Analyst. Investor. Early where the market is late.
Joined December 2025
- Tweets6.2K
- Following739
- Followers17.7K
- Likes10.5K
Someone who needs to raise enormous amounts of debt is now warning about his competitors’ debt risk?
Interesting timing, Elon.
Another shot at neoclouds.
Make investors nervous, make funding harder to secure, slow down competing capacity. A great way to keep compute scarce and prices high.
I’m glad $IREN has financing secured. Even less worried about $NBIS. Smaller players dependent on external funding could have a rougher time.
Meanwhile, his own plans call for another 10-20 GW of compute in 2027, with ambitions stretching even further.
That won’t come cheap.
Dirty game, Mr. Musk.
Pretty much.
As of the past few weeks, debt markets have almost entirely shut off access to money for any company that doesn’t have substantial revenue independent of AI data centers.
SpaceX closed $40B of debt for GPUs, but we’re investment grade and have Starlink & launch revenue.
That means the 200+ neoclouds are going to see guppy-level death and consolidation. Some will make it through, but not many.
15 executives across the AI supply chain, all pointing to the same thing: demand continues to outpace supply.
But there’s no single bottleneck. Constraints vary across compute, memory, optics, power, and capital.
And adding capacity in one area doesn’t mean the rest can keep up.
That’s what makes scaling AI infrastructure so difficult.
15 business leaders on AI infrastructure demand and capacity in the last ~month or so
>@ajassy: “2027 is largely sold out. … A good chunk of 2028 is sold out.”
>@sundarpichai: “We are capacity constrained on a daily basis”
>@JensenHuang: “Demand is extraordinary. Capacity is the constraint.” On future optics needs: “several orders of magnitude more than we currently ship today”
>Matt Garman, AWS: “the demand is just massive”
>C.C. Wei, TSMC: “we did not see any sign of slowdown”
>@LisaSu: “There’s very, very high demand for the next several years”
>Sanjay Mehrotra, Micron, on DRAM: “we do not have line of sight to when supply and demand will return to balance”
>Christophe Fouquet, ASML: Planned 2027–28 supply “will not fully meet the demand of our customers”
>Michael Hurlston, Lumentum: “completely sold out”. Bloomberg reports the optical-component horizon extends through early 2029
>@finkd: “we have not yet found the limit” of applying compute to Meta’s core business
>Josh Payne, Nscale: “unprecedented global demand”
>Tim Archer, Lam Research: “AI driving higher semiconductor demand and greater device complexity”
>@satyanadella: “I am confident in the long-term returns on our investments.” Based on demand and efficiency gains
>@LipBuTan1: “We don’t see any slowdown”
>Hock Tan, Broadcom: “The thing you hit on that I think might be the biggest constraint is capital”
Finn Stockinger retweeted
Hello everyone, our $NUAI deep dive is now live on Substack
A tenant + financing is the major catalyst. Which we expect soon. But what would the deal actually be worth to shareholders? That’s the groundwork we’ve laid in this report.
We start with management hires and the roles they play, followed by a full site breakdown, power availability, and the sector’s tailwinds and headwinds.
We walk through permitting: what has been secured, what remains outstanding, and how requirements differ across the development phases. This ties into @Mark_AAIG technical series, which I’ll also link in the comments.
We then examine different financing scenarios for Phase 1 and Phase 2, showing how funding structures, ownership and potential dilution change the economics for shareholders.
Together, this builds a foundation for understanding the full campus and its planned 1.4 GW of gross capacity. As tenant and financing details emerge, we can quickly reassess the investment case: how much capacity is contracted, on what terms, what it costs to deliver, and how much value ultimately accrues to NUAI shareholders.
Ask us questions, challenge our assumptions, or discuss the report in the Substack subscriber chat. You can also connect to our Discord and join the dedicated NUAI deep-dive channel.
Write-up and technical series in the comments 👇
Team AAIG
Elon Musk recently shared a very interesting take. He believes we’re about to see a lot of “stranded compute in partially completed data centers.”
I’ve already seen plenty of posts interpreting this as another reason to write off neoclouds.
I think, as usual, the market is jumping to conclusions a little too quickly.
To me, this is primarily about a mismatch between GPU delivery schedules and the timelines for getting power, cooling, and the rest of the infrastructure ready for deployment.
And this isn’t just a neocloud problem. Hyperscalers face exactly the same risks.
Of course, the consequences can be much more severe for smaller operators. GPUs sitting idle are expensive assets generating zero revenue, while financing costs keep piling up. That alone could put some companies in serious trouble.
So what actually causes a data center to remain partially completed instead of becoming operational?
I’d break the risks down into two stages.
1. During construction
This is potentially the most dangerous stage.
Delays in grid connections, transformers, switchgear, power distribution, and cooling infrastructure can push deployment timelines back by months.
And unlike many other issues, these bottlenecks aren’t easy to work around.
2. During deployment and commissioning
Even with the building and power infrastructure ready, there’s still plenty that can go wrong.
Rack installation, liquid cooling, network configuration, cluster stability, software integration, performance testing, and final customer acceptance.
The difference is that many of these issues can be addressed through engineering changes, troubleshooting, or phased deployments.
A networking issue might delay part of a cluster. But if you don’t have the power to run it, there’s not much you can do.
That’s why I believe data center sites with secured power, and especially those already energized, have a significant competitive advantage.
Vertically integrated operators may also be better positioned because they have more control over construction, infrastructure, and deployment timelines.
But there’s another question I’ve been thinking about.
What happens to neoclouds relying on colocation providers?
Imagine you’ve committed billions to GPUs, but your colocation partner misses the delivery deadline.
You’re left holding expensive hardware that can’t generate revenue, while your financing obligations continue.
Will contractual penalties and compensation from the colocation provider actually cover those losses? Or does most of that risk ultimately remain with the GPU owner?
I think this is something investors should be paying much closer attention to.
And hyperscalers aren’t immune either.
Look at Amazon.
Business Insider reported just last week that AWS has been receiving GPU systems before the infrastructure needed to deploy them is ready, leaving some hardware waiting in storage.
That raises an interesting question.
Was this part of a deliberate procurement strategy to secure GPU supply ahead of deployment? Or did Amazon simply move faster on hardware purchases than on data center readiness?
Those are two very different situations.
Either way, it shows that even the largest players aren’t immune to these bottlenecks.
I still believe neoclouds can be an incredible business.
But only if execution is handled with surgical precision.
Having GPUs isn’t enough. Having secured power isn’t enough. Even having a completed building isn’t enough.
You need to turn all of that into operational, revenue-generating compute, and you need to do it on time.
Execution is the key. Power is the key.
I’m curious where everyone stands on this.
Are you investing in neoclouds, or do you see the sector as a trap?
And which neocloud do you think is best positioned to execute over the next few years?
$SIVE is an interesting company, but it was a much better opportunity at $0.40, when I first brought it up.
The problem is, it was never going to live up to the ridiculous expectations the market had for it.
Remember when people were comparing it to $LITE? Funny how nobody brings that up anymore.
There are plenty of better opportunities in the optical space, with actual fundamentals to back up their growth. $SIVE just isn’t one of them.
I’d even go as far as saying I don’t see this stock getting anywhere close to its ATH for a long time.
It’s a dead end IMO.
Yes, I like $SIVE. It's a fun company to cover, since there's a lot of supply chain alpha here.
But the company completely Tall Poppy Syndromed itself last earnings.
And it's clearly shown in the drawdown.
Bit of a longer consolidated thought and some scattered bits of supply chain info:
- Sivers has 100M CW DFB laser capacity coming online Q4 2027 during an industry shortage.
And very large allocations from foundries online (they were doing these quals with Win Semi + 2 others back in 2023/2024), so they can actively monetize the bottleneck today.
The other interesting reason why bottlenecks are important (aside from ASP hikes), is because qualifications convert much faster + more likely given incentives around supply scarcity.
-> As one example, module makers have expanded the amount of suppliers, eg. (lightcounting cited companies expanding to 5-7+ suppliers from 2-3).
Innolight, for example last month responded to a question addressing tight 1.6T DFB laser supply in 2026–2027:
- Innolight was actively introducing new suppliers last month and are planning to sign 2-3 year-long contracts to lock upstream capacity
-> This is while your biggest merchant laser companies making up the world's 600m+ EML/CW laser capacity like $COHR have redirected lasers to their own programs. $LITE sold out into 2029, Landmark also locked a lot of their capacity into 2027-2029. $AAOI was an expected player, but ended up redirecting lasers into their own transceiver programs too.
And this coincides with times $SIVE has tremendous capacity available now + actively engaged with Chinese pluggables players (not saying it's Innolight, but just shows Chinese makers like Cambridge/Eoptolink are actively trying to secure laser capacity).
- Fun to note, before $JBL was announced, Sivers had and has ongoing direct engagements with hyperscalers as well.
We saw that hyperscalers approached $MTSI and $AAOI for laser capacity from their earnings transcripts (and hilariously got turned away from AAOI due to shortages).
So the odds pluggable makers or hyperscalers end up converting to capacity agreements with $SIVE seems a materially higher.
Just a little rant: for all the sellside/retail trying to model revenue, you can't just slap on a stupid 25-35% CAGR on NRE revenue since it's close to 0 revenue now.
Especially when they scale up to maybe 300M laser capacity (extrapolated from 2:1 external/internal capacity).
It's also not the same as $POET having OE packaging capacity at scale, since lasers are in a genuine shortage. So capacity is more likely to convert like how $AAOI is sold out, $LITE is sold out into 2029
We've already seen what revenue that laser capacity brings in since $LITE cited $5B in laser capacity from just one 2028 fab alone in Greensboro (on higher end ranges).
And whatever lasers Sivers sell, they're targeting ~50-60+ blended gross margins long term on maybe $3-6 for lower range lasers to $6-$12 for high end custom deployments. But no disclosures around capacity -> revenue conversion yet from Sivers.
Now here's the thing people don't really understand, "why no revenue?" for current opportunities?
$SIVE pre 2024 focused on custom development for future architectural shifts like CPO.
For custom deployments:
- We finally got conversion from 1 of their 4 or so lidar customers this week with ~ $AEVA.
- ~ $APPL was confirmed ongoing earlier this year with an explanation that "It takes tremendous amount of trials and algorithms way beyond just what we supply to them".
- All their customers from 2023-2024 like Ayar and 3 others, probably achieve volume in 2028 as CPO takes off.
Custom high end lasers programs are extremely sticky but take many years and are mainly NRE. And as seen with CPO scale up, faced delays with commercialization over the years.
So $SIVE from the CEO shift in 2024 realized "okay" let's ramp revenue with mass commercialization products today like 70mW/100mW/200mW lasers for 800g/1.6T pluggables.
And that's where they found $JBL and all your other large pluggable players today during the shortage.
Now... the work Sivers did in 2023-2024 earlier: Such as $AAPL, Ayar, $AEVA, other lidar/CPO players.
And their new mass commercialization opportunities like $JBL, $GFS, and many pluggables/NPO suppliers in engagement now.
Have all their inflection points converges 2027 onward, which is why I find the company compelling.
Where I think management went completely wrong was messaging and near-term execution:
- Deciding to focus on current tiny contracts like Seminex, AllSpace, Tachyon orders felt like the entire focus recently in earnings, shareholder letters, and media commentary
-> Brutal honesty to $SIVE management, majority of Western audiences literally don't care and it's immaterial to the valuation.
$SIVE moved 0% on all of the Seminex/AllSpace/Tachyon order announcements, and even dropped in price. But rallied like 50% on $GFS and something similar on $JBL announcement.
The reason why most took long positions in the first place was CPO.
So if $SIVE wants to go answer questions to Swedish media about "Why are you bleeding out" and defending that with cellular orders explaining pipeline conversion.
That's not going to help people stay long during drawdowns.
-> Second part was not committing to NASDAQ listing in a certain timeframe and actually executing.
- If there's delays, sure, just say it, but non-committal lawyer-like language is infuriating to hear, especially from the CFO.
I think they'll have that opportunity next month earnings to correct everything.
So wrapping up:
- There's nothing I see remotely close to $SIVE, in terms of % of global CW laser capacity they're coming online with relative to MC.
- They have a ton commercialization paths, from O-Net, $POET, $JBL, Ayar, $GFS, $AAPL, $AEVA, and many potential hyperscaler/pluggable engagements.
Even if one or two fall through, the breadth of ongoing quals is so deep + current laser bottleneck is severe, that it's more likely than not many will convert to production.
It's clear $SIVE wants to focus on dominating the laser chokepoint first, and ended up using their 700m SEK fundraise for capacity expansion.
But they should focus on communicating the billions in potential economic scale from that laser capacity conversion.
And execute on it.
The most absurd thing about AI is that AI develops AI.
You might think, what is he even talking about?
It's a big simplification, but it's precisely thanks to the development of artificial intelligence that we can develop better and better LLMs, design new hardware, improve testing, optimize production processes, and accelerate development and production ramp-ups.
This, in turn, allows us to build more data centers, faster and more efficiently, with even more computing power, which only accelerates AI development further.
And the cycle keeps accelerating. More data centers, more compute, and an even faster pace of development.
Of course, we'll eventually reach certain limits, but right now, we're riding a rising wave, and I believe we're still at the very beginning.
Sit down for a moment and ask yourself the question: where will we be in a year or two, where will we be when we have 2x, 4x, or 8x more computing power?
Why is $NVDA x $NOK AI-RAN so important?
This isn't a technology that's going to generate meaningful revenue today or next quarter.
But I believe it's a critical piece of the Physical AI revolution.
I try to explain you why.
Autonomous vehicles, robots, and drones need more than just powerful AI chips. They need reliable, ultra-low-latency connectivity and the ability to process data and make decisions in real time.
And not everything can be sent to a distant data center.
AI-RAN brings AI computing closer to where the action happens, turning traditional cellular infrastructure into distributed AI platforms.
That means networks won't just carry data anymore.
They'll increasingly become part of the infrastructure that powers Physical AI.
And that's why $NVDA x $NOK and major telecom operators matters far beyond telecommunications.
Regardless of market sentiment, demand is still there, and GPU rental prices are not going down.
Neoclouds 🚀
IREN’s @danroberts0101 speaks to @ArjunKharpal about AI financing and demand on CNBC.
Watch the full interview.
We should think about data centers as prime real estate. Supply is limited, and there are far more companies looking to lease capacity than there are available sites.
Whoever is willing to pay more gets the keys.
It’s really that simple.
Market selloffs are hard to get through. And they get harder with every passing day.
Every little bounce gives you some hope, and then the market drops again, leaving you feeling even worse than before.
That’s completely normal.
I get more frustrated every day. I hate losing. I’m extremely competitive. If I take 100 shots, I want to make all 100. But I’ve learned how to deal with it.
Life has taught me patience, and when there’s nothing I can do about something, I just keep doing what I do.
I analyze earnings reports, do research, read articles, and keep learning more about technology.
When sentiment turns, I’ll be ready. That’s why I’m not going to stress about it, and I’d suggest you do the same.
Just keep doing your thing and working hard.
Even if you believe this is Dot-Com Bubble 2.0, sooner or later the market will bounce back and another bull market will come. We might lose a ton of money along the way, but you need to be ready to make back as much as you can when that time comes.
Just to be clear, I don’t believe we’re in another dot-com bubble.
Not every company that’s been hyped to the moon will become profitable or live up to what the market has priced in.
But there are plenty of companies in the AI infrastructure space that have simply been caught in the crossfire and will keep posting record results quarter after quarter.
Remember, be ready.
Finn Stockinger
How Nebius’s dynamic GPU pricing works $NBIS
@stevehou Silicon Data looked at the first nine days of Nebius’s dynamic spot pricing.
Here’s what caught my attention.
Nebius can update prices every 15 minutes as demand and available capacity change.
You can follow the rate or set a maximum price. If the rate exceeds your limit, your instance stops.
From October 8–16, H100 spent most of the period at its $0.79/hour floor.
B300 started below the market median and finished above it on the final three days.
For me, the useful question is what happens around new capacity additions. Does spot pricing fall when GPUs come online, then recover as customers absorb them?
Tracking those two things together could tell us more than a rental-price screenshot. That’s something to test over time; nine days cannot answer it.
And these are spot quotes for interruptible workloads.
They don’t tell us what Nebius earns on long-term contracts.
Thanks to Steve Hou for the analysis 👊
Today everyone is wondering whether to implement CPO or not.
And @JensenHuang talks about how they started preparing a new process with TSMC already 5 years ago.
Lower signal loss, significantly higher data rates, lower bit error rates, higher quality, and much lower energy consumption.
5 years ago was a year before the launch of ChatGPT.
There are no coincidences here, just pure execution and roadmap delivery.
Where we are today was planned in advance.
Listen carefully to what He says so you won't be surprised by what happens in a few years.
Follow the leader.
H/T @benitoz
Last week Morgan Stanley cut its $AMD GPU demand forecast by 43% on 2026-2030.
Today $NVDA secure supply chain of the substrates, advanced packaging and CPO engines.
Do you connect the dots?
When @nvidia decides to speed things up, they take AI factory deployment from 6 months to 14 days.
How is it possible?
Meet DSX POD: a Vera Rubin AI factory built and tested in modular containers, delivered ready to connect at a prepared site.
Jensen: “An AI factory creates value when it starts producing tokens - not when construction is complete.”
You spend billions on compute, then wait months before it earns its first dollar. Cutting that wait is a pretty big deal.
Wondering how Jensen plans to deliver 70% growth in 2027? DSX POD is your answer.
They’re going to bring a massive amount of compute online next year.
Are you ready?
NVIDIA is scaling its supply chain fast: 2.5x substrate capacity, 4x+ advanced packaging, and 15x CPO optical engines.
The scale of this buildout is insane.
Source:
nvidianews.nvidia.com/news/n…
What if you could add memory without buying more GPUs?
$MRVL is showing a module combining HBM and DDR5, enabling 33TB memory pools connected by optics up to 50 meters away.
Memory capacity can scale independently of compute.
$PENG was already developing its Photonic Memory Appliance with Celestial AI before Marvell acquired it.
IMO, that puts Pengs in pole position to bring these systems to customers.