MU is down $90 today after CXMT announced a tremendous breakthrough in HBM3E. Again, this is tech but it's an interesting turn. Let's run this through the BS meter. Here's what we know: - CXMT struggled with HBM3 yields as recently as 2 months ago (rumors of 25% yield) - CXMT HBM3E is currently in risk production - TrendForce has not picked up this story or made any mention of it - Micron has said it can only produce 1/2 to 2/3 of demand for HBM3 - HBM requires 3x the wafer area of DDR5, CXMT promises 4x density improvement of HBM3 which would make it cheaper than DDR5 - If the CXMT 4x density improvement announcement is remotely true, it will be a massive step forward, even if their parts only meet HBM3 spec and even if their yield does not top 50% - Based on the description provided by CXMT, new dies can be stacked 12 high for 36GB of on complex HBM3/E, just like any other HBM3/E compatible part - CXMT started shipping HBM3 at the end of 2025 so it's been nearly a year of extremely limited quantity production - Micron has shipped over $1B of HBM4 (they don't share die or capacity count) and countless HBM3/E BS Meter = 20% likely to be sufficiently true as to be a market factor Likely reality = They are probably onto something and it could easily bear fruit but I'm extremely skeptical of their timeframe given they were struggling with yield 6 weeks ago and also because of past ridiculous and unsubstantiated announcements. Narrative: China has made so many announcements that turned out to be complete crap that it's difficult to take anything they say seriously. The retail investing market, of course, does not take the time to consider credibility and has forgotten the endless stream of past lies. Has everyone forgotten when they announced they had 2N figured out and ready for production, ahead of the rest of the world? That was two years ago and a different company but Chinese semis are still struggling with 7N and even 10N yields. This lack of credibility does not eliminate the fact that China is working hard and making good progress with semiconductors. HBM3E is generally suited to 10N, so there is some plausibility to the CXMT claim. Micron remains a better bet than ever, IMO. It's not all clear sailing for Micron. Micron advises roughly 2x cost increase from suppliers that will impact FY27. The next three Micron Fabs to come online are: Tongluo, Virginia (significant expansion), Taiwan. These three fabs are built and currently doing early qualification work. Production will from these will not affect Micron's bottom line this year and they won't be fully online until roughly the end of next year. This slow (to us but the Fab itself is working hard) ramp process is one of the reasons CXMT's new process announcement is unlikely to be true. Again, I'm not declaring it a hoax but the odds of being able to buy reasonably priced HBM3E on January 1, 2027 is zero.
Here's a media report card based on the above post. Reuters headline "China's CXMT makes breakthrough in advanced memory chips, the Information reports" Thoughts - There is a modest hedge in the article but Reuters is basically parroting the CXMT announcement in full detail Yahoo headline "China memory leader CXMT begins small-batch production of HBM3E silicon - report" Thoughts - Similar small hedge but essentially presents the announcement as fact Tom's Hardware headline "CXMT president says company’s first 4F² DRAM will be a server DDR5 RDIMM due by year-end — vertical transistors cut each memory cell’s footprint by a third on paper, but no speeds or capacities yet" Thoughts - The article goes into some detail on VSG and 4F2. It's actually quite informative. To be honest, I'm surprised at the quality of the article and the perspective it brings. While not particularly skeptical, it does mention the CXMT struggle with 10N and their struggle to be taken seriously by the semi industry. Kudos to Tom's Hardware on providing a bit of detail and some perspective. Reuters and Yahoo phoned this one in, exactly as expected.
Here's a note that is directly about tech. For those who think the cheap, Chinese, sky is falling, ASML has said the semi industry is going to be supply constrained through 2028. Believe the news at your own risk. [edit: This post does not say that all market sectors and all semiconductor categories will remain at peak price. It simply means the industry appears poised to remain at elevated price to the end of 2028.]
More tech to consider. TSMC reported $47B earnings in 26Q3. That's not for the FY. That's actual earnings for Q3. Tech is going to be a wild ride but there is an entire refinery of gas being poured on this fire.
The upside of the Tesla / SpaceX merger is the news cycle will have less time for lies for the next week. lol!
nVidia is openly talking about $2000 per year per GPU subscription fees for AI users. Is this a genius move or a death knell? I don't know. If anyone is still here and cares enough to post an opinion, I would enjoy reading it. I'll tell you this. I've recently watched interviews with some people attending an AI data center conference. There are people controlling billions of dollars of AI spend who are dumb as shit. There's simply no nice way to downplay their lack of basic understanding of the technology, the industry, or basics of how a data center works. The position of the people I watched is crystal clear: They know that they can build an AI DC and make a ton of money. They don't know anything about how DCs are built, how to implement anything, or how they're going to make money but they are confident they will get even richer by building an AI DC. An engineer was talking about ideas to mostly remove network latency from the inference pipeline. The next interview went something like, "Me have money. Me like money. Me build data center and get even more money!" There are still plenty of smart people out there but I consider it a fact that CEOs and entrepreneurs are getting dumber. There are way too many people who don't know anything and they don't want to know anything. They just want to make money by any means necessary and talk about how great they are. That's fine but I will have zero sympathy for these people when they fail and I will hate the world even more than I currently do when the government bails them out for being complete morons and losers.
I wish I knew enough about the details to contribute or even better understand a lot of what you talk about. Lol I do enjoy popping in for your take as I respect your thoughts/opinions a lot. I’m just here to hopefully ride the wave as a mostly index investor with some shares of individual tech to double down on what “looks” to be companies with some runway. Now on oil and refiners (which I don’t really buy due to their historically boom/bust) I do think this wave of new record profits is here to stay for 5+ years. I feel it’ll be a LONG time to go back down to what was also record profits 15years ago
Thank you for the post and the kind words, Blake. I appreciate it. I will reveal that Oil is on my list of sectors that are likely to be profitable well into the AI age. Eventually, everyone will go broke doing everything but I hope that won't happen until I'm done with this mortal coil. Thanks for the take on profit levels. You would know better than I would. My take is less authoritative but... I see a massive solar/nuclear build-out. Oddly, I think wind is likely to take a diminished role. Despite this, I expect oil will also grow over the next decade, at a minimum. I don't understand why people are so ignorant but, even the hemp trouser wearing oat latte crowd would to cry if petro chemicals were to take a supply hit. There is no solution to oil. We're starting to peel layers off that onion but there is no freedom from oil in sight, even if thorium were to make electricity free tomorrow. It's super useful stuff.
After the previous post, I wrote a post regarding a few things I consider germaine to this moment in the markets. I ended up deleting it because it was so negative. It was overwhelmingly "sky is falling" in tone so I assumed it was a symptom of a bad head space. After thinking about it for a few days, I think it was specifically accurate but generally more pessimistic than my actual feelings on the market. Here's a few thoughts on random topics (not so negative). These things are important to think about, from an investor point of view. A little knowledge will be worth a ton of money if/when events happen. These days, there is 100% chance events will happen. Supply chains are manipulated to create crisis for the purpose of money harvest; it's an epidemic. 1) RAM The YouTube talking heads are talking about a flood of cheap memory. Even channels I consider to be objective and immune to hype, are declaring the AI bubble to be over and brace for an avalanche of cheap DRAM. Meanwhile, AI data centers are absolutely crying for DRAM. They would prefer HBM3 or 4 but will take anything from DDR5 (desktop RAM) to LPDDR6 (leading edge laptop RAM). They would undoubtedly take LPDDR5, if they could get hold of it. So, forget about an avalanche of the cheapest DRAM in history. No doubt, that will happen at some point but that is not going to happen in the next 12 months. Prices are moving down as supply increases, at least for DDR5, but there will be no RAM oasis in this desert. From what I can tell, demand will consume supply for the next 24 months at an absolute minimum. MU is currently trading at $885. In two months, Micron will post their 27Q1 earnings and people will be blown away by the income. "But, that YouTube guy said the memory shortages were over?" Reality is exactly the opposite of the talking points. 2) Steel US Steel is primarily made from iron ore that comes from Minnesota and Michigan. The vast majority of ore is processed in Indianan, Ohio, Pennsylvanian with some in Michigan. Nearly all of this ore transits the Poe lock on Lake Superior. It's a bit more nuanced than that. There is a second, much smaller, lock on Superior but it's a lot smaller and the big ore ships don't fit. As a consequence, 90% of ore goes through the Poe lock. The Poe lock is in desperate need to be shut down for maintenance. That event will cause tremendous supply chain problems for American industry. A project to twin the Poe was approved and funded in 1986. Cost was under estimated by a factor of 100, although much of that comes from delay and inflation. As I understand, the project may be back on track to complete in 2030 with the hope the Poe can remain in service until then. The moving pieces in this situation are this: - If things remain smooth with China, America will simply buy cheap steel from China and there will be no issue. - If China embargoes steel, this situation will become an emergency. - If China embargoes steel and the Poe needs to be taken offline before the new lock is commissioned, this will be a medium crisis (there is some production in Utah but it will not sustain the country) 3) Energy The future is energy. Our AI overlords will require it so we will be forced to provide it. A) Solar remains a strong participant. There is a whole lot going on in this. 1) Solar data centers in space make a lot of sense. It will come down to service life. Space has the overhead of rocket launch and zero serviceability but solar power will be 99.5% available so energy cost becomes non-recurrent. If service life is long enough, reduced energy cost will pay for the orbital insertion. Ignore talk of cooling problems. Cooling is a problem but it's solvable and not the biggest problem. Reflect Orbital was an interesting idea that appears to be dead. They were going to use satellites to reflect sunlight onto solar farms. Zero new solar deployment cost. Zero new power transmission cost. If you own a solar farm, you could just pay to have more sunlight and drive generation well into the evening peak. It could not be more ideal. I believe the idea is now dead but you never know. There is quite a bit of work being done on orbital energy harvesting systems that would capture solar energy with a large array and then retransmit that energy using either infrared laser or microwave. IR laser has the problem of being blocked by cloud cover. Microwave has the danger of a large concentration of microwave energy. Still, there is a good chance we will see these things which are made viable by the work SpaceX is doing to lower the cost to orbit. Back on earth, dual mode solar cells have taken efficiency beyond 30%. We're at the leading edge of this change but solar arrays will capture double the energy in the next 12 months. I expect 18% efficient panels will be around for the next 20 years but they are on the way out, albeit slowly. B) Wind is a strong energy provider today. I believe it's role will diminish, in time. I don't believe it will go away, just become less desirable in a cost/benefit analysis. There is a ton of research being done on wind energy, much of it highly interesting, but it just isn't turning into practical production. I hope one or more of these ideas succeed but I am betting against it. C) Tidal energy makes more sense than anything but operational issues have put it out of reach, so far. If we can get some production online, it will generate data and improvements that will undoubtedly make tidal a primary energy source. It's predictable and it's massive. Right now, I see no chance to invest in tidal energy but the day is coming. D) I believe the next generation of fission reactors will be substantially more viable than current generation. This isn't fact; it's an opinion. To this point, promises of SMR efficiency advantages have not been fulfilled. I believe this is the technology to bet on. Thorium is the most promising step forward in nuclear. Thorium isn't a fuel so the name "thorium reactor" is misleading. Most, nearly all, uranium is not fissile. It can be upgraded or fissile material can be curated from uranium deposits. Thorium and fertile but non-fissile uranium can be combined to create fissile uranium; hence the name "breeder reactor". A slow thorium breeder reactor will be a massive efficiency step forward from current nuclear energy production. This is likely the future but it's not a slam dunk, to say the least. Fusion is a bit of a religion. You either believe in it's potential viability or not. Fusion is real. Making an operational fusion reactor is not real, at this time. I'm hoping for fusion in my lifetime but betting against it. E) Oil is one of the most misunderstood energy sources we have. Everyone knows oil comes from the ground. Everyone knows you can boil it, extract gasoline, and burn it in a vehicle or to make vulgar patterns in your neighbor's lawn. Well, that view is so simplistic as to be wrong. Earth needs petro chemicals. The average person has no idea how desperately these chemicals are needed. Meanwhile, we have advanced extraction to an incredible level. Refining has been energy intensive but we're making advances on this, also. Pressure is used either instead of or to supplement heat in many steps. Current efficiency levels have us consuming roughly 5% of the product in the refining process. Efficiency has roughly doubled in the last 50 years. It's extremely difficult compete against an industry that is as evolved and efficient as our current hydrocarbon players are. Long term, I wouldn't bet against oil. It will stop growing. It might even shrink but it's not going away. For me, investing in oil revolves around gluts and shortages. Gluts happen when a major oil producer is desperate for money. At some point, Iran is going to rebuild their rape palaces. They are going to knock down the price of oil to do it. Shortages are essentially only caused by war or embargo (can be considered an act of war). So, Iran is blocked from selling it's oil. Price goes up. Iran acts like an asshole, and gets pummeled, At some point, Iran will need to sell a ton of oil to rebuild their country. That will drive the price down. Price increases tend to be short term, from days to perhaps a year. Price decreases tend to be longer term, from several months to a couple of years. If you can't figure out how to profit from obvious oil price swings, you need to stop reading anything I write and simply buy an index.
I just want to mention, there is anecdotal evidence Intel's second gen 18A has slightly higher yields than TSMC's N2 suite of nodes which are also entering a second stepping. I suspect this to be real. Assuming so, this is an industry changing moment. Intel needs to get it's Ohio fab back on track. I'm confident they can. Intel is looking extremely good, right now. I'm not saying they're worth the spot price. Intel is hemorrhaging money. Intel may not survive. All I'm saying is: they have a very strong chance of turning the ship around. At a better price, I would own Intel. Given the shares have a very real chance of becoming worthless, I'm going to need a great price but I'm keeping an eye. Lip-Bu Tan is just what Intel needed. The other thing Intel needs is a stronger technical lead. Instead of complaining about the limited supply of sulfuric acid (they need purity levels that make this a rare resource), they should be refining it like Elon would do. Of course it's better to have a competent supply chain. You can't do everything but if nobody else will do it, sometimes you have to do it yourself.
This is tech but it's more about analysis and how to have a clue about an industry. There is some weird stuff going on. This is going to focus on nVidia but this can/should be done for a number of companies. I will only share nVidia thoughts so do your own analysis. Consider nVidia supply versus demand. N4 wafers per month: 3.7M (and scaling with yield gains and TSMC build-out) N3 wafers: not significant so excluded for brevity Supply: Given pessimistic yields, nVidia builds 1M Blackwell core GPUs per month. Given pessimistic yields, nVidia builds over 2M Blackwell AI DC GPUs per month (GPGPU). Demand: The largest, global scale, data centers use 200~250K GPGPUs. There are only a couple of these. A $1B USD AI DC would use about 8000 GPGPUs. That means nVidia has enough supply to build out 260 units of one billion dollar data centers per month or 3100 AI data centers per year (each worth 1 Billion US dollars). Where are the GPGPUs going?: We aren't building that many data centers. I've been trying to track DC builds and it looks like we are well under $10B per month (I've estimated 8.6B) of new data centers or data center expansions. That's total spend, not GPGPUs. Assuming 60% of DC CAPEX spend is GPGPUs, that comes to about $5B of GPGPUs per month and that's a high estimate. What happened to the other 98% of GPGPUs? Some will be deployed globally but there are a lot of embargoed countries and the US AI fetish consumes the vast majority of GPUs. Some parts are dual die so my number's are undoubtedly inaccurate. I could see 50% error in my guesstimations but I don't believe I'm off by an order of magnitude. There are two possible targets for the unaccounted for GPGPUs. 1) Dark projects. 2) Smuggled to embargoed countries (predominantly China). Given how obvious GPGPU smuggling is, the US government has to be OK with it happening for some reason that is in no way clear to me. Consumer GPUs: Where are the 900+K Blackwell GPUs going? Some of them are being smuggled to China where the GPU parts are extracted from the card and transplanted onto cards with massive memory for AI use. This makes for a surprisingly effective single core AI GPU. This likely explains why it's extremely difficult to find supply of 5080 product at $2000~3000 per copy. Rehosting these GPU cores should have close to 100% yield and is a relatively quick and easy process. I would be surprised if they damaged 1% of GPUs on removal. This is a significant factor and makes even more AI data centers unaccounted for. More Data Points: The numbers are undoubtedly wrong but they aren't that wrong. Any suggestions the production numbers are misleading and causing gross over estimation of product is a non-starter given the global DRAM consumption rate. In other words, dark projects and smuggling account for nearly all GPU consumption.
I just read a report from an independent lab that had a Panther Lake Intel ThinkPad battery lasted 30 hours in the web browsing test. I'm not sure anyone is independent anymore and we live in a post truth age but Panther Lake appears poised to push AMD mostly back out of the laptop space. Sorry Intel, your 18A kung fu is not going to regain desktop leadership but there is a chance to close the AMD leadership gap. The next generation of compute will likely be determined by whomever has the best reliability. This is the age of chiplet bonding. Chiplets warp both naturally and under heat. The warpage isn't much in real terms but lithography is literally down on the molecular level these days so it won't take much to break a CPU complex. Something like AMD's core die on memory controller was a massive step forward but is one bond layer (memory controller with one or two cores bonded for a maximum height of two die). Zen 6 and Panther Lake are going to have lots of layers so they better get it right or it could be an existential problem.
China is aggressively working on a thorium reactor powered container ship. The reactor is specified as having 10~25 year service life between refueling. This could be significant to the shipping industry but I present it as a data point on the chart of how important China sees Thorium breeder reactors. If China manages to containerize a thorium MSR, it will change the world in a profound way.
When I wrote this, INTC was at $66. Right now, INTC is at 52.25 (Market cap $278B). Intel financials the company look like a LearJet 60 flying through a Florida storm. Too big to fail? Intel is the 14th most valuable semi manufacturer by current market cap. Rank Company Market Cap (Approx.) 1 NVIDIA $4.98T 2 TSMC $2.14T 3 Broadcom $1.56T 4 SK Hynix $1.11T 5 Micron Technology $1.01T 6 AMD $837.5B 7 Samsung Electronics $754.6B 8 ASML $612.5B 9 China (Unnamed) $443.5B 10 USA (Unnamed) $351.1B 11 USA (Unnamed) $312.4B 12 USA (Unnamed) $300.0B 13 USA (Unnamed) $276.0B 14 Intel (Hypothetical) $278B 15 UK (Unnamed) $218.7B * Note: The unnamed companies are not listed but will be owned by big players. They are undoubtedly (in order): Owned by CXMT, AMD, Owned by TI, Owned by Applied Materials, Owned by TI, Owned by Arm Major semi charts show Intel at 9 with a market cap of roughly $0.5T. When semi rags reposition Intel, I suspect the blood bath will increase. Let's consider an Intel failure. I doubt Intel will be allowed to fail, although it's not outside the realm of possibility. The government is already propping them up heavily. nVidia was also asked to throw some cash at them, which they did. Retail investors would be the ones who would take nearly all of any failure hit. Banks will be OK. Equity partners will be OK. Market cap of $278B is still too high but it's starting to get very interesting. This is a situation in which I would not place a long term buy order, hoping to snipe a great price. There is a very real possibility of an announcement that would cause the order to execute and the value to be eliminated. I'm going to hang in there and watch for a couple of specific pieces to fall in place before pulling this trigger but we are inching closer to an entry point.
