El desplome del momentum: por qué Nvidia cae, Google sube con fuerza y los VLM lo cambian todo

[00:00] All right, gonna try and squeeze a quick one out in the day after Thanksgiving. Uh you guys can read the list. There's a lot of stuff that actually happened this week. So, uh S&P finishes up close to 4% as we get close to the all-time highs. Uh NDX up almost 5% for the week. Biggest weekly move since May. Uh MAG 7 kind of led the way up 5.4% despite Nvidia being down, taking us back to the high. uh best weekly uh performance
[00:33] since August and it was again led by small caps or at least this time by small caps which put in their biggest week uh since last year in November and if you want to take it back further I mean it was the third biggest week in almost two years. Uh so again small caps there's a story going on here factor- wise this is for the month momentum was crushed uh it was down 12.2 2 all the retail pain that I've been talking about, momentum was a driving factor and it was the long side of MO which was
[01:04] down 10%. Uh, and if you look at it relative to IWM, so if you try to take the beta out and you look at it down 11.4%, the worst weekly performance in terms of the long side of MO relative to small caps since the great financial crisis lows. >> [snorts] >> uh S&P divergence with Momo. Uh this is just before the close uh today. Uh almost at the all-time highs in the S&P MO broke down.
[01:35] This is takes us all the way back to 2023. So um S&P breaking away from MO. I think this is a warning sign of of more to come, but I'll go through that. Pure momentum did not have as big a move. The difference between these two, this is the Bloomberg pure momentum factor which strips out uh the sector biases in momentum. Uh this basically will just show you that if you think about this is kind of the am a uh AI thematic trade unwinding pure
[02:05] momentum did not unwind as much. Uh the importance of the long mo side uh a long side of MO is that it's very correlated to retail performance. If you look at here to here, basically you have a direct overlay with the white line here, which is Ethereum. Uh retail definitely uh bailed out of stuff during the past month. Uh they will jump back into whatever the new winners are. Size factor again. I thought and talked about this and said this was starting to roll over. MACD divergences. Well, we have
[02:36] size breaking down again. We'll see if this one goes. I think it will. size has been the dominant factor since the chat pt launch and since the Fed basically stopped rate hikes uh or pivoted and the reason this important is because this is also the time period that PMIs have been below 50 factors are very important to PMI shifts below here we have the PMI new orders component inverted uh so down here you're getting into the 60s and these lines show when MO has had a break
[03:07] you've seen new orders at each of these points when it's headed to 60 before. Uh we only got to 55 here at the beginning of the year and then we reversed up. I believe as I've said we're going to get into the 60s in the new orders category. We've got MO that looks like it's moved over. But it's not just MO the size factor very correlated. So if size all of a sudden starts to go this is inverted. This is a sector specific size not the pure size factor that I uh pure size factor which I just
[03:37] showed. Uh growth verse value also very very driven by what happens in PMI cycles. This is why PMI is so important because you get factor shifts. It means the winners next year are going to be different than the winners this year. I believe that's going to be the case because I think we're going to see a broadening out as PMIs work their way higher. The earnings the global earnings revision ratio sits near a 4-year high. upgrades continued out number downgrades I picked the points where we were above one uh that's where we are right now 1.01 01. I picked all those points where
[04:07] we broke out above after being below. And again, this happened when PMIs were up in by approaching 60 in every case. We'll see again if we get if the earnings revisions are showing us. We also have an important thing. Ryan Dietrich highlighted this. Uh the 493 finished their earnings at 4.7% year-over-year versus 5.9. This is the Mag 7 at 18.4. This has been uh major differences between the two. So, here's the expectations for MAG7. They beat by
[04:38] 3.7. Expectations for the 493. You can see basically where we're going. You're getting a uh a broadening out right now. And I think this is going to continue especially as the Fed cuts rates. Another sign that PMIs are heading higher. This is the Goldman Sachs current act activity indicator. Uh they break it down by different components. The red bar or the red component here is the manufacturing which would be the closest thing to the PMI. We are currently at the highest level uh going back over the last two years.
[05:08] Dallas Fed came out their outlook for uh shipments uh just made new highs going back to early 2022 verse PMI. Then you've got the capital goods new orders non-defense X aircraft. Uh again I've showed this before. It's been tracking higher. It continues to move suggesting PMIs are going to go higher. And then you have a similar type um this is probably the most important one. This is the IP diffusion for 6 months. So this measures within the IP
[05:39] uh the categories that are now positive uh moving higher and what you have is we're up near 60% on this. Historically when you've been at these levels you've had PMI levels which are above 55. So uh we'll see if this carries out into next year. If it does all those factor shifts are going to happen. Uh, and remember we are getting Fed rate cuts as this goes on. And this week we did see the move back up to about 80% for that. Jensen Yuang came out basically said something which I'm trying to get across to
[06:10] everyone here. You're insane if you don't use AI to do literally everything. The reason this is important again as I go through the AI component of this week holiday week uh webinar if you haven't heard about Genesis mission which was announced uh an an executive order signed by Trump this week. The plan is part of the Trump administration's aggressive low regulation strategy to boost big tech's race to stay ahead of China on artificial intelligence and cement US dominance in the fast expanding field. This is a major
[06:41] announcement. Uh I didn't see a lot of press on it as the week went on. There was a little bit more in case you missed it. I'm not going to go through and read all of these. I'm buzzing through this. Uh but the importance of this is amplified by the involvement of the DOE's 17 national laboratories. These labs house some of the world's most powerful supercomputers. Opening these assets to private sector AI development creates a computing backbone that no longer no single company could replicate. Basically, what they're doing is a Manhattan style project for AIdriven scientific discovery.
[07:12] This is important in the fact that it's taking us on the next stage of AI and the government is basically backing it. So, I just want to remember the tape is incredibly strong. If you're bearish, you're fading the tape. The Fed is cutting rates. Don't fight the Fed. Don't fight the tape. But when you add in that the government is going to make sure that this is a national importance to the degree that this executive order does which is to try and get around states from doing things. It is basically putting in a too big to
[07:42] fail for the frontier models. This is a critical component of of the public sector and the private sector coming together. The focus is to is in domains called out biotech fusion energy and quantum areas that could rewrite the rules of life and energy of cracked. I've talked about this relentlessly that over the next 5 years you are going to see dramatic changes as we go through the exponential point of AI not the LLM part not the text part we're going to get into the physical the visual all of
[08:12] that's coming and now we're combining and this is a DOE which controls many of the world's largest supercomputers and is central to energy policy critical because AI progress is power hungry this is to make sure that all of this work that will be done gets done Now, I've also mentioned that Eric Schmidt is the key person to listen to on AI. So, when you hear people say, and I'm not going to go through the call out again, but when people say AI is overblown or they say it's a bubble, they better have some knowledge of what's going on. If they're just saying
[08:42] this because it reminds them of the dot bubble, you're making a huge mistake in terms of pretending uh that this is uh going to be something that you can just call a bubble and actually go through it that way. So, here's what we got. um he wrote a book with Henry Kissinger titled Genesis, artificial intelligence, hope and human spirit. It was all about AI. It was about the biology and physical that the human mind physically cannot grasp on its own. So he's going through the second creation, a way to discover
[09:13] all types of things within reality and life. It's basically decoding the source code of nature to solve energy and material science problems. AI is not a bubble. AI is not overblown. The AI revolution is underhyped. If you don't want to keep swimming upstream or swimming against the stream when it comes to artificial intelligence, every single decision being made by governments, by monetary policy, and by the stock alpha
[09:43] is being driven by what's happening in artificial intelligence. This is going to be a theme going forward, especially as as I call it, the intelligence will effectively be what QE did. If you fought QE, you suffered. If you're going to fight AI, you're going to suffer as well. Um, Eric Schmidt is the shadow architect of this whole thing. He has effectively acted as the bridge between Silicon Valley and national security state for decades. Henry Kissinger's contribution to the book was geopolitical realism and the fact that we need to dominate this or China's
[10:13] going to go ahead. So regardless of what you think this is a necessity for biology, chemistry and physics to spot pardons uh patterns humans m miss. If you are interested in reading more about this the special competitive studies project with Eric Schmidt set up which is uh also something that Henry Kissinger had set up in years past a similar type approach to make sure that the government had a think tank uh that was helping them on the world's most important not only issues with inside the country but also geopolitically. uh
[10:44] does this relationship between the government and leading LLM companies put them in a too important to fail state? According to uh Gemini, the genesis mission effectively formalized the transition of leading AI companies from consumer consumer tech startups to national champions Lheed Martin, JP Morgan. Uh by weaving private A models into the fabric of national energy, defense and scientific infrastructure, the government is creating a mutual dependency that makes these companies too important to fail. That's why this is so critical. If you're going to fade the bubble, you're fading the
[11:14] government. The government is going to make sure at least for the next three years that this happens. But it's really a too big to fail moment. You can go through this is just continuing on it. If you have the view that like critics do that big tech capture the company will take over the government. It's kind of happening. It's like believing that your privacy is safe. There is none. Um this is all one thing happening. Uh yes, once the government declares AI is the foundation of US scientific and technological leadership, it becomes impossible for the US to let its frontier model LLM models fail. Is this
[11:44] order a way to stop states from being able to regulate AI? Technically, no. But in practice, yes. Again, this is going to make it very difficult because to fight it as a CER or included in the National Defense Authorization Act, regardless again of your viewpoints, this is trying to make kill switches and state laws at least get them to the point where they have to go through the courts. Uh this makes it very challenging. This was one of the risks that I highlighted in prior videos that I did believe the government at the state level would make it very
[12:15] difficult, particularly what's happening in Colorado and what's happening in California. This is allowing it to kind of go. The recent Genesis mission announcement from the government for advancements in science through the AI collaboration between the private models and the government data. Will this take advancements in the models to VLM from LLM? So this is why I want to introduce for the first time visual language models as opposed to large language models. This is going from text LLM. So the training that you've seen, we're now getting into what Alon Musk has talked about, what I've gone through, which is
[12:45] the ability to train models based on what they're seeing and the language. So visual language model that leads also into visual language action. Vision language action is getting you into edge devices and the real-time movements of robotics. So we have to do VLM models to allow VLA to start. So think of VLMs similar to LLMs, cloud-based sending stuff down, but the VLAS are the brain inside the machines now doing to be able to do things on their own. These are the two phases. These are far more compute
[13:16] heavy. That's why as we accelerate through there'll be different winners and losers and it will be very hardwarebased as we go through it. Uh I'm not going to read all of these but these are the things that are necessary for it to visualize to be able to make things. You need VLMs, vision language models, increasingly vision language action models for robotics. This is why Genesis practically guarantees a pivot towards multimodality because the problems being funded are multimodal problems. This is a critical movement and that's why I will be doing for people who want it uh either from my
[13:47] directly from me from a consulting basis or through 22V. This is what we're moving into. That's what we're going into next. All of this stuff is where I believe that the companies that had been winning, which were part of the chat GPT three-year movement, all on LLMs, we're now moving into this. There will be different winners in this. This is one of the reasons why you're starting to see Nvidia, uh, Coreweave, OpenAI, all of this stuff in there. As we move into this side, I'll go through why Gemini 3 was a critical moment, why TPUs were a critical moment. All that stuff fits
[14:17] into these three categories. If you want more, call the sales team at22V uh and we can talk about how we can work on that together. AI drug discovery. I sent this out this week. Again, biology is part of this. Helping people live, yes, forever, but at least expending lifespan for the next 20 to 50 years is going to be a goal of this. Drug discovery is just one part. Uh they are a critical compartment on it. Again, accelerating drug discovery needs VLM. Robotic lo uh lab automation, experimentation is VLA. You're going to
[14:48] have all of this stuff going on. The broader implication, AI is collapsing the cost of medicine. This is important for the inflationary component. Obviously, this is important for uh Medicare, Medicaid, the entitlements, the debt, the deficit. All of this stuff matters, which is why again the government is heavily getting involved. We entering a period where AI and biotech begins to reverse aging, cure disease, and collapse healthcare costs. The next decade will see the confluence of AI gene therapy, robotics, and diagnostics accelerating at least as fast as AI itself. Think about how fast
[15:19] it's already gone. Anthropic is now actively hiring life science researchers. Dario Moda has said repeatedly that he believes most of disease biology and medicine can be solved by the end of this decade. Um, again, this was in a moonshot episode this week. David Sinclair, who I've referenced his book in both writings and I think earlier things this year. I'm a big follower of his work on reversing uh aging. He's made some advancements worth looking up the details that I have on
[15:49] that slide. Uh and again, the market has spoken. So I've highlighted this the last couple weeks. Pharma relative to the S&P 1500 had its best month in the last 30 years. XBI trending higher. This has huge implications for small cap. Beautiful looking chart, but you can also see how it's moving with pharma. So, it's not just large cap pharma. Biotech's going as well. Here is the relationship between IWM or size relative with XBI over the NDX. So, XBI is outperforming
[16:20] the NDX. I believe that's going to continue. Size factor has moved this direction. This one is just IWM versus S versus SPY. Uh like I said, the pure size factor is doing better. Now the transition to VLMs and VLAS highly significant for advancements in longevity and research medical imaging and diagnostics in particular. Uh so VLMs have a huge part of it. If you I'm going to go through another couple. There were two moonshot podcasts. Definitely watch those this week. I'm
[16:51] going to highlight another one here. But the artificial intelligence show went through Gemini 3 Nano Banana. This is really what starts kind of the Gemini importance that went on this week. Gemini 3 demonstrates major upgrades across math, multimodal reasoning, and especially visual understanding, signaling that Google is now pushing the frontier in tasks that blend language, perception, and logic. Paul Ritzer tested Gemini 3 by feeding it a chaotic whiteboard photo from an October strategy meeting. This level of reasoning and vision coherence is not
[17:22] incremental. It is a shift. Gemini is quickly approaching OpenAI in terms of the numbers of users. Sam Alman privately acknowledged OpenAI staff that Google has temporarily pulled ahead. This is a big deal. Um they cover NanoBano Pro. The visuals that you saw with the uh uh VLM to VLM that was made with Nano Banana Pro. Everything I'm doing in terms of the visuals now is using Nano Banana Pro. Uh it's basically unbelievable in terms of what it can do with just words. So if you get the
[17:52] chance, play around with it. Um why the whiteboard example matters. The whiteboard test of Gemini 3 is far more important than a cool demo. It is proof point that vision language models are beginning to outperform traditional LLMs on real worldwal side take a photo of something particularly something that is on a whiteboard that's maybe unreadable in language and just watch the way it goes
[18:22] through and it connects it back to actual things. Um, VLMs are the bridge to real autonomy. This is really again the critical part. This is what we're entering next year. Autonomy becomes the story. Autonomy is hardware. The implications for compute. If the VLM race has become compute requirements explode five times to 20 times, you're going to hear even bigger numbers on this. It doesn't really matter. We are going to be training on images, video, spatial data sets, multimodal embeddings, multi, not just words anymore. We are moving on to
[18:53] visual plus words again. The computebound electrons copper latency limits. This is it. This is where the bottlenecks are. The VLMs the bandwidth explosion occurs. So anyone that doesn't believe semiconductors are going to continue that micron's not going to continue. VA's will just continue it going forward. You'll get into real-time video like I said. So this is training on videos that exist. This is real time video capturing and going through it. Think Tesla, think Whimo, think whatever you want, but you're going to be learning on videos. Uh, I
[19:24] want to remind people QE for the mind, how artificial intelligence is flooding the economy. If you fade AI, you are fading QE. It is not a good bet. The moonshots episode, this one in particular, Nvidia's record revenue, Elon's data centers in space, Gemini 3, insane performance. They go through the same thing. Talk about the visual reasoning breakthroughs. AI is beginning to understand the physical and visual world not just text. Um this led to the conversation of this this was in substack the chip made for
[19:54] the AI inference or the Google Google TPU. The systolic array is the key differentiator. So this is now going through the Nvidia chip and the TPU for Google. You saw the meta announcement it Google has been flying higher. Nvidia has been going down. There's clearly a ch a shift that's going on. Part of the reason is to understand that the chip moves data back and forth between the memory and the computing units for every calculation. The constant shuffling creates a bottleneck, the voyman bottleneck. In a TPU, data flows through the chip like blood through a heart,
[20:24] hence systolic. I'm bringing this up again because as you start to see these things in terms of a massive architectural leap for the TPU, this doesn't kill Nvidia in any way, shape, or form. So, don't get into this bearish thing. But if you want an explanation as to why Google's doing well and Nvidia is not, this is starting to go through the stuff. This is going through what is going to be needed. Memory becomes a major issue. But this is about the flowing between the chips, the TPU. So Nvidia's monopoly is finally getting real competition. This TPUs, but then
[20:56] you're getting all of these pieces. This is why Nvidia doesn't trade at five times the level it is out a few years in terms of its earnings per share because as I've mentioned the expected revenue out 5 years is about $420 billion yet we're expecting $5 trillion in capex and if there's 5 trillion in capex and Nvidia gets their normal cut 40% say that would be $2 trillion they're not getting that so it's already discounted that they're going to be losing their monopoly this is just making sure that
[21:27] people realize it is happening. It still doesn't mean that they won't be seeing the revenue grow. And remember, the revenue was just up 62% year-over-year and at a faster pace quarter over quarter and they have the blackwell really going out. So compute supply growth is accelerating faster. The model architecture, but user demand and new modalities are growing even faster. This is the whole point is that even with the bottle, even with what's going on, the more chips, this this is not going to take it down. Um optics is the next memory if you want more details on that.
[21:58] reach out to 22V. Um, here is this was before uh Friday's data. Uh, I just wanted to highlight that Google's up 70% year to date. I think the critical thing Nvidia and and Alphabet are up the most, but below there, none of the other ones are up as much as the Mag 7, and none of the other ones are up as much as the [snorts] S&P. So, it's actually been an underperformance year already this year for the MAG 7. I just bring that up because I think it's going to continue in the years ahead where you're going to
[22:29] have maybe a winner or two each year like Tesla next year, Google next year part, but you're going to have losers too. They're going to start uh eating each other. Open AAI versus Google. Why Samman fears ChachiBT might be losing the race. This is becoming a bigger story. And here we have the chart. So everyone who believed that this was the beginning of an AI bubble. You have the open AI complex which includes SoftBank, Oracle, uh AMD, Microsoft, Coreweave, they all got hammered this week, but the Google side, they don't even include Eli Liy went up dramatically. So again,
[23:01] you're just getting a shift. This has gone on continuously. But we have entered the stage where computer scaling faster the models can consume it and hyperscalers turning to the physical world in the AI supply chain. Nvidia is not the key anymore. or the AI industrial revolution building its infrastructure. It's in full mode with this. If you need a way to understand this, I put this together an analogy. You've got the highway, the gas station, and the car. Take all of these different components and break it down. The highways are the data centers. The cars are the AI models. Here are the models.
[23:31] The gas stations are the things that are necessary to keep it going. We have so many users and we don't have enough gas stations built right now. So, all of the models are there. the data centers are trying to be built. We don't have the ability to provide the inference necessary for everyone that wants these. And this is why we still have bottlenecks. This is why there will still be revenue growing continuously. And as people be able to use it more, you're going to see more profit margins spread out. Uh, so you can read through this on your own, but the next five years shift from more highways than ever
[24:02] to gas stations finally catch up, ending in com computational abundance around 2028, 2029, just as VLMs, VAS, and robotics hit scale. So, we have 3 years of a fairly continuous thing of the compute and power needs and then we're going to start to get where we get some kind of a a cliff that comes in where we have enough of this stuff. Uh, important just to make sure you have this in your mind. And again, they go through the Genesis mission on this one, but they also go through Claude 4.5. The reason this one is important, they call it the
[24:33] onset of recursive self-improvement. I've talked about how important this is, but again, this lines up again with the VLMs and the VLAS, it just means computers are learning on their own, uh, which means we accelerate even faster and faster. If sustained uh timelines are collapsed for robotic automation. So recursive self-improvement is critical. Um implication 90 to 100% of software development becomes automated. By next year most software apps, workflows and systems will be trivally generated often with a one-s sentence
[25:03] description. I've talked about software being the short side. It remains the short side. Yes, you can get some bounces maybe even next year, but I do not see them outperforming semis. expect the semis to broaden out. If you've got a cap weighted one where Nvidia is the the biggest one, maybe they suffer for the first six months of next year as people continue to move them lower in fears around TPUs and other things. But the reality is semis there's so many smaller ones that are necessary for the vision side. Think about the hardware side. So, uh just kind of going through
[25:34] it. Claude is not a VLM. It's a multimodal VLM that can do things, but it is not taking in there. I'm only bringing that up because this is not the moment where the VLMs accelerate via Claude. This is really more on the software side, but we're getting closer and closer every single day to that. Okay? And on the Claude code side, you go through it. Uh I'm just bringing this up because I've mentioned this on a couple podcasts, but this is the first actual news story I saw. Uh in Indiana, Northern Indiana and affiliate expect to
[26:04] spend about $7 billion on 2.6 six gigawatts of gas to be paid by for by Amazon. They will be paying not the customers in terms of electricity costs. Another thing I'm doing for clients is basically the power side where all of these red side are are the bottlenecks that are in there. Bottlenecks mean you can both make money and lose money. Uh it's the place where the PMIs are going to spread out. So if you want to call you call 22V on that as well if you want to get that uh finish up here last few
[26:36] slides. This is a great uh Substack done by Michael Green. Uh basically he's talking about how a broken benchmark which would be the poverty line uh and he's going through this. I think this is important just to read about because what I did is say okay he writes about it the poverty line which is quoted as 31,000. He's arguing it's really 140,000. Uh he makes a very compelling argument and goes through a lot of things including housing inflation, child care costs, health care burdens, dual income necessity. Um what
[27:08] I said is poverty trap meets the AI disruption wave. It's the way I look at everything. It's the reason I care about Bitcoin is because this is only going to get worse. AI is about to break the gradient. The government will be forced and rates will continue to need to be moved down despite what everyone says about the inflation side. We have we have a country where the real threshold equals 140,000 for a family four due to and you can go read all the different components. This is what is this is how AI makes all of it worse and how AI
[27:39] eventually could make it better. Finally, Bitcoin verse PMI. I'm just going to show this because as I've said, I believe PMIs are coming up. Here is the chart of Bitcoin both on the upside and downside. As PMIs go up, as they peak, as they go down, there has been a relationship even over the last two years on all these false hopes breaking above 50 or going up and then coming back down. If we are going to go up, I think Bitcoin will benefit significantly. And the final chart just to show you another factor to be watching. This is the pure trade in the
[28:11] US. Pure trade means trading activity. Uh the white line has bounced up significantly which means the most heavily ones that are seeing trading activity, Bitcoin is there. I hope everyone had a great Thanksgiving. Remember to subscribe. It helps me out in terms of continuing to be able to do this. Uh, I appreciate everyone and I hope you had a great Thanksgiving and I will see you next