Paradi Lab

$Nebius(NBIS.US) will become a "hyperscaler" very soon:

Via two axes:

1. Cost of capital

2. Cost per token

Everything else about a hyperscaler is downstream of those two variables.

Looking at cost of capital alone:

- Neoclouds are basically a spread business where they raise capital at one rate + earn a contracted return above it.

- Most neoclouds fund expansion via equity (v. expensive) and enter into a loop of dilute -> deploy -> depreciate -> repeat.

- Nebius's new $775M debt facility from last week breaks that loop.

- Their debt facility is collateralised against contracted cash flows from an "investment-grade" customer, so the lenders are essentially pricing off $Meta Platforms(META.US) / $Microsoft(MSFT.US) receivables rather than Nebius's own credit.

- That's v. bullish for Nebius if lenders price them in that way. Especially since early GPU collateralised neocloud debt was at double-digit rates.

- The structure repeats. Management says it will replicate the facility against >$40 billion of contracted backlog.

- Every new "investment-grade" contract therefore becomes a collateral factory and not just revenue. It manufactures cheap borrowing capacity -> which funds capacity -> which wins contracts.

- Customers also pre-fund the machine: deferred rev rose $3.2B in Q1, driving $2.3B of op. cash flow on $399M of revenue.

- Ofc that's a delivery obligation and not free money. But it's an obligation funded interest-free vs funding the same buildout w/ debt.

Despite what Burry will say, this isn't financing circularity:

- $NVIDIA(NVDA.US) $2B equity stake is a rounding error against a broader $20B+ capex programme.

- And the backlog is cash contracts paid out of $Meta Platforms(META.US) / $Microsoft(MSFT.US) opco's and not roundtripped semis money.

This "new" loop is essentially how AWS became AWS where they funded at bond rates while everyone else funded at much higher equity rates for like 15 years.

The deadliest mountain in the world is K2 in Pakistan.

But how deadly will the new K3 model by Moonshot be?

In my view, this is not another "DeepSeek" moment. Kimi K3 is net-positive for the AI trade and compute demand, while being a net-negative for Western fronteir labs.

After reading many reviews on X, my understanding is that Moonshot are just a few months behind the fronteir, rather than a full generation behind.

I will caveat that I am far from being an AI "expert". However, I believe I understand the commercials at a high enough level to comment on and summarize.

So from a commercial point of view, K3 comes out at roughly half the cost of Opus 4.8 for a slightly better model overall. Moonshot is charging ~$3 per million input tokens and ~$15 per million output tokens, the same rate card as Anthropic's mid-tier Sonnet and nearly 4x its own previous model.

In my opinion, these pricing dynamics are what makes the DeepSeek comparisons I've seen on X to be very lazy.

DeepSeek in Jan 2025 was a deflation event where they had near frontier capability built very cheaply and effectively given away at such low prices which implied that Western AI capex was excessive.

Moonshot is doing the opposite. It's charging Western prices because it believes it is selling capability, not cheapness, most likely because Chinese labs are starving for compute and actually need the revenue to continue building and serving demand.

The business model for fronteir labs like OpenAI and Anthropic is essentially the monetisation gap between their own leading models and everyone elses. Both of those companies are raising money at valuations that assume that gap stays wide and highly monetisable for years to come (at growing rates).

However, many AI experts on X say that the Kimi K3 gap can now be measured in months, and that the tier just below the frontier, where the bulk of real commercial workloads actually run, is potentially about to be served up at very low commodity prices.

As a result, the labs' defensible ground shrinks to the true frontier, to enterprise trust and security, and to distribution. Revenue can keep growing while pricing power erodes, and in my view, it is the pricing assumption (not the demand assumption) that carries Anthropic's and OpenAI's lofty valuations.

Using SoftBank as a proxy for OpenAI, we saw today that their stock price fell a huge 9% after Kimi K3 was released. I believe, now, that this valuation-digest was and is warranted.

I know many of my followers will care about the AI trade in upstream names though. However, I do not view the latest Kimi K3 model as having a fundamentally negative impact on those companies.

Every K3 token will still consume chips, memory and power that remain effectively sold out into next year. Moonshot will certainly need to secure additional compute to facilitate growing demand for their models.

As of right now, it does seem like exposure to model-layer margins, lab valuations and their listed proxies looks worse.

Conversely, exposure to the infra in upstream supply chain names that serves tokens regardless of whose model wins looks cheaper, all things considered, for no fundamental reason.

However....to be slightly more contrarian - you could argue that if cheaper Chinese models come to market at a rate of knots and leads to margin compression for the likes of OpenAI and Anthropic ---> will that not have a negative effect on their spend in the long-term? As a result, a negative effect on the upstream AI supply chain?

I am yet to think through this in much detail, but it certainly deserves exploring.

I want to finally mention some of the technical experts I have enjoyed reading since Kimi K3 came out, as a non-technical person:

- @zephyr_z9

- @scaling01

- @FundaAI

If you want to be someone who is constantly learning about AI, these guys are must-follows in my opinion.

Of course, as always, I would encourage anyone to correct me or debate me if I am wrong in any area.

Oof, $SpaceX(SPCX.US) Starship launch aborted just now.

Per Elon: "Next launch attempt hopefully in a few days."

Just highlights how risky the space business is.

Goes hand in hand with stock price volatility across the entire sector.

Appetite for the memory trade is still super high:

Record inflows for the South Korea ETF $EWY this week.

Tuesday: $814 million (previous record high)

Wednesday: $1.1 billion (new record high)

Investors are back to using $EWY as a proxy for the South Korea SK Hynix listing, just like in Q1.

Rather than paying the ~25% premiums for the US listed SK Hynix ADR.

And for added context on the demand for memory stocks:

The memory ETF $DRAM is currently at record-high AUM levels at ~$25 billion, despite only being released three months ago.

AI and semiconductor names are down significantly over the past couple of weeks.

However, what matters right now is for me to provide some useful context and "advice" on constructing your own investment philosophy.

There's no better time to learn about yourself than when times are rough.

Personally, my portfolio has taken a huge hit for a few reasons. Primarily because I pivoted around this time last year from a slightly more diversified base of stable technology compounders into a slightly higher concentration of higher beta and smaller market cap AI names.

However, if you're an inexperienced investor, what I can tell you is that market drawdowns are normal and expected (even though we can never time them exactly).

So if you plan to invest for decades, you should certainly expect to sit through many bear markets and many corrections.

The first thing to internalise is the arithmetic of losses where a 20% loss needs a 25% gain to get back to even. Similarly, a 50% loss needs a 100% gain to recover your losses.

This is why position sizing matters just as much as stock picking, and is a key reason why I'm against risky tools such as options or leverage for *most* investors.

Your primary goal when playing this game is to keep playing. A blown-up account via excessive leverage or backfired options trading is the worst case scenario, no matter how attractive or easy "gurus" online might make it seem.

That said, I do firmly believe in running a concentrated book of high conviction names that you can reliably track daily.

The specific number of positions varies person to person depending on a few factors like your individual ability to keep on top of latest events for each holding, including second and third order effects from other company's news.

However, every legendary investor's track record was built on a handful of high conviction positions. Sure, spreading capital across many names you barely understand can in theory reduce risk just because of the law of large numbers.

But in my opinion (and experience), it just guarantees mediocrity.

That said, concentration is a discipline of its own and is something I could write books on.

With the aim of keeping thing consise, concentration only works if you can name, very precisely, what you're concentrated in.

In simple terms: have you researched the company, sector and market to the point where you have utmost confidence in the trade?

But with concentration, you *need* diamond hands if all things are equal with the company and nothing has changed.

For example, $NVIDIA(NVDA.US) fell more than 50% in 2018 and ~65% in 2022. Anyone who capitulated in Oct 2022 sold one of the greatest companies of all time, probably because they lacked conviction in the trade.

And in turn, did not have any sort of investment philosophy.

So, you need to really sit down and talk to yourself right now, during this current AI drawdown, and ask yourself what your investment philosophy is.

You could probably run the exercise with Claude or ChatGPT and get some pretty enlightening outcomes for yourself.

But you must be honest for the good of your future self.

There's no better time to learn what kind of investor you are than during red days and weeks.

$IQE has unfortunately not been performing well recently.

However, we've only had net-positive news since May:

- $Tower Semicon(TSEM.US) multi year InP epiwafer agreement in June w/ minimum purchase commitments from year two

- $14M multi year PO from a "strategic global technology leader" for AI/datacentre applications this week

- $Macom Tech(MTSI.US) board involvement after anchoring a £81M raise to strengthen balance sheet

That said, it does seem like IQE is simply resetting (similar to other high beta names like $SIVE) after going up parabolically in a few months.

And Chinese export curbs on gallium + InP feedstock are inflating input costs with management also saying that they're sharing that pain with customers.

I'm also looking at the fact that their balance sheet almost ruined them a few years ago - which is now solved thanks to £45M from MACOM as a strategic investor tied to LTAs.

IQE is one of the only Western epi houses that can supply InP epiwafers at a scale that actually matters.

And InP sits under every optical interconnect going into datacentres.

Tower didn't sign a multi year supply deal and settle a long standing patent dispute in the process because it had alternative routes to source supply...

So currently, you've got:

1. A repaired balance sheet.

2. An anchored strategic partner w/ Macom.

3. Contracted InP volumes.

4. An option on photonics too.

However, risks are definitely real and please do not act on my words alone...

They've had execution issues in the past e.g. too much internal resource in their wireless segment which has been dying a slow death over the past couple years. (They're correcting that now just by going from their investor decks).

Plus of course a long dilution history. And China risks impacting input costs.

Drawdowns like the one we're in currently are all part of the game.

Where highly volatile businesses = highly volatile returns.

HUGE $Nebius(NBIS.US) news today, showcasing why they're the premier neocloud:

-> Nebius announces a new business model that lets infrastructure partners deploy Nebius’s AI cloud platform in their own AI data centers.

This is so f*cking good.

Simply:

- Nebius will let partners (e.g. DC developers, infra funds, sovereign AI projects) build + pay for data centres.

- Then pair that partner capacity with Nebius's systems, software stackm and customer book.

- Meaning that Nebius can expand their capacity "pool" even faster.

- With no capex or financing risks like debt/dilution.

Also, for Nebius: they get a very high margin rev stream w/ minimal financing requirements.

Since the partners will fund all the expensive stuff like the actual building / power / GPUs.

And then Nebius supplies genuinely scarce stuff like systems architecture / $NVIDIA(NVDA.US) supply chain access / software stack. And ofc, the GTM strategy w/ access to Nebius's customers.

Meaning that Nebius would sell a partner's DC capacity via their own in-house Nebius sales org. With identical service levels to their owned sites.

Then, Nebius takes a revenue cut / licensing fees / commissions / committed capacity deals.

Unbelievable from Nebius lol.

Nebius sells out capacity every quarter and sits on ~$46B of contracted backlog (mainly $Microsoft(MSFT.US) and $Meta Platforms(META.US)).

So their main constraint is nothing to do with customers or tech.

Rather, their limiting factor has been capital + energized power where every GW costs upwards of billions of dollars.

But...this announcement removes the capital constraints to crazy high expansion.

Wow, unreal $Aehr Test(AEHR.US) Q4 earnings and FY2027 outlook, +26% after-hours.

I have high conviction that Aehr is well positioned for many years of very strong growth:

- Multiple customers entering/expanding production

- Huge backlog

- Opportunities in new markets e.g. NAND and HBM applications for their WLBI system roadmap.

-> Q4 2026 earnings:

Revenue: $18.8M vs. Est. $18.7M

EPS: $0.11 vs. Est. ($0.01)

Bookings: $61M

Backlog: $101M

-> FY27 Guidance:

Revenue: up to $150M vs. Est. $85M (+200% YoY)

- WLBI demand from AI applications is expected to accelerate.

- AEHR also engaged with additional AI processor suppliers that are evaluating WLBI to improve product reliability and reduce yield loss from production burn-in.

With qualification testing for their WLBI exceeding all of their major customers' expecations.

With their major customers ultimately saying that they want to buy even more systems from Aehr.

"The potential revenue opportunity from any one of these devices is significant to Aehr in terms of near and long-term revenue streams related to the WLBI systems."

Then on their package-level burn-in systems:

- Record follow-on orders recently from Aehr's lead hyperscale customer.

- Customer wants to buy more PLBI systems.

- Aehr also engaged w/ multiple new customers for PLBI qualification incl. AI customers and robotics.

So overall, very significant TAM expansion if they can capture new customers across AI and robotics. And since they pass qualification cycles with ease, I have high confidence that they'll capture these newer markets also, driven by years and years of in-house process knowledge + patents.

I've mentioned a few times, but Aehr's lead SiPho customer is also ramping w/ follow-on orders over the last year. And more orders coming soon from them for additional WLBI systems.

"Our newest major silicon photonics customer...has provided us a forecast for additional systems this calendar year as it ramps capacity to support next-generation hyperscale data center deployments."

All basically meaning that SiPho + optical test and burn-in market has huge potential to grow significantly and be a major long-term growth driver for Aehr.

And very interesting to see the CEO mention the memory market roadmap for their WLBI systems also!

"We continue to work with multiple memory suppliers to align our solutions to meet the production needs of these companies' new capacity coming online."

Very exciting times for Aehr, and they remain one of my higher conviction names all throughout the AI supercycle.

High demand for Aehr solutions + Great execution = Bullish

Nanya Q2 Earnings highlights:

Revenue: NT$82.5m (+68% QoQ)

EBITDA: NT$63.8m (+93% QoQ)

EBITDA Margin: 77.3% vs. 67.2% last Quarter

But of course, there's a lot of retail worry w/ the memory trade rn, mainly thanks to volatility with SK Hynix and Samsung.

From Nanya, their DRAM market outlook/commentary looks like this:

- AI driven structural change is mitigating memory market cyclicality.

- Supply tightness is expected to persist over the next several quarters.

- Multi-year LTAs aligning supply-demand expectations.

Overall, aligned w/ SK Hynix CEO comments post-US IPO, confirming lack of supply for the forseeable future across the market.

With Nanya, you do get some drag from legacy rev lines though like from smartphones/PCs/consumer elecs. Where AI is only ~20% of revenue rn which will be higher in the future.

But they've confirmed that the structural shift towards AI is "intensifying memory shortages", shown by their shipments being flat for the quarter but ASPs increasing more than 60%.

Which looks set to continue for the coming quarters with continued lack of supply since their new fab only starts ramping in 2028.

This is a shocking statistic:

SK Hynix & Samsung shares/leveraged ETFs make up over 70% of trading volume in South Korea.

I think the memory concentration is rational.

But over 70% of daily volumes with ~15-20% in leveraged ETFs just reeks of a gambling mindset.

I can also guarantee that most of those "investors" don't know what SK Hynix do, and probably think that Samsung are going up cos of phones/TVs.

Then there was that WSJ article a few weeks ago that interviewed people who full-ported their life savings into 2x leveraged SK Hynix.

All paints a pretty stark picture around the state of Korean investment culture imo.

AI is not a bubble.

Lazy AI bears point to $NVIDIA(NVDA.US)'s market cap and through PTSD, claim it resembles $Cisco(CSCO.US) in March 2000.

However, any useful bearish analysis should look at what actually made the 1990s market a bubble from a macro sense, and whether those conditions exist today.

I now refer you to the attached chart.

In the late 1990s the two lines veered apart. Tech investment went vertical toward ~4.5% of GDP, while the economy-wide profit share rolled over from its 1997 high and fell hard into 2000.

Investment surged while profitability eroded...bubble!

Today, the lines rise in tandem.

Tech investment has pushed to roughly 4.9% of GDP, above the dotcom peak and climbing more steeply, while pre-tax corporate profits sit near 14% of GDP.

Meanwhile leverage has (mostly) stayed contained and the US current account deficit is shrinking. However, bears point to record levels of investment in isolation, choosing to ignore the growing profitability in addition.

Are they dumb, or are they ignorant?

Probably both.

In the run up to March 2000, share prices rose and multiples exploded. The market paid more and more for each dollar of invisible earnings.

This time, forward P/Es have barely moved even as share prices rocketed, because earnings expectations rose alongside them. For example, the Nasdaq 100 trades around 23x forward earnings, near its own 10Y average versus ~60x in March 2000.

But...but....the market is so concentrated!! 1! 1!

Yes, the ten largest S&P 500 companies account for ~40% of the index, above the dotcom peak.

But those ten companies contribute around 30% of total market earnings, compared to under 20% in 2000, and trade at roughly a 50% premium to the rest of the market against a premium north of 100% at the prior peak.

Now, bears will say: "If the rally is earnings driven, everything depends on whether the earnings persist!"

Correct.

But unfortunately for the bears, this is where things get uncomfortable.

AI type names have added on the order of $27 trillion in market value since late 2022, up from roughly $19 trillion just seven months earlier. Set that against any weak attempt to discount the additional profit streams AI can plausibly generate for US companies (estimates cluster in the trillions) and the market has capitalised a multiple of the realistic prize.

Not all of that $27 trillion is AI (the hyperscalers run enormous non-AI businesses), and more aggressive assumptions on adoption and productivity can lift the number.

But closing the gap requires increasingly heroic assumptions:

- that recent shifts in earnings shares are highly persistent

- that the boom's suppliers capture an outsized slice of AI's total economic gains

- that the economy-wide profit share keeps climbing indefinitely

Alright, cool. But what about all the circular financing?!

- Nvidia has committed tens of billions to OpenAI while remaining its primary chip supplier

- OpenAI has signed a cloud commitment with $Oracle(ORCL.US) reported around $300B

- Oracle in turn buys from Nvidia

- $Microsoft(MSFT.US) is simultaneously OpenAI's largest investor and one of its largest vendors

True, this somewhat resembles the dotcom vendor financing where Cisco booked loans to cash stricken carriers as revenue (roughly a tenth of sales at the peak), much of it later written off.

However, today's arrangements are mostly equity stakes in counterparties with genuinely fast growing revenue rather than disguised loans to fund purchases, and Nvidia has lately been unwinding parts of its ecosystem book.

And according to analyst reports, even the AI labs like Anthropic have now turned profitable. A feat many thought would be impossible only a year ago.

Personally, I treat this circularity as risk rather than a point to build a bear case around since it's all ultimately leading to greater earnings across the board. Even for fronteir labs. Shock!

This then leaves the one key question:

Will barriers to entry protect today's profits from erosion?

This entirely depends on each company's position in the AI supply chain.

- At the model layer, barriers are relatively fragile where frontier models will almost certainly converge longer-term, and open source alternatives have the ability to reset price floors. However, AI soverignty will ultimately result in the likes of OpenAI/Anthropic winning.

- At the hyperscaler layer, $Amazon(AMZN.US), $Alphabet(GOOGL.US), $Meta Platforms(META.US), and $Microsoft(MSFT.US) are set to spend (currently) $750B for 2026 AI capex where falling behind is not an option. These companies are led by people smarter than you or I - do you think they'll risk their entire business collapsing for AI? No. In fact, you can already see that AI is boosting their earnings measurably in recent earnings.

- Going further down, you've got irreplaceable companies such as $ASML(ASML.US) (EUV machines), $Taiwan Semiconductor(TSM.US) (CoWoS packaging), and HBM with $Micron Tech(MU.US), SK Hynix and Samsung who are gated by long qualification cycles and multi year LTAs where demand > supply up to the 2030's.

The risk of a 2000 style valuation bubble is massively lower than the bearish consensus believes.

The world is revolving around AI, and that'll continue for the forseeable future.

Memory shortage to last until 2030's:

According to SK Hynix's CEO today at the Nasdaq IPO press conference.

“The rate at which demand is increasing far exceeds the speed at which we can increase supply.”

Customers are signing long-term contracts because “they believe that the shortage situation will last for longer.”

Coupled with the fact that there are significant lead times and contraints involved with expanding/building fabs.

I think it's evident now that a structural change has already taken place in the memory market.

Which is confirmed by the Chairman also today:

“It’s not a cyclical business anymore. Even in a downturn, the long-term agreements help maintain volumes and memory prices.”

Yep, LTAs with minimum volume commitments and price floors.

So, for now, there won't be significant margin compression back down to historic ~50-60% levels like I've seen proposed recently.

Not to say that this will last forever, cos it won't since it'll all boil down to demand/supply gaps - which are currently expanding at a rapid rate.

The U.S. needs to figure this out asap:

A new study led by McKinsey finds that there will be a shortage of engineers and technicians in the U.S. semiconductor industry.

Which may stall expansion plans for:

- $Taiwan Semiconductor(TSM.US) ($265B in Arizona)

- $Micron Tech(MU.US) ($100B in New York)

- Samsung (Texas)

- $Intel(INTC.US) ($28B in Ohio)

All driven by only ~3% of U.S. engineering students who go on to work in the semis sector since AI labs etc poach them all with better salaries/benefits/equity.

It's absolutely critical for the U.S. to figure this out some way or another to claw back capacity from Asia.

Especially with Korea just announcing last week a near $600B investment w/ Samsung/SK Hynix to build more fabs.

Which they'll easily fill with skilled technical workers since there seems to be some cultural prestige there attached to careers in semis.

In the U.S. though, if shortages force fabs to run below planned utilization / push ramps out 1-2 years:

The US just ends up w/ expensive + underutilized capacity while TSMC Taiwan & Samsung Korea keep compounding process learning at home.

In turn, further strengthening their talent pipeline.

Current hyperscaler 2026 capex: ~$725B

They all report earnings at the end of July, where they'll update 2026 capex forecasts.

Current 2026 capex:

$Amazon(AMZN.US): ~$200B

$Microsoft(MSFT.US): ~$190B

$Alphabet(GOOGL.US): ~$190B

$Meta Platforms(META.US): ~$145B

With soaring memory costs as a key driver, surely all have to revise up capex for the rest of this year.

Just a quick summary on $Churchill Capital XI(CCXI.US) (Agility):

- Going public via $Churchill Capital XI(CCXI.US) SPAC -> $AGLT in around Q4.

- Roughly $640M private funding including backers like $Amazon(AMZN.US), $NVIDIA(NVDA.US), SoftBank, DCVC.

- Around 100 Digit humanoids deployed at $Amazon(AMZN.US), $GXO Logistics(GXO.US), Schaeffler, Toyota, $Mercadolibre(MELI.US) (just 9 facilities in total).

- Around 98% accuracy at Schaeffler (8 deployments) and GXO (3 booked deployments).

- Agility claims around 1.1 year payback period on owning a Digit humanoid.

- Over $300M Digit v5 orders (release in 2026).

- Pipeline includes 30+ customers.

- First to adopt $NVIDIA(NVDA.US) Halos for full-stack robotics safety.

- More safety means higher TAM for Agility in areas like elder care, retail, and last mile deliveries - i.e. areas where human interaction is needed.

- 10,000+ units/yr capacity at RoboFab in Oregon.

- Using ~75% US sourced parts.

- As production ramps to 10k+ annually, BOM will drop from $125k to ~$15-20k.

Imo, they're the furthest ahead out of any Western humanoid OEM when you combine tech + commercials:

- Figure has some controversial deployments at BMW, no idea if their runtime data is correct or not.

- $Tesla(TSLA.US) selling zero units externally rn.

- Apptronik have no deployments rn (according to sources) despite $GXO Logistics(GXO.US) and Mercedes rumours.

- Boston Dynamics have a genuinely great robot but v. slow to market w/ no customers yet until at least 2027.

$Churchill Capital XI(CCXI.US) (Agility) have a huge lead in humanoid safety thanks to $NVIDIA(NVDA.US):

1. Humanoid safety is a critical barrier to mass adoption.

2. $NVIDIA(NVDA.US)'s Thor and Halos for Robotics is the industry’s only full-stack robotics safety system.

3. $Churchill Capital XI(CCXI.US) (Agility) are the first humanoid maker to use Halos.

Agility has a very clear first-mover advantage over the other humanoid OEMs in bringing responsible automation to factories + warehouses.

Agility CEO: “Partnering with NVIDIA to implement and optimize the Halos for Robotics system extends our leadership in responsible automation, which is a nonnegotiable requirement for bringing humanoids safely into industrial workflows.”

For some of Agility's current customers like $Amazon(AMZN.US), $GXO Logistics(GXO.US) and Toyota.

We're just at the start of the robotics/humanoid adoption curve.

Which looks set to ramp up very soon given that safety hurdles are being overcome rapidly, as well as the actual humanoid itself improving with each iteration e.g. Agility's new Digit 5 humanoid.

Looks like both the safety element and the immature tech element are simultaneously improving at the same time and rate.

Which is probably why we'll start to see wider mass adoption in the coming years once the likes of Amazon/Toyota can prove the payback far outstrips that of an actual human worker.

Need help building some conviction in your AI names during this market downturn?

Read this post. Bookmark it. Come back to it.

Let Paradis guide you:

In my view, AI names are broadly down for a few reasons:

1. The market viewed $Broadcom(AVGO.US) earnings + guidance as decelerating growth.

2. Hawkish turn from the Fed.

3. A cluster of AI bubble commentary at the end of June.

Please note what's not on that list:

1. Hyperscaler capex cuts.

2. Cancelled orders.

3. HBM oversupply.

4. Collapsing GPU rental pricing, etc etc.

Broadly speaking, absolutely nothing negative has happened on the demand side.

If you strip the whole AI trade down to first principles, it mainly boils into two layers that the market always confuses:

1. Financial stuff e.g. equity valuations, financing loops, neocloud debt stacks, where scepticism is warranted and where some pullbacks are healthy.

2. Physical stuff e.g. wafers, HBM, CoWoS capacity, optics, where demand > supply essentially everywhere.

Nothing's changed on the physical layer. So in my opinion, this sell off is just a reset before hyperscalers guide capex in their upcoming earnings.

But, for the case of this post - let me just stress test the bear case for the AI trade:

Basically everyone's desperate for compute rn.

Total hyperscaler capex guidance for 2026 is near $725B across $Microsoft(MSFT.US), $Alphabet(GOOGL.US), $Amazon(AMZN.US), $Meta Platforms(META.US) and $Oracle(ORCL.US). Plus the OpenAI and Anthropic compute commitments layered on top.

I think that capex guidance has been revised up by every credible institution over a dozen times this year so far. From Goldmans, to MS, to BofA.

And I am sure that during earnings coming up soon, pretty much all of them will revise their capex forecasts up for 2026 remaining.

It was just a week or so ago, but remember how everyone felt about the $Micron Tech(MU.US) earnings? Pretty good, right? HBM sold out...LTAs...growing volumes of locked in demand/pricing.......I could go on and on.

But I feel like the bear case I see most often lately is to do with $NVIDIA(NVDA.US)'s ecosystem commitments. E.g. their ~$124B spider web of investments + POs across areas like LLM labs and neoclouds.

I'll look at that a bit below, but they exist because fronteir demand is outpacing everyone's balance sheet + ability to fund it.

But first, you MUST realise that the structural driver has shifted in a way that "bubble" doom posters miss/ignore:

- The 2023/24 phase was a training-led capex cycle with lumpy and cancellable revenue.

- The 2025/26 phase is inference and agent led with reliable recurring and usage based revenue.

Inference demand compounds with deployment = so do financials.

This is the transition dark fibre never made in 2000 where telecom infra was built ahead of a demand curve that took something like a decade to actually arrive.

Compared to AI compute that is being built behind a demand curve aka shortages. Shortages are not bubbles!

In 1999-2000, telecom capex was funded by companies burning cash. In 2026, the marginal dollar of AI capex is funded by 5 of the largest cash generation machines ever. They're so well insulated from the capex they're spending that it won't matter in a few years time.

Now, I think we all know that AI names are super crowded.

I've noticed that since joining X where I get people commenting things like "What happened to $Applied Optoelectronics(AAOI.US)?" and "Do you like $SIVE?" when the names drop 10% in a day from volatility.

But on top of that, you look at data like BofA Fund Manager Survey that shows >80% of managers naming long-AI/semis.

That's a record high for any trade or thematic in the survey's history.

So crowding of this level mechanically deepens drawdowns because everyone's stop is behind everyone else's...it can become a vicious cycle on red days. And it explains why an >10% air pocket can arrive on no company or sector news at all.

Then you've got idiots posting online and going on podcasts saying that the AI trade is just like the internet trade back in 1997.

But as we all know, those internet names like Pets dot com had literally no cash flows and most had fugazi financials.

You just have look at the financials for companies like $Sandisk(SNDK.US) which people also say is a bubble. They seem kinda fine to me? And will be fine going forwards. (Understating it slightly, their earnings will largely all be fantastic)

That being said....

Let me actually give you signals to look for that'll actually change my mind on whether I have conviction or just blind faith in the AI trade:

1. Capex cuts from hyperscalers rather than a deceleration in growth.

2. GPU rental spot pricing declining sustainably rather than plateauing.

3. HBM contract pricing rolling over while supply agreements get renegotiated.

4. Neocloud credit spreads gapping to distressed levels.

5. Fronteir labs like Anthropic/OpenAI seeing MRR's decline or stall.

These are probably the big five "tripwires" + none have been triggered so far. I don't see any being triggered for the forseeable future.

But all you need to know for a TLDR bull case for the AI theme is:

- demand for compute remains supply constrained.

- hyperscalers funding the capex are cash flow generative machines.

- valuations for high quality names from $Sandisk(SNDK.US) to $Micron Tech(MU.US) to $Broadcom(AVGO.US) to $NVIDIA(NVDA.US) are stupidly low.

On top of everything else mentioned here (I could keep going but I'll save you from reading more of my ramblings/rants).

That being said, in my view, there's certainly a case to rotate into higher quality names rather than random upstream shitcos.

I've mentioned this many times over the past couple of weeks.

From 2025 to early Q2 2026, you could basically long almost any name that had some sort of ties to AI and it'd go up 100% in weeks. $POET Tech(POET.US) and $IREN(IREN.US) come to mind.

I personally think being a little more selective going forwards will be more crucial, and financials will matter a lot more than going long a name that just falls in the AI bucket.

Paradis Stock Report: Harmonic Drive Systems (6324)

For upstream humanoid exposure. (Disclosure: I hold a position).

I have mentioned HDS a few times in my humanoids article and in various comments.

-> Here's my technical research article that I wrote a few months ago for a few buy-side institutions.

(With updated financials and amended/shortened for my anonymity and integrity to paying firms. Some customer mapping etc. may need updating also):

Report:

Every humanoid demo you have seen - Figure pouring a coffee, Optimus folding a shirt, a Unitree robot doing a backflip - leans on a very specific component almost nobody outside the motion control world can name.

Before the model decides to move the hand and before the motor delivers torque, something has to convert a low-torque rotor into the precise, high-torque rotation a wrist or finger actually needs.

That something is a precision reducer.

And in the highest precision joints, the reducer of record for 40 years has been the strain-wave gear.

Invented, branded, and still dominated at the premium end by one company: Harmonic Drive Systems (6324).

First, a piece of housekeeping:

- The entity that matters is the Japanese parent, listed on the Tokyo Prime Market.

- It owns 100% of Harmonic Drive SE in Germany (the former Harmonic Drive AG) and a majority of Harmonic Drive LLC in Beverly, Massachusetts.

- It is not a Nidec company. Nidec-Shimpo is a separate competitor, and Harmonic Drive actually sold its policy stake in fellow reducer-maker Nabtesco during the year to March 2025.

Their flagship product is a simple three-part assembly:

1. a wave generator

2. a thin walled steel cup called a flexspline

3. a rigid circular spline

[I've removed some technical stuff cos it can get boring quickly]

Harmonic Drive sells this as component sets (the CSF, CSG, SHF and SHG families), as integrated AC servo actuators bundling a frameless motor, the gear and an encoder (the SHA and FHA lines), and as lower-ratio planetary units for less demanding joints.

In a six-axis industrial arm or a cobot, this is the gear in the wrist, forearm and elbow.

The heavy base and shoulder axes belong to RV cycloidal reducers, where Nabtesco holds something like 60% of the world market; harmonic and cycloidal are complements far more than competitors.

The same strain-wave gear shows up in semiconductor wafer-handling and lithography stages, in surgical robots. In aerospace, Harmonic Drive units have flown on NASA's Curiosity and Perseverance rovers.

The customer map broadly reads like a census of the automation industry: FANUC, Yaskawa, ABB, KUKA, Kawasaki, Universal Robots, plus a long tail of semiconductor equipment and medical device OEMs.

[I've removed some speculative numbers and customers]

For analysts considering the bull-case optionality:

The moat is in the process, and specifically in one component: the flexspline. That thin steel cup is elastically deformed millions of times over its life, and it is the part that fails.

Patent and engineering literature is blunt about this...the flexspline fatigues, and the cheap routes to making one wreck its cycle life. Bulk-metallic-glass flexsplines crack at far lower cycle counts because their fracture toughness is a fraction of forged steel's. 3D printed metal flexsplines carry porosity and surface roughness defects that, in the words of one patent filing, dramatically reduce their lifetimes.

[Patent/IP stuff is boring, so I have removed]

Harmonic Drive's proprietary tooth profiles (the "IH" generation and its successors) spread load across many teeth meshing at once, which is how you buy fatigue margin without buying weight.

It is an empirical feedback loop: decades of field-return data from satellites, wafer handlers and arms running under extreme duty, fed back into design margins.

Layered on top is qualification lock-in:

In aerospace, semiconductor and medical applications, the reducer is designed into a system that then takes 12 to 24 months to re-certify if you swap it. Once a Harmonic Drive part is flight or fab qualified, the switching cost is the re-qualification programme.

That locks revenue for years, and in aerospace it is close to permanent.

Now moving onto the core demand vector: Humanoids

The narrative premium is humanoids and maths can be seductive.

A humanoid uses somewhere between 20 and 40 reducers across its joints, a mix of harmonic and planetary.

At a Harmonic-grade ASP frequently cited near $800 a unit, 100,000 humanoids is on the order of $1.6B of harmonic-reducer demand which is a 5-10x TAM expansion against today's industrial base.

From supply chain mapping, unit numbers are finally inflecting off zero:

Harmonic Drive has started to book real, if small, humanoid orders. Roughly ¥1.3B in a single quarter, around ¥2.5B guided for the year to March 2026, with management telling analysts the following year could double or triple that, and the count of potential humanoid customers climbing past 15, concentrated in Japan and North America.

In my view, this is the closest thing the listed world has to a pure play on the actuator content of a humanoid joint.

For clients, there are naturally some coverage concerns however:

Coverage is thin and almost entirely Japan-desk - Goldman, Morgan Stanley MUFG, Jefferies, Macquarie, Citi, Nomura, Mizuho.

For a generalist global fund the name is genuinely hard to own:

1. A Tokyo Prime listing

2. Japanese-GAAP disclosure

3. No quarterly order detail

And a business whose irreplaceability lives in segments (semiconductor, aerospace, premium humanoid) that headline market share data (which shows Harmonic Drive losing share) actively obscures.

However, in my view, the cyclical recovery is underway:

This is not a story stock with no business. For the year to March 2026, Net sales rose 7.0% to ¥59.6B.

It is important to factor in the quarterly cadence too:

Revenue climbed ¥13.5B, ¥14.3B, ¥14.3B, ¥17.4B through the year, and EPS went from minus ¥0.40 in the first quarter to ¥8.86 in the fourth as factory utilisation recovered and the CoS ratio fell.

This is textbook fixed-cost absorption, with incremental margins reportedly above 50%.

Guidance for the year to March 2027 is for sales of ¥68.0B, up 14.2%, operating profit up 141% to ¥6.2B, and net income up 180% to ¥4.5B.

Orders are corroborating:

- full-year intake of ¥61.6B was up 16.2%, with North American reduction-gear orders up 31.3% and Japan up 26.9%.

The balance sheet removes any solvency question from the discussion:

- Roughly 72% equity ratio, net cash, operating cash flow of ¥6.4B, and enough firepower to fund a 33% expansion of the US Beverly plant by December 2026 without touching shareholders.

In May the company set a fresh five-year plan:

- ¥100B-plus in sales and a >15% operating margin by the year to March 2031.

On review, this is a fortress balance sheet attached to a genuine moat in a market that is inflecting.

Risks:

China.

Harmonic Drive's global unit share has fallen from north of 70% historically to an estimated 20% today. In China (the largest robot market) domestic suppliers now lead outright:

Leaderdrive and Zhejiang Laifual held roughly 27.5% and 21.4% of the country's robotic harmonic-reducer market by 2025 shipments.

Leaderdrive grew revenue 47% to 570.7M yuan in 2025 and more than doubled net profit, and has already fulfilled humanoid R&D orders for Tesla and Figure.

The pricing gap should be emphasised at this stage:

Chinese Tier-1 units land at circa 40-60% of Harmonic Drive's price, and that gap is widening.

Laifual's average reducer ASP fell about 30% in two years, which its own IPO prospectus framed as a deliberate share grab. Independent stress-testing puts the best Chinese parts at 70-85% of the Japanese benchmark on torque density and lifespan, but at backlash parity.

Based on discussions with [famous robotics company], for most of the joints in most of the robots, "good enough at half the price" is the whole ballgame.

Management's own long-term framing has overall strain-wave share declining toward 6% in the long run.

[Some analysis on valuations and financial modelling, removed cos it's boring]

Further, the largest humanoid programmes may design around Harmonic Drive:

Tesla designs its actuators in-house and multi-sources the reducers, including from Chinese suppliers, with teardown chatter pointing to a meaningful and rising Chinese share of Optimus content at 30-40% lower cost.

Harmonic Drive is, at best, one of several reducer sources for the biggest program and not the sole-source chokepoint the current valuation and our forecasting implies.

[More boring bits on depreciation, yen strength, cyclicality, sales geographies]

To conclude:

1. Harmonic Drive owns a qualification locked chokepoint in premium precision reducers.

2. Sits on a strong balance sheet.

3. Is one year into a credible cyclical recovery.

4. Holds the cleanest listed exposure to humanoid actuator content that exists.

If a large Western humanoid program (Figure, Apptronik, even Tesla) confirms primary source Harmonic Drive content shipping at premium prices and scale, the structural story has potential to re-rate.

You wanna invest in humanoids but you've never watched Ex Machina?

Fun fact: this film sparked my interest in AI and humanoids when it came out in 2014.

This is basically a must-watch if you're investing in $Churchill Capital XI(CCXI.US) and upstream humanoid supply chains.

To see what the future of humanoids *could* look like with bad actors at the helm.

South Korean humanoid supply chains will be crucial.

For the upcoming humanoid production ramp.

GS forecasts that 30% of humanoid components will come from South Korea by 2035.

Due to their automotive industry expertise translating over into producing humanoid parts like actuators.

Just for a quick tldr on actuators:

- they essentially mimic human muscles/joints

- actuators currently make up ~50% of the BOM for a humanoid.

- Korea has an advantage in actuator production since they're similar to auto parts tech e.g. power steering, braking systems and autonomous vehicle tech.

And "several Korean auto parts makers have already leveraged this know-how to gain a position among the leading non-Chinese suppliers of critical actuators for robotics." (GS)

So for upstream humanoid exposure...

It does seem like South Korea will have a massive role in being the key supplier for non-China humanoids.

On top of their crucial position in AI supply chains with memory from SK hynix and Samsung...

Robotics is moving so fast.

I expect deal value to 10x in the upcoming years.

With private capital flowing into companies that can manufacture at scale + win paying customers.

Some of the biggest deals so far this year include:

-> Saronic (US) w/ $1.75B Series D in March for autonomous naval vessels.

-> Skild AI (US) w/ $1.4B round for humanoid software.

-> Apptronik (US) w/ $0.94B Series A for industrial humanoids.

-> Mind Robotics spun out from $Rivian Automotive(RIVN.US) and raised $0.9B YTD.

There'll be plenty of public investment opportunities though.

Either by investing upstream in the suppliers that make up humanoid's BOM e.g. in actuators or perception systems.

In a similar style to the AI supply chain, going upstream to find where the winners are.

Or via pure-play robotics companies selling B2B and B2C once production ramp hits in a year or two.

Three random themes & analysis:

1. Offshore drilling

2. Salmon

3. Airlines

These aren't multi-year stories like AI, humanoids or space. But are instead catalyst driven, short term trades.

High level notes:

1. Offshore drilling

Offshore drilling equities trade like distressed cyclicals even as they strengthen contracts.

$Transocean(RIG.US) for example trades at P/B of ~0.93 & P/S ~1.8x.

So sentiment is extremely bearish even while they've been adding backlog. Generalists have abandoned oilfield services since the industry is relatively small + scarred by the 2020-21 bankruptcy wave.

In terms of potential catalysts:

- Transocean-Valaris merger (Feb 2026) needs shareholder/regulatory approval which would be a potential re-rating event.

- Management guides utilisation above 90% by late 2026, approaching 100% in 2027.

- $Transocean(RIG.US) added ~$1.6B backlog since April + recent fixtures span ~$550k/day.

Oil is obviously the key risk factor though where an oil price collapse (e.g. WTI below ~$60) triggers capex cuts + contract deferrals. That margin compression would offset any backlog gains made.

No positions here yet for me, but I do have positions further down in the Airlines section.

$Transocean(RIG.US) does look kinda compelling at these levels tho w/ their backlog growth. But do I wanna hedge the AI trade more? Not sure just yet...

2. Salmon

This one's fun.

2025 was a record supply year for salmon w/ industry supply up 18%. Which ultimately crushed spot prices.

Aggressive harvesting in Q2/Q3 2025 depleted biomass, setting up a sharp H1 2026 slowdown.

Looks like a pretty nice setup though:

- niche Oslo listed cluster of companies

- undercovered by US investors

- punished through 2025 on oversupply

- strong NOK

- Norway's resource-rent tax

Catalysts:

- Norway's salmon production fell 3% in Q1 and will fall another 2% in Q2

- which leaves global supply growth of just 0-2% in H1 2026

Prices already seem to be turning too where the benchmark salmon price was ~74 NOK/kg on June 19. Which is up ~13% over the prior month. But ofc if biology recovers fast, then that'll be the key risk.

Names if you're curious: Mowi, SalMar, Bakkafrost

I probably won't take positions here, but it's been a fun area to research recently as a break from AI/semis lol.

3. Airlines

Boeing + Airbus delivery delays are artificially capping industry capacity which protects airfares even as consumer sentiment softens.

So the airlines have a cap on capacity growth which has handed them decent pricing power e.g. fares are running ~25% higher yoy + staying sticky even as fuel falls.

So investment-grade balance sheets + collapsing fuel costs into H2 should drive margin expansion.

$Delta Air Lines(DAL.US), $United Airlines(UAL.US), and $Alaska Air(ALK.US) are the core names where you'd wanna opt on the loyalty/premium oligopoly rather than seat sales. Even w/ Spirit gone, that's why you'd want to avoid ULCCs.

If I was to summarise each name:

- Delta is the quality leaner but looks a lot more priced in at these levels.

- United is probably better risk/reward than Delta rn.

- Alaska is highest risk but highest reward.

Southwest ( $Southwest Airlines(LUV.US) ) is another name but not looked too much into them in the theme.

Ultimately all are turnaround plays but similar to offshore drilling, oil's the key risk which obvs acts as headwinds to margins. V. macro driven w/ Hormuz etc opening.

And if Spirit routes get backfilled from July (?) then fare premiums could erode quickly.

But $Delta Air Lines(DAL.US) earnings are July 9 and $United Airlines(UAL.US) mid-July so could see a re-rate if sticky fares + falling fuel costs = good guidance.

Just as a disclosure, I did take positions in $Delta Air Lines(DAL.US) as a hedge when Iran/US deal was signed - think I commented it a few times on my posts.

---

Just for a list of fun themes I'm looking at outside of AI rn.

Missing a lot of notes here tbh and it's just a spew of thoughts/info, but just don't wanna bore you guys with all the nuts and bolts lol.

NFA ofc, these are just short-term investment areas if you're wondering about ideas.

Ok guys, these $Micron Tech(MU.US) gross margins are scary high:

Q3 Estimated: 81.9%

Q3 Actual: 84.9%

Your margins wouldn't be this high even if you robbed your local liquor store.

Helping you understand today's market sell-off:

Paradis Macro Report [June 23]:

-> KOSPI closed -9.99%:

Second worst day this year, and fifth worst ever. With SK Hynix and Samsung both down over 12%.

A few drivers:

1. MSCI kept Korea in Emerging Markets which killed the passive inflows were estimated to be $24B (UBS).

2. There was also forced deleveraging of retail leverage.

3. Profit taking mainly driven by SK Hynix overtaking Samsung's market cap (viewed as a market top).

4. Profit taking ahead of $Micron Tech(MU.US) earnings tomorrow.

Retail investors tried buying the dip ($5.6B), which is the largest single-day retail net buy ever. But that wasn't enough to overcome the sell off.

Looking at current volatility levels, the drawdown in Korea looks set to continue tomorrow as more retail investors are forced to unwind positions.

-> The sell-off in Korea extended into US semiconductors:

Nasdaq fell 3.2% and the SOX fell ~7.8%.

With names like $Micron Tech(MU.US), $Sandisk(SNDK.US), $Applied Optoelectronics(AAOI.US), and $Coherent Corp.(COHR.US) all down over 10% today.

AI stock price moves are largely memory-driven. If memory goes down, so do other AI adjacent names (usually).

However, it is crucial to remember that the memory supercycle is still in tact w/ sold-out capacity all through 2026-27.

Plus hyperscalers forecasted to spend up to 50% of 2027 capex on memory alone as per SemiAnalysis.

-> Micron earnings tomorrow:

I'm personally expecting Micron to report sublime earnings and guidance tomorrow.

However, do not be surprised if the market sells the report just like last quarter.

And if you've read my previous reports, you'd note that I have been firmly against using leverage for the past few weeks.

Today's events are proof.

New $Sandisk(SNDK.US) HBF Patent:

Analysis in plain English (hopefully):

The patent is about training models, which flash is supposedly bad at.

The reason they could file it is that they found the loophole where it doesn't cover full training. Rather, it covers fine-tuning via a method called LoRA which keeps almost everything static.

Instead of retraining the whole model, you freeze it & train only a tiny add-on bolted to the side. The frozen part never changes during the process.

Sandisk's insight is if the big part never changes, you don't need expensive memory to hold it....you just need to read it, over and over again.

And reading is the one thing that cheap flash is perfectly fine at.

So the patent parks the frozen giant model on cheap flash + keeps only the tiny changing part on the expensive fast memory.

Quite elegant imo since it doesn't try to fix flash's weakness. It just makes it irrelevant.

Imo the accurate way to frame this patent is that it widens HBF's tam from pure inference into training.

And the encouraging thing for Sandisk is that the programme behind this looks real rather than theoretical e.g. partnership w/ SK hynix to turn hbf into the industry standard.

Ofc, patents aren't revenue. What'll actually matter is qualification/shipments. I.e. roadmaps.

From my understanding though, first hbf chips sample later this yr w/ first products using them arriving some time in 2027.

---

Patent number: US 2025/0390733 A1