Notes that optical tech has been heavily trending and serves as the primary moderator.

• Arvind (Retail Investor): A retail investor with a day job and a decade of investing experience. Converted from an AI/semiconductor skeptic a year prior to a dedicated bull due to undeniable anecdotal evidence.

• Yuri Khodjamirian (CIO, Tema ETFs): Manages an institutionally focused ETF provider with roughly $4 billion in AUM. Tema launched the LAZR ETF (focused on photonics and optical tech), which partners exclusively with SemiAnalysis as an independent research provider.

InP (Indium Phosphide) Supply Chain & Bottlenecks

• Indium Metal Realities:

  • ~70% of global indium is produced in China as a byproduct of zinc refining (China controls about 50% of global zinc-smelting capacity).

  • However, indium metal itself is not where the true bottleneck lies; it is a relatively low-value commodity mostly consumed by touchscreens, with optical substrates using only a tiny fraction (a few grams per wafer out of ~1,100 tonnes of global output).

• The Real Crystal/Substrate Choke Point:

  • The actual bottleneck is in InP crystal and substrate growth, controlled globally by essentially three players: AXT (~40%, an American holding company with furnaces/producers in China), Sumitomo (~40%, Japan), and JX Advanced Metals (~10%, Japan).

  • Japan represents roughly 50% of supply, serving as a critical Western-aligned fail-safe if geopolitical export controls tighten. JX is scaling capacity significantly toward 2030, and SemiAnalysis projects China dependence will drop sharply post-2028.

  • InP has no viable substitute at scale because it uniquely emits the exact wavelengths required to minimize signal loss in fiber optics.

Co-Packaged Optics (CPO) & Adoption Timelines

• The Limits of Copper: Copper interconnects are hitting physical limits (at roughly 448G per lane) due to resistance, heat, signal degradation, and lack of pin/connection space at the edge of high-performance AI chips (like NVIDIA GPUs).

• The Transition Roadmap:

  • Near-Packaged Optics (NPO): Serves as an initial bridge technology.

  • Scale-Out CPO: Expected to ramp around 2027 alongside NVIDIA’s Rubin Ultra architecture for rack-to-rack connections.

  • Scale-Up CPO: Expected around 2028 with the Feynman architecture for inside-the-rack connections.

• Adopter Breakdown:

  • Meta and NVIDIA are expected to be the earliest industrial adopters of CPO.

  • Google and Microsoft are anticipated to stick with pluggable optics longer.

  • Amazon may leverage NPO as a hybrid stepping stone.

• Timing vs. TAM: Minor timeline shifts (e.g., pushing timelines out by a year) do not materially hurt long-term discounted cash flows (DCFs) given Goldman Sachs forecasts projecting a roughly 15x increase in the optical Total Addressable Market (TAM) for AI.

Execution Risks & Market Pricing (The AAOI Case Study)

• Execution Hurdles: Ramping production by multiples of current capacity creates severe operational challenges across MOCVD manufacturing equipment, lasers, and wafers. Companies like Lumentum report being completely sold out through 2028 despite quadrupling capacity.

• Market Inefficiency & Mispricing:

  • Applied Optoelectronics (AAOI) needs to triple capacity by the end of 2026 and double it again in 2027, and it is currently tracking ahead of schedule.

  • Despite this, the stock market heavily discounts the company (trading at roughly 10x forward estimates looking to 2028) due to a backward-looking reputation marred by historical cyclical downturns (past transceiver booms, COVID, etc.).

  • This disconnect presents a classic market mispricing opportunity where institutional and retail investors fail to price in a fundamentally changed operating environment.

Geopolitics & Macro Industry Shifts

• Geopolitical Pressures: Export licensing hurdles (such as AXT’s permit requirements) and transceiver bans create recurring friction, making Japanese supply and domestic U.S. production (e.g., AAOI) strategic assets.

• Order of Magnitude Shift: Legacy telecom markets historically bought tiny handfuls of lasers, whereas modern AI data center infrastructure requires hundreds of millions of lasers, changing the fundamental demand profile of the entire photonics sector.

Questions:

1. In terms of optical supply and demand — InP specifically — how tight are conditions? I have sources that say the supply of indium outside China is enough, but the real bottleneck is crystal growth. Is that the case?

2. What do you see the adoption rate looking like? It seems late 2027 or 2028 could be high volume, while CPO is still low volume today — especially for scale-up CPO. Who do you see as the early adopters when that time comes?

3. You mentioned that demand isn’t really in question, but timing may be the key variable. Is it just a matter of when? In other words, do you see execution risks in CPO or NPO? For what it’s worth, they haven’t yet been adopted at industrial scale. What kinds of execution risks do you foresee?

4. Circling back to the first question: I understand that indium metal itself may not be as big a bottleneck as it’s made out to be, but reliance on China is inescapable. Considering geopolitical tensions, is there a way around it? Are there alternatives being pursued — through agreements or technological redesign, designing around it? Is anything happening at an industry level, and how advanced is it?

[04:35] P Equity Research — Welcome

Welcome to our Space, everyone. I think more people will join as we continue. I’m PJ from P Equity. We have Arvind and Yuri from Tema ETFs, and we’re going to talk about photonics today. We’re going to ask some questions and basically develop a base-level knowledge regarding photonics.

We’ll start with introductions, briefly explain what we do, and then go forward with the questions. You obviously know me — you’ve attended my Spaces and follow me. I post a lot about semiconductors and information in general. Optical has been trending a lot, so I’ve been posting quite a bit about optical. That’s my side — I just post and share information, and I’m excited to see what kind of information we can generate from this Space.

Arvind, maybe you can go next, and then we’ll hand it to Yuri.

[05:36] Arvind — Introduction

My name is Arvind. I’m relatively new to Spaces. I’m a retail investor who has a day job, but I’m really passionate about investing, and I’ve been investing for the last decade or so.

On AI and semis, I was one of those people who was skeptical a year ago, but the anecdotal evidence became too hard to deny. So here I am, one of those people converted into an AI/semi bull. I love investing because it’s a hobby and a passion for me. I’ll hand it over to you, Yuri.

[06:25] Yuri — Introduction

Thanks, everyone, for joining. My name is Yuri Khodjamirian, and I am the Chief Investment Officer at Tema ETFs. We are an institutionally focused ETF manager, and we recently launched our photonics and optical ETF. The ticker is LAZR.

We’ve been researching and investing in the photonics space for quite a long time. I even remember looking at some of these companies well before the revolution in AI.

Tema manages several different ETFs and about $4 billion under management. We have ETFs focused on memory and space, and we pride ourselves on building good universes, taking a thoughtful approach to the companies within them, and doing deep fundamental research — especially in some of these more technical areas. I’m excited to get started and dig into the topic.

[07:25] PJ — Question 1: Where is the real InP bottleneck?

Let’s get into the questions. I’ll begin with mine, then hand it to Arvind, and we’ll deliver closing remarks at the end.

A lot of people have been talking about indium phosphide, including me. In terms of optical supply and demand — InP specifically — how tight are conditions? I have sources that say the supply of indium outside China is enough, but the real bottleneck is crystal growth. Is that the case?

[08:02] Yuri — Indium metal versus InP crystal/substrate supply

It’s a really interesting question. For people newer to the space: to create photonics and optical equipment, you need materials different from the silicon used to create other semiconductors. One key material is InP — indium phosphide.

Right now, we have one of the most exciting booms our generation has seen in the AI capex buildout. All of that will require a lot of optical equipment, and that equipment has to be supplied. Suddenly, you’re seeing a surge in demand and a tighter supply situation.

If we break down the supply situation: around 70% of indium is produced in China. It’s a byproduct of zinc refining, so you need zinc smelters, and not all zinc smelters have an indium-recovery unit. About 50% of global zinc-smelting capacity is in China, and China effectively controls around 70% of indium supply.

But indium is a relatively low-value product. The amount that goes into substrates to produce InP components is basically a rounding error: a few grams in a wafer, versus perhaps 1,100 tonnes of global output. So the 70% China share people talk about sounds scary, but that’s not really where the bottleneck is. Indium is mostly used in touchscreens, where it has useful properties.

As you said, the real bottleneck is in the crystals. I’ll use “InP crystals” and “InP substrates” somewhat interchangeably. That market is controlled by roughly three companies: AXT, which we own in our LAZR ETF; Sumitomo in Japan, a large conglomerate; and JX Advanced Metals. Between them, they roughly control the market: around 40% AXT, 40% Sumitomo, and 10% JX Advanced Metals, with the rest held by smaller suppliers.

AXT is interesting. It’s an American holding company, but it owns an InP producer in China. Its furnaces used to grow these crystals are basically in China. Japan also has about 50% of supply through Sumitomo and JX. The key issue is this supply chain, and there are risks if China starts limiting exports. The more Western-aligned supply comes down to the Japanese; Japan is a key part of this.

Another way to think about it is pricing. Substrates are a small part of the total average selling price of transceivers. If substrate prices rise, that can incentivize more supply — and customers such as NVIDIA will be willing to pay more.

The other important point is that there’s basically no substitute here. InP produces exactly the wavelengths of light needed to minimize loss in fiber optics. This is what goes into these data centers. There may be other types in some cases, but there really isn’t a substitute at scale.

One major optical/photonics bull case is that we’ve mainly used copper for these data connections, but copper is starting to hit its physical limits. We’re trying to transfer so much data, over distances that make optical essentially the only solution. Demand is growing.

So where are the supply choke points? Crystals are one, but there are sources of supply. Even if China restricts InP exports, around 50% of supply is still in Japan. The way we think about it is: the bottleneck is real, but there are many other bottlenecks down the supply chain — wafers, equipment used to make the lasers, and other stages. You have immense demand growth, and when you analyze the supply situation it becomes precarious. That’s one reason we’re so excited about the space.

[12:36] PJ — Question 2: CPO timing and early adopters

That makes a lot of sense. So it isn’t just the raw substrate; it’s also the different parts of the supply chain — the equipment, wafers, lasers, and so on.

My second question is about co-packaged optics. What do you see the adoption rate looking like? It seems late 2027 or 2028 could be high volume, while CPO is still low volume today — especially for scale-up CPO. Who do you see as the early adopters when that time comes?

[13:18] Yuri — CPO, NPO, Rubin Ultra and Feynman

It’s an interesting question and one of the big debates in optics. It illustrates why this theme has long structural growth ahead.

The computing power of NVIDIA GPUs has increased so much that all the surrounding components must be upgraded to keep up. A key component is how quickly data can be moved around within the data center. We’re trying to push copper to its maximum, but it’s hitting its limits.

When you increase bandwidth and the number of lanes, copper can’t keep up. There are many technical reasons. Even literally at the edge of a chip, there aren’t enough connections to connect all these things.

One solution is to bring fiber-optic cables and optical components closer and closer to the chip: first onto the circuit board, and eventually integrated into the chip itself. That’s what co-packaged optics, or CPO, is.

It’s already being used in scale-out. Broadcom and NVIDIA have products on the market — switches using CPO. But what you mentioned, scale-up, is where the roadmap debate becomes interesting.

Scale-up CPO is a Rubin Ultra and Feynman story — the next generations on NVIDIA’s roadmap. The roadmap is first CPO around Rubin Ultra for connections between racks, a 2027 story, and that is a powerful growth story on its own. Then you get to inside-the-rack scale-up with Feynman, the next generation, which — as you said — is a 2028 story. This is where copper faces challenges at around 448G per lane.

NVIDIA hasn’t said anything definitive about the architecture, so much of what we’re discussing is still market conjecture. What we like about the photonics chain is that there are multiple bridges. With Rubin Ultra, you have NPO — near-packaged optics — as a bridge. Even if CPO arrives later because of manufacturing challenges, there are other growth drivers in the meantime as data centers integrate optics as closely as possible.

As some listeners may know, we partner with SemiAnalysis on the ETF. They’re an exclusive research provider and research partner for us. They’ve pushed back some of their CPO timelines.

Rather than focusing only on exact timing — and I’ve been investing for many, many years — investors should focus on the size of the opportunity. If you look at the overall DCF, a one-year shift doesn’t matter much when the addressable market is this large. For context, Goldman’s forecast has the optical TAM for AI going up something like 15 times. There will be debate about an architecture that isn’t fully defined yet, but it doesn’t affect near-term DCFs as much as people think.

We think the earliest adopters of CPO will probably be Meta and NVIDIA. Google and Microsoft will probably stay with pluggable optics for a while. Others such as Amazon may start by using NPO as a bridge between the two.

[17:13] PJ — Handoff

That’s excellent — a great explanation. Arvind, I’ll hand it to you for your questions.

[17:27] Arvind — Question 3: Is adoption only a matter of timing?

Thanks, PJ, and thanks, Yuri, for answering those questions.

You mentioned that demand isn’t really in question, but timing may be the key variable. Is it just a matter of when? In other words, do you see execution risks in CPO or NPO? For what it’s worth, they haven’t yet been adopted at industrial scale. What kinds of execution risks do you foresee?

[18:01] Yuri — Execution risk, pricing and AAOI’s capacity ramp

A lot of these bottleneck or AI-buildout themes require understanding demand: how many lasers will go into data centers, getting a real understanding of overall demand, and thinking far into the future.

The second part is supply, and the two interact. They interact through pricing, but also through execution, because the companies we invest in ultimately have to deliver and execute on their targets.

Ramping supply is always challenging — especially when we’re talking about multiples of current capacity. Investors may see execution challenges only as a problem, but if the constraint exists across the industry, it can drive pricing and make the environment better for all companies.

We invest in companies where execution is very important. The demand picture is there, but they need to execute. Take AAOI [Applied Optoelectronics], which we own in the fund. It needs to increase capacity by more than three times by the end of 2026 to hit some targets, then double again in 2027. It’s ahead of schedule.

The interesting thing — and where the stock market can be such a great place to invest — is that the market isn’t pricing it to succeed. Despite strong demand and the fact that it’s ahead of schedule, the market is effectively saying, “No, these guys aren’t going to succeed.” Looking to 2028, the stock is on around 10 times, and the estimates themselves are conservative.

The market is offering an opportunity but handicapping the company because it is backward-looking. This company has a checkered history of delivering into previous transceiver booms — the earlier cloud boom, COVID, and so forth — and it has a bad reputation because of that.

But there is now evidence that it may be able to deliver. For us, the questions are: Does the company show evidence of delivery? Is the environment different this time? What is the strategic value of the business?

This is a vertically integrated company with U.S. production. All the recent noise around China — which I’m sure we’ll discuss as well — creates value for it; it is a major beneficiary. So the market is already looking at some of these companies and pricing in execution risk.

[21:00] Arvind — Question 4: China dependence and alternatives

Even yesterday, Coherent was saying it plans to double its InP capacity twice over the next 15 months. As an observer who isn’t part of the daily grind, it’s difficult to absorb all the numbers being thrown around for these ramps, so naturally these questions come up.

Circling back to the first question: I understand that indium metal itself may not be as big a bottleneck as it’s made out to be, but reliance on China is inescapable. Considering geopolitical tensions, is there a way around it? Are there alternatives being pursued — through agreements or technological redesign, designing around it? Is anything happening at an industry level, and how advanced is it?

[22:04] Yuri — Export controls, Japanese supply and broader bottlenecks

This is a very interesting situation, and I touched on it a little already. China does control some of these things. Take AXT: it has to apply for export permits every time it tries to get crystals out.

We covered that indium metal itself isn’t the problem. The problem is crystals and substrates. But around 50% comes from Japan, and Japan is capable of delivering these materials, so there are fail-safes.

I don’t think anyone can say China won’t use this geopolitically. It has happened in the past and can happen again. China can restrict exports, but 50% of supply remains in Japan, and capacity is expanding. JX is looking to expand capacity by something like 20 to 30 times by 2030. When we talk to SemiAnalysis, they think dependence on China will be much lower after 2028 than it is today. That should ease some of the situation.

But it is a risk. A more recent China-related issue has been the ban involving transceivers, which is quite disruptive — I’m sure we’ll talk about that as well. There are geopolitical angles to understanding supply and demand.

As I said before, bottlenecks are everywhere. There are bottlenecks in machines and MOCVD equipment. I was listening to Lumentum, as I’m sure many people were, and the message was basically: “We’re sold out through 2028 despite quadrupling capacity.” That’s the kind of market we’re in with optics.

[24:02] Arvind — Closing handoff

Thanks. One key question has always been whether the constraint is so intense that it hinders photonics implementation, or whether it’s just one more manageable issue. Now that you’ve clarified that the situation may improve toward or after 2028, that’s reassuring.

I’ll hand it back to you, Yuri, for final remarks, and then to PJ for closing.

[24:28] Yuri — Final remarks

We’ve dived quite deeply into several core areas, so it may help to step back and think about the overall photonics and optical opportunity and how different companies might benefit.

I highly recommend an interview with the Lumentum CEO; he laid out the different opportunities ahead very well.

This industry historically served industrial and telecom markets that didn’t grow very much. Over the last couple of years, enormous growth in AI — and the need to move data quickly — has caused the optical industry to take off.

You have supply-and-demand imbalances benefiting different companies in different ways. This is a genuine bottleneck: supply chains are difficult to ramp at this pace, creating positive pricing and significant potential for these businesses.

It’s often difficult for investors to grasp the scale. The demand picture is there. The Lumentum CEO described old telecom customers as buying only a handful of lasers, whereas now, looking at AI racks, the potential is in the hundreds of millions. That’s the order of magnitude we’re discussing.

As you move through the technology evolutions toward CPO, optics will become embedded in the chips and change the industry’s dynamics. That’s why this space is so interesting. It started from a weak demand picture and therefore a weak supply picture. Now supply has to ramp, and these cycles often last much longer and become much bigger than people expect.

There are also other businesses in the LAZR portfolio that we didn’t discuss today and that stand to benefit: companies making the equipment used to make lasers, companies selling transceivers, and the entire supply chain that has to mobilize to deliver this — just as it does for other components in the AI capex buildout. For us, it’s a very interesting time to be investing in the space.

[26:38] PJ — Closing

That’s awesome. I think we’ve had a strong conversation about InP, and hopefully people were able to learn something. It’s definitely a trending topic and will be discussed for the next few years as we transition away from copper.

Thank you, everyone, for joining. We intend to do more Spaces like this in the future. We’ll see if we can potentially get SemiAnalysis on as well, which would be great.

Thank you for your time, everyone, and thank you, Arvind and Yuri. It was great. I think everyone got to learn something from this.

[27:25] Arvind — Thanks

Thanks, PJ, for hosting. Thanks a lot, everybody, for joining.

[27:41] Yuri — Article and fund link

Thank you so much. If people want more information, we recently co-wrote an article with PJ that’s published on our website. If you go to temaetfs.com/insights, there’s more information there, and you can also find information about the LAZR ETF and its holdings.

Thanks very much, everyone. Have a nice rest of your day. Bye.

[28:28] End of recording

Speedrun of the things to watch out for this week.

Tomorrow is Monday! How is the market looking like right now?

Keeping today unusually short…I don’t see much to cover this week.

Investors should be focused on:

• Iran Conflict Update

• Earnings Calendar

• Market Update on Memory

• Economic Calendar

Iran Conflict Update

Lol. No comment…just watch the development rest of the week…im impressed more BS wasn’t stirred up over the week. So far so good…

Earnings Calendar

• Nebius

• Super Micro Computer

• Cisco

• Coherent

• Cerebras

• Lumentum

• CoreWeave

• Applied Materials

A lot of big names in the AI trade are reporting this week. Do not miss out!

Market Update on Memory

I think I have to cover this briefly, because of how many are invested in memory and how popular Jukan’s tweet on optics vs memory got over the weekend (3M impressions lol).

1. With Korean leveraged ETFs effectively dead, LPs are in a redemption rush, which could bring out additional sell on flow.

2. Nvidia is nerfing Rubin Ultra's HBM and responding with optics, tying multiple racks together, so that even if Rubin Ultra's per rack performance is not superior to Rubin, at the cluster level optics let the Rubin Ultra cluster hold an edge over the Rubin cluster. This holds even if Rubin Ultra's HBM nerf is a supply problem rather than a demand problem.

3. Consensus is forming that memory prices will peak within the next two quarters.

I think he is right. The market is sentiment driven and the so called “efficient market” really doesn’t exist in the short term, only in the long term.

A lot of investors aren’t buy the LTAs signed by these memory makers, and don’t believe margins will remain stable until they see it in future quarters themselves. This to me is one of the reasons why memory has faced selling pressure with rest of tech & semis has been climbing higher the past 2 weeks since Leopold liquidation called the bottom.

Analysts clearly got ahead of themselves 2 months ago with some of the DRAM ASP increases they forecasted, and this is typical of analysts…and several have started pulling back such robust estimates and being slightly more conservative which I believe also plays a role in how memory is moving.

That said, I don’t think memory is a dead trade but this rotation out of them into optical may last for several more weeks, and when people start selling optical again…I anticipate that is when the money starts moving back into memory. It is either that or hyperscalers start correcting again 👀

Economic Calendar

A lot of reporting this week.

• July CPI Inflation data - Wednesday

• July PPI Inflation data - Thursday

• July Retail Sales data - Friday

• August MI Consumer Sentiment data - Friday

That is your speedrun for the week.

Disclosure: My portfolio is my personal risk management. This doesn't mean you should follow me. It is to give you an idea of how I am positioned and where my money is. Not financial advice. Do your own research.

Welcome August!

Read more

Speedrun of the things to watch out for this week.

Welcome to August. Tomorrow is Monday! How is the market looking like right now?

Investors should be focused on:

• Iran Conflict Update

• Earnings Calendar

• Market Update

• Economic Calendar

Iran Conflict Update

Trump was saying he will strike Iran right at the close of the market on Friday (how convenient 👁️) and then he announced they have cancelled the attacked.

He also stated “perimeters of a deal has been agreed to” which Iran followed up with saying that no such deal exists. Ok…

What an excellent thing we have going on for almost 6 months now…

Now Iran and Oman are negotiating over transit through the Strait of Hormuz. Contents of this? Unknown. Iran claims this does not involve in the closing or reopening of the Strait.

This has been a very exhausting war. I truly hope it comes to a conclusion sooner than later. We have lost lives of many innocent people…653 U.S. military casualties and counting…

Keep a close eye on this…in my discord, we have a bot that shares live news all the time and half the messages are related to this conflict…🤣

All of this boils down to the way gas prices move. After rising from the lows from weeks ago, gas is now back to trending lower, at $3.11/gallon compared to $3.50 just a little while ago.

A lot of inflationary concerns with gas prices elevated. One of the reasons why yields are moving higher and the Fed may have to hike once before the year end. A rate hike can be detrimental to the AI trade.

Oh would you look at that…right as i was typing this up.

“there is an agreement on Hormuz and there will be a pact on denuclearization”

Such a good agreement that you have to follow up with “prepared to act when desired” clearly…

Earnings Calendar

Hyperscalers and friends finished reporting last week, but we still have many market moving earnings to be reported this week.

Monday:

• Palantir - has been delivering high growth, will it continue? How are their AI platforms doing?

• On Semi - what is demand visibility on power semis? Supply/demand dynamic?

Tuesday:

• AMD - sales growth? guidance? How is Helios doing? CPU demand?

• SpaceX - don’t have much to say

• Arista Networks - demand on networking

• Astera Labs - growth, guidance, supply/demand dynamic

• Caterpillar - sales of generators? how constraint is power?

Wednesday:

• Sandisk - NAND demand, margins, outlook, and supply/demand dynamic

• Western Digital - margins, outlook, and HDD demand

• TTM Technologies - PCB demand for AI

Thursday:

• Applied Optoelectronics - transceivers demand, supply/demand dynamic, capacity expansion

• Himax - sales growth, margins, outlook, CPO ramp?

Friday:

• ACM Research - outlook of WFE spend in China?

• Oklo - don’t have anything to say, maybe outlook on AI power needs

• Vistra - don’t have anything to say, maybe outlook on AI power needs

Many other economic indicator names reporting like etc.

In my opinion, the market moving reports will be from Sandisk, Caterpillar, and Western Digital.

Market Update

Leaving this short. Got nothing new to share…two days of green in the market, are we back? SOX still 19% off its peak, no reason to FOMO buy - anything can change in an instant.

A lot of names look far better in R/R now then they did in April and May.

Semis are now the most oversold since bottom in late March.

Economic Calendar

Friday, August 7th, the jobs payroll and unemployment numbers comes out.

Nonfarm payroll is expected to be 91K, up 34K from prior.

Unemployment rate is expected to be 4.3%, up from prior 4.2%

Why does this matter?

Strength in labor market = reason for Fed to consider a rate hike in the future and could potentially lead to yields rising.

That is your speedrun for the week.