How compute and AI will create next-gen superapps

Danny Crichton:

Dylan, thanks so much for joining us.

Dylan Patel:

Thank you for having me.

Danny Crichton:

So when I think about what your beat is with semi-analysis, it has to be one of the most exciting places to be in the world right now. There's so much news going on, most prominently probably the US government's investment in Intel now two, three weeks ago. But that's not just it, right? Everyone from the large cloud hyperscalers, to the materials companies, to US, China, new types of architecture, everything is coming to a head right at the peak here of the compute era. Give us, from your perspective, what are the top priorities that you are seeing right now? What are people not paying attention to?

Dylan Patel:

I think the biggest thing that people aren't paying attention to is, as revenue ramps for these labs and they continue to complete these massive rounds, whether it be OpenAI's $30 billion round, or Anthropic's most recent round, which is well north of $10 billion, is when you combine those two, it tells you what their spend is going to be for compute.

And so, in the case of OpenAI, if they're hitting $20 million error by the end of the year, if Entropic is hitting $8 to $10 by the end of the year, that plus whatever they raise indicates actually their exit rate for a compute spend is that number plus capital that they've raised that they're spending going out the door. So that also means that hey, by the end of '26 OpenAI could be spending anywhere from 50 to 70 billion on compute. Anthropic can be spending $30 billion on compute by the end of next year, which is a humongous step-up, right?

Each of those companies will have more infrastructure deployed obviously through their cloud partners than the clouds themselves had deployed as a whole when ChatGPT was released. So the level of spend here, the level of infrastructure build out, it really is mind-warping, but you get a lot of folks saying, "Oh, there's no ROI here." But actually, the ROI is clearly there in terms of look at the revenue curve and the inflection of it, the run rate that the inference is actually profitable. So whatever dollars they're making from services and inference are actually being spent on compute of course, but a big portion of that is actually being round tripped to a training compute.

Not all of it. A lot of it is being spent on the inference compute as well, but if you've had access to say Anthropic or OpenAI's fundraising rounds information, you can see that their inference margins are positive. So there is an ROI here on inference spend. I think this is what a lot of folks are missing is that it's not just a purely speculatively fueled bubble. It's actually a lot of it is based on there's ROI here.

Shahin Farshchi:

I think Dylan is spot-on. I think there's definitely part of the investment that's going to be going into these companies going out the door to build the infrastructure that's required for this, not only for the training and the inference. I think it's important to understand that ultimately is going to be unity economics these companies will live and die by. And so the more efficient their spend on the compute, the more capable they are in being able to predict their compute needs and to be able to buy from the right vendors, and we'll talk about that later, then the better position that they're going to be in for in a very competitive environment.

Danny Crichton:

One of the questions I have here, obviously we just two, three weeks ago we had GPT-5 go out. And one of the most important facets of GPT-5 is it's really not just one model. It's really a router across multiple models, ones that are very compute efficient, some that are sort of in the middle of the performance curve, and then Deep Research and others that are extremely performance intensive.

One of the interesting dynamics here I think is to what degree do you think chip efficiency drives the cost curve down for OpenAI, Anthropic, the other LLM models versus algorithmic improvements, not just to the inference model, but to these routers being able to route these queries to the most efficient and best performing pieces of their LLM stack?

Shahin Farshchi:

I can take a stab at that. Ultimately, it's hardware and software that go together for building these kinds of advantages. Mosaic, which we funded back in 2020, they had the assumption that if hardware moves forward with Moore's law and creates efficiencies on the order of Moore's law, algorithms can provide orders of magnitude improvement. And so the expectation there was hardware and software working together to build these improvements.

I think ultimately you should be thinking of the problem holistically. You made a comment, Danny, on chip efficiency. I would just take that a step further and focus on system level efficiency. You can make the chips super efficient, super fast, but if they're not being utilized, then you're not getting a return on your investment.

And so a lot of the more recent innovations are not only around chip performance but also memory performance, system level performance and ultimately, chip utilization or investment utilization, and I think there's a lot of opportunity there still to go. NVIDIA has done a great job with Blackwell, but I think there's greater improvements to be made. And there's a lot of great startups out there, a few of which that we've backed that are going to help achieve those efficiencies.

Dylan Patel:

I think the other aspect of this is it's really hard for these companies to balance their infrastructure today because of how bursty the traffic is, whether it be kicking off Deep Research queries, whether it be Claude code usage being so high for certain hours or GPT search or whatever product it is. Volumes of all things on the internet undulate a lot throughout the day, but AI is perhaps even more spiky than many other parts of the ecosystem.

Usage of say a YouTube application or Netflix is far less spiky than even AI usage because a lot of it is used in productivity. And so this spikiness of the utilization is extremely impactful for the infrastructure needs. And then when you layer on the fact that because you're constrained, especially on service style plans, say $200 a month or $20 a month type plans for monetization or even free plans, you have a double whammy there in terms of, well, if you're not paying per use, and then in addition you have all the spikiness of traffic, how do you deal with it?

For OpenAI, it was adding the router and that way they can route people to high compute, low compute based on the type of query they're asking, but also based on the infrastructure needs. Or Anthropic, if you're using Claude code at night, they actually upgrade you to Opus all the time, but if you're using it during the day, they actually make you use Sonic. And they have these hourly rate limits for that specific reason, is so they can try and balance the infrastructure.

Another way of doing this is, Hey, OpenAI offers batch query where you can get half price of the API if you batch it. All of these things are focused around let's try and reduce the compute intensity of the highest traffic hours and let's buffer up the compute utilization when it's nighttime. And so that's a big challenge as well.

The nice thing about AI is that a lot of it is not... Ultra low latency is sensitive, right? It's not like search. It's actually, hey, if I use a thinking model or if I'm using these sorts of applications that aren't immediate response, actually have some time to ping a server on the other side of the nation or in West Texas where the internet's not even that great, or all the way to another country, in fact.

So that's another aspect of this infrastructure build out that's really noteworthy. It doesn't need to be in US-East-1 in Virginia servicing, pinging directly to your other servers in US-East-1, right?

Danny Crichton:

Well, and if you go back in history, I mean I don't know if it's apocryphal or not, but the argument for Amazon Web Services was essentially Amazon had this massive surge of traffic during the Christmas holidays between Thanksgiving and Christmas, where they almost had to double and triple their infrastructure. And then it was just vacant for months at a time and what do you do with all this hardware that you're not using for the other 11 months of the year?

The argument was, well, we can make that a platform, we can sell it off. We can balance the traffic out there. Don't know if that's a real story or not, [inaudible 00:07:15] so to speak, but I do think you're getting at something that's really tough, which is real time. And I think we haven't had a good conversation about this in a while.

But one of the arguments for one of the Grok and others who have been trying to focus on this is how do you balance between the kinds of AI algorithms that you can batch, you can do overnight, "Hey, go through my data, I want to report the next morning. It's okay if it takes a couple of hours"? Versus I'm on a call with a customer support agent, I need an answer in the next five to 10 seconds or I might hang up and not make the trade, not make the buy, lose it as a profit center. How do you balance across what needs to be real time, what can be batched and afford that within software and API layers?

Dylan Patel:

I think that's a critical question that's hard to answer. The OpenAI had to do so much work to get their voice mode to work. If you look at other voice modes out there, for example, Grok's, Annie is another one, right. One way they solved it is they make the beginning section of the query actually be a dumber model. They can respond with faster, so you immediately get a response, right?

A lot of times, at least even with humans, I throw ums and yas in the beginning of a conversation, which lets me, if it's a voice agent, it's the same thing. Throw some ums and yas and, "Oh, that's a good point," so that you can actually think about the answer and that's the sort of intelligence that needs to be built into these models.

Just like we're talking about with the router for OpenAI, send it to the thinking model when you have low traffic, send it to maybe even nano if you're really constrained on traffic or they've hit rate limits. It's got to be the same thing with these real-time applications. But then at the same... If it's a direct voice agent, then maybe you route it to the dumbest model possible or on device model initially before you get the really high intelligence.

If it's a robot, maybe the high intelligence is doing the route planning, but the on-device model for the robot is actually doing the fine motor feedback. If it's a customer service agent, there's got to be other intelligent ways to do this where you're actually combining multiple models, multiple types of infrastructure so you can get that latency down.

Because you're never going to get, let's call it Grok 3 or Gemini Ultra or GPT-5 Pro level intelligence in real time. There will always be that cusp of max intelligence where a use case can get unlocked and then you want to be able to layer it in with how do I achieve the minimum latency?

Danny Crichton:

Let me pivot the conversation because I think GPT-5 is opening a new era of development for these companies. OpenAI is almost hitting 10 years. It was founded back in 2015. It's crossing the chasm from this research lab to now an extremely popular consumer technology company. They have a product that's used by hundreds of millions of people, arguably a billion if you include a lot of very soft users. That's the fastest growth of any consumer app in history. Facebook took probably 15, 16 years to get to that sort of point. Now another doubling, Google similarly.

And so there's a huge question of as these companies are optimizing the models for performance and efficiency, trying to drive actual EBITDA, how do they start to adapt the consumer applications? Because we saw a new CEO come into OpenAI who is going to bring the consumer DNA into those systems.

And I'm curious, how do you think the company adapts over the long haul? We haven't seen, for instance, advertising that comes into this. And I know this is something that Michelle has brought up a little bit in our own internal conversations, but do you see an evolution of the business models here or is this subscription of it's free, 20 bucks, 200 bucks locked in people's heads?

Dylan Patel:

I think you can already see for certain stuff it's changing. So for example, Cursor had to pull back on their subscription and push forward paper usage because the variability of users was really high. Anthropic, same thing. They had to introduce all these... Initially I thought, "Oh, $200 a month. Easy." Then people used it way, way, way more than they thought. There were people on Reddit showing $20,000 worth of usage in a single month versus the 200 for the plan. And so they had to introduce these rate limits and they've had multiple iterations of introducing more rate limits to bring down the cost.

So obviously pay per usage is really, really the hope that you can get to. You pay per usage, you get some margin on that, but at the same time, the subscription model is the one that customer understands more. "I get access to it for X price," rather than, "I don't have no clue what I'm going to end up paying. I'll just pay whatever."

So I think the monetization method is there's a lot of work to be done here, especially in the value capture side where people aren't capturing the value. So I think it's truly an open question as far as $20 a month, two hours a month, they're going to release a $2,000 a month thing and I'm going to buy it for every single person who works for me. I just know that's going to happen because it makes sense, right?

Cloud code, we have the subscription plus the usage model attacked on both. So if you get the rate limit you go for the subscription, which is then you go to the usage-based. Probably would just use usage-based except that the subscription is them subsidizing me. It's the venture investors in Anthropic subsidizing me. So it's like I love it, even though their cloud code is a very profitable for them. Anyway so I think it's tough to say exactly where it comes from and where it goes, but I think this is a key aspect of it. Yeah.

Michelle Fang:

To double-click on the business model, Dylan, do you see a world where AI interfaces like GPT-5 become super apps, that end up both serving ads but also becoming that platform that facilitates purchases? We see most of this coming out of China for super app somewhere that's being generated, but if that's going to happen in the US, what would be privacy UX have that platform aspect that's going to evolve for users?

Dylan Patel:

Yeah, yeah. I think that's what we made a post about a week and a half ago, two weeks ago, which is GPT-5 sets the stage for ad monetization and super app. And then it was clickbait because it sets the stage for ads, but we're actually not talking about them doing ads.

Michelle Fang:

It works. Yeah.

Dylan Patel:

But it's not them doing ads, it's actually just them creating a super app. And so many people saw that title and angrily DM'd me like, "They're not going to do ads, we're not going to do ads." It's like, no, but that's how you clickbait people to reading it. And then you explain the super app, which is their current CEO of applications built AI monetization via purchasing on Shopify. Add AI purchasing on Shopify.

And what is a super app if not, it can do everything for you? So you open up WeChat and it actually lets you book an Uber and it lets you book a nail salon appointment and it lets you get a haircut and it lets you also pay people and it lets you buy things and it lets you chat with people. Obviously, it lets you do everything and AI, you're going to do everything through AI. That's the hope. That's what Meta wants to do. That's what OpenAI wants to do. That's what you're going to, whether it be contact your friends, send emails, buy stuff, be influenced, whatever that outcome is, is everyone wants a super app.

That's the reason why Zuckerberg is building, he built Meta Reality Labs, right? He's tired of, and he invested so much into it is because he doesn't want to be at the whim of anyone else's platform. He wants to control the platform. Think maybe he didn't recognize it was AI that is the interface between... Is the next sort of platform. It was terminals and it was AD and GUIs, or mouses and keyboards and GUIs and then it was touch. And what that device is has also transformed, but the next stage is AI as the interface between humans and machines.

Basically that AI is a super app at that point because it does everything. It sits between you, whether you're booking a flight or booking a nail salon appointment or buying something new or researching something. It is a super app in its own. And so then you can take monetization on take rates. I joked with one of the co-founders of xAI at a dinner the other week. There was a bunch of lab people and they were grilling this one specific co-founder, like, "Why are you doing this? Do you really think this is good?" Blah, blah, blah, blah, right? Obviously all the criticisms against Annie's AI companion is what we'll say. And their whole point was the AI has to be good at emotional intelligence and empathy as well. And I'm like, "Okay, that's what Annie's doing."

I think if you gave Annie the option of delivering something to your home, it's like, "Oh, hey, I bought you a pizza or I bought you something even more explicit," the user would love it and then you take a rate on it. I think this take rate model, has to be the easiest way to monetize the user. You give them your credit card information, you authorize them to take X percent of any purchase and then that's it, right? Just like an Uber does, just like Airbnb does.

Danny Crichton:

May ask you, I mean obviously we've had a huge amount of focus on the application layer. We're talking about the infrastructure layer. There's also just an innovation layer that goes across all of them, which is all these new chip companies. So we're in a world in which NVIDIA right now, I'm looking at the market cap is $4.35 trillion, one of the most valuable companies that's ever been created in the history of humanity. But at the same time, you see this massive expansion of investments into a whole slew of additional companies that are trying to specialize in either parts of the iStack, trying to compete directly within NVIDIA or just find their little niche.

Michelle was formerly at one of those prominent companies that has been in the news quite a bit. I'm curious, when you look at the startup landscape right now, obviously semiconductors has traditionally been very hard as an onboarding on an ramp. It takes a lot of money going through tape out, getting into its factories, getting it produced, but there's so much money and so much focus here that it seems like some of this is getting upended. What are you seeing from the innovation layer?

Dylan Patel:

So I think the big challenge is like, hey, people want to do everything and capture the entire value of say an NVIDIA and compete directly with them. But at the same time, if the infrastructure spend as a whole is skyrocketing, right, the majority of that goes to chips. Over time the hope is that a lot of this chip spend moves away from NVIDIA to other vendors, whether it be Google's own chips, Amazon's own chips, Meta's own chips, et cetera. You do start to get more proliferation, OpenAI's own chip, they have a chip team, this proliferation of different types of chips. And then you could be a vendor into them as well.

It doesn't necessarily mean that you have to be directly selling to the biggest users of compute, right? If you're a chip company and your goal is to try and sell to OpenAI, compute to them or Anthropic, then it becomes a challenge where OpenAI is their own chip team and Anthropic has really deep collaborations with Amazon's chip team. Why not try to sell directly to these customers as well? That spend is going to be huge there.

And so there's a lot of technologies that are on the cusp there as well, whether it be in networking technologies, memory technologies, different types of power delivery. There's all sorts of different hardware that is a supplier into this ecosystem as well. Now obviously NVIDIA has massive gross margin dollars and the other hyperscalers have less. But as that percentage of spend grows to other vendors, non-NVIDIA, whether it be the hyperscalers or AMD or maybe even a startup, although doubtful, the proliferation downstream to other components also means you have a lot more customers to sell to.

And the sales cycle, while still long, is much easier because your customer is very sophisticated. Of course, it's still the same problem of concentrated sales. It's at least somewhat reasonable to have many customers there. And it is a smaller surface area of innovation that you need to... You're going to break through innovation in one small vertical area and sell to many customers rather than having to innovate across the entire frontier that NVIDIA is on.

No one starts a new company and says, "We're going to do search." Actually Perplexity did, but it's a different topic, right? No one says, "We're going to make a new..." Actually, there's a lot of companies doing social media too. I don't know, I'm trying to think of something that is realistic. Okay, no one makes a new company that says, "We're going to make the Amazon," right? That is just not a thing people do.

Danny Crichton:

Or you say as you do a Shopify to Amazon. So Amazon is a one-stop shop, and then you create a platform that empowers a whole ecosystem of new shops, or something like that. You take a different model, a different approach. I'm curious-

Dylan Patel:

Correct. And so Shopify was very successful, so why don't you do this with chips, right? There are many customers. NVIDIA can be your customer, AMD can be your customer, as can all these other hyperscalers. And even Sovereign for projects and efforts, right?

Danny Crichton:

Shahin, I'm going to direct this question to you, but I will point out because it just came out a couple of days ago with NVIDIA's results that NVIDIA's top two customers represent 39% of its revenues according to its latest financials, which is an unbelievable level of concentration. But Shahin, when you think about the startup ecosystem here, obviously customers are scaling up around NVIDIA, around CUDA, some of these built-in lock-in technologies, but there does seem to be some opportunity just given the level of spend and the fact that customers do want to diversify. They don't want to have a monopoly provider in their data centers. They don't want to be holding to only one architecture, only one firm. How are customers picking from a wider ecosystem of companies?

Shahin Farshchi:

Well, customers are spreading the net very, very wide. In the case of the Mag 7 or the Frontier Labs and the Hyperscalers, they have their internal efforts and they keep talking about those. They constantly evaluate any startup that moves just to be ahead of everything that's coming out of there.

But one point to keep in mind of the previous conversation is that a lot of startups like to talk about unseating NVIDIA, but you don't necessarily have to unseat NVIDIA to build a massive company. Go back 10 years ago, AMD went from a single-digit billion-dollar market cap company to a double-digit billion-dollar market cap company just by taking single-digit percentage market share away from Intel.

And so, you can very much build a massive business today just by taking some market share away from NVIDIA. And the way you do that is by understanding where there could be potentially gaps in the market and also obviously, become a viable second source to NVIDIA for a lot of these customers that don't want to have a lock in effect with NVIDIA.

Now, that said, it doesn't mean that if you offer a product that's significantly cheaper to own and cheaper to operate that you can't take serious market share away from NVIDIA. And there's a handful of companies that are trying to do that. A lot of folks need to be mindful that there's only so much you can do to keep in mind that the ultimate performance of a transistor is not something that's dictated by a fabulous semiconductor company, is dictated by the foundry and the technology and the process.

And what you're doing as a startup that's building on traditional processes is simply optimizing something at the expense of something else. And so if you can predict where the market is going, if you can predict where workloads are going, the nature of those workloads as you guys were talking about earlier, then you can potentially build something that is significantly better in terms of ROI, ultimately because that's what customers ultimately are trying to solve for.

Now, there are a handful of companies that are using quantum, that are using photonics, that are using analog compute, that are trying to fundamentally change the math behind how transistors behave and getting significant efficiency gains. But those are longer term projects. In most cases they're science projects, although some of these companies have raised many billions of dollars and have higher likelihood of succeeding.

But to the points Dylan made earlier, there's many ways to play this game. It can be a vendor into NVIDIA or into any of these companies that are building accelerators. I still build an interesting business. In fact, packaging and test is an area that people haven't been paying attention to that is now becoming extremely important. In fact, if you look at the AI chip scarcity, it's driven by packaging and integration moreso than the semiconductor itself. When people talk about the scarcity of memory, I mean memory is the ultimate commodity. It's the packaging of that memory into high bandwidth memory that is the limiting factor here.

So there's a lot of ways to skin the cat. There's a lot of games to play here. The sexiest game obviously is going directly after NVIDIA and unseating NVIDIA. I think realistically, obviously at Lux, we are all about the super ambitious founders that want to do that, but realistically, you still can build a really interesting company simply by taking market share away from them.

Danny Crichton:

Dylan, you and I, we obviously have talked a bunch on the technical and business. We overlap on a lot of this, but you and I have a massive disagreement, which is the purchase of USG into Intel, 10% shares. From my understanding, based on the TBPN stack chart, you are among the most in favor of the US government investing in Intel, protecting this asset for the us.

I think it's a catastrophe. And I won't speak for Shahin or Michelle, both of whom have their own experiences here, but I have never seen state capitalism in the United States prove successful. You are a little bit more optimistic. Tell us why.

Dylan Patel:

Okay, so I think it takes a bit to sort of get to why, because I don't like state capitalism either. I think that's a terrible idea. We win by being an economy that's competitive and an economy that is capitalistic. And there's certain ways to do state investment into industries that actually improves that, right? Look no further than Chinese auto. It is the most cutthroat competitive market in the world. Yes, there are some SOEs, yes, there were a lot of subsidies, but actually it's extremely competitive, which is why prices are falling so fast and which is why they're out-innovating any of the traditional auto OEMs in the world. And that's why they're taking all the share in foreign countries.

Yes, it's partially, but a lot of it's because they're out-innovating and out-competing because they have such fierce competition domestically, much more fierce than any other auto industry in the world. And so when you look at fabs, obviously I would love for that to be the same case, but we've got some different realities. A is the CHIPS Act, it was a... I don't know. What's your opinion on CHIPS Act generally, right? Subsidize fabs a little bit and get them to be built here. Subsidize them enough to make it a little bit better than building fabs elsewhere. What is your general thought on that one I guess? Let's start-

Danny Crichton:

Look, I mean in some ways, Chris Miller, George Schneider and I wrote a paper pre-Biden administration, which in some ways became the CHIPS Act and parts of that legislation were directly from that paper. I don't think our number was chosen, but the regional innovation hubs was pulled from our paper. So I mean generally in favor. Generally the idea is moving the economics around, but that was not about ownership of companies.

Dylan Patel:

Yeah. So I think when you take a look at the CHIPS Act and how it was administered, there were certain things that had to be done. Intel has to ramp to a certain level of volume on their fabs to get the money. And likewise, Samsung, same. Likewise TSMC. There's no question TSMC is going to ramp to that volume in their fab as much as they kick, moan and scream about it being too expensive. That's really just to get subsidy dollars.

At the end of the day, the cost differential between Taiwan and the US is not that large. And so a pretty small subsidy was able to build, get a pretty large additional investment in the US, much larger than was required or was granted to them. Right. Now with Intel, it's slightly different. Intel economically did not make sense for them to build Ohio Fab, but they lobbied for it, got the grant. The grant was the largest of all the companies. Partially it was debt, partially it was a grant. Guaranteed loans is what I should have said. But Intel will not ramp to that volume because they didn't end up getting an 18A customer.

Now, 18A is not a complete failure. They're actually manufacturing some chips right now. New laptop chips will launch with it in the next six months. From everything the OEMs are saying, the DELs, the HPEs, the Lenovo's of the world, Asus, et cetera, the chip actually looks okay, right? Looks better than the last generation. Looks competitive with AMD, who's on TSMC process technologies. It's not a bad note, it just wasn't able to win customers because a lot of the stuff that's been leaked through the media, for example, one media outlet was saying, "Broadcom mostly just hated that it was hard to design with."

Others have said that the yields are not as good as expected, but it's not completely behind, right? Intel is only a few years behind TSMC at most, and Samsung's even further behind based on all information out there. So Intel is a number two, but this industry is very, very capital intensive and very, very difficult. And the switching cost is absurdly high.

If you're in this race to build whatever chip in whatever market, do you just dedicate engineering resources to switching foundries to use a slightly worse node with higher risk on yield or do you invest on improving the architecture? Because you have these stacking S-curves, Moore's law and the slowdown of it. You only get X percent improvement from node, but then you get a larger improvement from architecture and then you get an even larger improvement from software. There's these stacking curves and it's like you might as well focus on whatever's further up the curve, as much investment on architecture as possible. And then the next for the chip rather than switching design processes and saving a little bit on wafer cost.

And so Intel has this really hard problem of they must leave for TSMC to get any customer business, otherwise it makes no sense. People don't care about diversification really, even though it's just too much. You're not going to tie a hand behind your back because of supply chain diversity. Companies like Apple love diversifying their supply chain, but they single source literally just one component, which is TSMC. Everything else they've diversified, even if it means they pull back a little bit on performance, but not that component. Screens, audio chips, all these things are diversified and it is like this with fab.

So then you're stuck. You're like, "Okay, Intel is the only American fabulous. It's the only company in America that's doing process R&D. It's the only company out here." They are behind, I would say not absurdly so, but they are behind. It is extremely strategically important that there are fabs in the US. Taiwan invasion is a true risk and a true possibility. There are many parts of the US government that believe that it happens this decade, but many folks are outspoken into thinking it'll happen in the next five years, whether it's like-

Michelle Fang:

People think 2027, 2028. Right?

Dylan Patel:

A lot of people think this, right? Whether it's a full on invasion or it's a lot of pushing KMT and doing that kind of... How would I say this? Informational attack or it's just a blockade until they subvert to some demands in China. How hot it is a question, but a lot of people think it's happening in the next five years.

So if TSMC in Taiwan is no longer under control of Western markets, i.e. most of the stock ownership is by American investors actually. And no longer what happens now if some disaster happens. TSMC in Arizona is actually stuck at a certain level. They're not going to continue to advance and process technology. Within four or five years, China's own fabs will be ahead of TSMC if TSMC just disappeared from the world. Within four or five years, Intel and the Samsung will be far beyond where TSMC was.

So what do you now? Because the process R&D only happens in one place in the world for each of these companies, it only happens in Shenzhou, Taiwan for TSMC, it only happens in Pyeongtaek and then it's Hillsborough for Intel. If Taiwan is gone, TSMC's fabs are broken. Best case they freeze, even though they get a lot of materials from Taiwan still actually. But at best case, you still get those. They're frozen at a technology level. Intel's faltered already. Now what? By 2028, if Intel does not get more money, I think they're bankrupt.

And so you sort of have this really challenging problem, but effectively you must subsidize them for geopolitical reasons. Now, how do you subsidize them if the method that we created the CHIPS Act, they're not going to meet the obligations? They're not going to ramp Ohio in time. In fact, they can't even afford to ramp Ohio because their business has lost share faster than they expected, their design side of the house.

I personally don't think the design side of the house should even be subsidized. There's no reason for that. There are American companies that do better than them on every vertical, whether it be CPUs and AMD, accelerators in NVIDIA, networking hardware and Broadcom and Marvell. And you go down the list, automotive, Qualcomm, NVIDIA, et cetera. In terms of self-driving cars, Mobileye, FPGAs, AMD, every part of the company has competition on the design side.

But the fab side, this needs to be subsidized somehow when the fastest mechanism was converting the failed obligation on the CHIPS Act to ownership stake. And a lot of this ownership stake only happens if they sell the foundries, by the way. So it's a little bit, the media reporting is not exactly nuanced enough to capture all these. But it is a very difficult thing to get away with, but you have to subsidize them. Or geopolitically America is really screwed in, even if it's a 10% chance that that Taiwan gets taken over and I think or some sort of turmoil happens in Taiwan in the next five years. Because otherwise Intel is bankrupt/has cut so much that they can't continue process development in the next two years or three years.

Danny Crichton:

Shahin, Michelle, I know you have comments, so what are your thoughts?

Shahin Farshchi:

Listen, this is all symbolic. This is all political. Ultimately, Intel doesn't need to be owned by the US government. There are many smart investors out there with tons of access to capital. And there is a strategic reason as to why we're doing this for all the reasons that Dylan said.

If you look at TSMC or these other great institutions, a lot of them have been supported by governments. There's a lot of these things just take a lot of time and a lot of capital. In the case of manufacturing chips in America, it makes absolutely no economic sense to do this. It's all for geopolitical purposes and it's a cost-benefit question. We're basically purchasing a very, very expensive insurance policy here.

And the question becomes, okay, how much are I willing to spend on this? So far, the amount that we've spent so far and that people have in mind have seemed to justify the cause, but this is not going to be limitless. At some point, the benefit will not outweigh basically the premiums that we're paying for this insurance policy.

Danny Crichton:

Michelle, you were on Capitol Hill for a period of time, AI advisor in the Senate. What's your thoughts?

Michelle Fang:

I would echo what Dylan has said and throw a question back, but I think for CHIPS and Science when it came to be, I think back in August of 2022, 2023, when they're rolling it out and actually setting up the chips in the chips office, the investment office actually look at these applications, I think it's because it's really hard to actually build fabs at a competitive timeline here. I think there were a couple of reports from different think tanks that's like if you make a fab and in Japan it's actually cheaper and you can build in half the time roughly.

And so I think when TSMCG tried to build a fab in Washington, I think in 1997, they built it and it was called WaferTech, but they actually weren't able to make it competitive and then actually grow that business. And so I think when they thought about coming back, I think the only way it would make sense for them would be if the US would actually subsidize them.

And so maybe that would be one component for TSMCG and that's not enough for Intel, so we have to come up with a custom package for Intel. I think I'm aligned that I think we actually do need to have a fab domestically here. I think that is something that the US needs, the geopolitical cost of that may outweigh the cost and the competitiveness. But I think the question that I have is that's like CHIPS and Science 1.0. We saw the $52 billion package play out. $39 billion for manufacturing, 10 to 12 billion for R&D.

I don't know if the R&D parts actually going to be allocated, Dylan, but question for you is do you think that there is enough political will for a CHIPS and Science 2.0 what's happening to the R&D side? A lot of those is actually supposed to be allocated for packaging and whatnot, but do you think that there is going to be an evolution of this? And if so, what is roll of the state in the next couple of years?

Dylan Patel:

Yeah, I think talking about Shahin's point of run, what is the ROI here? It's an expensive insurance policy, is his statement. And I guess again, how do you economically think through this, right? If there is a 10% chance that Taiwan is no longer US influence and is now completely influenced by China, and they can start to do what we did to them, i.e. cut off Huawei, instead of cutting off Huawei, they decide to cut off NVIDIA. What is the impact to the American economy? Right?

Holy shit, it is way, way more than 10% times NVIDIA's revenue, which is $200 billion. It's a lot more than that even, right? NVIDIA's revenue is 200 bill. It's actually like 240 bill this year or something like that, right? Somewhere in that range, 10% of that number is more than we gave Intel already. And then you add on Google and all these other companies, you add on the knock on effect of all the AI spend, right?

Actually AI infrastructure spend is not just the 240 billion NVIDIA revenue. There's also the data centers and the CapEx for all these other things. Laying fiber, all this. Actually, oh, the AI impact to the economy. The GDP is four or $500 billion this year because of all the capital investment. Okay, 10% of that? Well then in one year it's 40 billion, $50 billion.

Our expensive insurance policy is actually quite cheap relative to... Again, 10% is a number I just made up. Government people know way better than what the risk is of losing Taiwan. I have no clue what that risk number is, just that people treat this very seriously. So you stack that on, and it's obviously this administration versus last administration, they would have very different ways of implementing this insurance policy. But I think both sides of the aisle agree this is a challenge.

Trump's method for getting a TSMC to invest a lot more is threatening tariffs on them. There is his method for getting Apple to invest a lot more in supply chain here and for NVIDIA to invest a lot more in supply chain here is threatening tariffs on them. And so you see this already. You see Foxconn, who makes all of NVIDIA's AI and boards in Taiwan, or they make all of the ones that they make for NVIDIA in Taiwan is making massive facilities in Guadalupe, Mexico as well as multiple facilities in Texas and elsewhere in the US.

Same with all the other ODMs that work with NVIDIA. You see this with SK Hynix making a memory facility for packaging, HBM and in Illinois. You see this for TSMC in Arizona. Their second announcement is not tied with any CHIPS Act money. It's just, "Let's go to the White House and let's make this announcement."

Now, is it all going to come to fruition or not? Will Trump be out of the... Will those policies still be there? Will these threats of tariffs still be there? Is the timeline for manufacturing semiconductors fast enough or manufacturing the fabs or building a facilities fast enough to where the threat of tariffs is there? Will you have actually spent all this money? Is TSMC just going to make a bunch of empty factories, which is a very small portion of the cost relative to the full fab and then never fill out the fab because the new administration is in power now? These are all questions that matter, but his method is threatening tariffs. His method is other things like this, whereas the last administration was more about direct money. But there is a sovereign wealth fund that he keeps talking about.

So whatever the mechanism is, I think there's a lot more appetite for investing in companies domestically. I just don't know if... And I think the ROI calculation, if it's 10% is obviously worth it to subsidize AI infrastructure industry in the US. Part of that can be through subsidizing in this insurance policy, whether that's through threatening people to build factories here with tariffs or it's giving them subsidies to build factories here for assembling servers of building boards, building power modules, building fabs of all kinds.

These are all... Building memory. All of these things are part of this insurance policy that we need. And so I guess I don't know the mechanism because politics are impossible to understand. If I spent time trying to understand them, I wouldn't understand politics and I definitely wouldn't understand semiconductors either.

Danny Crichton:

So that's why everyone in DC feels so stupid. I will say on a more positive note-

Dylan Patel:

Feels?

Danny Crichton:

... I think that-

Dylan Patel:

No, I'm kidding. I'm kidding.

Danny Crichton:

I'm trying to be gentle to many of our good friends, but I will say I do think the conversation has gotten much more robust over the last year. I remember during CHIPS and Science and that discussion was always about leading edge chips. A huge part of even our report and others three, four years ago was to say, "Look, you have legacy chips. You have nodes that are not on the cutting edge, not the leading edge, but are critical to radar, that are critical to cellular communications, that are critical to all these other applications, particularly in areas like defense, aerospace, et cetera." And none of that ever came up.

And now I hear when we go to DC in 2025, people have a much greater understanding of, "Look, there's a lot of ingredients that go into chips. There are a lot of different use cases and applications. You want a robust supply chain around all of those. We do need to focus on packaging." 2, 3, 4 years ago, you could not get anyone to care about packaging. It was the lowest price margin part of the supply chain. It's very labor-intensive. It's a very unhappy part, which is why it tends to get outsourced even in the 1970s to Hong Kong and Singapore and other parts of Southeast Asia. It's still being outsourced. It's just now being outsourced from them to other places.

But now I do see a dawning realization that there's a lot of complexity here. You want more of that supply chain locally and domestic. And it's not just leading edge. And so that to me is significantly more robust. And then Shahin and I, we were just talking to one of the major defense primes just last week, and that was one of their huge concerns as well, is how do you build a full soup to nuts durable supply chain within chips, particularly for aerospace applications, where in many cases, those parts ultimately are getting sourced from China? And those parts would be used in any conflict whether China or not.

Shahin Farshchi:

So you touched on a lot of things there. I just wanted to emphasize a couple things. First of all, last time I checked, there was plenty of capacity on these legacy nodes for analog, for power, for these other applications outside of cutting edge compute, which is what we're talking about here that's relevant to AI.

On the packaging side, you're absolutely right that packaging has not been sexy historically. What we're observing now is that there's a lot of IP that's going into packaging that's enabling the manifestation of this next tier of compute. Because you have chiplets, you have high bandwidth memory. You have these reticle size chips that are stitched together and packaging becomes a real bottleneck. And again, a lot of IP goes in there, which is why suddenly people are paying a lot of attention to packaging where they didn't before. But that's for these particular use cases.

And so it's really important for folks to divide the world of semiconductors and the related supply chain into legacy semiconductors and supply chain, which in my opinion is pretty well multi-sourced. And then there is this next generation of AI where there is a heavy dependence on TSMC and this surrounding ecosystem of companies doing packaging and tests and HBM and whatnot that can create serious bottlenecks there. And that's a relatively recent phenomenon. Let's not forget that GlobalFoundries is the former IBM and they're all over the world, and there's plenty of domestic production, but they are just not equipped to be able to deliver tomorrow's AI chips or whatever we're working on today.

Danny Crichton:

Well, we've gone around the world and across the stack, but Dylan, you have now created one of the most popular publications on substack and around the world around semiconductor. It's a topic that I think was in arcane circles of academia and a couple of engineers somewhere around Sunnyvale. Now more than 150,000 subscribers, a hugely read publication. I'm curious, just as we close out here, in the last couple of weeks, the article that you wish more people had seen, even though you're already very popular.

Dylan Patel:

Yeah, I would say we left substack nine months ago, so there's nine months of growth from there.

Danny Crichton:

There you go. There you go.

Dylan Patel:

Yeah. Yeah. But it's tough to say what I think more people saw. But I think there's, for example, I think the GPT-5, super app, all that conversation we talked about. But I think the other one that I think people haven't really seen enough, I think is on robotics, right? There is a lot of hype around robotics. Everyone's just like, "Oh, humanoid," blah, blah, blah. It's devoid of information at the level of hype with Figure and all these other companies that don't deserve to... Whatever.

Anyways, there's a progression that's going to go through on robotics, and that's because hardware as well as software are developing at different paces and rates. And I think that one, I think is the one that I would have other people read as this robotics level of autonomy. We had a lot of help from some external collaborators as well, such as Nico Cimelli and Rob and Joe, and also we have an internal analyst who's focused only on robotics.

But I think that's the one that's like, let's tame down some of the super hype. Let's also at the same time talk about how robotics is going to be massive, massive, massive for many different classes of jobs. And at what point those sorts of levels of autonomy kick in based on hardware readiness, based on software readiness. And what are those things? Because the term for humanoid robotics, obviously you can put infinity, but how do you actually chart to get there, right? And what's missing, what's there already, et cetera.

Danny Crichton:

Well, Dylan Patel, founder, CEO, and Chief analyst at SemiAnalysis, the must-read guide for everything going on in semi-conductors and chips, thank you so much for joining us.

Dylan Patel:

Thank you, Dylan.

Shahin Farshchi:

Thank you.

‍