NairoLB @NairoLB242

Very interesting what's happened to AI infra and semis. Forward EPS estimates are up, but multiples are down. Either the buy-side is expecting less earnings than the sell-side numbers suggest, or people simply don't believe this earnings power can continue... and we're seeing the cyclicality of semis come back into play, something many believed would no longer be the case with AI in the picture. Or the market is simply pricing in too much macro risk, and this is actually a great buying opportunity...(?)

“A man may rise to high success in the world, and even to lofty altitudes in the spiritual realm, and again descend into weakness and wretchedness by allowing arrogant, selfish, and corrupt thoughts to take possession of him. Victories attained by right thought can only be maintained by watchfulness. Many give way when success is assured, and rapidly fall back into failure.” from As Man Thinketh I think this is an interesting example of how circumstances can influence your thoughts, and why you cannot simply assume that reaching a better state means you will remain there. When you are trying to become stronger, the thoughts and behaviours that create weakness are exactly the things you need to move away from. Discipline, humility, effort and the right thoughts are what allow you to rise in the first place. But once you succeed, the environment changes. Arrogance becomes easier when people respect you. Selfishness becomes easier when you have more power to act on it. Corruption becomes more consequential when you actually have something worth corrupting. The same characteristics that would have prevented you from becoming strong in the first place can therefore re-enter once strength has already been achieved. And I think that is what makes maintaining success so difficult. The state changes, and therefore the test changes with it. You cannot just turn on autopilot once you reached the top because the things pulling you towards weakness are now different, more accessible and often more rewarding in the short term. Success does not remove the need for discipline. It increases it if you want to stay there. Because the thoughts that got you there still need to be protected after the circumstances around you begin making their opposites easier to hold. Of course you can be arrogant, selfish or corrupt while you are weak. But when you are already on the floor, you can't go any lower. When you are at the top, you can either stay there or go back down. Humans often desire what they don't have and when you are at the top, the things you don't have are the most undesirable...yet they still look more appealing than before. NLB

x.com/i/article/2112…

This might be a canary in the coal mine... first signs of hyperscalers being able to get out of their compute agreements with neoclouds, something that becomes very important if we start to see an oversupply of compute capacity at some point. The site was expected to have 23k NVDA B300s, and the reason for cancelling is a lack of power supply pre-2030. As the other parts of data centres beyond the chips themselves are starting to show, one can see that not everything in the economy moves at AI speed... data centre projects will get delayed, which could actually help the AI trade by avoiding a "supply outpacing demand" scenario. NLB

Bloomberg @business

4 days ago

Microsoft has pulled out of a UK data center deal with British neocloud Nscale at a critical time for the startup’s initial public offering hopes, the Telegraph reported. bloomberg.com/news/articles/…

66 96 622 316K 119

I find the argument that rising GPU rental prices somehow invalidate concerns around neo-clouds vulnerable. It's like arguing a few months ago that memory stocks must be great investments just because memory prices are at their highs. Or that oil companies are great investments because Oil prices are high. Yes, demand is strong and supply is constrained. But what happens when new capacity comes online? That's precisely the question here. Today's GPU rental prices reflect a shortage of available compute. But the capacity neo-clouds are borrowing billions to build isn't necessarily available to rent at those prices today. By the time it comes online, the supply-demand balance could look very different. If demand doesn't outpace that additional supply (of which there is a lot coming, with BofA projecting 212 GW of incremental capacity through 2030), prices come down, and suddenly the returns on those investments look quite different. And this is where the massive debt that neoclouds have taken on becomes important. Even if AI demand eventually materialises, debt has a timeline that revenue cannot simply ignore. Construction delays, slower deployment or lower rental prices can leave companies needing additional financing before the cash flows arrive. We saw something very similar during the fibre buildout around 2000. Demand for bandwidth was real, and companies like Global Crossing had enormous backlogs. The internet was being built, and the underlying demand ultimately materialised. Yet that didn't stop Global Crossing and other heavily leveraged infrastructure builders from going bankrupt. They were financing enormous capacity expansions with borrowed money, expecting future revenue to arrive quickly enough to service that debt. And that's precisely the risk with CoreWeave. Its backlog may be substantial, but that backlog only becomes cash as data centres are built, connected to power and brought online. Meanwhile, interest payments are due regardless of whether construction is progressing according to plan. If those projects are delayed, the company may need to borrow even more just to bridge the gap, potentially at increasingly expensive rates. Creditors are becoming increasingly concerned about the financing risk, even while AI demand remains strong. I'm not arguing that AI demand isn't there. I'm arguing that pointing to today's GPU rental prices tells us very little about the returns neo-clouds will earn on capacity they haven't even built yet -- CSCO served a shortage and went from H1'01 rev growth of +24% to -29% H2'01...and ended up writing off $2.5bn in inventory. Things can turn quickly. NLB

𝐀𝐠𝐫𝐢𝐩𝐩𝐚 𝐈𝐧𝐯𝐞𝐬𝐭𝐦𝐞𝐧𝐭𝐬 @Agrippa_Inv

5 days ago

While neocloud investors got absolutely wrecked last week, GPU rental prices skyrocketed. B300 rates are now up >20% over the past month. Know what you own... H/T: @matthew_sigel $IREN $NBIS $CRWV

80 120 1K 108K 152

There's a lot of anticipation around $META's earnings this week. Muse's monetisation and costs and whether growing demand requires another capex step-up or whether existing capacity is enough and can now be put to work. But something else caught my attention: the timing of their earnings. Over the past four years, Meta has consistently reported Q3 earnings on the same day as, or one day after, Google, Microsoft and Amazon. This year, Meta reports on October 21, a full week ahead of both. Even more interesting, Meta only announced the date on October 12, just 9 days in advance, compared with 23–29 days in previous years (2022–25). Google and Microsoft had already announced their reporting dates on October 6 and 7. Why the sudden desire to go first? Some possible explanations: Set the tone: Get ahead of the other hyperscalers and shape the narrative around AI returns and capex. Confidence: Perhaps management likes the numbers and wants to get them out ahead of everyone else. Own the spotlight: Give investors time to digest the results without competing hyperscaler earnings. Avoid getting overshadowed: Let the stock react to Meta's own results, rather than being influenced by what Google or Microsoft report. Could just be scheduling, of course. But given the significance of this quarter, it's a curious break from a remarkably consistent historical pattern. NLB

Quality Equities @qualityequities

5 days ago

$META Meta reports earnings this Wednesday. 4 things that matter more than EPS: 1. Ad pricing and volume: Advertising is about 98% of Meta’s revenue. Last quarter, the number of ads shown rose 14% and the average price per ad rose 12%. If both keep growing, that is the clearest

3 2 52 5K 12

You can now add files and tabs to normal Google Search...which is how I always thought Gemini should be used by Google. Not as a separate product to sell to consumers, but as part of the Google ecosystem the company has worked so hard to build. I think this also helps ease some of the worries that Google Search is dead. Google recognised the threat and took the necessary steps to counteract the change. This gives Google a reason to keep pushing Gemini and improving the experience to the point where people stop using other LLMs for googling and just go back to the originator of search... If Google manages to integrate ads properly into AI-generated responses, it can still monetise the same search intent, just in a different format. The real threat isn't people no longer clicking on blue links, but them leaving Google's ecosystem altogether. In fact, if Gemini better understands what users are trying to accomplish, ads could potentially become even more relevant and valuable...though whether that translates into better economics remains to be seen. Yes, agent interactions can still bypass Google Search, but there are some things I would simply rather google myself instead of having to wait for an AI to answer and then go back and forth. As you know, I think people forget in this whole AI agent debate that humans are simple creatures of habit and that people -- despite many thinking otherwise -- still want to stay in control. I always think of it this way: questions are for LLMs, but if I already know what I'm searching for and where to find it or have something specific I need to do, Google Search often still dominates.

Something I've noticed in recent in the AI sell-off on Thursday (after Ant Rev numbers disappointed a bit) is how software has split into two groups when it comes to reacting to AI news. Names like SNOW and DDOG have integrated AI so successfully that it has become a major part of their growth story. But that also means they're increasingly tied to the AI trade itself. Concerns around AI demand, adoption and returns now hit these names as well. Meanwhile, software companies that have been slower to integrate AI have continued to be less exposed to these sell-offs. Their businesses might not be doing as well in the long-run if they can't figure out AI, but they're also not as dependent on AI continuing to deliver for their short term share price movement.

Shanu Mathew @ShanuMathew93

5 days ago

Databricks describes AI as the “fastest growing part of our business” in a posting for a Product Finance Director covering AI products and GPU investment strategy. In August, it reported a revenue run-rate above $7B, growing >80% YoY. Its AI products had already crossed $1.4B in

0 4 31 24K 17

I think this is also why the market doesn't seem to care all that much about some of the strong supplier earnings we're seeing. As you rightly point out, much of that spending has already been committed. But I think even those commitments are becoming less reassuring. If the companies making them start to realise that the returns aren't coming through as expected, commitments may not be worth as much as we think. Companies could try to get out of contracts, even if that means paying penalties. Especially with the neoclouds, where financing is already becoming an issue, a reduction in commitments as existing contracts roll off could ripple through the entire supply chain. I think that's what the market is now nervous about with the supply-side names. Not just whether the returns on AI infrastructure will materialise, but how much confidence we can thus really place in the spending commitments underpinning the supply chain.

“The universe does not favour the greedy, the dishonest, the vicious, although on the mere surface it may sometimes appear to do so; it helps the honest, the magnanimous, the virtuous.” I think this is a particularly interesting point because I feel like a lot of people will believe the exact opposite; that the universe does favour the greedy and dishonest. You see people who, in a relatively short amount of time, manage to get to places you think you want to be. They attain things you desire, most often money, and it may seem like their greed or dishonesty is being rewarded. But I think the important distinction here is that the state they have attained is not necessarily the state you actually want to be in. Money is great, of course. You would rather be rich than poor. But money without richness of mind is not going to solve your problems either. Life is not a sprint I think this also goes back to the idea that life should not be viewed as a sprint. In fact, it shouldn't just be viewed as a marathon either, but rather as a never-ending race. Those people may appear to be rewarded because they chose the quicker, easier path. But the important question is whether that path actually leads to the same destination. I'm a big believer that you don't need to take the hard way if there's an easier way leading to the same goal. The goal is what matters, not how difficult the journey is. Of course, great goals will naturally involve difficult paths, but there is no reason to deliberately make things harder if an easier route gets you to the same place. And I think that's the distinction here. The greedy and dishonest might have chosen a quicker path, but they're ultimately arriving at a different goal, or at least a different state of being. The same circumstances don't mean the same outcome This is where Allen's choice of words, "on the mere surface," becomes particularly interesting. On the surface, it looks like these people are being favoured. They get the money, the success, and the things others want. But underneath, they may simply be becoming victims of short-term thinking and immediate gratification. They aren't developing their minds. They continue to have bad thoughts and, consequently, continue to bear bad fruit. And even though their circumstances might look desirable to someone with a pure mind and good thoughts, those exact same circumstances can be much worse for someone whose thoughts are bad. Think of a gambling addict who suddenly comes into a lot of money. On the surface, that seems like a great thing. But for someone with a gambling addiction, the money might simply allow them to gamble more, become even more addicted, and ultimately make themselves more unhappy. Compare that to a poor gambler who doesn't have the financial means to keep feeding the addiction to the same extent. The additional wealth, which would be a positive thing for someone with a different mindset, can actually make the gambler's situation worse. And I think that's the broader point here: circumstances are not one-size-fits-all. Their value depends on the mind of the person experiencing them. The positions attained by the greedy and dishonest may look like exactly what you want, but those same positions do not necessarily represent the same outcome for them as they would for a righteous person. What looks like success from the outside might not be success at all when you consider the state of the person's mind. NLB

If you can look beyond the privacy concerns, I feel $AMZN is showing the correct way forward for AI agents. Messaging your agents on WhatsApp and Telegram is nice for sure, but if it can really be as easy as just pressing a button on the back of your phone and talking to it, or having it pick up your real-life activity on its own...that's where it gets really interesting. You suddenly have actual use cases that matter and that you experience in everyday life, rather than things you, in one way or another, have to tell your agents about, whether through them reading your notes, calendar or you messaging them. That's what makes AI agents effortless enough that using them actually becomes worth it. Will the wristband work? I doubt it...just too much competition for that area on your body. But attaching something to the back of your phone seems perfect. Again, big big questions around privacy here (esp. in Europe, as you can imagine), but I think this shows that the current ways we interact with AI are likely not the ways that will ultimately make AI work for consumers -- and probably not just the consumer but having this in the office, as it is shown in the launch video is such a good application of AI in enterprises. And Amazon is under-the-radar one of the best-equipped companies to bring AI to consumers, given its existing consumer ecosystem, Alexa, shopping infrastructure and distribution. Add the potential Decart acquisition to that, and it is interesting how Amazon seems to be putting more and more pieces together on the consumer AI side. Still doesn't change my concerns around the AI agent model behind it making mistakes and the very tight margin for error. But this definitely reduces the friction I feared and can probably expand the use cases as they just become more practical and easier to spot in real life -- so we might see something other than restaurant reservations or flight booking. NLB

bee @bee__computer

7 days ago

Introducing Bee Frontier and Bee Eclipse, two personal AI devices designed to remember the moments that matter most. One wears on your wrist. One attaches to your phone. Both keep you present. ↓

53 37 336 140K 158

x.com/i/article/2110…

Memory stocks were getting hit hard today, partly on the back of developments around China’s CXMT and the prospect of more supply coming into the market. And I think there’s an interesting read here for the broader AI trade. For all the talk about memory somehow being different this time, at the end of the day, it appears to still a cyclical business -- or at least have such qualities. We had a massive shortage, prices went through the roof, and everybody started building out capacity. Eventually, that capacity comes online/tech improves, supply catches up with demand/demand decreases due to better tech, prices fall, and all that investment starts working against you through negative operating leverage. So why should AI compute be fundamentally different. Look at the amount of capacity being planned by hyperscalers, neoclouds, Oracle and everyone else trying to get a piece of the AI market. Individually, everybody believes their investment makes sense because they expect demand to keep growing. But collectively, they could end up building far more capacity than the market actually needs. And importantly, demand doesn’t even have to stop growing for that to happen. If compute supply expands faster than demand, particularly when you factor in technological improvements that make each new generation of chips more efficient, you can end up with an oversupplied market despite continued growth in AI usage. That ultimately means pressure on compute pricing, utilisation and returns on invested capital. Just as with memory, a business can be growing while its economics deteriorate. What’s also interesting is that some of the bottlenecks currently slowing the data centre buildout might actually be protecting the market from this outcome. Memory shortages, power constraints and deployment delays are all limiting how quickly new capacity can come online. As those constraints ease, the buildout could accelerate, potentially bringing forward the very oversupply investors aren’t particularly worried about today. Now, I’m not saying we’re at that point yet. The broader AI trade is selling off a bit in these last few days, confidence is clearly weakening, and concerns around the growth of AI labs are becoming more prominent. But I would be surprised if this were already the big dot-com-style unwind. For that, I think we’d need to see actual evidence of capacity exceeding demand, growth slowing and the economics of these investments deteriorating. We aren’t really there yet. Still, memory is an interesting reminder of how these cycles eventually work. The question isn’t whether AI demand keeps growing. It’s whether demand can grow fast enough to absorb everything that’s being built to serve it. And when everyone is building capacity under the assumption that they’re going to be the winner, somebody eventually has to be wrong. NLB

And who is using the smarter models? Do they just work and make decisions what to work on by themselves such that every human gets replaced? Bc if the users of these smarter models are humans, then humans are the user of dumber models also, as without humans using smarter models, the dumber models wouldn’t be used

Okay so…since when is a guy worth $200bn and the owner and biggest promoter of the product any indication for consumer adoption I love zuck but i hope this doesn’t make people more bullish

Shay Boloor @StockSavvyShay

a week ago

Mark Zuckerberg says his AI agent took him from ordering Instacart “twice a year” to “twice a week” which is ~50x more often just by making shopping easier. Thats what makes agent commerce so interesting because removing effort of buying things could dramatically increase how

keyboard_arrow_left Previous keyboard_arrow_right Next

51 34 459 62K 39

Yes. It’s almost like a reverse of the early days of Facebook. Back then there was a customer-investor mismatch bc the investors weren’t on FB so they didn’t understand it -> didn’t invest. I think what we have now is that investors think this is what customers want/will use given the current version of it. An investor/product/expectations-customer mismatch -> they invest bc they and their friends want to use it, not the average person

Vu Tran @vu0tran

a week ago

Having worked on Muse, I believe widespread agentic AI adoption is going to go slower than we think. The people who build it, and the investors who fund it are all on the upper end of conscientiousness and openness to new experience. It's a problem that we often project this on

99 39 693 80K 321

x.com/i/article/2108…

This is pretty much what I've been trying to build with my own "second brain" The structure is relatively simple. I consolidate all my research, thoughts, trade observations and frameworks into one central "brain" - a structured Obsidian vault. Everything is classified so agents can distinguish between what I actually believe, what is external research, and what I believed six months ago but might no longer agree with. Then I give different Hermes agents access to that same brain, each with a specific job. For example, an agent checks new thoughts against existing research, looking for connections, contradictions or ideas that reinforce previous theses. Another monitors financial news and proactively messages me when something relevant comes up. You can extend exactly this structure to earnings. Give an agent the task of monitoring releases, retrieving your previous thesis, assumptions and relevant historical data, and identifying what has actually changed. Then have an agent cross-reference those developments with other companies, industries or ideas already sitting in the brain. The important part is that the agents aren't starting from zero every time. They have access to years of your actual thinking, with the context to interpret it, and can proactively bring things to your attention rather than waiting for you to ask. I wrote about it here, admittedly in a more fun way, but the applications are very real: x.com/NairoLB242/sta…

NairoLB @NairoLB242

3 weeks ago

A Day in the Life of My Second Brain 4:30 AM My first brain wakes up. My second one has been working the entire night. This is already an improvement on most junior analysts. The second brain is a few years of notes, ideas, research, trades, journal entries, frameworks,

1 0 3 821 5

The potential Amazon acquisition of Decart looks quite interesting, especially because it could benefit two very different parts of Amazon's business. On the retail side, Decart's real-time generative video technology opens up some interesting possibilities for how we shop online. Imagine having a virtual dressing room where you can try on clothes in real time, move around, switch between outfits, or see how different products look in your home. Now combine that with an AI shopping agent. Instead of browsing through hundreds of products yourself, you could simply tell the agent what you're looking for, have it find suitable options, and send them straight into your virtual room for you to try out. Clothes, furniture, accessories, you name it. You could even ask it to suggest different combinations or alternatives and see them immediately -- essentially you are going shopping with your personalised stylist without leaving your house (which obv. kinda takes the fun out of it but adds ton of convenience imo) For Amazon, which already has the customer base, product catalogue and shopping infrastructure, this seems like a very natural addition. The other angle is AWS. Decart isn't just a real-time video company. It also develops technology to improve AI compute efficiency, something that becomes increasingly valuable when hyperscalers are constrained by power and available compute capacity. We've been seeing this race towards efficiency accelerate across the hyperscalers. Microsoft's latest announcement around its own silicon, reportedly offering c.40% better power efficiency, is another example of how important getting more compute out of the same power budget has become. If Amazon can integrate Decart's optimisation technology into AWS and potentially improve the economics of Trainium, that alone could make the acquisition strategically interesting. So you have two quite different applications of the same acquisition. On one hand, a new layer for Amazon's core retail business that could change how customers discover and try out products. On the other, technology that could help AWS become more efficient at a time when every improvement in compute economics matters. Pretty interesting combination, and one where both sides fit naturally into businesses Amazon already operates. NLB

Shay Boloor @StockSavvyShay

2 weeks ago

$AMZN is reportedly paying ~$7B for Decart after the AI startup pushed real time video from 26 fps on GPUs to ~101 fps on Trainium. Amazon is effectively buying software expertise to improve Trainium utilization and make its custom silicon more attractive as AI workloads expand

64 56 633 76K 35