Paradi Lab

Kinda crazy how no one cares about frontier pacing now. That was ridiculous from Dario and co.

Markets have since digested the fear from last Monday quite nicely and now seem excited with $Meta Platforms(META.US) Muse which is a massive efficiency unlock + HUGEEE compute multiplier. Every lab will build something similar eventually = CPU and cloud compute trade keeps going nuts.

(Wish I had access to Muse in the UK!)

Plus there's strong rumours of OpenAI releasing new models soon -- GPT-6 Sol. And new Anthropic models in the pipeline.

Semiconductor fwd P/E has been tracking inverse oil prices over the past year. Almost perfectly.

⬜️ $SOXX Fwd P/E Multiple 🟦 Inverse oil price

Since the Hormuz shock in Feb:

As oil prices drop, semiconductor valuations rise.

As oil prices rise, semiconductor valuations drop.

AMD, TSMC and OSAT Channel Check:

1. AMD's AI GPU production (MI455/MI450) is expected to remain 100% at TSMC CoWoS in 2027. Production estimates land at a mid-point of 600k units for both SKUs next year per recent MS forecasts.

2. While ASE and Amkor still lead OSAT capacity, their capacity ramps are rumoured to be progressing slower than expected. This means that TSMC may be required to support CoWoS-L wafers for AMD's Venice CPU prodction which would further constrain internal capacity for MI455/MI450 production. Per MS, this would equate to TSMC supporting ~80k CoWoS-L for Venice production.

$AMD(AMD.US) $Taiwan Semiconductor(TSM.US) $Amkor Tech(AMKR.US) $Advanced Semiconductor Engineering(ASX.US)

$Nebius(NBIS.US) INCREASE ON-DEMAND GPU RATES EFFECTIVE OCTOBER 1.

H100: $3.85 per GPU-hour -> $4.50

H200: $4.50 per GPU-hour -> $5.40

B200: $7.15 per GPU-hour -> $8.50

B300: $7.85 per GPU-hour -> $9.50

AMD EPYC Genoa CPU and memory rates will also increase.

Intel Ice Lake CPU pricing remains unchanged.

Honestly - the FOMC decision / rate hike means nothing until the Iran war ends.

After all, the Fed can't reopen the Strait of Hormuz...

Pacing the frontier is ONLY about Anthropic's IPO.

Investors are being buttered up for fewer future model releases.

Why? Margins.

I don't think margins are "bad" per se, but there are always deeper questions on whether they last/grow longer-term.

Ultimately, shipping a new model with better capability at current price points hurts the labs. From a unit economics perspective, that's a price cut per unit of intelligence every time a new model is released. That compounds super fast at the current, highly frequent release cadences.

So slowing model releases means Anthropic can slow price cuts per unit of new intelligence. I think that this is their only clean lever left to preserve margins unless they jack up prices to unforeseen levels which no customer will readily entertain.

This is why we saw Sam and Elon agree with Dario yesterday. It only works if everyone else slows down too. As a margin lever, pacing has to be done industry-wide, and an industry-wide agreement not to compete on cadence is basically an agreement not to compete on price.

Looking at costs: the labs expense training as R&D as they go, so nothing gets amortised in the accounting sense. But the economics are the same. Every frontier model is superseded IMMEDIATELY after the newest one is released. Fewer models being shipped means they each earn more for longer, which means training cost per revenue $ drops without the labs needing to spend anything less than they already are.

This is an investor's dream lol. Again, no coincidence that this is all coming out so soon before Anthropic IPOs...

With OpenAI, I think it's the same logic but from the opposite direction. They don't have the profitability Anthropic has so it makes more sense for OAI to wait to IPO as pacing takes effect. That's why Sam agreed with Dario.

Elon only agreed with Dario because xAI are just so far behind the frontier lol. Also...Elon clearly wants the US to win above all else...so how exactly does pacing the frontier benefit the US in the race against China?

So many tech valuations hinge on forced customer retention scores. I hope the market wakes up and realises that there's absolutely zero moat in usage streaks, notification loops and decent UI. These just end up with low-quality network effects that tech companies harp on about to save face from a relative lack of innovation.

For this reason, Oura's IPO will be tragic. They'll push subscription prices even higher to satiate shareholders, which will work in the short-term, but eventually customers will fall off the edge of a cliff once they've reached the peak of the price curve. My guess is that that's coming soon based on how they've hiked prices over the years.

I genuinely feel like the world needs to wake up to these fads lol. Why do you need a ring or watch to tell you how you feel? I understand tracking steps or calories burnt, but stuff like sleep scores are probably the most useless sales feature that's been fabricated out of thin air over the past decade. Like, do people not know if they had a good nights sleep once they wake up? Who ACTUALLY cares if your sleep score was 87% in March and then 82% in April???

The answer: people who PAY for the hardware/subscription.

It's all backwards. You pay for this thing which means you're forced to care about what it tells you. Obviously there's a tiny subset of the physical elite who do care, but for most (relatively sedentary) customers, why does this stuff even matter other than sybolizing status / fashion accessory.

Whoop are another (probably worse) example with an even worse product because their marketing/hype has been the only thing that's kept them functioning. Them IPO'ing would be genuinely hilarious because the stock would pump when they announce a new sportsperson sponsorship rather than anything to do with product innovation...which would never happen btw.

I hope Apple takes the entire health hardware market by just acquiring Oura lol. That'd be huge aura (not Oura) points for John Ternus.

🚨Leopold/Situational Awareness returns?

Since last Friday, there's been a string of significant Flex calls bought in concentrated AI names: $315M total options premiums / $1.1B of delta / $5.8M of vega.

Positions:

- $Sandisk(SNDK.US): Jan 2040/2200 calls - $57M prem, $198M delta, $1.04M vega.

- $Bloom Energy(BE.US): Jan 250/310 calls - $48M prem, $140M delta, $565k vega.

- $Intel(INTC.US): Jan 105/115 calls - $48M prem, $185M delta, $870k vega.

- $Coreweave(CRWV.US): Jan 105/115 calls - $43M prem, $160M delta, $770k vega.

- $DRAM: Jan 65/70 calls - $43M prem, $173M delta, $855k vega.

- $SK Hynix(SKHY.US): Jan 190/210 calls - $39M prem, $148M delta, $710k vega.

- $AMD(AMD.US): Jan 540/580 calls - $36M prem, $193M delta, $1.08M vega.

It'll be confirmed as Leopold if we start to see some big FLEX calls in $Micron Tech(MU.US) / $Nebius(NBIS.US) / $Taiwan Semiconductor(TSM.US) soon.

Data via CBOE. NMR + GS desks aligned on "single mystery buyer" narrative...

Just some notes from $Nebius(NBIS.US) fireside chat at Goldman Sachs Tech Conference:

TLDR: Yep, it's bullish.

1. Order book extends into H1 2028 which is 2 quarters further out than at Q2 ER.

Arkady said "people are demanding tens of thousands of vGPUs and GPUs now. So we see demand today as unlimited."

Bro...what?!!

At Q2, they said that all of 2027 could be sold today. But Nebius having visibility ~1.5 years away massively de-risks their ~5GW contracted power target and their ~$25B of FY26 capex.

On demand, Arkady also said that demand visibility now extends to 24+ months (vs. 18 months previously).

This is HUGE because demand duration is a huge crux for neocloud bears (kinda makes sense why). Even $Coreweave(CRWV.US) CEO said at the GS Conference that they are "struggling to meet demand everyday."

2. "We do not pre-sell much. We are focusing on free capacity, which we will be selling later"

I think two things are happening here with Nebius

First - this is the opposite of $Coreweave(CRWV.US)'s model where their ~$104B backlog is take/pay dynamics. Nebius are instead choosing to sit on uncontracted 2027 capacity so they can sell into rising prices, kinda like how $Micron Tech(MU.US) and co. were doing pre-LTAs.

Second - keeping some spare capacity keeps room for longer-term strategic partners arriving into 2027 (which will be extremely supply-constrained). I.e. enterprise names coming via the $Palantir Tech(PLTR.US) partnership.

I think both reasons make sense, I'm fairly confident that 2027 pricing will be higher.

3. "We actually have a list of new customers that are looking for any of the older generation chips that come available."

This lines up with $Coreweave(CRWV.US) disclosing at Q2 an A100 contract priced out to 2029 and completely guts the residual-value thesis bears like Burry lean on.

4. $Shopify (SHOP.US) "used open-weight models, trained it with their own data repetitively, and they achieved the quality which is higher than they had with GPT-5 and 6."

Feels like that's the whole enterprise adoption thesis summarized....narrow domain, pvt data, repeated loop, open weights > fronteir, at a fraction of the cost.

Probably also why the token factor and the Tavily acq. matter.

So yeah, Nebius' entire infrastructure goes kinda crazy. More than just GPU rental.

5. Contract mix

Marc: 3-6 month short-term deals go out "at a multiple of the ARR per megawatt" of the core. 1-3 year medium-term deals are "the lion's share" and 5+ year hyperscaler deals were done "with the explicit intent that we are looking for the capitalization benefit."

- So the longer duration $Microsoft(MSFT.US) + $Meta Platforms(META.US) contracts are lower risk collateral for financing the build

- the opportunistic short-duration surge contracts at materially higher pricing are the top-ups.

- and the 1-3 year book with AI natives and enterprises (priced above hyperscaler deals + prepaid) are the core + fastest growing segment.

Just for a summary of the points I found most interesting / different from Q2 earnings.

Cool to also see the "the vision that Arkady has is us becoming a hyperscaler."

I've been sharing the same vision for some time now, and is why Nebius ultimately deserve to trade multiples higher than Coreweave for example.

Someone check-in on Burry!

How can you not love the $Nebius(NBIS.US) partnership with $Palantir Tech(PLTR.US)?!

Palantir naming Nebius as their "preferred sovereign AI infrastructure partner" is so cool because it effectively hands Nebius enterprise distribution on a plate.

It also diversifies future revenue away from $Microsoft(MSFT.US) and $Meta Platforms(META.US) which forms the basis of many Nebius bear theses.

Incredible growth expected for $Lumentum(LITE.US), $Coherent Corp.(COHR.US), and $Applied Optoelectronics(AAOI.US) with new GS optical transceiver forecasts.

- Global optical transceiver TAM revised up by 33% / 81% / 115% in 2026-28E.

- In 2026/27/28E: optical transceiver market to reach $68B/131B/148B.

- Driven by (1) higher rack-level AI server and ASIC AI server shipments and (2) higher estimates on the usage of optical transceivers per Nvidia GPU in rack-level AI servers.

- 800G and above segment to increase at a +69% CAGR to reach $45B/108B/130B.

- 800G/1.6T shipment volumes at 45M/33M units in 2026E, growing to 49M/71M in 2027E, with 3.2T shipment ramping up to 23M/68M in 2027/28E.

- Silicon photonics will account for 60%/80%/80% of 800G/1.6T/3.2T optical transceivers.

Just insane growth on top of what most people had forecasted a few months ago.

Just a quick macro note.

Feels like the broader market will move sideways for the next 2-4 weeks, but it wouldn't surprise me to see some pressure on high beta names.

That said, fundamentally, nothing's changed w/ solid earnings across most sectors especially AI/semis names.

Stocks are in a huge bull marke, even factoring in the July dip in semis (useful to zoom out here w/ loads of names up 2-4x YTD). History tells us that this kinda bull market ends with a Fed hike / recession. Recession won't happen imo, but it's currently unclear if the Fed hikes...and to what extent? I ultimately feel like it needs to happen at some point - Sep or Dec. Get it done with lol.

Seasonality could come into play too w/ midterms usually being bad for equities pre-election, then good for equities post-election. Historically, Sep is a down month before Q4 and pre-Xmas period being positive. Lots of data around this.

In terms of catalysts, Iran/US always the hottest topic these days. Oil price impacts and second-derivative equities impacts.

For AI/Semis, you've also got the GS Communacopia + Technology Conference soon w/ OpenAI etc all doing fireside chats. On top of $Micron Tech(MU.US) and $Broadcom(AVGO.US) earnings this week.

Time will tell whether MU earnings can re-rate memory names, or AVGO earnings have the same impact for the broader semis ecosystem. $NVIDIA(NVDA.US) didn't really have that effect, so I kinda doubt AVGO will.

Silicon Photonics' optical transceiver market share is estimated to exceed 50% by 2027.

$Taiwan Semiconductor(TSM.US) VP of Advanced Packaging said today that silicon photonics is becoming the mainstream form thanks to the structural transformation of optical transceivers.

He also stated that the real bottleneck for large-scale deployment lies in lasers, optical fibers, fiber optic connectors, and product testing.

If I was to pick the "main" players instead of just name-dumping:

- $Lumentum(LITE.US) / $Coherent Corp.(COHR.US): InP lasers - CW for SiPh, ELS for CPO, 200G EMLs

- $Aehr Test(AEHR.US): SiPh WLBI testing

- $Corning(GLW.US): PM fibre, fibre arrays & MPO/MTP connectivity

- Sumitomo Electric (5802): InP lasers + substrates

Obviously you've got a ton of other names from: $Applied Optoelectronics(AAOI.US) for InP lasers, $IQE for InP base epi, $BESI for hybrid bonding and $Keysight Tech(KEYS.US) / $Viavi Solutions(VIAV.US) for optical test.

Feels like TSMC have been running the same narrative alongside the COUPE roadmap for a while now though.

But always good to get some re-confirmations on where everyone's favourite bottlenecks sit / narratives coming from TSMC themselves.

With $SIVE, you're paying EXTREMELY premium 2027 multiples for a "what if" scenario.

That is: What if Sivers' $1.2B opportunity pipeline converts into *actual* sales by roughly H2 2027.

Whereas with $Lumentum(LITE.US), $Coherent Corp.(COHR.US) and $Applied Optoelectronics(AAOI.US): you've already got revenue that's sold out, LTAs to 2028-30, shipping today, on 30-50% GMs. No "what if" needed.

So either:

1. $Lumentum(LITE.US), $Coherent Corp.(COHR.US) and even $Applied Optoelectronics(AAOI.US) are VERY cheap right now [5-14x 2027 sales].

2. Or Sivers is VERY expensive right now [~35x 2027 sales based on my current modelling].

I'll share a post later/this weekend for more detailed comments.

$NVIDIA(NVDA.US) generating $96.2B of highly profitable revenue, while also growing 106% annually is genuinely insane at their scale. People need to let that sink in.

Combine that with expectations to "grow revenue by approximately 70% in fiscal 2028" under current supply constraints...

Makes me think that $NVIDIA(NVDA.US) is perhaps the cheapest and highest quality growth stock we'll see for a long time.

Thoughts on $NVIDIA(NVDA.US) later today:

Prediction:

- I don't see a revenue beat being a catalyst for NVDA.

- Everyone knows that revenue will be ~10-15% ahead of estimates thanks to GB300 (and GB200) volume expansion.

- Moreover, looking historically, recent beat & raise quarters have not been a positive re-rating catalyst for NVDA's share price... perhaps the quarter's already been priced-in due to a +12% move in the last month.

- On the flip side, maybe this time will be different thanks to the stock currently trading at extremely low valuations... I actually think broader institutional sentiment improved as of yesterday too with significant short covering which drove the share price up. Time will tell.

For next quarter:

- No surprises - guidance will of course increase thanks to additional ramps in rack volumes, plus first Vera Rubin orders filtering through.

- I expect gross margins to remain stable thanks to ASP increases offsetting continued memory price surges from $SK Hynix(SKHY.US), $Micron Tech(MU.US), and Samsung.

In terms of "qualitative" catalysts, there are a few areas Jensen and co. need to touch on though (common threads across the street):

1. HBM:

Per JPM/MS/GS (and wider industry comments), NVDA have made changes to HBM content for both Rubin and Rubin Ultra. This is of course due to memory supply being extremely scarce currently - but what does this mean for NVDA exactly? On the flip side...what does this mean for the memory players? Can NVDA keep passing on price increases to customers if memory prices continue soaring? I hope this is all expanded on on the call in some detail.

2. Competition such as $AMD(AMD.US), $Cerebras(CBRS.US), OpenAI Jalapeno:

Around ~40-50% of NVDA's revenue comes from four hyperscalers that are all deploying their own stuff - $Alphabet(GOOGL.US) TPU, $Amazon(AMZN.US) Trainium, $Meta Platforms(META.US) MTIA, $Microsoft(MSFT.US) Maia. Plus OpenAI / $Broadcom(AVGO.US) revealing benchmark results for their in-house Jalapeno chip which shows 1.5-1.9x more throughput per kilowatt + 1.7-3.6x lower end-to-end latency against NVDA's GB200 and GB300 rack systems. All of these factors have been working against NVDA recently, even just from a sentiment perspective. Jensen's updated views will be appreciated!

3. Funding partnerships:

A few weeks ago, NVDA announced a partnership with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR "to establish AI compute infrastructure financing platforms to mobilize over $500 billion of third-party capital." Again, what does this mean for NVDA? They quite clearly have sight over demand longer-term, but "circular financing" is still something that still comes up a lot these days (despite Jensen outlining why it's not in his X article).

Big day ahead, expecting some volatility thanks to last minute positioning adjustments.

Just some of my notes from $Semtech(SMTC.US) Q2 earnings:

TLDR: Like $Applied Optoelectronics(AAOI.US), it's all about Semtech expanding capacity - CEO said the capacity they've secured "may not be enough" for FY28...

1. Q2 upside is from 1.6T qualifying early

- Pretty shocking (in a good way) how compressed the qualification timelines have become.

- Hyperscalers and everyone upstream are pulling timelines forward so quickly rn.

- We saw with $Lumentum(LITE.US), $Coherent Corp.(COHR.US), $Applied Optoelectronics(AAOI.US) especially that demand is just bulldozing any qualification barriers. Now seeing the same with Semtech.

2. FiberEdge is quite underappreciated

- I agree w/ mgmt that people are "overindexing on CopperEdge" - the better business is probably TIA/driver.

- 800G TIA share has gone from ~18% two years ago to well over 50%, and they expect >50% share at 1.6T by January.

- Impressive...

- Industry 800G units: they entered the year on a ~50M forecast and are now hearing 80-90M vs. ~20M two years ago.

- Impressive again...

3. Content per transceiver

- goes from high single digits to $80-90 at 3.2T.

- Quite funny - one analyst assumed that "high double digit" content meant teens. CEO corrected him with $80-90 lol. That's ~10x!

- Photonics fab capacity goes 3-4x by year-end (they picked up a fully facilitated fab next to the existing one).

- I don't think the market has modelled any of this. Even at half the claimed content, DC revenue stops scaling w/ transceiver units + starts scaling w/ units times content.

- And every merchant InP line being tripled is another pointer that InP demand is way ahead of supply.

- Which is the same signal $Applied Optoelectronics(AAOI.US) sent by clearing its HQ building for InP wafer expansion.

4. More capacity needed (obviously)

- Secured capacity "may not be enough" especially 2H FY28.

- Semtech are negotiating prepayments + joint capex with front/back-end partners, and qualifying additional OSATs to spread geopolitical risk.

- Pricing: no erosion expected near term and none in the booked backlog. Cost increases are being passed through. Just what you wanna see :)

5. Gross margins are pretty insane

- 54.5% in Q2 -> 58.3% guided -> 63.9% excl. the cellular module business being divested (closes Q4).

- CFO framed ~64% as the post-close starting point.

- I actually think 64% is the margin floor, not the target. Pretty sure they know they can do more lol. No reason to send out such a high target to the market otherwise.

CEO also said that "we have the financial capability" to fund the FY28 capacity push":

Looks like that's the case based on quick napkin maths:

- FCF was $61M in Q2

- Q3 EBITDA guided to $134M

- I'd say roughly $300M+ annualized FCF exiting the yr

- Add $204M cash + $62M coming from Compal for the cellular module unit closing in Q4

So would be surprised to see any more dilution on top of what they already did a year ago.

Overall though - pretty good earnings.

Today seems to be a red day for semis and memory.

I will be pleasantly surprised if we end up green - that'll be a huge/positive sentiment shift towards semis.

Traditional defensive stocks are all showing some strength, as well as SaaS which I personally class as "defensive tech". That's the < long SaaS, short SaaS > trade. Moreover, if $Apple(AAPL.US) is up in the pre-market, then high beta semis tend to see a negative hit.

Yields for 30Y treasury is at the highest point since pre-GFC at 5.32%. Similarly, the 10Y is at 4.74% - also at pre-GFC highs.

Seems like it's due to a few reasons: rising oil prices (new highs for Aug) and renewed fears of rate hikes from the Fed.

I don't think a hike makes sense right now, but time will tell what happens with Iran. The Fed drops their meeting minutes from the July meeting this week - will be an indication of where rates head.

I don't see this as a long-term issue though. Just feels like traders are unlevering positions after a strong few weeks of momentum across semis and memory. Many institutions are now overweight high beta tech and are sitting on (relatively) huge gains MTD. Would make sense for them to unwind some of their positions / rebalance amid an uncertain macro environment.

Fundamentally, I wanted to emphasise that the AI trade is stronger than ever after Q2 earnings. Semis and memory stocks get thrashed around from pillar to post when macro is so volatile.

Like I mentioned the other day, if names like $Nebius(NBIS.US) or $Sandisk(SNDK.US) can have 10% green days, they can certainly have 10% red days.

Overall, I am very bullish long term - just good to be aware of macro overhangs which drive short-term prices.

Just need to ride into Jackson Hole next week.

$SOXX $SPY $QQQ $SMH

GPU fleet unit economics.

TLDR: Tough to be bearish on $Nebius(NBIS.US) or $Coreweave(CRWV.US) right now.

Some maths, but it's important to understand.

Using $Nebius(NBIS.US) 1-3yr contract economics:

- $20-25M ACV per MW at a payback of 1 year 10 months.

- Take $22.5M midpoint ACV/MW at a ~55% cash margin ( $Nebius(NBIS.US) core AI cloud runs ~50%, $Coreweave(CRWV.US) 59%) and you get ~$12.4M/MW of annual cash flow.

(Payback runs against cash flow, not revenue)

- 1.83 years of that implies "all-in" capex of ~$20-23M per MW (GPUs included).

- Which ties to Nebius' $20-25B capex guide for 2026 against ~1GW.

- Over a 4-5 year life that's roughly $50-62M/MW of cumulative cash flow against ~$22M of capex.

- NPV is +$20-25M per MW even discounting at 12-15% and IRR near 50%.

Then years 1-2 of DCF nearly cover the entire capex on their own. Meaning that the GPU could go to $0 residual after year 2 and the deal basically breaks even. That is insane.

So really, the bear case doesn't need slow decay to be wrong. Rather, it needs the contracted years themselves to fail.

And those years are effectively walled off rn.

- Around 70% of $Nebius(NBIS.US) deals carry prepayments covering 50-60% of the capex, on take/pay terms.

- $Coreweave(CRWV.US)'s $104B backlog is take/pay with ~21% of it recognized more than four years out.

As we now know, $Coreweave(CRWV.US) signed an A100 contract running into 2029 (a SKU introduced in 2020) "at or above where it was years ago", and CEO told CNBC a batch of H100s coming off an expired contract was re-booked immediately at 95% of the original rate. $NVIDIA(NVDA.US)'s CFO also said that A100s shipped 6 years ago are still running at full utilization.

And $NVIDIA(NVDA.US) is now willing to underwrite residual value itself, up to 25% per project in the new financing platforms. I don't think they'd be willing to backstop if they expect GPUs to be worthless in around 3 years or so.

For the neoclouds themselves though, depreciation is the biggest cost driver e.g. Nebius D&A was larger than their adj EBITDA last Q.

If economic life actually extends to a new base case scenario of 5-6 years rather than the old 3-4 years...

Earnings power re-rates higher across $Coreweave(CRWV.US), $Nebius(NBIS.US) and even the hyperscalers as ultimate beneficiaries.

Photonics notes after $Lumentum(LITE.US), $Coherent Corp.(COHR.US), and $Applied Optoelectronics(AAOI.US) earnings:

TLDR: Earnings were bullish for photonics. The entire sector is sold out and LTAs are removing cyclicality.

1. Everything is sold out

Every single CEO highlighted that demand is strong and growing, with current revenues/margins mainly gated by limited capacity.

- $Lumentum(LITE.US): “Lasers will remain effectively sold out for the foreseeable future despite our rapid capacity expansion.”

- $Coherent Corp.(COHR.US): “The demand is robust across almost every single product across data center and comms, and it is really just a matter of whether we can sell as fast as we can ramp production.”

- $Applied Optoelectronics(AAOI.US): Revenue “bounded almost entirely by our production capacity and key component availability”

Very nice and very expected. But good to have confirmations from the horse's mouth.

I also wanted to remind everyone that historically, optics has been a deflationary business mainly due to ASP erosion over time. Similar to memory.

However, everyone being sold out completely flips that with Lumentum saying that they’ve “been surprised at our ability to price up” on CW lasers, holding a premium on spec tightness because of better transceiver yields.

Filtering down onto margins, when ASPs go up, gross margin rebuilds organically without any product mix help at all.

For example, Lumentum crossing to over 50% non-GAAP GM is a pricing power story instead of one “artificially” inflated by selling more pricey products.

At a high level, that’s a good sign of further margin expansion further down the line.

2. LTAs

Lumentum has customer LTAs extending through to 2030 with minimum demand guarantees and set pricing. AAOI is similarly negotiating three year laser LTAs with 2-3 large customers.

TLDR: These LTAs are critical for the longer term durability of the optics sector.

Ultimately, customer-funded capacity is negative working capital expansion, which highlights that the optics buyers are the ones bidding up for scarcity.

The Lumentum and $AXT(AXTI.US) deal from a couple of weeks ago is a pretty clean example where $87.5M of Lumentum deposits applied as shipment credits, reserving substrates through to 2031.

We're now seeing this pattern is everywhere e.g. hyperscalers pre-funding supplier capex in exchange for allocation, transferring volume risk onto the buyer. This is the same dynamic at a higher level with $NVIDIA(NVDA.US)’s $500B financing announcement to support “independent compute financing platforms”.

So, the whole AI buildout is shifting capex risk onto customers and capital markets. In optics that means suppliers get to expand capacity with someone else's balance sheet and a contractual demand floor.

That’s the best terms any capital intensive company can get and, as we're seeing, acts as a huge incentive to ramp capacity as fast as possible.

3. 1.6T ramp

Lumentum: 1.6T is “ramping incredibly fast. In fact, we have seen the 1.6T ramp only pull in, be stronger, demand increase. That ramp is even faster than what we thought, say, three months ago.”

With ~80% YoY laser output feeding it, Lumentum guided data center growth “to exceed 80%” this quarter. They initiated 1.6T production as planned and the CEO stated that “we appear to be first to market in many instances, ahead of larger competitors” with a signal-integrity team “widely acknowledged as the best in a competitive field.”

AAOI is the fourth supplier qualified for 1.6T at a major hyperscaler, qualification finishing “within the next couple of weeks” and a >$200M order book which looks to be just the start with the $471M monthly transceiver revenue by mid-2027 maintained from last quarter.

At 800G the laser was the binding constraint. At 1.6T it's the 3nm DSPs and the 200G TIAs and drivers. AAOI's management specifically said that the bigger constraint now is DSP and TIA supply, not lasers, because they make their own lasers (“good news, they are making our laser. Otherwise, laser is the biggest bottleneck”).

Does this then make $MRV and $Broadcom(AVGO.US), plus the TIA/driver suppliers $Semtech(SMTC.US) and $Macom Tech(MTSI.US), the kingmakers of the 1.6T ramp??

I think that'd make sense because allocation at the DSP and analog front-end seems to be the new share battle given how constrained the laser makers are. And it then decides which module houses ramp on schedule and grow earnings fastest.

So in my opinion, if you want to know who ships 1.6T in volume in 2027, DSP and TIA allocation is the real signal/alpha.

Looking at AAOI specifically now:

- Q2 800G revenue was $12.8M.

- The guide is ~5x sequential in Q3, then more than $70M of 1.6T in Q4.

- Building to roughly $471M/month month by mid-2027 - of which management pegs $217M monthly from 800G and $164M from 1.6T.

- That’s a roughly 12 month path from a ~$36M monthly DC run rate to a $471M one.

- Requiring capacity to climb from ~200,000 units/month now to >650,000 by end of 2026 to >930,000 by end of 2027, every qualification to land on time, and DSP/TIA allocation to be there at each step.

Essentially, that’s several consecutive flawless quarters of capacity, qualification and third-party allocation with zero slip ups. Personally, I’d be modelling the mid-2027 exit meaningfully below $471 million/month and treat the target as the ceiling rather than the base case.

Whether you want to trust AAOI’s management is up to you, but so far so good.

And just touching on the CW/EML mix:

This is basically a margin event for the laser owners whichever way the pendulum swings.

Lumentum is closing the CW vs. EML margin gap by shrinking CW die, 200G EMLs are already over 25% of EML revenue heading to a majority by mid-2027, with over 50% EML unit growth targeted by the December 2026 quarter. The CEO expects SiPh viable at 1.6T but EMLs “come back in a meaningful way” at 3.2T.

Ultimately, in a constrained environment, wherever a laser can be sourced will be used to support the buildout.

While everything is sold out, customers take whichever laser they can get, so EML vs. SiPho is a margin-mix question for the supplier, not a demand question.

4. NPO = TAM expansion

Lumentum’s CEO framed near-packaged optics as pure additive TAM: “The NPO opportunity is completely additive for us, significantly increasing the optical TAM” and noted even “our largest CPO customer is looking at NPO for specific new use cases, further increasing the optical TAM at that account.”

I actually think “additive” TAM holds quite well here with NPO:

- NPO and CPO first displace copper scale-up links inside and across racks, not front-panel pluggables on scale-out.

- That means pluggable/scale-out demand and integrated-optics/scale-up demand stack rather than substitute, at least through 2028.

- The TAM claim stops being true only if scale-out itself ever moves onto the substrate which doesn’t seem to be the case for the next few years at least.

In my opinion, this is a rare case where the bullish management framing survives various mechanism checks.

This is why the content-per-XPU maths is the entire point, and it is why the laser makers sound indifferent to the specific form factor of CPO or NPO etc. Moving from a shared front-panel module to per-package light sources multiplies discrete laser count and optical power per accelerator.

That’s the mechanical reason Coherent can say NPO and CPO carry “comparable content” - in both, Coherent is selling the laser, the external laser module, the isolators, the polarization-maintaining fiber it makes in-house, and the fiber-attach.

On standards: optically these are converging on OCI/MSA-based interfaces at 200G/lane, but the form factors are proprietary per customer (Wupen Yuen). That cuts in the incumbents' favour, because it raises qualification lock-in per socket even as the optical layer standardizes.

Ultimately, the CW laser supply names and VCSEL optionality matter here. Coherent is pushing a 200G VCSEL for NPO-type applications, which widens the playing field. In terms of who loses - retimed pluggable-only franchises see their share of the optical $ shrink in relative terms as integrated optics adds a new dollar pool on top.

After looking into it a bit more over the past few months, I do believe that NPO wins the first wave because it trades power and cost for simplicity around time to market.

That, in the end, doesn’t matter to whoever owns the lasers because they’re selling comparable content into whichever one customers pick.

—

Just a few of my personal notes which I filtered down into some of the main points.

There’s other things to be aware of too like the ongoing InP chokepoint, memory bottleneck, scale-across at Lumentum, and OCS type stuff.

But don’t want this post to become longer than it already is.

$Micron Tech(MU.US) comments from KeyBanc conference:

1. Micron are supplying less than half of data center demand.

2. Micron's customers are insensitive to memory prices.

3. Micron highlight multi-year demand with some customers signing LTAs to 2030+.

4. Micron says it will be the only company producing memory in the U.S.

This commentary highlights that there is zero cyclicality. No one's signing LTAs years out under inelastic pricing conditions if you think supply catches up next year.

Useful to remember also that the end of prior memory cycles came because supply > demand, where capacity was built ahead of committed demand.

However, in my opinion, the market has re-rated memory back down towards cyclical valuations.

Regardless of continued price increases and GM expansion, surely Micron etc. deserve to be valued higher purely based on the supply/demand gap that's there to be filled over the coming years?

Notes on $Lumentum(LITE.US) ahead of earnings tomorrow.

On the Q3 call (May), the CEO put the demand/supply imbalance at “somewhere greater than 30%” and said the company is “having to make choices across the company” to service demand, noting that significant customers are looking to up their orders that “we simply can't service”.

Then on pricing, the CFO said “prices are holding or they're increasing for what customers want”.

When you combine supply constraints with relative pricing power, you get a pretty powerful company...

Ultimately, Lumentum can't be competed away and forced out of the market mainly because of their 200G EML volume production. Mitsubishi, Sumitomo, Furukawa and $Coherent Corp.(COHR.US) can all do 100G, but no one can do 200G like Lumentum, and at their volumes since they're the only supplier shipping 200G EMLs in volume today.

On capacity, Lumentum's EML output is up ~8x since FY23 and management guided to roughly 50% more InP output and more than 50% EML unit growth from the December quarter last year to the December quarter this year. All while the demand/supply gap has extended. Their fabs are clearly straining, with the Japan wafer fab already “fully allocated”.

And the new Greensboro InP fab won't come online until 2028 mainly because of reactor installs and cleanroom work. Even then, any incremental output would still need to pass customer qualifications before hitting the topline, so it's still some time before seeing any sort of material benefits.

Honestly, my first instinct on any “sold out for years” claim is skepticism, however with Lumentum, their output went up ~8x and the demand gap still got wider.

That's pretty insane, and is why the stock trades at relatively lofty valuations today. YTD, a ton of pricing-in has already happened, but I do see continual growth vectors incoming:

Looking at industry numbers, 800G+ goes from ~20% of transceiver shipments in 2024 to over 60% in 2026, and EML/CW laser capacity roughly doubles this year to ~50M units a month (TrendForce puts the 2026 figure at 50.7M). Yet lasers are still a bottleneck, so if supply doubles and the shortage doesn't close, the demand curve only steepens.

- For context, TrendForce has the top three EML makers (Lumentum, $Broadcom(AVGO.US), Mitsubishi) holding ~72% of the market between them, and only Lumentum ships 200G in volume.

With CPO, $Lumentum(LITE.US) is the named laser supplier on $NVIDIA(NVDA.US) Spectrum-X/Quantum-X photonics switches (Nvidia has signaled it wants more than one laser source over time, but Lumentum is the one in the design today). The “CPO kills the optics suppliers” bear case has always struck me as an argument made by people who haven't looked at where the light actually comes from: the transition moves the laser into an external module and makes it more important…not less.

Personally, I estimate hundreds of CW lasers per 512-port switch. Lumentum already has a multi hundred million $ CPO/ELS PO on the books which gets delivered in H1'27. You can link this to $SIVE because their InP lasers going onto $GlobalFoundries(GFS.US) SiPh platform proves that SiPh still needs an InP light source, regardless of whichever packaging approach dominates at the end of the day.

There's also a vertical integration story to Lumentum that mirrors $Applied Optoelectronics(AAOI.US)'s but it runs in the opposite direction. $Applied Optoelectronics(AAOI.US) is a module maker that built its own laser fab; $Lumentum(LITE.US) is a laser maker that bought a module business (Cloud Light) and is closing the loop since its own CW lasers are set to go into ~20% of its own transceivers this quarter, and 200G EML revenue more than doubled sequentially last quarter.

So in a laser-short world, the module makers who own their laser supply actually ship, and the ones who don't just sit on their hands. Both $Lumentum(LITE.US) and $Applied Optoelectronics(AAOI.US) worked out the same thing from opposite ends, and the fact that a module maker and a laser maker independently converged on full stack integration should tell you where the bottleneck is…

Just touching on the proposed FCC China transceiver ban again. This would actually cut both ways for $Lumentum(LITE.US). Cloud Light builds in Thailand and benefits directly from that shift. On the flip side, Chinese vendors make roughly 60% of datacom optical component revenue, with InnoLight and Eoptolink taking around 60% of the high-end 800G+ orders — and $Lumentum(LITE.US) sells EMLs to them.

The problem is that $Lumentum(LITE.US) reports revenue by shipping destination, not end customer: China plus Hong Kong was ~25.5% of last quarter's revenue.

That figure overstates true China exposure, because much of it is chips delivered to contract manufacturers building for US hyperscalers, but it also hides how much of the EML business depends on Chinese module makers specifically.

Honestly, nobody outside the company knows the real split. But if Chinese customers' orders fall before Western module makers scale up to replace them, Components revenue could dip for a quarter or two, and it's quite hard to forecast accurately ahead of any earnings releases. (Assuming the ban actually goes ahead).

Even so, I think the ban would be a net positive because it just changes who assembles modules rather than impacting the demand side. Chinese module makers switching to domestic EML suppliers would still face a multi year qualification cycle at 200G that they can't shortcut. Upstream, China's export controls on InP substrates have pushed 6-inch wafer prices up ~250% (roughly $1,400 to $5,000 since February 2025), all while $Lumentum(LITE.US) locked their own substrate supply a couple weeks ago with a capacity reservation at $AXT(AXTI.US) running through to 2031, ahead of anything peers have signed.

Obviously, that supply still sits inside China's export permit regime (AXT ships from China and says it can't predict permit timing) but in my view, Lumentum being first in line is worth more than the risk it carries.

For a TLDR on key watchpoints from the Lumentum earnings call:

- 200G EML trajectory within datacom lasers (it more than doubled sequentially last quarter) and the overall EML unit ramp.

- ELS volumes and any second 1.6T Cloud Light customer.

- Greensboro InP timeline and whether first output slips or gets pulled forward.

After the July selloff, it does seem like a lot of froth has come out ahead of this ER compared to a couple of weeks ago.

Disclosure — I have a position in $Lumentum(LITE.US).

Recommended reading for the weekend:

1. Aswath Damodaran (@AswathDamodaran) — Information Timing and Release: The Gaming of Guidance!

2. Dwarkesh Patel (@dwarkesh_sp) — Why compute might get 10x+ more expensive in coming years

3. J.P. Morgan — Semiquincententacles

4. @citrini — Protection Matters: Cybersecurity’s Winners and Losers

5. Neil Shah (@neil_shah) — Innolight, Coherent, Lumentum: Who Wins and Loses in the Proposed FCC Ban on Chinese Transceivers?

Just some key points from $Applied Optoelectronics(AAOI.US) earnings call today.

TLDR: Q4 is very important as management have informally guided ~$330M of 800G/1.6T revenue.

1. Demand is 20-40% above what $Applied Optoelectronics(AAOI.US) can supply. Customers keep pressing them to accelerate delivery.

- Like I said earlier, $Applied Optoelectronics(AAOI.US) is now a manufacturing story, not a demand story.

- I.e. can they produce enough to fulfil demand?

2. Q3 guided to $255-$290M which is +130% YoY and +42% sequentially at the midpoint.

- Management stated ~$330M of combined 800G & 1.6T revenue in Q4 alone.

- Implies Q4 revenue to be ~$480-500M, a potential 2x sequentially.

- I can't remember a Western optics company ever reporting growth at that cadence. Which makes Q4 the most important quarter, above Q2 and Q3.

- $Applied Optoelectronics(AAOI.US) stuck to their estimated $471M monthly transceiver revenue in 2027, which they mentioned last quarter.

I was expecting 60-80% sequential revenue growth, but $Applied Optoelectronics(AAOI.US) were hit by 100G revenue declining $20-25M because a customer couldn't source enough 100G switches due to memory shortages. If you add that "lost" revenue back, you get to a more reasonable ~50-55% growth.

3. 800G revenue

- $12.8M in Q2 which is >2x since Q1.

- Guided to grow ~5x in Q3 to ~$62M-$65M

- I mentioned earlier that "800G is a forward event that has to show up starting in Q2 and really in Q3" — great to see it happen on schedule.

4. 1.6T revenue

- $Applied Optoelectronics(AAOI.US) will be the fourth qualified supplier at one major hyperscaler.

- Full qualification expected within weeks w/ first shipments late Q3.

- >$200M order in hand with the bulk delivering in Q4.

- Q1 2027 1.6T revenue to double Q4 revenue.

- Customer demand indications of >500k units/month = $300-350M/month by end of 2027.

5. Capacity for 800G and 1.6T

- End of Q1 = 100,000 units

- Current = 200,000 units

- End of 2026 = 650,000 units

- End of 2027 = 930,000 units

So a ramp from Q1 -> Q2 as promised back in Q1.

New Texas facility for 800G/1.6T production starts up in late Q3, which frees their HQ building for InP wafer capacity expansion. Shows that modules and lasers are expanding in a clear/deliberate chain.

6. Capex

- Q2 capex was $565.5M incl. $280M of equipment prepayments.

- H2 capex intensity guided higher.

- No C Suite / Board prepays that much for equipment against weak demand signals.

7. CPO / ELSFP

- External laser source modules ramp late 2026 into 2027, targeting ~400k pieces/month by 2028.

- Total laser fab capacity +~300% by Q3 2027 plus a second Houston fab planned at ~4x the current facility's size.

- "Several major CPO customers love our laser, we just can't make enough"

Ofc, 400k/month in 2028 is two fabs and two years away. But it materially extends $Applied Optoelectronics(AAOI.US)'s moat duration as merchant EML scarcity eases in 2027-28 and deflates the original captive laser edge.

8. Chinese transceiver ban reports

- U.S. manufacturing is "the most important element of our appeal"

- Early customer feedback is that they'll "give us much more share" for U.S.-made supply.

- $Applied Optoelectronics(AAOI.US) expects to remain the largest U.S. AI transceiver manufacturer.

- With capacity booked through to Q2 2027, the response is clearly to add capacity more aggressively from ~Q3 2027.

9. InP substrates

- supply secured into end of 2027 across ~5 qualified suppliers (Europe, Japan, China).

- partnership/JV discussions with substrate suppliers for 2028–29 volumes.

- signals intent to integrate further downward as laser volumes 10x for CPO

Just initial thoughts on the things that stuck out to me most on the earnings call. I'm sure there's more stuff I'm missing rn.

I do believe though that the whole story now compresses into Q4's ~$330M of 800G/1.6T delivery.

For $SIVE, Q2 earnings from $GlobalFoundries(GFS.US) yesterday is strong third-party confirmation of Sivers' 2027-2028 photonics ramp.

Reminder: In June, Sivers signed a collaboration with Global Foundries that puts their lasers into GF's SiPh reference designs and CPO platform.

Read-throughs from $GlobalFoundries(GFS.US) earnings transcript:

1. $GlobalFoundries(GFS.US) silicon photonics revenue to "more than double in 2026". NPO ramp in 2027 / CPO ramp in 2028.

- Important external validation of $SIVE's entire photonics thesis.

2. $GlobalFoundries(GFS.US) 2027 NPO / 2028 CPO cadence lines up with Sivers' own guidance.

- "Production ready status in 2027".

- Multiple ramps in 2027.

- Sivers estimated CPO SAM of $1.0-1.4B by 2028.

- So GF is effectively narrating the demand curve into which Sivers has positioned their Glasgow ops + Win Semi capacity.

3. $GlobalFoundries(GFS.US) CID grew 20% sequentially and 62% YoY which is the fastest quarterly growth since 2022. Growth guidance raised to 50-60% YoY.

- Direct demand validation for the platform Sivers plugged into, from their collaboration in June. Growing demand.

- GF's acceleration is the leading indicator Sivers earnings in 2027+ with funnel filling upstream before Sivers seeing the units.

4. $GlobalFoundries(GFS.US) CEO: "We now expect our silicon photonics revenue to more than double in 2026 compared with the year prior"

- Clean revenue funnel logic if GF SiPho more than doubles and Sivers is the partnered external laser route into GF reference designs.

- Sivers has guided its own laser qualification to "production ready status in 2027" so Sivers captures this as design-in/NRE in 2026 and unit revenue in 2027+, not coincident with GF's 2026 doubling.

5. $GlobalFoundries(GFS.US) secured 7 new optical networking design wins with customers from transceiver suppliers and hyperscaler players.

- All are a potential downstream pull for external laser content.

- Sivers has been building this coverage independently: $Jabil(JBL.US) 1.6T LRO, $POET Tech(POET.US) ELS, O-Net/Enablence ELS, and legacy Ayar Labs.

- Looks like Sivers' Q1'26 disclosure of an opportunity pipeline up 77% YTD to $799M is the mirror image datapoint to GF's design win momentum.

6. $GlobalFoundries(GFS.US) have "more than 40 customers today in SiPho" with customers telling GF "I don't have enough secured."

- This just corroborates Sivers' narrative of an InP laser chokepoint and supports their pre-emptive capacity build via WIN Semi.

- If GF customers are scrambling, the matching external laser capacity is just as scarce = pricing power for Sivers.

7. $GlobalFoundries(GFS.US) could "10x photonics capacity"

- GF ramping PIC output 10x at existing fabs implies a proportional pull on external light sources.

- For Sivers, they need to scale InP output in lockstep or lose allocation to $Lumentum(LITE.US) / $Coherent Corp.(COHR.US).

- Sivers' answer is the WIN outsourcing + Glasgow expansion (historically framed as targeting >1,000 wafers/week).

- Interested to see if Sivers report capex/capacity commitments scaled to GF's 10x narrative.

Just as a list of the main read-throughs you can draw from Global Foundries' comments to Sivers.