It is really good to see you getting such high view counts for these videos I really enjoy them. Especially the one on ASML. Which was the first one I saw.
Look pal... the supply is about 2-3 time lower than what the market needs today. Not only this, but this situation lasted for a year already and will tend to expend until 2022 making a year and half of shortage to coup over. In some market, we are talking about 10 times the demand over the supply, like GPUs. The pandemic proved that capacity is not enough and that the semiconductors industry is not just an commodity business, it is now a requirement for almost everything. You are making your analysis with the idea that the demand for semi will stabilize... when it is not. Until these fabs are built, supply will be an issue with backorders of about 2-3 years... while demand rising during that time because technologies like AI and data-center are going to keep growing. Telework is going to stay, there is no putting the genie back in the bottle here. Chips are going to be more and more integrated into vehicles. 5G will push IoT to a crazy amount of chips for intelligent cities... and let's not talk about electronics and industrial automate. You are missing the point that the pandemic shown us how much we relate on semi and this is no turning back. The matter of fact is... TSMC and the other companies, are going to sell everything they can produce. We are not in a normal cyclical episode, it is a never seen situation.
You forget about Quantum Networks. It started 2018, and yet starts to build out. IoT requers a lot of chips too. Europe invest because of reliability and legacy systems.
From a Mechatronics engineer in a poor country with zero high tech industry (Colombia): thanks a lot for your analysis of the semiconductor industry, it absolutely blows my mind!! Much respect for Taiwan and South Korea, amazing people!!!
So start some. My brother in law from Colombia is high tech, but had to go to Texas to get a good job, even though he still lives in Medellin. Your country has the talent. Just stop exporting it.
Despite the incoming flood of new fabs and semi's, it's important to note that there's an arms race right now for 5, 4, 3, 2 nm production. A lot of these new fabs aren't all producing the same products - for example no matter how much Intel invests in 7, 5, 3 nm they won't be competitive against TSMC for the next 3 years at minimum. Until we get a clear picture of who is pumping out what chip, what contracts have been signed by who, I don't think we can declare a bust as incoming.
I don't even need to look at any data. That a bust is coming is a given in an industry that is constantly at an arms race to the bottom. Every major player in this industry is constantly operating at that fine line between profitability and bankruptcy. The tricky part is to predict what straw will break the camels back - it could be anything any time. In 2005 it was the housing crisis in the US - houses didn't even use chips back then, they just bankrupted a critical portion of the market with a domino effect. It could have been the pandemic in 2020, but the industry went up instead. Could be the end of the pandemic when everybody decides to go outside instead of buying gadgets. Could be a minor skirmish between American and Chinese ships with just 2 shots fired and no casualties in the chinese sea and everybody loses confidence in the international market. Could be anything...
@@KonradTheWizzard Well, that is true for anything. There could be a bust sometime in the next 10-100 years. But, I think we're speculating the next 5-10 years, if we are over-investing and ramping up supply to dangerous levels.
@@vincent4696 I see that as opportunity to figure out how to use that capacity. I really wish there was a service I could buy custom batches of 1K chips of my own made chips (I have some ideas on how to side-step the costly engineering process, I only actually need the physical production, not the IP, I'm working on my PhD thesis on it) , which is a tiny batch at that scale, but they can make it by putting different silicon from different people, even if we have to wait months, at such low batches, we can afford it, and the factories would be over-invested and needing jobs to keep running anyway. I wanted kind of a ChipWay, like a PCBway of sorts.
Intel has lost its way, as ARM gets into more and more WinTel products Intel will continue to see its cash cow shrink. AMD is doing a great job also shrinking their market share. Intel is still creating some good tech but not on the scale they used too.
@@shenjiejason8350 This world is becoming more modern all the time. The number of server farms is growing. The number of super computers is growing, the number of compute devices in people's homes is growing. AI is growing rapidly. Businesses rely more and more on computers. What is meant? Bust cycles don't last long, and this video is almost negligent for underestimating the amount of growth there is. And then I can move on to other fields. Medicine. There's more medical equipment that's controlled by processors or at least needs a small IC in them to run it. Transportation. There's more ICs going into vehicles than ever before, and as vehicles get smarter and smarter, it means more ICs, or ICs with more power. Either way that uses more die. Robotics. Yeah that's growing like crazy. Almost any manufacturing is mostly automated. And then there is the effect of companies like TSMC simply coming out with a new node. I can use AMD as an example and how it's effected the server world. Zen 3 EPYC, which is AMD's latest server CPU is the best X86-64 server CPU on the market. Lots of companies are ignoring their regular life cycle for replacing servers (a server, being a single circuit board (motherboard or mainboard), with either one or two sockets for mounting a server CPU, and this board slides into a server rack, and therefore a server rack holds a few servers). They're ignoring their normal life cycle plans because of how efficient Zen 3 EPYC is. The thought process is, you can't afford NOT to replace the older servers, because of the power savings) So this is the main part where this video has no understanding about. AMD Zen 2 and Zen 3 EPYC CPUs are made using TSMC 7nm node, along with GlobalFoundries 14nm node. The CPU is what's called an SoC (System on Chip). The processing cores are made with TSMC 7nm. The I/O portion of the CPU is made with GloFo 14nm, or maybe 12nm for Zen 3, I forget which. For the next generation coming from AMD, Zen 4, EPYC processors will be made using TSMC 6nm and 5nm nodes. The cores will be TSMC 5nm and the I/O will be TSMC 6nm. And then you move to the next generation, as long as plans stay on track, Zen 5 will be made using TSMC 3nm. So I threw a bunch of numbers at you in the above par. But what does it mean to companies that have larger server farms or any company that has a few servers? Power efficiency. Companies are moving to CPUs made with TSMC 7nm because of power efficiency. And, the way this is typically evaluated is perf/watt. How much work does the CPU per watt consumed. This is a big deal for server farms, because it's not just about the server being able to do more given a certain amount of power, but it's also about air conditioning. These server farms have HUGE electric bills. So now the great news for these companies, moving from TSMC 7nm to TSMC 5nm gives about a 25% boost in power efficiency. Moving from TSMC 7nm to TSMC 3nm gives about a 40% boost in power efficiency. So, you have a server farm with 1000 CPUs in it, let's say, and you currently are on TSMC 7nm with Zen 3 EPYC. You decide to wait for Zen 5 EPYC so you get a boost of roughly 40% in performance, while maintaining the same power envelope. That's an almost OMG!!!!!!!! kind of figure for a server farm. So, the turnover of parts, simply because Samsung, Intel and TSMC are coming out with new nodes that consume less energy to do the same amount of work, or give you a big boost in performance for the same amount of energy cannot be underestimated. Because there are so many server farms around the world now, this is huge business for chipmakers like AMD, Intel, companies that makes processors using ARM, etc...., and that business gets back to the fabs, TSMC, Samsung and Intel mostly. In this video the dude said he wasn't going to make predictions about this year or that year, but then basically claims there's going to be a big bust. There isn't going to be a big bust for fabs. There's going to be big upswings, then equilibrium, then MAYBE some small downswing, and then another big upswing. THAT has been the cycle with fabs. If you're talking about a specific sector such as CPUs, GPUs, SSDs, DDR memory, etc.... they can have bigger swings, but for fabrication, the downswings are small because there's always a need for ICs of some type or another. You can compare the trend of CPUs for instance to the trend of fabs, and you see even when sales have slowed for CPU sales, it doesn't affect fabs. If you understand the business of fabrication, you should understand WHY this is the case. Companies make contracts with fabs months or even a year, maybe even up to 18 months in advance of the fabrication company making products for that company. Apple already has a contract with TSMC for their 3nm node, and that node isn't even OUT yet. That's the way it works. Apple wants to make sure they're on the most advanced node. That's why their products tend to be more efficient with power. Intel has a contract for TSMC 3nm for products they'll come out with at the end of next year. AMD has contracts for their GPUs and CPUs. Nvidia has contracts with TSMC for their GPUs for their next generation. This is to say, I wouldn't watch this guy and make stock decisions based on it. He doesn't study markets well enough to have ANY understanding of what's happening. Go watch people who make tech news instead. What this guy has done is dig into the past to give information about how these companies have evolved over time. While this is interesting, it doesn't predict the future. But, take all information with caution. Just because TSMC may have a big upswing in business doesn't mean that Samsung will, or Intel will (as a fab). You have to actually watch videos and read about what these companies are doing. Success for TSMC could mean the demise of another fab. At some point companies making older nodes are going to fade away as prices come down for newer nodes. It doesn't matter what equipment you make. If you can get a smaller IC that uses less energy to do the same thing your older ICs are doing, and it costs you less money, you're going to move to a newer node.
@@johndoh5182 Very good comment. Lots of good information. A thorough understanding of the company and its competitive environment is the only "edge" investors have over other investors in finding reasonably valued stocks. (Peter Lynch)
One caveat in this is that you treat all semiconductors as commodity parts. There are MANY semiconductor manufacturers out there, especially defense, automotive and aerospace which don't participate in commodity markets. This is a subset of the semi industry which is isolated from the rest which won't be affected by the bust because they can't get their parts from the likes of TSMC and Samsung.
I've heard from the news that automakers having problems because of a chip shortage. Each car probably has hundreds of different chips, can't just a few missing chips stop the production?
@@Maxim6431 I agree. The automotive sector is now becoming a part of commercial market, commodity. Defense/aerospace is still an isolated market, meaning a small piece of pie.
Well technically all semiconductor are commodity items. Some are used in the public/private sector or the military sector. There are minor differences in the public and private sector as the private sector is usually more demanding of their products while willing to pay good money for it. The military sector on the other hand is vastly different compared to the public/private sector. The ones used by the military usually undergo hardening and resistance processes to make them more resilient than their civilian counterparts. Not only that, there are minor differences such as material swaps for 1 small component of the whole piece but that swap makes a difference. You can't have chips coming loose at high stresses due to massive increases in g force due to sudden turning nor can it come off at really low temperatures(underwater or in really high altitude). You use special soder to make sure that the chips and other components will stay on the circuit boards. In one of the recent vids by Asiannometry, he goes into this a little bit. But know that the companies that supply parts for the military are capitalist run so expect then to stop production if there isn't as good as a profit margin as it used to have.
@@fuckedupbody4194 "Well technically all semiconductor are commodity items." No, this is not correct. There are MANY semiconductor components which are NOT commodities.
Brilliantly described and well done, this is a good example of market capitalism at work. As a Malaysian working for a electronic component supplier, we experienced boom last year for the demand of OEM equipment used to support these fabs overseas. Now I can foresee the potential for a slowdown in the medium term.
Your opinions would be more convincing if backed up by some analysis of chip consumption growth projections in absolute terms versus production capacity. The graphs you presented were only relative production increase and decrease. As mentioned in other comments growth in autonomous technologies will really drive up chip consumption for example.
besides that, datacenters want absolutely every single bleeding edge node they can get. TSMC could triple 7nm capacity tomorrow and it would still probably be completely sold out.
Extremely unlikely to happen in near future because it takes minimum 5 years to get a Fab operational to a satisfactory degree. Most of the fab plans we hear are recent or just planning stages.
But we didn't have a shortage until recently, so if those capacities go to work fully or just partially those fabs don't need to have high productivity to create too much in sum
@@tomlxyz There was a shortage incoming. There hasn't been enough capacity provisioning to meet demand medium term. Even without the bursts in demand due to work-from-home and crypto boom(s), the capacity would run out shortly anyway, maybe not in late 2020, but in just a couple years. Every year, people need slightly more flash storage, more cloud servers, more this and more that, now fridges have computers too for some unfathomable reason. Overprovisioned capacity is, to an extent, a good thing. Sure it just costs money when it's not needed, but not having capacity when it's urgently needed can be even more devastating. Whole industries depend on this capacity being present. Besides, the big growth cycle right now isn't actually that big relative to the existing size of the industry.
What expert forecasting shows (real data backed by extensive research) is in part what's driving all the FAB build outs. Do you think that because it's 'hot and in the news' that a bunch of investors run off to invest $20B in a FAB. These investments are huge and not taken lightly and without a wealth of information backing them up. What would be better is for you to figure out 'why all these investments are needed' rather than 'why all these investments are not needed'.
Ummm, no. It takes a long time to develop a new node. It's expected that the factories that are being built by TSMC and Samsung will be operational in about 3 years, because none of these plants are being used to develop new nodes. In AZ for instance, that will start out as a 5nm fab. That node is already in mass production in Taiwan and it's the node being used for Apple products. Once a plant is built, it takes weeks to get the equipment working and tuned properly. Not years.
all sources i can find see it happening no later than 3 years from now, with most saying closer to 2... and pretty reliable sources see it happening as early as late next year.
It is likely that the subsidies will continue to maintain technology parity since the current build-out is based on a strategic national security interest.
With regards to fab expansions, unless you have equipment planned in the build pipeline already, it can take you over a year to get a slot, depending on general business conditions. With how it is now, if you haven't planned it already, you may well hit the next downturn before the tool is delivered. Saying that, a downturn right now may just be a slackening in a generally increasing demand (hopefully :-) ). If you are looking at building a fab, you are not looking at current conditions at all, but extrapolating out a long way - state of the art fabs take 3-5 years to build and kit out, with a likely 1+ year to bring online. If everything aligns it will be a minimum 3 years from decision to build to first silicon.
that is why with more advanced nodes there is hope that instead of die-optimised chips, there will be more fpga and neuromorphic , universal chips possible. universality is both bad and good. ability to have fpga for beta products at much smaller node than predicted final product (f.e. microcontroller) is something everyone looks forward to.
I was discussing this issue with some bitcoin miners just earlier today and I pointed them to your channel as they were sure high GPU prices would carry on for at least two years. Based on what I have seen on your channel I didn't buy it given the massive capital expenditure by TSMC in 2021. Your writing is so good and well informed on these complex topics melding geopolitics, technology and economics. Are you a trained journalist or journalism student?
Crypto GPU prices and supply ... who knows. Whole new set of dynamics. I am neither a journalist nor a student. My day job is entirely different from what I do here. And that’s good because it helps blow off stress
Fabs take years to develop, and years to build, and years before they start making chips. 3 years minimum for TSMC, see how long it took Intel to pump out their 10nm chips (close to 10 years, and not even at full capacitty) It will be a long while before this big subsidy has any affect on prices
@@TheMasterofComment There's too much variable in saying how long it takes a fab to start producing products on a large scale. It only takes 2 - 3 years to build the plant. The amount of time it takes to start producing product could be a few months or many years. For instance the plant being built in AZ is a 5nm node factory. TSMC staff along with the makers of the equipment could come in and probably have it producing at full capacity in less than 6 months. This isn't comparable to what happened to Intel and 10nm. Intel was the company that developed and refined the processes to make their 10nm node, even though they didn't make all the equipment. But once again, for the AZ 5nm plant, TSMC 5nm is already being produced in mass. TSMC is already producing it for Apple products so it will be a VERY mature process even by the time the AZ factory is built. So, barring equipment failure, it shouldn't take long to start producing. I know there is a time period to get the equipment working properly, but that's not years, it's weeks.
google up for barkhausen noise effect in economy. Trying to predict noise profile of something rensembling spin glass is futile task, especially that butterfly effect causes such complex system to be susceptible to external influence. F.e. you could say you can try to create very detailed model of magnetic core so you think barkhausen noise will become predictable, and then your core gets struck with three cosmic rays, and few atoms internally decay due to natural radioactivity and bang - that is why it is called noise. Economics have very similiar problems, though ofc. because domains can be much better artifically aligned there is more room for microeconomy to do it's job. Still whole system blows into chaos on any possible excuse.
The only the thing I feel that was missed here as part of the demand, is the increase in performance smaller architectures are bringing to the table. On paper it is fairly easy to assume that we most likely did not create this huge amount of demand for the upcoming fab capacity in just one year, if you consider the "normal" growth rate of demand. However, performance and efficiency gains are getting much larger as we progress to smaller and smaller architectures, and this where I feel the trend in demand is not so "normal". On a smaller scale we can consider just the newest gpus, as an example of a very large performance increase over past generations of cards, and the demand for them proves that people now see upgrades as worth while. This is new territory, since for many years only small performance increases were made, which meant people held on to old equipment to realize more of its value and investment. Now we are seeing 30-50% improvements which has triggered a massive amount of demand for not just gpus but cpus as well. People now see the larger advances in performance as good reason to finally make upgrades on new generations. And with this same incentive, it is easy enough to assume that large data centers and all kinds of chip using facilities will be planning large overhauls of old equipment, in addition to their normal growth rate and expansion. Again this will be due to the fact that newer processes will offer better efficiency and higher performance that save on cost for compute power. This could be a one off event as "everyone" moves to smaller architectures then waits till the next notable increase in performance, but I think it is unlikely. Smaller processes have a lot to offer, and generational changes will continue to impress. And I feel it could very well snowball into an even more technology dense population as we begin to utilize more and more compute power. Just some food for thought! Enjoyed the video. I hope to have some good discussion in the comments. :)
These are all good points and some which I considered myself recently. However, TSMC's new fab is not going to bring "hundreds of thousands of wafers coming to the market over the next 1-2 years". That's not even close. Samsung is simply adding some capacity in Austin. They aren't building a new fab.... yet.... and TSMC's new fab isn't going to bring full output to the market for another 3-5 years!! So even if the cycle plays out like you say it will, that puts the bust out into 2024-2026... AT THE EARLIEST! And people will in fact go nuts once the pandemic lifts. Well over a year of pent up demand FOR EVERYTHING! will be unleashed on the world economy later this year, which is before the chip shortage will even be worked out. Semiconductor prices will have a chance to boom massively well before they inevitably bust again.
There is no pent up demand for high end chips I think, if anything, people will probably spend less time on their computer/phone and go out instead. If less people work from home and less data is generated from online activities there will be lessened demand for data center services too (probably less growth not an actual decrease). IoT devices like smart mirrors and wifi armchairs don't require the same kind of chips, these might remain in high demand but it's easier to scale production of these chips so I think there would just be a short boom as restaurant owners and such start buying new stuff followed by a steep increase in production, they are just commodities after all.
@@semkjaer3581 It will probably bounce back, but there was already a trend for home-working. It will never go back to the previous level. Even if the demand decreases, it wont decrease to the same level as before.
@@monad_tcp well this guy was saying people would 'go nuts' on semis after the pandemic and I just don't see that happening, I'm not saying the semi industry will do bad, I just don't think the current excesses will hold.
A bit late to the party. As a retired EE I have not been following fab technology very closely, thanks for the timely update on the situation. It is always interesting observing the intersection of technology and economics. The rapid advance of semiconductor technology has been incredible, it never ceases to amaze me how much functionality can be built into small battery powered devices. Having said that as a geezer I find that the technology in many new devices actually reduces usefulness and customer satisfaction. My wife and I recently purchased a 2020 Toyota Corolla. The difference in electronic content is in stark contrast to our previous Corolla of a decade ago. Some features are very useful however, many tend to impose the "will of the car" on us because the car decides what is best and refuses to operate in the way we desire. Other examples are making what was once a simple operation: such as ignition switch assy mode or parking brake activation activation (we have manual transmission cars). The keyless ignition now requires numerous operations to put the car into assy mode to simply listen to the radio without the engine running. So many steps in fact it requires consulting the 600 page owner's manual to do it. The sound system included a short term free subscription to cloud based services that we have no interest in using. Like most modern electronics the system requires a non-ignorable firmware update so each time we start the car and have to click thought the nag scree to access other functions.
This is fabulous education - you are teaching us about business cycles while talking about the most complex machines built by mankind Would be wonderful to understand these cycles in other industries as well- steel being one example.
@@GodzillaGoesGaga And yet the stupid Biden Admin came out recently and said it's leaving those tariffs in place and using them as one part of a strategy to deal with China. I would tell you that the disruption to American business and consumers isn't worth it. Go figure. Really I think the Biden Admin would get rid of part of the tariffs if Congress would get rid of the 2017 tax cut so there's more money coming into the govt., but until then the money coming in from tariffs are helping to fund a govt. that's spending a lot more money. It's just the burden has shifted from business and those making millions of dollars to consumers.
@@LuisVasquez-nu4hu considering that a place to live is by far more important then products with chips ... The core point is that long term and short term demand are 2 very different things, and can at a time be inverse.
THOSE were NOT "random thoughts." .... You KNOW your s**t, inside out and that makes you FASCINATING to listen to. ... I subscribed on the first video, the last one I watched, can't wait for the next one. ... (Your 53 000 subscribers will be 530 000 in no time).
Taiwan #1! Also, if there is a flood of supply, that means I might be able to get cheap and good PC components. May even be able to build up some server racks and storage lol.
you are ABSOLUTELY spot on. This goes to show Politics shouldn't get into commerce. Tears will flow like river in the next 5 years. All those tech will be obsolete before they can generate a profit.
i appreciate the angle. it has some sens in it BUT you left out real Boom drivers and that demand in this field will never go to 0 unlike in many other volatile markets
I got to work as an engineer in the industry for 20 years, visiting TSMC, Global Foundries, and ST fabs, very interesting technologies, incredibly complex. Apple's M1 5nm CPU is a quantum leap in semiconductor manufacturing. A good story is the billions spent on 450mm that all got cancelled 10 years ago as 300mm densities got smaller and went 3D with deep tunneling etching.
Overall I feel that the demand is going to surge further in the coming years with the surge in demand for electric vehicles, 5G, etc. TSMC is definitely going to benefit considerably from this. Looks like, TSMC is a great stock to buy in regular intervals.
And the huge turnover of parts. Think of all the server farms that exist. You're talking about millions of ICs for CPUs along with all the chips that go on server boards. Companies aren't going to keep older servers when newer ones pay for themselves in electricity cost savings.
Just had an meeting in my job at a semi conductor factory, at least they arent planning on building a new fab right now, but are trying to finish building their current new ones and squeezing out each and every drop of production out of the existing ones. Management/Cooperate certainly are all giddy and dont appear to expect a downturn at all. Thanks for the info, now I know I have to keep looking around for some second option fall back jobs.
Your analysis have some truth about the over supply situation within 2 or 3 years. Actually there is not much chip shortage for laptop and personal computer and smartphone chips. Most chips shortage are from EV cars, autochips and some consumer products, which are low end chips, 20 nm nodes.
@@slopedarmor That is the biggest bottleneck making it difficult to build a new gaming rig though, and the GPU shortage has more to do with crypto than anything else.
@@ManaDrain315 Crypto but the effects of material shortage cant be neglected. Supply chain disruptions have a far large impact, that crypto has contributed to but it is a minor contribution
A lot of system builders (PCs) can't build or sell products at the same level as they should be able to because GPUs are 1.5X to 2X the price that they should be at. This makes the cost of a new PC a LOT more than what it should be. So, parts like CPUs, RAM, NVMe drives, etc... are starting to drop in price because people have enough of these. People are holding off now. For laptops, they tend to be seasonal purchases with the shopping period around Christmas being the biggest selling period. I expect there will be a LOT of laptops sold at the end of this year due to big increases in performance and battery life. For smartphones, yes there's a slowdown because new products are about the same as older product. But his analysis lacks a LOT of understanding because it totally misses what's going to happen in the business world, and that's where growth for TSMC in particular is going to be massive. But as I said above, growth would still be large in the consumer space if GPUs weren't so overpriced. Expect huge turnover of all things computing when TSMC is mass producing their 3nm node. It will align with a lot of change in computing and graphics and the world will be salivating to get their hands on these newer parts. This is when he's predicting a bust cycle basically, and I say that's total BS. That's going to be a boom cycle like has never been seen. Power efficiency increases of about 40% from today's NEW products?? That's not a bust cycle. That's when server farms around the world will swap out parts.
Good morning from Hsinchu, drinking coffee opposite of TSM. I came to Taiwan 10 years ago, when revenue was 10 billions. They will break 50 billions usd this year. Bust will be temporarily.
This is such a good summary for the current situation, well done! I think the bigest question for the semiconductor industrie is the progress of autonomous driving and IoT. But if there is really to much capacity for chips coming in the next years, investors will have a hard time ;)
Even with autonomous cars, which have what maybe 3x the dies in them, how does that soak up all that capacity? Especially since the supposed goal of them is to reduce the number of cars needed to less than 1 per person. I think everyone has lost their damn mind over this chip shortage and the resulting government money, the capacity they're building won't be necessary until the 2030s I'd wager (with no data to back it up mind you lol)
@@musaran2 I rather think it will end up like any othe r "hot" booming topic in IT. Big Data levelled out FAR below what everyone expected, bicause a big pile of hay doesn't make it more likely to find needles. NoSQL has found its niche and is hiding in there while SQL is still going on. Let's go back a bit further: 5th generation programming languages / visual programming - nobody is even talking about it any more because it was a bad idea to begin with. AI makes sense for a limited set of problems which are currently dominated by humans because normal algorithms fail, but which are repetitive enough to automate. Like driving, image recognition, voice recognition - classic AI (neural net) topics. How many of those do you think you need? Of those: how many can you afford? Currently everybody believes that AI will solve all the problems that Big Data failed to solve - context sensitive classification problems - but those need not just data or a dumb learning algorithm - they need actual intelligence. Engineers. Knowledge. Experience. An AI has about as much experience as a new born ant. AI will not absorb the chip capacity. The pattern with all those hot technologies is that they are promoted as a "silver bullet" - the way to solve all your problems and free you of the obligation to think (or for managers: to hire people who can actually think, but cost money). After 70 years of IT we should now understand that silver bullets don't exist.
@@KonradTheWizzard AI can improve itself unsupervised, that is the game changer. AI feeds on data, if strives on underused big data. AI has already solved BIG challenges. I can't think of a trade fundamentally safe from AI takeover, not even art. The salary of a human worker can pay for a nice robot, and the salary of a human trade can pay for a lot of AI training.
@@musaran2 We are at least a few decades, more likely centuries, away from what you call AI. You will not meet Lt. Cmdr. Data within your life time. Not even his brother B-4. What we have right now stupidly learns whatever set of data you place in front of it. Your success depends on your selection of data as well as the exact layout of the network you chose. Even if it is an unmitigated success it can merely classify data that lies within its learning parameters (recognition networks) or make very simple decisions based on a (hidden) polynomial function (feed forward net or fuzzy logic) as long as you stay within the definition interval of that function that you know very little about. If you get outside the functions definition interval the AI will malfunction in surprising ways. It can not improve itself substantially. It may be able to consume more data and refine its internal function, but you have absolutely no clue whether the AI has the same idea about "refinement" as you do. You need to supervise it. You can play tricks in the lab by reversing the function and make your AI "dream" of weird shapes that are derived from the dataset it learned. But this is neither intelligence, nor creativity. It is mathematics playing tricks on you. I admit that the images released by Google are intriguing to look at. I don't want to take away from what we realistically CAN do with AI. Yes, we will be able to create automated image and voice recognition systems that help us with a lot of tedious and repetitive work - Alexa and Siri are just the beginning. Yes, we will be able to create self-driving cars. But we still have a lot of really hard work in front of us to get this done and AI is NOT able to magically find real patterns in data that human minds assisted by statistical software are confounded by - human minds are still tens of thousands of times more intelligent than any AI. And we have no clue how to get from current AI to brain - it will take a LOT of work and experimentation. You can not replace an engineering degree and human ingenuity with a bunch of data fed into a complex polynomial function with a name that sounds like magic. No matter how much you wish it to be true. Remember: we will have a functioning true AI in about 20 years - we've been saying so since 1950! And we are experts! We know. (Full disclosure: I did my degree on a neural network simulator system in 2000.)
Read a bunch of comments, that I agree with, you concentrate on the obvious users of chips, computers, cars and consumer tech, but miss the fact that all industrial machines from the largest to the smallest all use thousands of chips. And they are just the tip of the iceberg of other chip users. Nothing sold today that connects to mains or batteries is free of chips. Even the battery packs themselves have chips in them.
This channel is new to me. I'm impressed by both the range of videos it has on the semi industry and the detail it provides. It's boggling to think that about three decades ago, there was debate as to whether features could be constructed smaller than 100 nanometers. Yet now, it sounds like 2 nanometers is within grasp. Approximately 3 nanometers is the point at which some of those advocating nanotechnology (e.g. Eric Drexler) claim features could be atomically ASSEMBLED by migrating atoms/molecules instead of using ablation to selectively remove material from a bulk solid (like a wafer) or relying on a uniform vapor deposition
@@harshivpatel6238 I'm no expert in the field. But from what I've read, what is represented by the number (e.g. 5 nm) is the size of the smallest feature, not a complete transistor. The line width is perhaps the smallest feature. Wikichip says the following ..."TSMC started mass production of its 5-nanometer N5 node in April 2020. TSMC considers its 5-nanometer node a full node shrink over its 7-nanometer process. In early 2021 TSMC plans on introducing a second version of its N5 process called N5P which provides additional performance"
@@gregparrott while it's certainly exciting, you should add 6 months for a fab node to mature after it's announced, and 6-12 months for products made on that to reach various parts of world.
@@harshivpatel6238 As an end user, about the only significance I attach to it getting so small is the purported reduction in power consumption. As I understand it, Apple's new M1 chip runs extraordinarily cool, and is the result of both RISC architecture and small feature size. The M1 Max has 57 billion transistors and up to 64 gB of integral memory.
@@gregparrott M1 has a lot of accelerators for different work loads, CPUs & GPUs are more of a general purpose hardware, especially CPU. It's hard to be as good as M1 when you target general purpose use cases.
I doubt there will be a Bust anytime soon. I have been waiting 6+ months for Nivida 3080. At this rate I might as wait for Nividia 4080. Also on ebay Nivida 1080ti are still going for 500 dollars!
My old video card now is 2 times more expensive than 4 years ago. I don't think there will be a "bust" anytime soon. Also I see RAM and HDDs are vanishing from our hardware store. There are still some models but something is definitely happening. There are less and less models also the prices are going up.
Big difference between a cutting edge enthusiast GPU and chips used for say cars or drones or other more mundane things. That’s where the glut will be according to this logic. We’ve already seen cycles like this in memory - remember when ssds few years ago got up to USD 100 for 250gb?
@@suntzu1409 these things are cyclical, and older fans are being built in places like China and intel are building in Europe. In 3-4 years time there will probably be an oversupply, that’s how these things go
It takes 2 years min to build a fab. And I mean minimum. They started roughly 12 months ago. We wont see shit start to get better for at LEAST another year. A year after that we may see a 'glut'.
I’ve been thinking this for a while too. Each factory is also getting more and more productive with time. When new fabs are built on this scale those productivity gains will be applied to the new fabs and output will be insane. Market won’t know what to do with all those chips. If we have a war then all bets are off. War will likely break out in Taiwan and all that capacity goes to zero. Plus the need for chips for missile systems will send demand skyrocketing. It’s 50/50 at this point and all depends on what happens with the China/Taiwan situation.
It depends. If there are a lot of foundries going bust because of this likely overcapacity that may result in the survival of couple of them, which will results in duopolies if not monopolies. If that is the case, the consumer will not win in the long run.
Adding the 'smart' to clothing & jewelry & everything else, tools, guns, vehicles rolling data centers *autopilot* 5G, everyone including homeless people with smartphones & smartwatches // prices at every scale to keep everyone hooked on their twitter feed / like mainlining information hard smack style with a screen syringe // TH-cam videos of whatever hits the spot like movies or music streaming // more cloud compute & server action // cryptomining was only the start, the blockchain idea for security improvement coming to nearly every industry, just like NFT's are trending as non fungible tokens // the zeitgeist that electricity demand will steadily increase over time as more people get online // you know what I mean. Excellent analysis even if its just your thoughts! Your right, boom & bust, its not just chipfabs, its whole economies
If deep learning continues to deliver results while demanding ever more powerful hardware support, we will not see an oversupply of chips coming from the latest generation of fabs. If deep learning fizzles out, as artificial intelligence has done in the past, we'll definitely see an oversupply and crash in the semi market. It's good to understand the overall market fundamentals covered by this video as well as the specific differences in the current cycle which may result in a deviation from past behavior.
Deep learning cannot "fizzle out." That would be like the PC market fizzling out. It can mature, but even that is years away. Deep learning has way too many useful applications to fizzle out, just like the PC has way too many useful applications to go away. The market for the *current* applications of deep learning is huge. Most companies/industries that can benefit from deep learning haven't even started to implement it. And we haven't discussed the yet-to-be-discovered applications of deep learning....
Semis don't have to stand on their own two feet, that's a rather simplistic way of thinking. The cost of a shortage simply has to be greater than the cost of the subsidies. Semis are so integral across so many industries these days, that permanently subsidizing their production and ensuring a constant glut may actually be cheaper than dealing with natural shortages. You see the same thing with the energy industry. Energy is so important to the function of the economy that it is worth constantly subsidizing energy production rather than dealing with the economic fallout of energy shortages.
But we forget that use of electronics is getting increasing day by day...we can see how more electronic we are using than 5year back. Everything is getting smart...hence upward demand will absorb new capacity also....
Intel fabs in the US have stood on their own feet for more than 50 years. They were making silicon when most people in asia were still riding around on bicycles. I'm pretty sure others can stand on their own feet once the initial investment is paid for. Most will be put directly to work on US military contracts. They don't really need to be competitive in the consumer space. Much like steel, fabs are deemed a national security asset.
That make no sense, building a multi billion cutting edge fab just to sell chips to the US military makes absolutely no economic sense, unless is a small fab. The US military is a small buyer of chips compared to commercial costumers. By example Intel sales to china totaled 20% of their yearly revenue compare that with the sales to US military and will make more economic sense to build a fab in China.
@@elchippe The US government doesn't want its fabs in China. And they will pay any price to make sure they stay on US shores. It considers communist China an enemy. And Taiwan to be too close to it and in danger of invasion. They recently paid TSMC billions to construct a fab in Arizona. And have refused to let any Intel fabs relocate overseas. They don't care if it makes economic sense and consider it a national security issue. They plan to subsidize it with taxpayer money.
@@ComfyShortz They can put billions if they want but they will have a 10 billion dollar fab idle doing nothing, but at the of the end of the day the company will sell the fab and move on.
ALL the companies being subsidized to build new fabs in the US, which right now I know of Samsung, Intel and TSMC make billions a year in profit. There's no such thing as running subsidies just because. The subsidies help out with the HUGE upfront cost of building a fab. After that these companies make their money by producing products, via contracts or making their own products. This video was a propaganda piece, which is a shame because many of this guy's videos are very good. This was utter garbage. It was simply an attack against the subsidies, and using ZERO information to project future growth and then suggesting a "bust" will happen. But NO trend data supports this. If anything trend data shows the inverse, for fabrication. Now, what trend data WILL show is that some segment of electronics can have overproduction, such as CPUs, GPUs, SSDs, DRAM, HDDs, etc.... and this is because many companies make CPUs, GPUs, SSDs, DRAM, etc... But fabs don't make products just because. They make products based on the contracts they have to make products, and it's some other company who sells finished products that can go through boom-bust cycles. TSMC for instance has had revenue growth every year since 2009.
So if the biggest fab goes to subsistence level by overproducing and oversupplying the market, and that causes the smaller fabs to go under, I am guessing the big fabs don't shed tears.
There’s been a significant increase in at home schooling and stay at home work. Some may go back, but many businesses see the value of stay at home work. Why spend hundreds of thousands on office space and equipment when you can offload that responsibility onto your employees? More money to invest elsewhere in your business.
A lot of people are agreeing with you in the comments so I thought I’d give my two cents. Descent analysis but I’d say we’re far from seeing any kind of crash. GPU prices stayed crazy high during the three year slump in crypto when it basically wasn’t talk about (2017-2020). With online gaming, TH-cam only getting bigger, companies revamping to cut costs, etc chip prices and demands will stay for a long time. These badass 7nm chips are still brand new and 5-2nm chips are only going to make everything better. The leap will be huge and everyone will want what’s faster. Speed and efficiency is going to be the key to who is the next long term super power, China knows this which is why they’re working so hard to get more chips. Governments need faster and bigger super computers to stay ahead and not get hacked. I think (I have a PhD in physics and have been investing for a few years now) that we have at least another 10 years before the tech improvements become minuscule to the point where the demand drops off.
I forgot to mention cloud storage. That’s one of the biggest needs for chips that an average person doesn’t think about because it seems like magic. More and faster cloud storage will continue to drive the chip market forward for another 10 years if not longer. We’re only in the beginning of our generations industrial revolution.
@linkzable I don't believe anything a Chinese company says they can do without an insane amount of evidence (just look what happened to HSMC). Also, there's no way known to make those chips without EUV. Like the physics doesn't exist. So no US doesn't have to worry about China right now. TSMC is going to kill Intel not China.
Lower end fabs will definitely face a lot of stiff competitions soon, but I assume advanced chip fabs will take longer to build so they have more time to make their money back.
The lower end Fabs are not idle. They are busy making chips for lower demand product lines. Eventually, they will retire, but not until they extract as much profitable life as possible.
@linkzable Nothing in life is guaranteed, but doing something is better than doing nothing. Also, just having TSMC and Samsung engineers are not enough. There are required chemicals that only Japan makes which they can ban (that's one of the key reasons Samsung lagged behind, as Japan have rationed them against Korea due to their political issues). Required machines from ASML which are already banned. As long as the bans stay in place it will be near impossible for China to catch up.
Well done video I would definitely not be a long-term investor in any chip manufacturer. The fact that tsmc and Samsung don't have to worry about profit shows that subsidies can help you capture the market. But like you point out at a terrible price. But giving the current geopolitical situation of Southeast Asia. And recent years trend of manufacturers outsourcing chip production it is vital keep r&d and manufacturing ability within the United States. People often forget the greatest discoveries are made in the manufacturing process not in a University.
Market: "We just need to meet a temporary gap of about 30% due to COVID.” Chip Manufacturers: " DONT WORRY WE ARE INCREASING OUR FAB CAPACITY BY 100000% FOR THE NEXT 20 YEARS!”
I think the whole shortage was more like this: - Automakers: OMG, COVID is coming, we're doomed, cancel all chip orders. - Fabs: fine, we'll give the capacity to the next customer in the queue. - Automakers 6 months later: OMG, we have no chips, we can't build cars, CHIP SHORTAAAGE. Please give me some chips. - Fabs: stand there. - Automakers: PLEASE! - Fabs: there.
"are PC master race and miners going to buy all those chips?" Yes. Seriously, if nVidia and AMD magically found 10x the capacity just for GPUs they would still sell every_single_chip just like they are doing now. We are seeing markups greater than 100% to the consumer for these products and every GPU has an ROI of less than a year at current crypto pricing and we aren't even talking about current gen CPU which is basically vaporware for a majority of PCMR buyers. People are resorting to bundles and pre-builts just to get these components and willing to pay far over MSRP and demand is only increasing. So many people have skipped this entire generation and using chips from 2-3 generations ago waiting for the market to calm down and there is no sign of this happening until global fab capacity increases by an order of magnitude.
Strong info on foundry capacity, but not much on the demand trends. My main investment semiconductor company works very closely with their customers, monitors their actual inventory and manufacturing requirements, and is supplying most of their customers with 'just-in-time" chip supplies during the current shortage. This allows the end manufacturer to carry very light inventory. Industry evolution, enabled by global communication.
Great analysis. It's clear you were right What you didn't predict was the Ukraine war and it's absolutely massive insatiable demand for electronics. Each side is fielding tens of thousands of drones each month. Combined with sensors, terminals, headsets and a myriad of other electronics it's clear that if it wasn't for the war the industry would be in the middle of a bust. Although you have to wonder who had advanced knowledge on the Ukraine war..... Seemed like a pretty amazing bet to spend some $200-300b on fabs despite the collapse in crypto and the covid WFM situation.
Another great video. I really agree with your point on the CHIPS Act's most important element being the R&D fund. In this industry, if you're simply trying to build more of what is out there today, you don't differentiate yourself at all and won't generate any influence other than by brute force (subsidizing enough capacity to put others out of business). Instead, if you develop the technology first, you get a better hold of that technology in terms of developing the rules of the road across the world. A good example of this you pointed out in another video is how the US is able to 'influence' who is able to buy the EUV machines from ASML since they hold a foundational patent.
Well done presentation of your thought on the semiconductor cycles. I would add a couple of thoughts or questions really - - Why isn't the automation of the chip automation getting cheaper over time instead of more expensive. - Is there a scenario where demand may increase exponentially and so there is a law named after some jack-azz like "David A's law of AI chip demand doubling every 6 months"? - Also what are the economics involved in the machines that make these chip making machines.
I remember learning that most industries also have other industries which only exist to serve the other industry. In the semiconductor industry, these fab component suppliers are far less numerous and they know the trends and the industry’s willingness to pay. Because there are less players and capital expenditures are so high, there is almost no competition. Thus, the market does not behave as one would expect. They can charge whatever the market will bear, not what the customer wants to pay in a free market where they can always choose the lowest price.
"Why isn't the automation of the chip automation getting cheaper over time instead of more expensive." Not quite sure what you mean. There are different processes involved in making ICs. But, prices are going up because cost goes up. And, the smaller the process node, the more expensive the equipment is to make that node. It doesn't matter that the processes are automated. They've been automated probably as long as ICs have been made, going back to the 70s. When you fit billions of transistors on an area the size of a small fingernail, the equipment that's used to lay out materials and etch the circuits are incredibly expensive. And, it's not something where there's economy of scale. Only a small number of these pieces of equipment are made. "Is there a scenario where demand may increase exponentially" Yes, but only for a time period. The fact that TSMC 3nm is going to be about 40% more power efficient than their 7nm node means most companies that use a lot of chips, so think server farms, will replace out most of their equipment for the cost savings on their electric bill. So, just the fact that some of these new nodes are very power efficient means growth is going to be great, and the dude didn't think about this. Now, just because making a newer node becomes more expensive, the ICs made from them can drop in price, because there are many more ICs produced from a single wafer. So, pricing is not so predictable unless you study the economics of fabrication, and there are much better tech channels that do this. Also where he doesn't think so clearly about this particular story is on continued subsidies. Once a fab is built, the operational costs aren't going to be subsidized. The operational costs won't vary much between different fabs. This isn't human intensive work. So, continued subsidies? I think not. MAYBE if a plant gets upgraded from something like 5nm to 1nm, there will be a subsidy to invest in the new equipment. But not the factory since it's already built. But it would be an investment 10 years down the road, and it would be a smart investment to help with a subsidy.
You are not taking into account the ever increasing demand. Even if they hit an occasional glut in supply, extra capacity is always handy for times like now. Besides, the life cycle of most complex electronics is pretty short. Everybody are upgrading every few years. Phones, computers, gadgets etc.
If China finds a way to reverse-engineer chipmaking techs and create new ones, semiconductors will be extremely cheaper. Look at the solar industry, China entry caused solar panels becoming extremely cheap.
Not everything can be made through reverse-engineer, chip is one of them. beside chip is made with over 1000 companies work together, chip is in an idea of supply chain.
@@genuinennessbefitting4734 Again, you underestimate the Chinese Communist Party. People used to say the same thing when China entered the renewable industry, now just look how China completely dominates all renewable industries. China has already had an international supply chain within the ASEAN, so it can skid Western sanctions at ease.
i think demand will stay high, new technology and industries such as electric self driving vehicles, AI, cryoto data mining, solar and wind, smart fighter jets like F35s, smart grids ALL need semiconductors. And these tech and industries have recently gotten a boost with stimulus money from many various governments.
If you’ve enjoyed this video, watch some of the others in this playlist: th-cam.com/play/PLKtxx9TnH76QEYXdJx6KyycNGHePJQwWW.html
It is really good to see you getting such high view counts for these videos I really enjoy them. Especially the one on ASML. Which was the first one I saw.
Look pal... the supply is about 2-3 time lower than what the market needs today. Not only this, but this situation lasted for a year already and will tend to expend until 2022 making a year and half of shortage to coup over. In some market, we are talking about 10 times the demand over the supply, like GPUs. The pandemic proved that capacity is not enough and that the semiconductors industry is not just an commodity business, it is now a requirement for almost everything. You are making your analysis with the idea that the demand for semi will stabilize... when it is not. Until these fabs are built, supply will be an issue with backorders of about 2-3 years... while demand rising during that time because technologies like AI and data-center are going to keep growing. Telework is going to stay, there is no putting the genie back in the bottle here. Chips are going to be more and more integrated into vehicles. 5G will push IoT to a crazy amount of chips for intelligent cities... and let's not talk about electronics and industrial automate. You are missing the point that the pandemic shown us how much we relate on semi and this is no turning back. The matter of fact is... TSMC and the other companies, are going to sell everything they can produce. We are not in a normal cyclical episode, it is a never seen situation.
You forget about Quantum Networks. It started 2018, and yet starts to build out. IoT requers a lot of chips too. Europe invest because of reliability and legacy systems.
@@evalangley3985 True That
What I don't understand is don't you want the best wafers not the low teir or larger NM versions
From a Mechatronics engineer in a poor country with zero high tech industry (Colombia): thanks a lot for your analysis of the semiconductor industry, it absolutely blows my mind!! Much respect for Taiwan and South Korea, amazing people!!!
Thanks for watching!
You should visit Asia!!!
@@harrykekgmail i wish I could, I had the fortune to meet korean and Chinese engineers in my former workplace, great guys.
So start some. My brother in law from Colombia is high tech, but had to go to Texas to get a good job, even though he still lives in Medellin. Your country has the talent. Just stop exporting it.
🇰🇷♥️🇨🇴
I love your semiconductor industry analysis....no one else does it!
Despite the incoming flood of new fabs and semi's, it's important to note that there's an arms race right now for 5, 4, 3, 2 nm production. A lot of these new fabs aren't all producing the same products - for example no matter how much Intel invests in 7, 5, 3 nm they won't be competitive against TSMC for the next 3 years at minimum. Until we get a clear picture of who is pumping out what chip, what contracts have been signed by who, I don't think we can declare a bust as incoming.
I don't even need to look at any data. That a bust is coming is a given in an industry that is constantly at an arms race to the bottom. Every major player in this industry is constantly operating at that fine line between profitability and bankruptcy. The tricky part is to predict what straw will break the camels back - it could be anything any time. In 2005 it was the housing crisis in the US - houses didn't even use chips back then, they just bankrupted a critical portion of the market with a domino effect. It could have been the pandemic in 2020, but the industry went up instead. Could be the end of the pandemic when everybody decides to go outside instead of buying gadgets. Could be a minor skirmish between American and Chinese ships with just 2 shots fired and no casualties in the chinese sea and everybody loses confidence in the international market. Could be anything...
@@KonradTheWizzard Well, that is true for anything. There could be a bust sometime in the next 10-100 years. But, I think we're speculating the next 5-10 years, if we are over-investing and ramping up supply to dangerous levels.
@@vincent4696 I see that as opportunity to figure out how to use that capacity. I really wish there was a service I could buy custom batches of 1K chips of my own made chips (I have some ideas on how to side-step the costly engineering process, I only actually need the physical production, not the IP, I'm working on my PhD thesis on it) , which is a tiny batch at that scale, but they can make it by putting different silicon from different people, even if we have to wait months, at such low batches, we can afford it, and the factories would be over-invested and needing jobs to keep running anyway.
I wanted kind of a ChipWay, like a PCBway of sorts.
Intel has lost its way, as ARM gets into more and more WinTel products Intel will continue to see its cash cow shrink. AMD is doing a great job also shrinking their market share. Intel is still creating some good tech but not on the scale they used too.
@@monad_tcp Fabs need to produce billions of chips to be profitable. 1K chips is just not profitable especially at the sub 5 nm level.
From experience of working in the industry, the "Bust" is merely a pin prick in this massive industry
What do you mean?
There is always massive background capacity. New fabs pick up the crumbs around the edges.
@@shenjiejason8350 This world is becoming more modern all the time. The number of server farms is growing. The number of super computers is growing, the number of compute devices in people's homes is growing. AI is growing rapidly. Businesses rely more and more on computers. What is meant? Bust cycles don't last long, and this video is almost negligent for underestimating the amount of growth there is.
And then I can move on to other fields. Medicine. There's more medical equipment that's controlled by processors or at least needs a small IC in them to run it. Transportation. There's more ICs going into vehicles than ever before, and as vehicles get smarter and smarter, it means more ICs, or ICs with more power. Either way that uses more die. Robotics. Yeah that's growing like crazy. Almost any manufacturing is mostly automated.
And then there is the effect of companies like TSMC simply coming out with a new node. I can use AMD as an example and how it's effected the server world. Zen 3 EPYC, which is AMD's latest server CPU is the best X86-64 server CPU on the market. Lots of companies are ignoring their regular life cycle for replacing servers (a server, being a single circuit board (motherboard or mainboard), with either one or two sockets for mounting a server CPU, and this board slides into a server rack, and therefore a server rack holds a few servers). They're ignoring their normal life cycle plans because of how efficient Zen 3 EPYC is. The thought process is, you can't afford NOT to replace the older servers, because of the power savings)
So this is the main part where this video has no understanding about. AMD Zen 2 and Zen 3 EPYC CPUs are made using TSMC 7nm node, along with GlobalFoundries 14nm node. The CPU is what's called an SoC (System on Chip). The processing cores are made with TSMC 7nm. The I/O portion of the CPU is made with GloFo 14nm, or maybe 12nm for Zen 3, I forget which. For the next generation coming from AMD, Zen 4, EPYC processors will be made using TSMC 6nm and 5nm nodes. The cores will be TSMC 5nm and the I/O will be TSMC 6nm. And then you move to the next generation, as long as plans stay on track, Zen 5 will be made using TSMC 3nm.
So I threw a bunch of numbers at you in the above par. But what does it mean to companies that have larger server farms or any company that has a few servers? Power efficiency. Companies are moving to CPUs made with TSMC 7nm because of power efficiency. And, the way this is typically evaluated is perf/watt. How much work does the CPU per watt consumed. This is a big deal for server farms, because it's not just about the server being able to do more given a certain amount of power, but it's also about air conditioning. These server farms have HUGE electric bills. So now the great news for these companies, moving from TSMC 7nm to TSMC 5nm gives about a 25% boost in power efficiency. Moving from TSMC 7nm to TSMC 3nm gives about a 40% boost in power efficiency. So, you have a server farm with 1000 CPUs in it, let's say, and you currently are on TSMC 7nm with Zen 3 EPYC. You decide to wait for Zen 5 EPYC so you get a boost of roughly 40% in performance, while maintaining the same power envelope. That's an almost OMG!!!!!!!! kind of figure for a server farm.
So, the turnover of parts, simply because Samsung, Intel and TSMC are coming out with new nodes that consume less energy to do the same amount of work, or give you a big boost in performance for the same amount of energy cannot be underestimated. Because there are so many server farms around the world now, this is huge business for chipmakers like AMD, Intel, companies that makes processors using ARM, etc...., and that business gets back to the fabs, TSMC, Samsung and Intel mostly.
In this video the dude said he wasn't going to make predictions about this year or that year, but then basically claims there's going to be a big bust. There isn't going to be a big bust for fabs. There's going to be big upswings, then equilibrium, then MAYBE some small downswing, and then another big upswing. THAT has been the cycle with fabs. If you're talking about a specific sector such as CPUs, GPUs, SSDs, DDR memory, etc.... they can have bigger swings, but for fabrication, the downswings are small because there's always a need for ICs of some type or another. You can compare the trend of CPUs for instance to the trend of fabs, and you see even when sales have slowed for CPU sales, it doesn't affect fabs.
If you understand the business of fabrication, you should understand WHY this is the case. Companies make contracts with fabs months or even a year, maybe even up to 18 months in advance of the fabrication company making products for that company. Apple already has a contract with TSMC for their 3nm node, and that node isn't even OUT yet. That's the way it works. Apple wants to make sure they're on the most advanced node. That's why their products tend to be more efficient with power. Intel has a contract for TSMC 3nm for products they'll come out with at the end of next year. AMD has contracts for their GPUs and CPUs. Nvidia has contracts with TSMC for their GPUs for their next generation.
This is to say, I wouldn't watch this guy and make stock decisions based on it. He doesn't study markets well enough to have ANY understanding of what's happening. Go watch people who make tech news instead. What this guy has done is dig into the past to give information about how these companies have evolved over time. While this is interesting, it doesn't predict the future.
But, take all information with caution. Just because TSMC may have a big upswing in business doesn't mean that Samsung will, or Intel will (as a fab). You have to actually watch videos and read about what these companies are doing. Success for TSMC could mean the demise of another fab. At some point companies making older nodes are going to fade away as prices come down for newer nodes. It doesn't matter what equipment you make. If you can get a smaller IC that uses less energy to do the same thing your older ICs are doing, and it costs you less money, you're going to move to a newer node.
Yah. The bust is in bonuses. The businesses will March on, albeit without some of the weaker entities.
@@johndoh5182 Very good comment. Lots of good information. A thorough understanding of the company and its competitive environment is the only "edge" investors have over other investors in finding reasonably valued stocks. (Peter Lynch)
When you are shopping in the supermarket with a empty stomach, you always bought more than what you can eat.
This is no supermarket, semiconductors don't just grow on trees haha!
You mean all you cat eat buffet?
unfortunately, not true for me. fuck. I need to be buying more next time.
@@allentchang you mean all you can eat warren buffet?
Always eat before you shop.
One caveat in this is that you treat all semiconductors as commodity parts. There are MANY semiconductor manufacturers out there, especially defense, automotive and aerospace which don't participate in commodity markets. This is a subset of the semi industry which is isolated from the rest which won't be affected by the bust because they can't get their parts from the likes of TSMC and Samsung.
I've heard from the news that automakers having problems because of a chip shortage. Each car probably has hundreds of different chips, can't just a few missing chips stop the production?
@@Maxim6431 I agree. The automotive sector is now becoming a part of commercial market, commodity. Defense/aerospace is still an isolated market, meaning a small piece of pie.
Well technically all semiconductor are commodity items. Some are used in the public/private sector or the military sector. There are minor differences in the public and private sector as the private sector is usually more demanding of their products while willing to pay good money for it.
The military sector on the other hand is vastly different compared to the public/private sector. The ones used by the military usually undergo hardening and resistance processes to make them more resilient than their civilian counterparts. Not only that, there are minor differences such as material swaps for 1 small component of the whole piece but that swap makes a difference. You can't have chips coming loose at high stresses due to massive increases in g force due to sudden turning nor can it come off at really low temperatures(underwater or in really high altitude). You use special soder to make sure that the chips and other components will stay on the circuit boards.
In one of the recent vids by Asiannometry, he goes into this a little bit. But know that the companies that supply parts for the military are capitalist run so expect then to stop production if there isn't as good as a profit margin as it used to have.
@@fuckedupbody4194 "Well technically all semiconductor are commodity items." No, this is not correct. There are MANY semiconductor components which are NOT commodities.
I like how you are one of the few people I have ever heard that both understands moores law and medium to advanced economics at a rigorous level.
Thank you, it’s like to 90’s al over again. To Costa Rica!
Faith in the TH-cam algo restored, brilliant channel.
Every TSMC shareholder should follow your channel, greetings from the Dominican Republic!
greetings from somaliland
De lo mío 💪
@@JuanJDumeP Klk activoo 💪🏾
Do not invest in TSMC. They have a monopoly in the system. Instead invest in Intel.
@@mrshaun8812 Intel is yet to prove they can compete with AMD and TSMC. When they make some changes maybe then we'll see.
Brilliantly described and well done, this is a good example of market capitalism at work. As a Malaysian working for a electronic component supplier, we experienced boom last year for the demand of OEM equipment used to support these fabs overseas. Now I can foresee the potential for a slowdown in the medium term.
capitalism at work causing instability as usual
@@kefsound the continued innovation is due to capitalism
Your opinions would be more convincing if backed up by some analysis of chip consumption growth projections in absolute terms versus production capacity. The graphs you presented were only relative production increase and decrease. As mentioned in other comments growth in autonomous technologies will really drive up chip consumption for example.
besides that, datacenters want absolutely every single bleeding edge node they can get. TSMC could triple 7nm capacity tomorrow and it would still probably be completely sold out.
In my opinion, relative metrics make far more sense than absolute numbers.
Extremely unlikely to happen in near future because it takes minimum 5 years to get a Fab operational to a satisfactory degree. Most of the fab plans we hear are recent or just planning stages.
But we didn't have a shortage until recently, so if those capacities go to work fully or just partially those fabs don't need to have high productivity to create too much in sum
@@tomlxyz There was a shortage incoming. There hasn't been enough capacity provisioning to meet demand medium term. Even without the bursts in demand due to work-from-home and crypto boom(s), the capacity would run out shortly anyway, maybe not in late 2020, but in just a couple years. Every year, people need slightly more flash storage, more cloud servers, more this and more that, now fridges have computers too for some unfathomable reason.
Overprovisioned capacity is, to an extent, a good thing. Sure it just costs money when it's not needed, but not having capacity when it's urgently needed can be even more devastating. Whole industries depend on this capacity being present.
Besides, the big growth cycle right now isn't actually that big relative to the existing size of the industry.
What expert forecasting shows (real data backed by extensive research) is in part what's driving all the FAB build outs. Do you think that because it's 'hot and in the news' that a bunch of investors run off to invest $20B in a FAB. These investments are huge and not taken lightly and without a wealth of information backing them up. What would be better is for you to figure out 'why all these investments are needed' rather than 'why all these investments are not needed'.
Ummm, no. It takes a long time to develop a new node. It's expected that the factories that are being built by TSMC and Samsung will be operational in about 3 years, because none of these plants are being used to develop new nodes. In AZ for instance, that will start out as a 5nm fab. That node is already in mass production in Taiwan and it's the node being used for Apple products.
Once a plant is built, it takes weeks to get the equipment working and tuned properly. Not years.
all sources i can find see it happening no later than 3 years from now, with most saying closer to 2... and pretty reliable sources see it happening as early as late next year.
So great, my AMD photo (released under CC0) was used in this video 2:01!!
incredible video
i hope the algorithm picks it up and give it the attention it deserves
It won't, most people would rather watch dancing girls on tiktok.
Quality content man, quality! One of my favaourite YT channels!
Yep when you're putting Wi-Fi on your coffee maker, you're getting a little outrageous
We already have wifi coffeecups.
There are a lot worse examples haha
Wifi toasters are where it's at. No buttons, you have to use an app to use it.🤣
:D :D :D
How are we going to build our Mocha-Botte DOS attack network without all of those coffee makers?
A lot of wisdom for such a small TH-cam channel 👍 definitely hope you'll get big soon 😉
It is likely that the subsidies will continue to maintain technology parity since the current build-out is based on a strategic national security interest.
With regards to fab expansions, unless you have equipment planned in the build pipeline already, it can take you over a year to get a slot, depending on general business conditions. With how it is now, if you haven't planned it already, you may well hit the next downturn before the tool is delivered. Saying that, a downturn right now may just be a slackening in a generally increasing demand (hopefully :-) ).
If you are looking at building a fab, you are not looking at current conditions at all, but extrapolating out a long way - state of the art fabs take 3-5 years to build and kit out, with a likely 1+ year to bring online. If everything aligns it will be a minimum 3 years from decision to build to first silicon.
that is why with more advanced nodes there is hope that instead of die-optimised chips, there will be more fpga and neuromorphic , universal chips possible.
universality is both bad and good.
ability to have fpga for beta products at much smaller node than predicted final product (f.e. microcontroller) is something everyone looks forward to.
I was discussing this issue with some bitcoin miners just earlier today and I pointed them to your channel as they were sure high GPU prices would carry on for at least two years. Based on what I have seen on your channel I didn't buy it given the massive capital expenditure by TSMC in 2021.
Your writing is so good and well informed on these complex topics melding geopolitics, technology and economics. Are you a trained journalist or journalism student?
Crypto GPU prices and supply ... who knows. Whole new set of dynamics.
I am neither a journalist nor a student. My day job is entirely different from what I do here. And that’s good because it helps blow off stress
Fabs take years to develop, and years to build, and years before they start making chips. 3 years minimum for TSMC, see how long it took Intel to pump out their 10nm chips (close to 10 years, and not even at full capacitty)
It will be a long while before this big subsidy has any affect on prices
@@TheMasterofComment There's too much variable in saying how long it takes a fab to start producing products on a large scale. It only takes 2 - 3 years to build the plant. The amount of time it takes to start producing product could be a few months or many years. For instance the plant being built in AZ is a 5nm node factory. TSMC staff along with the makers of the equipment could come in and probably have it producing at full capacity in less than 6 months. This isn't comparable to what happened to Intel and 10nm. Intel was the company that developed and refined the processes to make their 10nm node, even though they didn't make all the equipment. But once again, for the AZ 5nm plant, TSMC 5nm is already being produced in mass. TSMC is already producing it for Apple products so it will be a VERY mature process even by the time the AZ factory is built. So, barring equipment failure, it shouldn't take long to start producing. I know there is a time period to get the equipment working properly, but that's not years, it's weeks.
google up for barkhausen noise effect in economy.
Trying to predict noise profile of something rensembling spin glass is futile task, especially that butterfly effect causes such complex system to be susceptible to external influence.
F.e. you could say you can try to create very detailed model of magnetic core so you think barkhausen noise will become predictable, and then your core gets struck with three cosmic rays, and few atoms internally decay due to natural radioactivity and bang - that is why it is called noise.
Economics have very similiar problems, though ofc. because domains can be much better artifically aligned there is more room for microeconomy to do it's job. Still whole system blows into chaos on any possible excuse.
Excellent observation, good to keep an eye on it
The only the thing I feel that was missed here as part of the demand, is the increase in performance smaller architectures are bringing to the table. On paper it is fairly easy to assume that we most likely did not create this huge amount of demand for the upcoming fab capacity in just one year, if you consider the "normal" growth rate of demand. However, performance and efficiency gains are getting much larger as we progress to smaller and smaller architectures, and this where I feel the trend in demand is not so "normal". On a smaller scale we can consider just the newest gpus, as an example of a very large performance increase over past generations of cards, and the demand for them proves that people now see upgrades as worth while. This is new territory, since for many years only small performance increases were made, which meant people held on to old equipment to realize more of its value and investment. Now we are seeing 30-50% improvements which has triggered a massive amount of demand for not just gpus but cpus as well. People now see the larger advances in performance as good reason to finally make upgrades on new generations. And with this same incentive, it is easy enough to assume that large data centers and all kinds of chip using facilities will be planning large overhauls of old equipment, in addition to their normal growth rate and expansion. Again this will be due to the fact that newer processes will offer better efficiency and higher performance that save on cost for compute power. This could be a one off event as "everyone" moves to smaller architectures then waits till the next notable increase in performance, but I think it is unlikely. Smaller processes have a lot to offer, and generational changes will continue to impress. And I feel it could very well snowball into an even more technology dense population as we begin to utilize more and more compute power. Just some food for thought! Enjoyed the video. I hope to have some good discussion in the comments. :)
Software Enginering from México here, you just got a new suscriber your content is gold
These are all good points and some which I considered myself recently. However, TSMC's new fab is not going to bring "hundreds of thousands of wafers coming to the market over the next 1-2 years". That's not even close. Samsung is simply adding some capacity in Austin. They aren't building a new fab.... yet.... and TSMC's new fab isn't going to bring full output to the market for another 3-5 years!! So even if the cycle plays out like you say it will, that puts the bust out into 2024-2026... AT THE EARLIEST! And people will in fact go nuts once the pandemic lifts. Well over a year of pent up demand FOR EVERYTHING! will be unleashed on the world economy later this year, which is before the chip shortage will even be worked out. Semiconductor prices will have a chance to boom massively well before they inevitably bust again.
But will the pandemic even lift ? or we will go into trade war with China.
There is no pent up demand for high end chips I think, if anything, people will probably spend less time on their computer/phone and go out instead. If less people work from home and less data is generated from online activities there will be lessened demand for data center services too (probably less growth not an actual decrease). IoT devices like smart mirrors and wifi armchairs don't require the same kind of chips, these might remain in high demand but it's easier to scale production of these chips so I think there would just be a short boom as restaurant owners and such start buying new stuff followed by a steep increase in production, they are just commodities after all.
Why would pandemic demand will stay after pandemic
@@semkjaer3581 It will probably bounce back, but there was already a trend for home-working.
It will never go back to the previous level.
Even if the demand decreases, it wont decrease to the same level as before.
@@monad_tcp well this guy was saying people would 'go nuts' on semis after the pandemic and I just don't see that happening, I'm not saying the semi industry will do bad, I just don't think the current excesses will hold.
A bit late to the party. As a retired EE I have not been following fab technology very closely, thanks for the timely update on the situation. It is always interesting observing the intersection of technology and economics. The rapid advance of semiconductor technology has been incredible, it never ceases to amaze me how much functionality can be built into small battery powered devices.
Having said that as a geezer I find that the technology in many new devices actually reduces usefulness and customer satisfaction. My wife and I recently purchased a 2020 Toyota Corolla. The difference in electronic content is in stark contrast to our previous Corolla of a decade ago. Some features are very useful however, many tend to impose the "will of the car" on us because the car decides what is best and refuses to operate in the way we desire. Other examples are making what was once a simple operation: such as ignition switch assy mode or parking brake activation activation (we have manual transmission cars). The keyless ignition now requires numerous operations to put the car into assy mode to simply listen to the radio without the engine running. So many steps in fact it requires consulting the 600 page owner's manual to do it. The sound system included a short term free subscription to cloud based services that we have no interest in using. Like most modern electronics the system requires a non-ignorable firmware update so each time we start the car and have to click thought the nag scree to access other functions.
Fine, I'll subscribe to your newsletter 🙄
I feel like you have a format that would translate really well into a podcast as well. Anyways keep it up 💖
Remember to subscribe to the newsletter!
This is fabulous education - you are teaching us about business cycles while talking about the most complex machines built by mankind
Would be wonderful to understand these cycles in other industries as well- steel being one example.
This is all great news for consumers. In a year or two we will finally have affordable electronics again :)
The USA needs to get rid of the import tariffs. It makes the USA exports non competitive. Stupid idea by the Trump administration.
@@GodzillaGoesGaga stupid biden keep that one brilliant Trump move
@@GodzillaGoesGaga And yet the stupid Biden Admin came out recently and said it's leaving those tariffs in place and using them as one part of a strategy to deal with China. I would tell you that the disruption to American business and consumers isn't worth it. Go figure.
Really I think the Biden Admin would get rid of part of the tariffs if Congress would get rid of the 2017 tax cut so there's more money coming into the govt., but until then the money coming in from tariffs are helping to fund a govt. that's spending a lot more money. It's just the burden has shifted from business and those making millions of dollars to consumers.
amd and nvid have realised they can charge anything. dont expect lower prices ever, will always be a new excuse.
@@stephenhartley2853 eh, price did drop to msrp
i've seen a few of your videos. despite the varied topics, can't find disagreeable fact or opinion. subbed!
The chips must flow !
They who control the chips, control the universe...
These are epic. So much information. Thanks.
There’s no crash coming buddy. Chips are gonna be in demand in everything from cars to cows.
Like there was no housing crash in 2006 because everyone lives in a house?
@@feuser1 I'm not sure, but I think houses and chips are different things... And their markets might just work very differently...
@@feuser1 comparing chips to a speculative housing market.....really...
@@LuisVasquez-nu4hu considering that a place to live is by far more important then products with chips ...
The core point is that long term and short term demand are 2 very different things, and can at a time be inverse.
@@fugehdehyou Stocks and companies very much can be and are speculative.
Super solid content. Thank you for creating.
Goood morning from australia!
great video btw I have thinking about this topic for a few weeks now.
Howdy!
THOSE were NOT "random thoughts." .... You KNOW your s**t, inside out and that makes you FASCINATING to listen to. ... I subscribed on the first video, the last one I watched, can't wait for the next one. ... (Your 53 000 subscribers will be 530 000 in no time).
Taiwan #1! Also, if there is a flood of supply, that means I might be able to get cheap and good PC components. May even be able to build up some server racks and storage lol.
you are ABSOLUTELY spot on.
This goes to show Politics shouldn't get into commerce.
Tears will flow like river in the next 5 years.
All those tech will be obsolete before they can generate a profit.
"Good enough" tech can(!) find buyers, at a reasonable price.
i appreciate the angle. it has some sens in it BUT you left out real Boom drivers and that demand in this field will never go to 0 unlike in many other volatile markets
I got to work as an engineer in the industry for 20 years, visiting TSMC, Global Foundries, and ST fabs, very interesting technologies, incredibly complex. Apple's M1 5nm CPU is a quantum leap in semiconductor manufacturing. A good story is the billions spent on 450mm that all got cancelled 10 years ago as 300mm densities got smaller and went 3D with deep tunneling etching.
The empire, long divided, must unite; long united, must divide.
It is a year later and I'm still waiting for a break. Chip demand seems to have a new lower bound - much higher than the last.
Well ASML must be very very happy right now ... just sayin' ...
@@happygimp0 :) that's the top irony of it all :)
UCTT too
Exactly, ASML has no competition, even in the foreseeable future.
I have absolutely no clue about semiconductors, electronics and their industry in general, but it was interesting to watch.
Overall I feel that the demand is going to surge further in the coming years with the surge in demand for electric vehicles, 5G, etc. TSMC is definitely going to benefit considerably from this. Looks like, TSMC is a great stock to buy in regular intervals.
IoT too
And the huge turnover of parts. Think of all the server farms that exist. You're talking about millions of ICs for CPUs along with all the chips that go on server boards. Companies aren't going to keep older servers when newer ones pay for themselves in electricity cost savings.
Just had an meeting in my job at a semi conductor factory, at least they arent planning on building a new fab right now, but are trying to finish building their current new ones and squeezing out each and every drop of production out of the existing ones. Management/Cooperate certainly are all giddy and dont appear to expect a downturn at all. Thanks for the info, now I know I have to keep looking around for some second option fall back jobs.
Your analysis have some truth about the over supply situation within 2 or 3 years. Actually there is not much chip shortage for laptop and personal computer and smartphone chips. Most chips shortage are from EV cars, autochips and some consumer products, which are low end chips, 20 nm nodes.
I dunno, gpu prices havent been cheap
@@slopedarmor That is the biggest bottleneck making it difficult to build a new gaming rig though, and the GPU shortage has more to do with crypto than anything else.
@@ManaDrain315 Crypto but the effects of material shortage cant be neglected. Supply chain disruptions have a far large impact, that crypto has contributed to but it is a minor contribution
A lot of system builders (PCs) can't build or sell products at the same level as they should be able to because GPUs are 1.5X to 2X the price that they should be at. This makes the cost of a new PC a LOT more than what it should be. So, parts like CPUs, RAM, NVMe drives, etc... are starting to drop in price because people have enough of these. People are holding off now.
For laptops, they tend to be seasonal purchases with the shopping period around Christmas being the biggest selling period. I expect there will be a LOT of laptops sold at the end of this year due to big increases in performance and battery life.
For smartphones, yes there's a slowdown because new products are about the same as older product.
But his analysis lacks a LOT of understanding because it totally misses what's going to happen in the business world, and that's where growth for TSMC in particular is going to be massive. But as I said above, growth would still be large in the consumer space if GPUs weren't so overpriced. Expect huge turnover of all things computing when TSMC is mass producing their 3nm node. It will align with a lot of change in computing and graphics and the world will be salivating to get their hands on these newer parts. This is when he's predicting a bust cycle basically, and I say that's total BS. That's going to be a boom cycle like has never been seen. Power efficiency increases of about 40% from today's NEW products?? That's not a bust cycle. That's when server farms around the world will swap out parts.
Finally! I was waiting for this video for weeks
Good morning from Hsinchu, drinking coffee opposite of TSM. I came to Taiwan 10 years ago, when revenue was 10 billions. They will break 50 billions usd this year. Bust will be temporarily.
Your analysis is so good! Thank you.
1:10 I thought I was listening to CGP Grey, that felt so weird... in a great way! That was great delivery!
This is such a good summary for the current situation, well done! I think the bigest question for the semiconductor industrie is the progress of autonomous driving and IoT. But if there is really to much capacity for chips coming in the next years, investors will have a hard time ;)
Even with autonomous cars, which have what maybe 3x the dies in them, how does that soak up all that capacity? Especially since the supposed goal of them is to reduce the number of cars needed to less than 1 per person.
I think everyone has lost their damn mind over this chip shortage and the resulting government money, the capacity they're building won't be necessary until the 2030s I'd wager (with no data to back it up mind you lol)
IMO AI learning will boom and absorb A LOT of chips.
I even suspect Elon's Dojo is planned only next year precisely to ride a favorable price wave.
@@musaran2 I rather think it will end up like any othe r "hot" booming topic in IT. Big Data levelled out FAR below what everyone expected, bicause a big pile of hay doesn't make it more likely to find needles. NoSQL has found its niche and is hiding in there while SQL is still going on. Let's go back a bit further: 5th generation programming languages / visual programming - nobody is even talking about it any more because it was a bad idea to begin with.
AI makes sense for a limited set of problems which are currently dominated by humans because normal algorithms fail, but which are repetitive enough to automate. Like driving, image recognition, voice recognition - classic AI (neural net) topics. How many of those do you think you need? Of those: how many can you afford? Currently everybody believes that AI will solve all the problems that Big Data failed to solve - context sensitive classification problems - but those need not just data or a dumb learning algorithm - they need actual intelligence. Engineers. Knowledge. Experience. An AI has about as much experience as a new born ant.
AI will not absorb the chip capacity. The pattern with all those hot technologies is that they are promoted as a "silver bullet" - the way to solve all your problems and free you of the obligation to think (or for managers: to hire people who can actually think, but cost money). After 70 years of IT we should now understand that silver bullets don't exist.
@@KonradTheWizzard AI can improve itself unsupervised, that is the game changer.
AI feeds on data, if strives on underused big data.
AI has already solved BIG challenges. I can't think of a trade fundamentally safe from AI takeover, not even art.
The salary of a human worker can pay for a nice robot, and the salary of a human trade can pay for a lot of AI training.
@@musaran2 We are at least a few decades, more likely centuries, away from what you call AI. You will not meet Lt. Cmdr. Data within your life time. Not even his brother B-4.
What we have right now stupidly learns whatever set of data you place in front of it. Your success depends on your selection of data as well as the exact layout of the network you chose. Even if it is an unmitigated success it can merely classify data that lies within its learning parameters (recognition networks) or make very simple decisions based on a (hidden) polynomial function (feed forward net or fuzzy logic) as long as you stay within the definition interval of that function that you know very little about. If you get outside the functions definition interval the AI will malfunction in surprising ways.
It can not improve itself substantially. It may be able to consume more data and refine its internal function, but you have absolutely no clue whether the AI has the same idea about "refinement" as you do. You need to supervise it.
You can play tricks in the lab by reversing the function and make your AI "dream" of weird shapes that are derived from the dataset it learned. But this is neither intelligence, nor creativity. It is mathematics playing tricks on you. I admit that the images released by Google are intriguing to look at.
I don't want to take away from what we realistically CAN do with AI. Yes, we will be able to create automated image and voice recognition systems that help us with a lot of tedious and repetitive work - Alexa and Siri are just the beginning. Yes, we will be able to create self-driving cars. But we still have a lot of really hard work in front of us to get this done and AI is NOT able to magically find real patterns in data that human minds assisted by statistical software are confounded by - human minds are still tens of thousands of times more intelligent than any AI. And we have no clue how to get from current AI to brain - it will take a LOT of work and experimentation.
You can not replace an engineering degree and human ingenuity with a bunch of data fed into a complex polynomial function with a name that sounds like magic. No matter how much you wish it to be true.
Remember: we will have a functioning true AI in about 20 years - we've been saying so since 1950! And we are experts! We know. (Full disclosure: I did my degree on a neural network simulator system in 2000.)
Great analysis and reporting. Very informative and much appreciated.
Good evening from Costa Rica
Good evening!
Read a bunch of comments, that I agree with, you concentrate on the obvious users of chips, computers, cars and consumer tech, but miss the fact that all industrial machines from the largest to the smallest all use thousands of chips. And they are just the tip of the iceberg of other chip users. Nothing sold today that connects to mains or batteries is free of chips. Even the battery packs themselves have chips in them.
Who else is watching and reading comments at the same time 😊
Bitcoin is The future, lnvesting in it now will be the wisest thing to do especially with the current rise
I bought my first car with the profit I made from it
2021 online stock market is just difficult and unbelievable, l rather invest my money on crypto
For real it's true, This could be the wisest word l've heard, you got my points 👍
Nice words, Yes especially with the current Rise 😉
This channel is new to me. I'm impressed by both the range of videos it has on the semi industry and the detail it provides. It's boggling to think that about three decades ago, there was debate as to whether features could be constructed smaller than 100 nanometers. Yet now, it sounds like 2 nanometers is within grasp. Approximately 3 nanometers is the point at which some of those advocating nanotechnology (e.g. Eric Drexler) claim features could be atomically ASSEMBLED by migrating atoms/molecules instead of using ablation to selectively remove material from a bulk solid (like a wafer) or relying on a uniform vapor deposition
It took 7-9 years for EUV, actual transistors aren't smaller than 40-50 nm. \TSMC 7nm is just a marketing term, it's bigger than 3x that
@@harshivpatel6238 I'm no expert in the field. But from what I've read, what is represented by the number (e.g. 5 nm) is the size of the smallest feature, not a complete transistor. The line width is perhaps the smallest feature. Wikichip says the following ..."TSMC started mass production of its 5-nanometer N5 node in April 2020. TSMC considers its 5-nanometer node a full node shrink over its 7-nanometer process. In early 2021 TSMC plans on introducing a second version of its N5 process called N5P which provides additional performance"
@@gregparrott while it's certainly exciting, you should add 6 months for a fab node to mature after it's announced, and 6-12 months for products made on that to reach various parts of world.
@@harshivpatel6238 As an end user, about the only significance I attach to it getting so small is the purported reduction in power consumption. As I understand it, Apple's new M1 chip runs extraordinarily cool, and is the result of both RISC architecture and small feature size. The M1 Max has 57 billion transistors and up to 64 gB of integral memory.
@@gregparrott M1 has a lot of accelerators for different work loads, CPUs & GPUs are more of a general purpose hardware, especially CPU. It's hard to be as good as M1 when you target general purpose use cases.
Good reminder to look to the future and not be too suckered into the flavor of the month, lol
Insightful, good support and reasoning. Thanks for this video 👍
I doubt there will be a Bust anytime soon. I have been waiting 6+ months for Nivida 3080. At this rate I might as wait for Nividia 4080. Also on ebay Nivida 1080ti are still going for 500 dollars!
My old video card now is 2 times more expensive than 4 years ago. I don't think there will be a "bust" anytime soon. Also I see RAM and HDDs are vanishing from our hardware store. There are still some models but something is definitely happening. There are less and less models also the prices are going up.
Big difference between a cutting edge enthusiast GPU and chips used for say cars or drones or other more mundane things. That’s where the glut will be according to this logic. We’ve already seen cycles like this in memory - remember when ssds few years ago got up to USD 100 for 250gb?
@@dugowf766 yeah but there arent lots of new ancient node fabs being built, and demand for them is (or will be) vastly more than supply
@@suntzu1409 these things are cyclical, and older fans are being built in places like China and intel are building in Europe. In 3-4 years time there will probably be an oversupply, that’s how these things go
Great report. Thanks.
It takes 2 years min to build a fab. And I mean minimum. They started roughly 12 months ago. We wont see shit start to get better for at LEAST another year. A year after that we may see a 'glut'.
I’ve been thinking this for a while too. Each factory is also getting more and more productive with time. When new fabs are built on this scale those productivity gains will be applied to the new fabs and output will be insane. Market won’t know what to do with all those chips.
If we have a war then all bets are off. War will likely break out in Taiwan and all that capacity goes to zero. Plus the need for chips for missile systems will send demand skyrocketing.
It’s 50/50 at this point and all depends on what happens with the China/Taiwan situation.
is Arizona a smart place to build one of these things considering how much water they use?
It will not matter to the people of Arizona until they’re told they’re not allowed to water their lawns. Then it’ll be pitchforks and torches.
Love ur videos bro.. love from India . these are real informative n helpful
Who will win? The consumer 😎
The crossover I did not expect.
It depends. If there are a lot of foundries going bust because of this likely overcapacity that may result in the survival of couple of them, which will results in duopolies if not monopolies. If that is the case, the consumer will not win in the long run.
@@IotaEtaSigma Just buy on sale or second or third newest of graphics cards released. Only fools preorder, and only fools spend more than msrp
Consumers have never been bailed out
@@IotaEtaSigma And a second order effect, this could reduce investment in future generations of chips too, also bad for the consumer
Adding the 'smart' to clothing & jewelry & everything else, tools, guns, vehicles rolling data centers *autopilot* 5G, everyone including homeless people with smartphones & smartwatches // prices at every scale to keep everyone hooked on their twitter feed / like mainlining information hard smack style with a screen syringe // TH-cam videos of whatever hits the spot like movies or music streaming // more cloud compute & server action // cryptomining was only the start, the blockchain idea for security improvement coming to nearly every industry, just like NFT's are trending as non fungible tokens // the zeitgeist that electricity demand will steadily increase over time as more people get online // you know what I mean. Excellent analysis even if its just your thoughts! Your right, boom & bust, its not just chipfabs, its whole economies
If deep learning continues to deliver results while demanding ever more powerful hardware support, we will not see an oversupply of chips coming from the latest generation of fabs. If deep learning fizzles out, as artificial intelligence has done in the past, we'll definitely see an oversupply and crash in the semi market. It's good to understand the overall market fundamentals covered by this video as well as the specific differences in the current cycle which may result in a deviation from past behavior.
Deep learning cannot "fizzle out." That would be like the PC market fizzling out. It can mature, but even that is years away. Deep learning has way too many useful applications to fizzle out, just like the PC has way too many useful applications to go away. The market for the *current* applications of deep learning is huge. Most companies/industries that can benefit from deep learning haven't even started to implement it. And we haven't discussed the yet-to-be-discovered applications of deep learning....
@@briancase6180 "If deep learning continues to deliver result...". IDK man, maybe try reading my comment again?
@@GeezerTuber ok. Whatever.
Semis don't have to stand on their own two feet, that's a rather simplistic way of thinking. The cost of a shortage simply has to be greater than the cost of the subsidies.
Semis are so integral across so many industries these days, that permanently subsidizing their production and ensuring a constant glut may actually be cheaper than dealing with natural shortages. You see the same thing with the energy industry. Energy is so important to the function of the economy that it is worth constantly subsidizing energy production rather than dealing with the economic fallout of energy shortages.
But we forget that use of electronics is getting increasing day by day...we can see how more electronic we are using than 5year back. Everything is getting smart...hence upward demand will absorb new capacity also....
As a holder of a whole lot of SOXL shares, this is very relevant to my interests. Going to watch the entire playlist now.
burrr
@@user-gu1hl2kx2k Hey, I'll take it however I can get it these days
I also hold soxl at a high price, is it gone and won’t come back?
@@evesolis6133 It'll come back. As long as people want more computers and phones, SOXL will inevitably rise.
@@evesolis6133 its gonna go back up
Intel fabs in the US have stood on their own feet for more than 50 years. They were making silicon when most people in asia were still riding around on bicycles. I'm pretty sure others can stand on their own feet once the initial investment is paid for. Most will be put directly to work on US military contracts. They don't really need to be competitive in the consumer space. Much like steel, fabs are deemed a national security asset.
That make no sense, building a multi billion cutting edge fab just to sell chips to the US military makes absolutely no economic sense, unless is a small fab. The US military is a small buyer of chips compared to commercial costumers. By example Intel sales to china totaled 20% of their yearly revenue compare that with the sales to US military and will make more economic sense to build a fab in China.
@@elchippe The US government doesn't want its fabs in China. And they will pay any price to make sure they stay on US shores. It considers communist China an enemy. And Taiwan to be too close to it and in danger of invasion. They recently paid TSMC billions to construct a fab in Arizona. And have refused to let any Intel fabs relocate overseas. They don't care if it makes economic sense and consider it a national security issue. They plan to subsidize it with taxpayer money.
@@ComfyShortz They can put billions if they want but they will have a 10 billion dollar fab idle doing nothing, but at the of the end of the day the company will sell the fab and move on.
ALL the companies being subsidized to build new fabs in the US, which right now I know of Samsung, Intel and TSMC make billions a year in profit. There's no such thing as running subsidies just because. The subsidies help out with the HUGE upfront cost of building a fab. After that these companies make their money by producing products, via contracts or making their own products. This video was a propaganda piece, which is a shame because many of this guy's videos are very good. This was utter garbage. It was simply an attack against the subsidies, and using ZERO information to project future growth and then suggesting a "bust" will happen. But NO trend data supports this. If anything trend data shows the inverse, for fabrication.
Now, what trend data WILL show is that some segment of electronics can have overproduction, such as CPUs, GPUs, SSDs, DRAM, HDDs, etc.... and this is because many companies make CPUs, GPUs, SSDs, DRAM, etc... But fabs don't make products just because. They make products based on the contracts they have to make products, and it's some other company who sells finished products that can go through boom-bust cycles.
TSMC for instance has had revenue growth every year since 2009.
Very interesting view and commentary. I have seen this in many industries over time.
So if the biggest fab goes to subsistence level by overproducing and oversupplying the market, and that causes the smaller fabs to go under, I am guessing the big fabs don't shed tears.
Hey! I can see my windowless yellow office in this video. That's fun.
There’s been a significant increase in at home schooling and stay at home work. Some may go back, but many businesses see the value of stay at home work. Why spend hundreds of thousands on office space and equipment when you can offload that responsibility onto your employees? More money to invest elsewhere in your business.
A lot of older employees are refusing to go back ‘in to the office’. So, yup.
Good video, never mind instagram and selfies. The chip market is the centre of the universe now. Thank you
A lot of people are agreeing with you in the comments so I thought I’d give my two cents. Descent analysis but I’d say we’re far from seeing any kind of crash. GPU prices stayed crazy high during the three year slump in crypto when it basically wasn’t talk about (2017-2020). With online gaming, TH-cam only getting bigger, companies revamping to cut costs, etc chip prices and demands will stay for a long time. These badass 7nm chips are still brand new and 5-2nm chips are only going to make everything better. The leap will be huge and everyone will want what’s faster. Speed and efficiency is going to be the key to who is the next long term super power, China knows this which is why they’re working so hard to get more chips. Governments need faster and bigger super computers to stay ahead and not get hacked. I think (I have a PhD in physics and have been investing for a few years now) that we have at least another 10 years before the tech improvements become minuscule to the point where the demand drops off.
I forgot to mention cloud storage. That’s one of the biggest needs for chips that an average person doesn’t think about because it seems like magic. More and faster cloud storage will continue to drive the chip market forward for another 10 years if not longer. We’re only in the beginning of our generations industrial revolution.
@linkzable I don't believe anything a Chinese company says they can do without an insane amount of evidence (just look what happened to HSMC). Also, there's no way known to make those chips without EUV. Like the physics doesn't exist. So no US doesn't have to worry about China right now. TSMC is going to kill Intel not China.
you are smart, I am subscribing to you man. Keep up the good work. ;)!
Superb insight, thank you! As an archivist, I’m an advocate of magnetic tape
Thanks asianometery! As an EE, I'm hoping for a surplus, instead of all the high-cost limited availability components right now...
Yes, it's infuriating! Run-of-the-mill always-available-cheap items have randomly dried up!
Lower end fabs will definitely face a lot of stiff competitions soon, but I assume advanced chip fabs will take longer to build so they have more time to make their money back.
I’ll have a video later about this but in general it takes 2-3 years.
Congratulations, you're actually following the subject and not drivelling on about bitcoin bullcrap!
@@Asianometry so 2023-24 would be the time so consider selling AMAT TSM AMD shares
The lower end Fabs are not idle. They are busy making chips for lower demand product lines. Eventually, they will retire, but not until they extract as much profitable life as possible.
@linkzable Nothing in life is guaranteed, but doing something is better than doing nothing. Also, just having TSMC and Samsung engineers are not enough. There are required chemicals that only Japan makes which they can ban (that's one of the key reasons Samsung lagged behind, as Japan have rationed them against Korea due to their political issues). Required machines from ASML which are already banned. As long as the bans stay in place it will be near impossible for China to catch up.
Well done video I would definitely not be a long-term investor in any chip manufacturer. The fact that tsmc and Samsung don't have to worry about profit shows that subsidies can help you capture the market. But like you point out at a terrible price. But giving the current geopolitical situation of Southeast Asia. And recent years trend of manufacturers outsourcing chip production it is vital keep r&d and manufacturing ability within the United States. People often forget the greatest discoveries are made in the manufacturing process not in a University.
Market: "We just need to meet a temporary gap of about 30% due to COVID.”
Chip Manufacturers: " DONT WORRY WE ARE INCREASING OUR FAB CAPACITY BY 100000% FOR THE NEXT 20 YEARS!”
I think the whole shortage was more like this:
- Automakers: OMG, COVID is coming, we're doomed, cancel all chip orders.
- Fabs: fine, we'll give the capacity to the next customer in the queue.
- Automakers 6 months later: OMG, we have no chips, we can't build cars, CHIP SHORTAAAGE. Please give me some chips.
- Fabs: stand there.
- Automakers: PLEASE!
- Fabs: there.
Enjoy this video very much. Indeed the over-capacity is a concern in the not too far future.
"are PC master race and miners going to buy all those chips?" Yes.
Seriously, if nVidia and AMD magically found 10x the capacity just for GPUs they would still sell every_single_chip just like they are doing now. We are seeing markups greater than 100% to the consumer for these products and every GPU has an ROI of less than a year at current crypto pricing and we aren't even talking about current gen CPU which is basically vaporware for a majority of PCMR buyers.
People are resorting to bundles and pre-builts just to get these components and willing to pay far over MSRP and demand is only increasing.
So many people have skipped this entire generation and using chips from 2-3 generations ago waiting for the market to calm down and there is no sign of this happening until global fab capacity increases by an order of magnitude.
I doubt it. Time will tell.
Strong info on foundry capacity, but not much on the demand trends. My main investment semiconductor company works very closely with their customers, monitors their actual inventory and manufacturing requirements, and is supplying most of their customers with 'just-in-time" chip supplies during the current shortage. This allows the end manufacturer to carry very light inventory. Industry evolution, enabled by global communication.
hello They seem to be pretty well thought out " random thoughts" m
Great analysis. It's clear you were right
What you didn't predict was the Ukraine war and it's absolutely massive insatiable demand for electronics.
Each side is fielding tens of thousands of drones each month. Combined with sensors, terminals, headsets and a myriad of other electronics it's clear that if it wasn't for the war the industry would be in the middle of a bust.
Although you have to wonder who had advanced knowledge on the Ukraine war..... Seemed like a pretty amazing bet to spend some $200-300b on fabs despite the collapse in crypto and the covid WFM situation.
Costa Rica in the comments right on! Loving asianometry
Buenos dias from Taiwan!
Another great video. I really agree with your point on the CHIPS Act's most important element being the R&D fund. In this industry, if you're simply trying to build more of what is out there today, you don't differentiate yourself at all and won't generate any influence other than by brute force (subsidizing enough capacity to put others out of business). Instead, if you develop the technology first, you get a better hold of that technology in terms of developing the rules of the road across the world. A good example of this you pointed out in another video is how the US is able to 'influence' who is able to buy the EUV machines from ASML since they hold a foundational patent.
Well done presentation of your thought on the semiconductor cycles.
I would add a couple of thoughts or questions really -
- Why isn't the automation of the chip automation getting cheaper over time instead of more expensive.
- Is there a scenario where demand may increase exponentially and so there is a law named after some jack-azz like "David A's law of AI chip demand doubling every 6 months"?
- Also what are the economics involved in the machines that make these chip making machines.
I remember learning that most industries also have other industries which only exist to serve the other industry. In the semiconductor industry, these fab component suppliers are far less numerous and they know the trends and the industry’s willingness to pay. Because there are less players and capital expenditures are so high, there is almost no competition. Thus, the market does not behave as one would expect. They can charge whatever the market will bear, not what the customer wants to pay in a free market where they can always choose the lowest price.
"Why isn't the automation of the chip automation getting cheaper over time instead of more expensive."
Not quite sure what you mean. There are different processes involved in making ICs. But, prices are going up because cost goes up. And, the smaller the process node, the more expensive the equipment is to make that node. It doesn't matter that the processes are automated. They've been automated probably as long as ICs have been made, going back to the 70s. When you fit billions of transistors on an area the size of a small fingernail, the equipment that's used to lay out materials and etch the circuits are incredibly expensive. And, it's not something where there's economy of scale. Only a small number of these pieces of equipment are made.
"Is there a scenario where demand may increase exponentially"
Yes, but only for a time period. The fact that TSMC 3nm is going to be about 40% more power efficient than their 7nm node means most companies that use a lot of chips, so think server farms, will replace out most of their equipment for the cost savings on their electric bill. So, just the fact that some of these new nodes are very power efficient means growth is going to be great, and the dude didn't think about this.
Now, just because making a newer node becomes more expensive, the ICs made from them can drop in price, because there are many more ICs produced from a single wafer. So, pricing is not so predictable unless you study the economics of fabrication, and there are much better tech channels that do this.
Also where he doesn't think so clearly about this particular story is on continued subsidies. Once a fab is built, the operational costs aren't going to be subsidized. The operational costs won't vary much between different fabs. This isn't human intensive work. So, continued subsidies? I think not. MAYBE if a plant gets upgraded from something like 5nm to 1nm, there will be a subsidy to invest in the new equipment. But not the factory since it's already built. But it would be an investment 10 years down the road, and it would be a smart investment to help with a subsidy.
You are not taking into account the ever increasing demand. Even if they hit an occasional glut in supply, extra capacity is always handy for times like now. Besides, the life cycle of most complex electronics is pretty short. Everybody are upgrading every few years. Phones, computers, gadgets etc.
If China finds a way to reverse-engineer chipmaking techs and create new ones, semiconductors will be extremely cheaper. Look at the solar industry, China entry caused solar panels becoming extremely cheap.
Not everything can be made through reverse-engineer, chip is one of them. beside chip is made with over 1000 companies work together, chip is in an idea of supply chain.
@@genuinennessbefitting4734 Again, you underestimate the Chinese Communist Party. People used to say the same thing when China entered the renewable industry, now just look how China completely dominates all renewable industries. China has already had an international supply chain within the ASEAN, so it can skid Western sanctions at ease.
Any particular chip that you were talking about? There’s a lot of different kinds.
With a lot of different techniques to achieve quality and yield.
Perfect and very informative! Thanks a ton !
i think demand will stay high, new technology and industries such as electric self driving vehicles, AI, cryoto data mining, solar and wind, smart fighter jets like F35s, smart grids ALL need semiconductors. And these tech and industries have recently gotten a boost with stimulus money from many various governments.
this is an amazing analysis. I am Subbed.
Good morning from SG!
@Junehong Tan lmqo
@Junehong Tan You're just jealous.
Excellent analysis.