Why He Sold Everything - Bubble Boi on Kimi, CXMT, and China's Chips
- Bubble Boi moved from a year-defining AI long book to roughly −15% SPX delta because Kimi weakened the memory-and-capex thesis behind his winners. His July 17 “I sold everything” post was not a permanent bear call: after gains of hundreds or thousands of percent, institutional managers had every incentive to protect bonuses, while he believed “you have to panic in time” when new facts reduce conviction.
- Kimi’s linear-attention architecture could materially reduce incremental DRAM demand by keeping most model memory fixed as context grows. Bubble Boi stressed that critics estimate only roughly 75% is fixed and, after further research, no longer sees the outcome as clear-cut; still, if one rack eventually serves 100,000 users instead of 1,000, another $500,000 rack becomes harder to justify. The paradox is that higher utilization may actually strengthen demand for premium Nvidia GPUs.
- Nvidia remains dominant, but its high margins create an irresistible opening for hyperscalers to vertically integrate through internal accelerators and interconnects. Bubble Boi expects much future compute to shift toward TPU, Trainium and other proprietary systems, though a growing market means Nvidia need not disappear: “your profit is someone else’s opportunity.”
- The sharpest long idea was CXMT at a sufficiently low IPO valuation because ordinary DRAM—not merely HBM—is capturing the AI memory squeeze. Bubble Boi argued that CXMT has no HBM product but can sell ordinary memory at premium prices to Chinese customers; he viewed 130 as extremely cheap. The exchange also mentioned a roughly 400 valuation or price, which Bubble Boi said did not seem as cheap, though the transcript leaves the exact asset and units ambiguous.
- He remains highly bullish on Intel because he believes its stated $20 billion capital budget is sufficient and 14A has a real customer commitment. Customer prepayments, 18A production and management’s refusal to add capacity without secured demand underpin a scenario that produces “ridiculous numbers,” potentially around 100% annual revenue growth once capacity converts into shipments.
- Bubble Boi expects markets to keep punishing hyperscaler capex until monetization becomes visible, but he does not think cheap Chinese models alone destroy the economics. Google can deploy AI throughout its existing products and, in his estimate, model-serving margins can exceed 90%; the true threat would be Chinese systems creating capabilities—such as drug discovery—that US frontier models cannot match. “The coding is already over.”
- His next watchlist shifts from crowded AI bottlenecks toward neoclouds, flash-based inference and a possible Qualcomm surprise. He is paying hundreds of thousands per month through Nebius to run and test Kimi himself, sees Phison-style flash controllers as a route to “DRAM abandonment,” and thinks Qualcomm’s extraordinary accelerator-throughput claims might be real despite calling the company terribly managed.
1. Product knowledge—not valuation screens—drove his best trading year
Bubble Boi called this the best trading year of his life: all his prior years in trading or finance combined did not equal what he earned this year. AI supplied the opportunity, but his technical experience let him distinguish “what was nonsense and what wasn’t” by studying products and their prospective users rather than sorting companies into statistically cheap and expensive buckets.
That technical edge remains most valuable when it creates disagreement with the market. Intel reports can take days for investors to decode because people less immersed in technical details assign different weights to the facts; the opportunity exists between the arrival of information and the market’s understanding of its engineering consequences.
The discipline behind his concentrated approach is revisability, not infallibility: “I am very wrong.” When he has researched a thesis deeply, he is willing to invest with confidence; when something genuinely new appears—as Kimi did—he reduces risk quickly instead of defending the old work.
2. Nvidia’s margins invite substitution while commodity DRAM drives memory profits
The host framed Nvidia’s margins as roughly 75%. Bubble Boi’s response was that current dominance is obvious, but high profits attract alternatives. Hyperscalers have powerful reasons to own ASICs and interconnects rather than depend indefinitely on one supplier; “your profit is someone else’s opportunity,” even if total AI demand grows enough for Nvidia to remain important.
He separated HBM excitement from the underlying DRAM squeeze. HBM consumes standard DRAM wafers that are cut and stacked, removing supply from a market that still serves PCs and other systems; consequently, ordinary DRAM prices have risen even faster while HBM pricing has been comparatively stable.
Micron illustrates the trade better than the technology story, in his view: he called its HBM the lowest-quality offering among Samsung, SK hynix and Micron, yet commodity scarcity still lets it earn exceptional profits. A less differentiated product can briefly produce profits above Nvidia’s, but “you can’t just rip people off like that forever.”
That setup makes CXMT compelling at the right price. It has no HBM product, but Chinese demand for ordinary DRAM is sufficient to support premium pricing; Bubble Boi viewed 130 as exceptionally cheap. The later discussion of roughly 400 was not clearly identified by asset or units: Bubble Boi said that figure did not seem so cheap, while the host said it could still grow to that point.
3. Chinese semiconductors may escape the policy trap that crushed other champions
The host raised the old rule “never trade Chinese companies.” Bubble Boi attributed it partly to government intervention: support can build an industry, then disappear when companies run into trouble, leaving participants in fierce competition. He cited electric vehicles as an example in which broad cheap capacity mattered more than creating one maximally valuable corporate winner.
Semiconductors may be different because globally competitive capacity requires too much capital for endless domestic fragmentation. His conclusion was explicitly selective rather than universally bullish: cooperation and continued state support make more sense here, which is why he personally owns some Chinese companies.
Alibaba is his broad Chinese-AI infrastructure expression despite losing money on previous BABA trades. He sees an inexpensive e-commerce franchise, a high-quality cloud available beyond the US, exposure to major Chinese AI labs outside perhaps DeepSeek, and prospective Huawei-chip arrangements; if Chinese AI usage expands, “this money will pass through Alibaba’s servers.”
4. Kimi changed the memory thesis, but the bearish conclusion remains conditional
Had Kimi merely matched frontier models using familiar infrastructure, Bubble Boi would only have trimmed risk. What changed the thesis was linear attention: earlier Qwen implementations had not looked competitive, whereas Kimi reportedly achieved “world-class results” while eliminating most of the growth in memory consumption as conversations lengthen.
The caveat matters: critics told him only about 75% of the model is fixed while the remainder still scales. After researching the trade-offs more deeply, he retreated from the cleanest bearish version—“I don’t think the situation is so clear-cut anymore”—but expects Chinese and eventually American labs to accept some version of the compromise.
The host’s car analogy captured the capital-cycle risk: if an innovation makes an existing car ten times faster, the owner benefits enormously but may not need another car. Bubble Boi agreed that installed servers may serve far more users profitably, while future racks become harder to justify; another innovation could move capacity from 1,000 users per rack to 100,000.
The Nvidia effect could run the other way. Inference often leaves expensive GPUs idle, whereas Kimi’s method can keep them busier, making high-end hardware more economical; Bubble Boi therefore distinguished a potentially negative DRAM-demand shock from a potentially positive utilization story for Nvidia.
5. Hyperscaler capex will be punished before its economics are proven
The host invoked the pyramids as an example of investment that takes time to repay. Bubble Boi compared AI infrastructure to Columbus’s investment and a laundromat: durable capacity must be funded before it earns. Once data centers fill, he calculates that model serving can carry “over 90% margins sometimes.”
Thread Guy expected markets to keep punishing hyperscalers while capex forecasts rise, with conditions potentially improving next year. Bubble Boi explained why Microsoft is more exposed: investors cannot identify the product beyond Copilot or understand why labs that build and diversify their own infrastructure should rent its capacity; Google can embed AI across Search and other products, giving its spending a clearer internal use.
The host’s pushback—worth keeping—is that cheap Chinese open models may drive token prices toward zero before Google recoups its investment. Bubble Boi’s answer was capability-based: commoditized coding is not the end market; value shifts toward medicine, law, aviation, movies and scientific discovery. If Chinese models make discoveries US systems cannot match, “sell everything” and “run away.”
Bubble Boi does not think the Anthropic and OpenAI IPOs are already doomed, though a prolonged bear market could change that. Elon benefited from moving first because “speed is very important for getting money.” Thread Guy suggested the popular hedge of buying Anthropic and selling OpenAI; Bubble Boi noted that this removes much of the nuance.
6. Intel’s 14A customer matters more than the market’s capex fears
Bubble Boi called Intel’s earnings excellent despite the sell-off. Investors suspect spending will exceed guidance, but management insists it needs only $20 billion; his own calculations say that is more than enough, especially because customers are making prepayments to secure future capacity.
The node progression anchors his confidence: 18A is already in production, while Intel is going “all in” on 14A because, he believes, a real customer exists. Management repeatedly said it would not expand capex without guaranteed demand, and he takes that seriously because fabrication equipment requires a continuous order stream rather than API-style utilization.
Translating $20 billion into tools, wafers and selling prices produces what he called “ridiculous numbers,” including an illustrative path toward roughly 100% revenue growth in a year. His preferred entry horizon is the next six to twelve months, particularly if capex rumors offer a cheaper price before reported results make the thesis obvious.
He rejected SemiAnalysis’s Intel 18A–SMIC comparison as selective and “disingenuous” because transistor density is one metric, not total node performance; he also questioned why it chose Intel rather than Samsung. 18A’s advantage is greater energy efficiency. China catching up is a question of when, not if—“the laws of physics are the same”—but he wants equivalent comparisons across Samsung, 18P and 14A before drawing conclusions.
7. Neoclouds, open models and flash form the next set of asymmetric bets
Thread Guy described an uncertain report that Shaw had a roughly 20% investment, perhaps around $1.5 billion, in an Australian neocloud. Bubble Boi defended Leopold Aschenbrenner’s position without disclosing his friend’s private thesis. He bought a neocloud early in his own reallocation and expects acceptable near-term results; the harder long-term question is whether these businesses become “real clouds, like real infrastructure” rather than one-cycle GPU landlords.
His Nebius usage makes that demand concrete: self-hosting or launching Kimi costs roughly hundreds of thousands per month, depending on contract. He wants direct evidence rather than anyone’s assurance—testing behavior, benchmarks, modifications and optimizations—and expects a race to operationalize Kimi once full weights are released because “it’s not that easy.”
On Chinese open-source restrictions, he sided with competition while acknowledging Jensen’s incentive to spread compute ownership. A government could sensibly bar sensitive agencies from using a foreign API, but firms would simply download Kimi and offer secured domestic hosting; a broad ban would not preserve an AI monopoly or duopoly.
The next hardware substitution is flash. Phison’s controllers can reduce DRAM dependence, and flash offers terabytes of density; the mistake is treating headline latency as decisive when AI access patterns determine what must be fast and when. His trigger is explicit: “When you start to see this happening—sell DRAM. We’re not there yet.”
Photonics already looks “blown away,” though he suspects a less crowded implementation may emerge without today’s expensive lasers and advanced TSMC processes. Qualcomm is the other wildcard: claimed accelerator throughput looks implausible to many investors, but it “might actually” be real; despite alienating customers and losing a lawsuit to Apple, the company is “like cockroaches—they never die.”
Full transcript
And boom, Mr. Bubbles, the guy with the bubbles, part two. How are you, dude?
Yo, how are you? I'm fine, friend. Nice to see you again.
Look, I want to start very quickly. First of all, thank you for stopping by so suddenly. Secondly, I want to pay tribute to you, because it seems that during the last Intel report—probably about a quarter ago—you came to us and became one of our most successful guests with your trading ideas.
Intel, and it seems SanDisk is another company whose stock has only gone up since you first appeared on the stream. So congratulations on that.
Yes, thank you.
Yes, please. Well, look, thanks for coming. We agreed that if Intel beat earnings expectations, you'd come, and Intel did impress with its report. But the market is in a bit of a strange position right now, so I think it's actually even more valuable for you to speak out now.
The situation looks like this. How are you doing now? How is your year going overall? How are you doing in trading on the market at the moment?
I think this is the best year I've had in trading or finance in my entire life. It seems like all the years I've worked combined don't compare to how much I earned this year.
All the years you worked together don't equal this year?
Yes.
Why do you think this was the year you finally got it all together and all the work paid off? Why do you think this happened?
I think part of it is definitely because of the AI trade, and also because of my career and experience. It was much easier for me to distinguish between what was nonsense and what wasn't, and who would succeed and who wouldn't.
I don't sit there every day looking at financial statements thinking, “Oh, this company is expensive, and this one is cheap.” I always think only about products and the customers who will use those products.
For me, it was just a matter of being in the right place at the right time. I've always been fascinated by AI and have thought about it and studied it for a long time. I just happened to be in the right place. It seems like all the research and knowledge I've accumulated over the years is incredibly valuable.
I know I asked you about this last stream, but first, congratulations. Are you still trading full-time, or do you work a job?
I guess I was more into consulting then, right? I was able to take on consulting projects simply because I had left my previous job and didn't want to go back to a large corporation or some very hierarchical place.
I used to do consulting, but I definitely don't do that now.
To your previous statement that you had the technical expertise to see through all the nonsense—what was real, what was fake, and what was valuable during this initial surge of interest in AI—do you think this advantage has disappeared and a new game has begun?
Does technical experience no longer help as much, or is it as valuable as it was 6 months or a year ago?
It's definitely just as valuable, especially if you disagree with the market. The market needs time to understand new information, or it may give certain data more weight than you do because it's run by people who aren't as immersed in technical details.
Intel reports are always a good example, because no one can understand what's what for a few days. But I also think an example of people overthinking things is when everyone thinks that whatever Nvidia does is the only thing that matters.
People don't realize that most of the computing in the future probably won't come from Nvidia. Most likely, it will come from the computing power of all the hyperscalers: TPU, Trainium, Meta.
This doesn't mean Nvidia will disappear. The market is growing. This whole class of market is growing. There is development in this area. But you have to understand that what one company says doesn't set the rules for the entire industry.
The industry moves based on what is the optimal choice from an engineering or finance perspective.
So what do you think this means for Nvidia's future moat and dominance? When you have 75% margins like that, it's like being in the right place at the right time, right?
As Uncle Jeff always says, your profit is someone else's opportunity. You can't get away with it forever. There's a reason why hyperscalers invest in their own ASICs for their own internal workloads, or invest in their own interconnects.
They want to own it, because with these costs, why depend on a supplier? Even if you want to run a business, it would be very unwise if you only preferred one microphone company for your podcast, wouldn't it?
You would think, “Well, I like this microphone, but I'm willing to change it if they start making too many corrections or stop shipping the product.”
I just don't know. I think it's clear that Nvidia has a pretty dominant position right now, but the incentive for labs and hyperscalers is vertical integration. That's the incentive.
This is similar to Trini's post. It seems he was writing about memory and Micron. After their financial reports, he said, “Okay, the profits are crazy. Their numbers are disgusting.” But the flip side is how long the market will let them rip off everyone else.
Of course. That's right, up to your point. Micron has higher profits than even Nvidia, and they sell a less differentiated product.
People are more willing to think about the value that Nvidia adds. They're more willing to pay a premium because you're managing the infrastructure, creating a unique, differentiated product, and taking on all these capital expenditures. People don't want to take that kind of risk all the time.
That's why people pay a premium. But Micron is actually probably the worst memory manufacturer in terms of quality.
Really?
Yes. In terms of quality, their HBM has always been of the lowest quality.
One more thing, I don't know if I mentioned this in the last stream, but if you look at the big three memory manufacturers—Samsung, SK hynix, and Micron—people think everything revolves around HBM because that's what GPUs use.
That's part of the story, but most of the growth is actually coming from commodity RAM. To make HBM, you take standard DRAM wafers, cut them into dies, and stack them. So they consume standard DRAM.
People still need DRAM for other things, but there's a big AI deployment happening where people are just going to start consuming it. The price of regular DRAM has skyrocketed even more than HBM.
HBM is stable or unchanging, maybe a little up or down, because they have power. Regular DRAM—the stuff you put in your PC—is skyrocketing. That's where Micron and everyone else makes their money.
In fact, Micron is the most glaring example, because their HBM has always been of the lowest quality. Their HBM is the lowest quality, but they just dominate at the lower end of the spectrum.
Exactly. Let's say Nvidia—I don't have those numbers in my head right now—but let's say Nvidia gets 10% of its HBM from Micron, 40% from SK, and 50% from Samsung, or whatever.
HBM is just a smaller part of the market than regular DRAM, and Micron is going to set a market price for it.
So what does this mean for CXMT's IPO?
It's funny, because you've already said, “Oh, I made all these great predictions.” I'll call you: if you get access to the CXMT IPO, you should invest in it fully.
I was thinking about pawning everything I have to get more money for it, because the price is extremely low. She is extremely short.
The thing is, CXMT, despite being Chinese, doesn't actually have an HBM product. The demand for regular DRAM is so high that they can still charge a premium price, and they sell mostly to Chinese companies.
There was a lot of talk that CXMT would disrupt the DRAM market. It's not profitable for them either, because they don't make good HBM, but they can sell to all the Chinese companies and make a ton of money.
If you look at what price Micron is trading at, what CXMT is IPOing at, and how many bits of DRAM they're shipping, CXMT is extremely cheap. That's partly because it's Chinese.
At what price does it still remain extremely cheap? I think Hyperliquid's valuation is now around $400 billion, give or take.
That doesn't seem so cheap to me. But I think the Chinese IPO may be at 130.
Yeah, that sounds pretty cheap. But even at 400, it will grow to that point. I don't think it's the end of the world, but if it were 130, I'd probably move to China to get access.
Of course. I don't have much experience trading in emerging markets, but there's an old rule: never trade Chinese companies.
Why? Where does this come from?
I think part of it is because of government intervention in markets. The government supports the industry, and then as soon as it gets into trouble, they say, “You have to figure it out yourself. Stop living off us.”
Chinese industry and the CCP prioritize certain industries. They did that with electric cars a few years ago, right? All the electric-car companies were a great choice, but now they're not, because they're in a state of fierce competition.
China is not about that. At least from my perspective, the Chinese stock market and the government there don't seek to have just one big winner. I don't need the company with the largest market capitalization. Rather, I want to grow the industry and have people compete to lower costs.
I want this to become a public good.
For example, electric cars—cheap electric cars—are great for the Chinese people and great for the world. Cheap memory is the same thing. But with these semiconductor companies, they can't really use that strategy. They have to support them because the capital required to play in this field, especially on a global scale, is so great that it makes no sense to force people to compete. It makes sense for people to come together and work together, right?
So I think it's a great opportunity for semiconductor companies, and I personally invest in some Chinese companies.
Speaking of investing in Chinese companies, I may be wrong, but I'm pretty sure it was you who tweeted about buying BABA. I have a feeling that every great investor, at some point in their lives, has tried to play BABA's stock price and lost money. They mostly lost money.
I guess every time I played BABA, I lost money. But they're in a crazy position right now, especially with the rise of Chinese open-source Kimi and Qwen. It feels like they're in—or own—all these Chinese AI labs. They own everything except maybe DeepSeek.
But if you forget about that for a moment and just look at Alibaba's e-commerce business and its cloud business, it's so cheap. It's very cheap, and yet it remains a high-quality cloud provider. If you were outside the U.S., in Asia, an emerging market, or even Europe, you could use Alibaba Cloud.
Now they have some of the largest and most powerful AI companies in the world there, and they will start making deals with Huawei to get special chips. So I think Alibaba is in a great position. If you believe that the Chinese AI market will grow, then that money will pass through Alibaba's servers.
Huawei is a private company.
Yes. I'm not sure. I think they have some kind of public listings. I could be wrong.
Interesting. Regarding BABA, I want to go back a little to the general market situation. I asked you about the results at the beginning of the stream. I don't want to make this a confrontation, but Serenity has posted monthly results probably since the beginning of the year, and it's up 49% in a month.
This is crazy because DRAM has fallen significantly, questionable assets have been hit hard, and market bottlenecks have caused a lot of things to collapse, while the SPX has only fallen a couple of percent—2% or maybe 3%. Do you think this is the general picture of profitability among the financial community right now?
I don't like to give Serenity credit, but you have to admit that most of these companies are down 50% from their all-time highs. It's a massive sell-off in the industry.
People see it and say, “Manage risks better. Do better.” I'm not saying that all of Serenity's choices are good. I'm just saying that these companies have grown by thousands of percent. Even if they fell 50% from the historical high you randomly chose for the day, they're still some of the most efficient companies in the world and will probably be the best for at least the next year, maybe longer. It depends on the company.
But when you see SanDisk or Intel changing by more than 10% in a day, you can't really be shocked that they can fall 20%, 30%, or 50%. It's just the nature of the game for something that has a high growth rate.
So here's what I would say about it: I don't think this is a sign that you should just sell everything and that it's the end of everything. It's a sign that you should probably take some profits and be careful, so you don't end up with zero. You just need to be aware of what's going on.
Let's say you think about dollar movements. SanDisk is probably down about 50% from its all-time high, but in terms of market capitalization, it's barely changed from where it was at the beginning of the year.
That's a good point. Let me see how far it has really fallen. I think it's somewhere around 40%.
37.4%.
Yes. Okay. Speaking of selling everything and panicking, I have a screenshot: July 17, 3:26 p.m., by Bubble Boi. “I sold everything.”
Yes. I sold everything.
Even this?
The next 2 weeks will be difficult for most of you. Good luck.
850,000 views. Wow. This was a real hit. I didn't even realize how much reach you had—850,000 views.
First of all, congratulations on this level of coverage. But where are we now with your market exposure? I mean, is it still positive?
My portfolio is clean. Well, I have long and short positions. I have no positive exposure to the SPX. On a delta basis, I have a negative position on the SPX, right?
But it's not too negative. It's not like I'm 100% shorting. I think it's about minus 15% of market exposure.
Oho.
That's simply because you have to be aware of some game theory in this. All of these assets have grown by hundreds of percent, in some cases by thousands of percent.
You have to understand that the institutional investors who bought them did so either at the beginning of the year or in the middle of the year. Their reasoning is this: if you bought Intel at the beginning of the year and invested 10% of your money in it, you probably got a 30% or 40% profit just by doing nothing. This is just one small position, and it's a great result for a professional financial manager—one of their best results of the year.
From their perspective, they need to lock in profits. Again, you have to understand that their goal is not to become the richest person on Earth. Their goal is to get their bonus. That's their main goal, right?
So it's kind of a positioning thing. We also got the earnings reports, the market took off, and everyone realized that infrastructure development is a reality. Now the questions are: Are the hyperscalers' capital expenditures sustainable, or have some of these companies even overinvested?
A big reason I tweeted about it was that I had positive exposure for a long time. If there were days when SanDisk fell by 10%, I could lose a maximum of 1% of my portfolio, even though SanDisk could be one of my largest positions.
The reason I acted quickly and panicked and sold is that you have to panic in time. There's been a lot of innovation, and it took me a while to sit down and read all the research on what Kimi and other AI models are doing.
I'm in the San Francisco Bay Area—not today, I'm in New York, but mostly I'm in the Bay Area—and I talk to people who are working on this research. I ask them questions: Do you think this technique is scalable? How soon will this start to spread to other labs?
I began to understand that new facts had emerged, so I could no longer hold these assets with the same confidence as before, based on the previous thesis.
Oho. What are your thoughts on capital expenditures after studying Kimi?
To me, the emergence of Kimi is a triple destruction of narratives. First, it's a Chinese open-source company. Second, the cost per token is much lower. Third, they allegedly did it on Huawei chips.
How do you think Kimi's emergence has changed your long-term vision for AI investments? If Kimi came along and was just a really good model but still used roughly the same methods or infrastructure as other AI models, I would probably reduce the risks a little bit, but I wouldn't change my position drastically, right?
I think the important thing is that Kimi uses a new technique that completely eliminates the increase in memory consumption. The memory size is fixed, and they got very good results. They didn't just shrink the memory and get acceptable results; they got world-class results.
The technique that Kimi uses is what we call attention mechanisms.
Attention mechanisms.
Yes. These AI models have this thing that people call “attention.”
Of course.
It uses a technique called linear attention. Linear attention was used in Qwen models back in the day, or even not that long ago, right? But Qwen wasn't that competitive, was it?
No.
So we didn't give it much importance and didn't believe it would scale. It is now becoming increasingly likely that this scales quite well. If that's the case, you have a fixed amount of memory. It's always fixed, right? Memory doesn't grow depending on how long you communicate with it, and the memory per user will be fixed.
Although it's not completely fixed. Many people objected to me and said that only 75% of the models are fixed. The rest is still growing to some extent.
My point was, okay, let's assume every model in the world switches to this. How much memory will we need?
Yes.
But since I've researched this more deeply, I don't think the situation is so clear-cut anymore, right? This doesn't mean that all is lost. It means that there is a certain compromise that, in my opinion, many Chinese labs—and later American ones—will begin to accept.
Theoretically, if every leading model switched to this, demand for DRAM would drop significantly.
Well, for DRAM, definitely. With HBM it's different, but for DRAM, exactly: you need to offload less data, so you can do without as much of it.
But you could say everything is fixed because they already bought the memory, right? They already have these servers. Everything is fixed; they just serve more and more users. So they'll make more and more money.
Not bad. Got it.
For them. They'll earn more. Imagine you have a car and you wake up and someone says, “Oh, I found a new way to drive, and now it's 10 times faster.” You would say, “Why sell the car?”
Yes. That's great, but I'm not going to buy a new car.
Yes. So I thought it was partly because of that. But I also think that we now need to reconsider the assumption that they'll spend money wildly for the next 2 years.
Who’s to say that another innovation won’t come along that simplifies things even further, and now you have one server rack serving not 1,000, but 100,000 users? How do you justify spending $500,000 on a new server rack? This is getting more difficult, although we still have 2 years until that actually appears.
This is definitely a topic of discussion for everyone. Even if that happens and they move on to Kimi, the money will just be reallocated to other things. As many have already said—and they were right for the wrong reasons—Kimi is actually very useful for NVIDIA, because NVIDIA’s main problem was that the GPU just idles during inference.
Mhm. You paid for that expensive GPU and you just don’t use it. Kimi makes the GPU busier, so it becomes more sensible to buy high-end Nvidia hardware.
Got it. But I think all these arguments are complete nonsense and lies. This is another point.
Going back to your point about hyperscaler capital expenditures, Google’s report came out on Wednesday. They announced a significant increase in their capital expenditure forecast. Google fell 20% from its peak. It fell $30 from the $350 level, where Berkshire—that is, Buffett—bought $10 billion of their apparently $80 billion [?]. The MAG 7 stocks were crushed yesterday.
So at what point do you, or the market in general, start buying Google stock? Or do we just hate this increase in capital spending on principle?
The market—financial analysts—hate capital expenditures because they reduce free cash flow, and for them, that’s the main thing. A company is worth as much as the total free cash flow it will generate from now until infinity, right?
If you don’t have free cash flow for the next year or 2, and you had to buy a company just to hold it for 2 years, you theoretically wouldn’t make any money if you bought the whole company and just held it for 2 years. Again, this is a theory.
Yes, yes. But the pyramids weren’t built in a day, right? It takes time to do the right thing. It takes money and investment, and I think the market will punish them for capital expenditures simply because there are old-school fund managers with finance degrees, and that’s all they know.
They don’t understand that the pie is bigger on the other side.
Exactly. What were Christopher Columbus’ capital expenditures when he discovered North America? Probably quite high. But when we got here, it was worth it, right? I’m probably applying that same model to this.
So I think the market will punish them as long as they continue to forecast higher capital expenditures. On the other hand, I don’t think they will. I think it was quite clear this year that they were going to finish the projections a little higher by next year.
I think next year you’ll start to see the situation gradually changing for the better. You mentioned Google, but it’s Microsoft that’s been hit the hardest, and that’s because Microsoft’s whole strategy is, “We’re going to spend capital to build these AI data centers.”
Investors don’t buy into that because they’re asking, “What for? What is your product?”
For example, Copilot. What are we doing? For Copilot’s sake?
And then they say, “We’ll rent it out to people,” and the response is, “Well, the big AI labs are building their own data centers and diversifying, so why should we believe you?”
Of course, there’s some truth to the idea that people will rent from anyone, but this is where the market will punish them. Google didn’t get penalized for a long time because Gemini was pretty good. They use AI for all their products, not just Gemini. It’s a normal workload, right?
I said Google Search. They have it everywhere.
Of course, but the thing is that they can only do this up to a certain point. Eventually, the market will give them an ultimatum and say, “Google, you can’t dilute my shares by another 20% this year. You can’t just do another issuance. You can’t borrow $1 trillion. Just calm down.”
But here’s the thing: when they put in that capital expenditure, the money starts coming in. It’s like—I’m trying to come up with a stupid example—like I’m opening a laundromat. I have to invest in washers and dryers. It might take me a year or 2 to make my money back, but it will start happening with these data centers.
Eventually, the money will start coming in, and by my calculations, it’s a lot of money. This is a high-margin business—extremely, extremely high margin. We’re talking about margins of more than 90% sometimes for servicing these models.
To a certain extent, it depends on whether Gemini is good or cutting-edge, which it is not right now, and for which they are perhaps being punished. But then there’s this gray cloud over the cheap Chinese open-source industry, which is causing the value of the tokens to fall to zero. What if Google’s benefit from capital expenditure is not as profitable as it seemed it might have been a year or 2 ago?
We’ll address the Gemini issue to some extent. The fact that Gemini is not great is not good. If they had an increasingly better model, it would fully justify the capital expenditure.
Yes. Let me explain, because it’s related to your second point. I think the emergence of Chinese models and their creation of good models is not such a big problem for capital expenditure.
That’s the thing. If the model starts to improve beyond the Chinese models, we evaluate them based on advanced capabilities, right? If the next Gemini model can do something valuable—like find new drugs that can treat people—the value of that Gemini model is based on how much those drugs are worth.
Imagine finding a bunch of drugs. Imagine it can make a movie. Imagine it can do all these other things. We’re on the right track, and we’ll get there. This will take years, but we will achieve this.
Chinese models can’t do that—unless Chinese models start creating new capabilities that American models can’t even match. This is scary. But it’s not a big deal that they can write better code, because the coding is already over. The coding disappeared.
The next step will be doctors, lawyers, and airlines.
So if Kimi releases something and makes a new pharmaceutical discovery, then sell everything.
Yes. Okay. Run away.
Got it. But until that happens, we’re in America, you know.
What do you think this means—I know I’m jumping from topic to topic—for the economics of the OpenAI and Anthropic IPOs? Did Elon just outsmart everyone? He was the first to start.
There will still be time for them. Elon was smart to go first, and he acted quickly, because 112 is SpaceX now.
Speed is very important for getting money, and the more money you have, the more you can fight. I don’t think the IPOs are “fried,” but if we’re in a real, prolonged bear market, then yes.
I don’t see that happening. Despite my tweet about selling everything, I want to buy it back, but I think there’s a time for that, and now isn’t it. Anthropic’s IPO shouldn’t fail. It’s still supposed to happen at the end of this year, I think, right?
We still need to see OpenAI, because they don’t want to go public and then have people immediately say, “The Anthropic model is better. Let me sell OpenAI.”
The most popular trade in the world would probably be buying Anthropic and selling OpenAI as a hedge, right? But again, that eliminates a lot of nuance.
So are you planning a re-entry based on price capitulation, or some news event about the development of AI?
I think the price should come down a little bit, but we also need clearer signals from AI labs and hyperscalers that they’re not worried and that everything is fine.
I’m doing my own research on what these new innovations actually mean. For example, what impact do they have on all parts of the supply chain? Part of it is about becoming more confident, because when I invest, I invest with confidence.
I’m not an index investor. I’m not trying to buy 1% of every company. I try to invest in a company I like, right? So I have a slightly different approach to investing.
Are you raising a fund?
What fund? A hypothetical fund.
Oh, I don’t know anything about this fund I’ve heard people talking about.
You know, there’s one more thing we haven’t talked about. I know you’re leaving soon.
You should leave in 5–10 minutes.
No, no, I have time. I have time.
Good. Another topic we didn’t really discuss is Intel, which is kind of funny. To start with, if I remember correctly, when you came in about 3 months ago, your prediction for Intel was based on its packaging capabilities and the new node.
What do you think about Intel as a company now and its future growth? Do you have confidence in it?
Yes. I’m very optimistic, by the way. The earnings report was excellent, despite the sell-off, because they provided capital expenditure forecasts.
The reason for the sell-off is that people think they’ll spend more on capital investments than they say. They’re very adamant that they only need $20 billion, and according to my calculations, $20 billion is more than enough.
Additionally, they mentioned during the call that they have prepayments from customers. Customers deposit money, saying, “I want to get the capacity as soon as you have it.”
Forget about 18A, which is already in production, and 18B. They mentioned that they’re now going all in with their next-generation 14A.
This is because there is a client. They have talked about it many times: they will never increase capital expenditures until they have a guaranteed customer. I take their word for it. Capital expenditures are no joke, especially in this field, because when you launch this equipment, you need a constant flow of orders. It’s not like an AI inference cluster, where it’s just an API. You need new clients constantly, right?
I definitely think they have a long-term commitment there. So I’m very optimistic about Intel. To be honest, I think the market is selling off the stock a little bit because of rumors about capital expenditures. People just don’t know what they don’t know, right? Like me when Kimi came out. I didn’t know what I didn’t know. I had to reduce the risk a bit here.
I think now is the best time. Maybe not right now, but over the next 6 months or a year, you could buy it even cheaper. By my calculations, 14A has a real customer. You can just do the math: $20 billion—how many tools does that allow you to buy? How many wafers can that turn into? How much will you charge for them? You start getting ridiculous numbers, like 100% revenue growth in a year. Disgusting.
I also heard that my name is now being mentioned in these investment funds because it’s become popular to bet against “Bubble Boi” ahead of the Intel report, I guess. But people don’t take everything into account until they see the results in the reports, right? When it happens, it happens instantly.
Like when Intel last time beat expectations and said, “We have positive margins. We will have higher growth due to processors and longer-term positive margins.” That’s all they needed. Then the whole movement happened instantly. We have a joke on stream: never bet against Bubble Boi.
I think you can bet against me. I’m not always right, but if I’m sure about something, I wouldn’t be too sure if I didn’t have a good deal of research behind me. I am very wrong, for example, when I know I haven’t done my research and something new happens, like Kimi. I revise my plan very quickly. I think that’s what should be done.
Okay, here’s another side of the market that I don’t understand at all, but I saw it and saved it just in case. SemiAnalysis did a post about SMIC and how they outperform Intel 18A. Can you explain what this is and whether China can catch up with them in this regard?
I think you’re asking the right person. Regarding the second part of your question, certainly China can catch up. I’m not one of those people who sees EUV as something divine that no one can replicate. We’re not idiots. The laws of physics are the same here in the U.S. and in China, right? You can only hide things for a certain amount of time. The question is more when they will catch up, not if they will catch up.
The comparison of 18A and SMIC, in my opinion, was very disingenuous. The comparison is based on transistor density and other metrics, but it is not performance. This is just one indicator. You look at one metric and try to determine the overall performance of the node. I could list a bunch of other metrics that suggest Intel is better.
One more point: 18A was in its early stages when they attacked Intel. Why didn’t they choose Samsung, which is doing even worse than Intel? Right? To me, it looks pretty purposeful. I don’t think they’re my close friends, so I think that’s why they’re trying to fool people.
But 18A is their first node that started to ramp up volume, and they just started to move their real capacity—their own internal capacity—from TSMC to 18A. The advantage of 18A is greater energy efficiency. It’s not so much about higher transistor density, is it?
I wish they would do the same comparison for 18P. For 14A, I would also like them to make the same comparison. If they were a truly reputable publication, they would keep updating the data and doing direct node comparisons, right? It’s perfectly normal for someone to want to point out flaws. I just think the way the information is presented already creates bias.
For example, the media is biased by the very fact of choosing what to report, right?
How much of an impact do you think they would have on the market if they went on a full-scale anti-Intel crusade? What kind of impact do you think they have on the market?
I think in certain situations it’s great, and in certain situations it’s not. Even if they started a whole campaign against Intel, it wouldn’t make much difference. Intel will spend money and make money. It’s not like they have any—after all, Google or Nvidia aren’t going to read SemiAnalysis reports and say, “Hey, it’s actually worth taking the order from Intel.” It doesn’t work that way. This is more about the investment community, right?
Yes. Yes.
Even then, everyone needs to do their own research. If you don’t do it, then what’s the point? I’ve heard from professional hedge fund managers that they forbid their employees from reading sell-side analysis. They forbid them. They simply say, “Don’t read it at all.” Sell-side analysis exists only to shed light on the situation. It’s not really there to tell you what to do.
Yes, I like this view. Another thing happened that I thought was pretty exciting. I think it was a month ago, but Leopold Aschenbrenner’s last report, Situation Awareness, was expected, I guess—that’s the word I’ll use. Later, it was revealed that Shaw had a 20% investment, which I now own, of, I don’t know, $1.5 billion or something, in an Australian neocloud.
What do you think about Leopold’s direction? It’s a bit like trading around AI bottlenecks. Like, okay, let’s do photonics. To hell with everything. What is cheaper than photonics? But now asset managers are doing it. What do you think about this?
I have to make an introduction. Leo is my good friend. I don’t give him any advice or anything like that, but we have a mutual respect for each other. I don’t think he’s that—he’s more experienced than he seems at first glance.
I think his investment in neocloud has a very specific thesis, and I actually don’t think it’s wrong. It depends on the specific neocloud, but he has a very valid point about it. Also, if you’re going to invest in flash memory, photonics, or any of these things, they’re cyclical by definition, right? You invest, you build, and then you wait until you have to do it again.
He has a very specific thesis, and I don’t want to voice it publicly because it’s better for him to say it himself when he’s ready. I don’t want to give away his story. But I don’t think it’s unfounded. I’m pretty optimistic about certain neoclouds. When I started reallocating my portfolio, one of the first things I bought was neocloud.
Therefore, I don’t think neocloud will have bad results in the short term. The only question is whether, in the long run, these neoclouds will turn into real clouds, like real infrastructure, right? There’s a lot of discussion going on, and I think a lot of smart people are following this. I wouldn’t say my entire portfolio is long on neocloud.
You should become a shareholder in Shaw with me.
Yes, I will look into this issue. I’m not really—I just saw the Situation Awareness report at 19.9%.
Okay. Another interesting thing that happened today is that Jensen created a Twitter account and shouted a battle cry, siding with the AI Obiotaur. It seems like we’re getting ready for this AI civil war, which is essentially Sam and Dario hating each other, maybe having common enemies, maybe Trump on their side, but they don’t have David Sacks against Chinese open-source companies, which is quickly becoming a major topic of discussion.
What do you think will happen?
I hate to agree with Jensen because he says what benefits him, right? This is good for Nvidia, and it loves neoclouds. He wants neoclouds to exist. He doesn’t want all the computing power to be in the hands of a handful of people while others can’t create models. For that matter, he wants everyone to have local computing power—every business. It benefits him, right?
I will say this: yes, there will be something like a civil war, as you say. Many people think we should ban Chinese open-source models, and they give all these nonsensical risks and reasons.
If you’re an AI lab, you’re a lab—the word “lab” is in your name. What do you do? Your job is to innovate, and if you can’t innovate, you won’t be a lab for long. I think you should focus on what America is best at. America is best at innovation. China is great at iteration. China is very good at taking something, making a thousand copies, improving it, refining it, and making it better.
One of my favorite things to do in my free time is watch YouTube Shorts. There are all these videos about ancient Chinese construction. They show how they built a bridge, how they first failed, and then, a thousand years later, they got the perfect bridge. It’s kind of built into Chinese culture. They’re so good at improving things over generations.
In short, I think AI labs trying to ban competitors is really pointless, and it’s going to happen with or without them. We will not have a monopoly or duopoly in the field of AI. This will not happen. There are many reasons I could name. One of them is data. You need certain types of data to take the next leap in capabilities.
How are you going to create a medicine if you don’t have data from biology and chemistry, right? You won’t be able to do that, will you? You understand that you can’t do much at home with textbooks, pen, and paper, right? I think they should realize that their task is not to fight for undemanding consumers.
Their task is to achieve significant results. If you can’t deliver results, then you shouldn’t be in this business at all. YouTube Shorts—I don’t want to single them out—but YouTube Shorts and the big 26 are a bit questionable. Yes, I got hooked on them in a strange way.
There’s an interesting, deep irony here: the US won the internet through openness, allowing everyone to build their own services on that base. Then, in the second round, it’s like starting over again, and the US is leaning toward closed AI, or at least pushing it, while China says, “No, we’re going to do open-source AI.” Xi gives a speech where he says, “No, we’re for open-source AI.” It’s deeply ironic how it all unfolds.
Many people tell me that they’ll ban Chinese models, but I don’t think that’s possible at all. Given the people in the administration who are advising Trump, and of course Congress, the Senate, and all that, it just doesn’t make sense. What do you think the strategy would be if this happened?
I spent a lot of time trying to think this through. At first, I thought the strategy was to invest in AI infrastructure—these hyperscalers, right? The question of external competition simply disappears.
They still won’t be able to ban it. We’ll just fly to China and come back with flash drives. We’ll figure something out. The real question is: if you’re working with the government, how do you ensure the security of these models?
If they ban it, they’ll say you can’t work with the government if you use their API. That’s normal. It makes sense. It’s quite reasonable. But there will be a bunch of companies that say, “Hey, we downloaded Kimi, and we’ll run it safely just for you.” This will become a new trend.
What’s your thesis on why there’s no chance the US will ban this? I don’t think that’s a generally accepted opinion.
You see, Jensen made his post. Do you see the people who signed up for it?
Yes. Elon’s greatest post for everyone.
Elon doesn’t mind. If you can beat Elon, go and win. He’ll try to become better. That’s the point. This is a competition. So why are you trying to use dishonest methods?
I think even Sam—it’s really just Dario, right?
Sam, yeah.
Yes, of course. Even Sam—I don’t think Sam hates it at all. I think it’s more like this: we’re focused on creating something big. We’re not focused on commoditizing some small Python scripts. That’s not what will make us the most money in the future.
Okay, I’ll give you one last question, and then we’ll wrap up. Even if you’re not investing in anything crazy right now, what new trends or deals do you find most exciting for the rest of the year?
I hope I talked about this on your last stream. I said that photonics was blown away. Photonics has blown away. I was absolutely right, and people continue to grasp at straws.
Although I will say that maybe there will be a new way to do “photonics” that is less blown out. But these won’t be the names everyone knows now, with all those expensive lasers and advanced TSMC technology. This will be something completely different. So I think that’s what fascinates me.
I also think neoclouds are in a very interesting position right now. I’m a Nebius customer myself because I want to launch Kimi and get this over with. The only way to really get Kimi up and running is to get one of those big servers.
How much does it cost?
I don’t know if I can say, because it depends on the contract, but it’s about hundreds of thousands per month.
So what the hell are you doing there?
We’re conducting research.
Why do you have to run this? We overlooked this. What are the advantages of this approach in terms of shelf layouts? Running it locally like this?
For me, it’s more that I’m not going to take anyone’s word for what Kimi’s impact will be. I’m going to launch it and say, “Okay, how does it work? What do the tests and all that say?”
Another advantage is that you launch it yourself. You can do a lot of things. You can change it and optimize it. I also think that when Kimi becomes really open, when they release the weights, there will be an interesting race to see who can make it work, because it’s not that easy.
We spent a lot of time on streams discussing local AI and the best way to run it, and I always say, “Damn, I wish Mac mini were a separate company.” By the way, the answer is flash. For AI, it’s flash.
Who exactly?
All of them.
All of them.
Yes. I think there’s a Taiwanese company your viewers should look at called Phison. Phison makes flash controllers and helps get rid of DRAM. I’m very optimistic about the idea of abandoning DRAM. Who knows? Maybe I’ll do it myself.
Could you give the best explanation of why flash?
Flash has density, right? You can store terabytes of data on a flash drive. So there’s that.
Besides, if you look at what you actually do when AI is running, all the naive engineers will say, “Oh, flash is slow.” Well, it’s slow if you use it for everything at once. But if you use it for AI, you need to know how to use it and how to interact with it.
AI doesn’t always need fast memory. You need to know what you need and when you need it, so the access pattern is more important.
Okay. Flash. Any other topics you like, or specific names?
Not this year, but next year, I’m very interested in what Qualcomm will do. Why are all the companies saying, “Oh, we’re building a Qualcomm device”? There was a SpaceX leak, and there was something about an open—what? They have some kind of secret ingredient. They just have this… What is this?
I was very skeptical, and I think everyone is very skeptical about Qualcomm now. We’ll have to see what they actually release, but there’s reason to believe they’re not completely incompetent or liars. They might actually, if you could somehow stream the data on Qualcomm’s data-center accelerators, achieve the kind of throughput they’re claiming.
They quote ridiculous throughput figures, and many people simply don’t believe they’ll achieve them. They don’t believe it’s real. It’s as if they think there are some minor terms of service or an asterisk in the text. I think there’s a chance they’re not really joking.
If they’re not kidding, do you think this will be some kind of breakthrough?
I think they have a huge advantage.
Exactly. Qualcomm is a terribly managed company. So let’s see. Why?
Just terrible management. They basically pushed every single one of their customers away. They pushed Apple away, they had a huge lawsuit, and they had a virtual monopoly on modems for 4G and 5G. Now they have to reinvent themselves.
But they’re like cockroaches—they never die.
I didn’t know the whole backstory of Qualcomm. So Qualcomm and Apple had serious friction and sued each other. Who won? Do we know who won?
Apple. Apple won. Of course.
Duh. Of course. Bubble, you’re in charge. Give me the final word. How low will DRAM fall? How low can it go? And what are you working on or passionate about right now that you haven’t mentioned yet, if there’s anything?
I’m waiting for flash to take over DRAM’s job. So when you start to see this happening, sell DRAM. We’re not there yet, but you can’t just rip people off like that forever.
Okay, last question. Give me some kind of goodbye. Name your favorite trader on Twitter right now—someone you think is top-notch, underrated, doing well, or whatever.
I need to get out my phone and see who the next “Bubble Boi” is. I want to know if that’s possible.
Unfortunately, I think Irrational Analyst is a very smart person. I think he’s a real photonics expert and engineer. I definitely wouldn’t advise taking everything he writes as absolute truth, but just read what he writes and talks about, and you’ll begin to understand this industry.
For example, in photonics, you’ll probably make better investments. And who knows, as I said, if photonics takes off? It’s going to be this guy who finds them, right? He’ll find you stocks, right?
He has a few recommendations, but we’ll see if they come true.
No, that’s not true. Wait, we hang out together all the time. We spend time with him all the time.
Are you wearing anything from Chrome right now?
No, no, no.
Didn’t you take anything from Chrome?
No, not today. I’m in the city. It’s a little hot here. I need to run errands.
It’s really a bit hot. I don’t know if Bubble is that bad. I’ve never seen him flaunt money in Chrome jeans.
No, with the Aul class in the club.
No, no, not yet.
Not yet.
Man, you’re the best. The second part was cool. Thanks for coming, friend.
I’ll say hello, because last time I tried to say hello to someone, I messed up. Congratulations to my guys from Through. Congratulations, Through. Maybe someday we’ll do a stream about what Through is. Let’s invite the guys from Through.
Robert Chang.
Yes, sir. Bubble Boi and Robert Chang, titans of the crypto industry—we’re in good hands, man. Congratulations to Through and Bubble Boi.
Congratulations, Bubble Boi. I’m looking forward to the third part soon. Maybe Intel for $200. Maybe I’ll see you in New York.
Yeah, sure, dude.
Okay, my friend. Big hug, dude. Thank you for coming. Goodbye.