Scott Caravello: Absolutely. You know, we have also just touched on it so many times on this podcast, right? Because it is a topic that hits on so many important parts about AI, whether it be geopolitics, national security, trade policy, industrial strategy, what have you. So, today we can devote a whole episode to the topic and cover the chip market itself, along with the policy landscape that sits beside it.
Katherine Forrest: Okay, I just want to say one thing, which is that this episode, I'm afraid, is not going to be the topic of really good cocktail conversation at your, whatever bachelor party you're going to tonight or wedding. Because while exciting to you and me, I'm not sure that the rest of the world really wants to hear about chips, except for our audience who is waiting for us to get going. So, let's do it. Let's do a quick level set on the basics of chips to remind our listeners. So, when we talk about an AI chip, people are usually, they're usually referring to something called an accelerator chip, which can either be a GPU, which is a graphics processing unit, or a more specialized type of chip or processor called an ASIC, A-S-I-C, which is an application-specific integrated circuit. So, regular CPUs, which are central processing units, which are the kind of chips that you have in your laptop or computer, they tend to take steps one at a time, or they process things sequentially for the most part. But AI models require performing lots and lots, sometimes millions of mathematical operations simultaneously, really in parallel. So, GPUs, which were originally designed for video game graphics, which also, by the way, require massive amounts of parallel processing. It turns out that those GPUs have the same kind of architecture that's very useful for AI.
Scott Caravello: Right, and so that's the GPU side of things. And then we get to the ASICs, which are those chips that are purpose-built for AI from the ground up, like Google's TPUs, which stand for Tensor Processing Units, or Amazon's Trainium chips that take it a step further by optimizing specifically for the type of math that AI models use.
Katherine Forrest: Right, so with the foundation out of the way, let's dig in a little bit more. And, of course, we have to talk about NVIDIA because NVIDIA is the most talked about player in AI chips. And data centers around the world, soon, as you know, I'm obsessed with orbital data centers. As soon data centers will be in outer space. They can't seem to get enough of NVIDIA GPUs.
Scott Caravello: Yeah, and so much so that Nvidia is estimating that it's going to sell a total of $1 trillion worth of chips based on a few of its latest architectures this year and next year alone.
Katherine Forrest: So, think about these numbers, a trillion dollars worth of chips. I mean, we didn't used to talk in those kinds of numbers, a trillion dollars worth of chips. Anyway, while Nvidia is strong, there are now other companies that are becoming strong as well. So, the landscape is shifting.
Scott Caravello: Yeah, and I think what's so fascinating about the whole thing is that the shift isn't coming from the usual suspects like AMD or Intel, who are the typical rivals in the chip space. It's the companies like I was just mentioning, Google and Amazon, who are actually Nvidia's own customers.
Katherine Forrest: Right, the hyperscalers, you know, those, like the AWS, Azure, Google Cloud, and the model developers, which also include Google, Amazon, Microsoft, Meta, they are all designing and deploying some of their own custom ASICs. And the reason for that is pretty straightforward. First, there is a desire to reduce dependence on a single chip source. And second, there is a ton of money to be made from selling those chips. Think about what we were saying just a minute ago about the trillion dollars in sales that Nvidia can expect or is expecting in '26 and '27.
Scott Caravello: Yeah, for sure. So, let's walk through a little bit of what they're actually doing. And so I mentioned Google's TPU, for example, which isn't really a new chip, by the way. They first launched the TPU back in 2016. And their latest generation, called Trillium, which is the sixth generation of the TPU, is available and deployed at massive scale across Google Cloud with over 100,000 chips deployed.
Katherine Forrest: Right, and the point that I made about revenue, so Google is no longer just using the chips for itself. Anthropic announced a deal to purchase up to a million TPUs from Google worth tens of billions of dollars to build one gigawatt of computing capacity. That's tens of billions of dollars for these TPUs from Google. So, think of that. Think of how much revenue Google will be taking in based upon chip sales. And Meta, reportedly, signed a multi-billion dollar deal to rent Google TPUs as well. So, there is a lot of money flowing to Google for its chips.
Scott Caravello: Yeah, and these chips that are being put out by customers and hyperscalers are really impressive and performing really well. They're manufactured at the three-nanometer process, which is where a lot of the cutting-edge AI chips still are, though two nanometers are being rolled out. And just to provide—
Katherine Forrest: Wait, I don't even know what a nanometer is…?
Scott Caravello: Oh, we’re going there.
Katherine Forrest: Alright, you're acting like the whole world knows what a nanometer is. I don't know what a nanometer is.
Scott Caravello: We're going there. Basically, the nanometer measure has historically described the distance between the transistors that are on the chip. And, right, those are the things that actually amplify and switch the electrical signals. So, when you start talking about getting the distance between them smaller and smaller, that allows the chip to run faster and more efficiently. Basically, the smaller the distances, the more advanced and complex the manufacturing process gets. But, and we'll get to this later, that all has an impact on the export controls that are applied to that tech.
Katherine Forrest: Right, and Microsoft is also making a chip. Let's talk about the Microsoft chip. I mean, it seems like we've now got these major hyperscalers also in the chip business, which I find fascinating.
Scott Caravello: So, Microsoft's Maia 200 chip was announced in January of this year, and that's focused specifically on inference, which is the workload that is actually serving AI to end users, making the models run, right? As opposed to training the models. And they've already deployed that chip in their Azure data centers, and it's being used to power certain OpenAI models, as well as Microsoft 365 Copilot.
Katherine Forrest: Right, and then there's Meta, and they made a very significant hardware move in March 2026, revealing four generations of its MTIA chips, the 300, the 400, the 450, and the 500. And what's interesting about Meta is that all four of those chips are built on, as is typical with Meta, open architecture, so there are no licensing fees. And, so, that is, I find, really fascinating. Those chips, the Meta chips, are being manufactured by the Taiwanese company that has the four letters TSMC. And TSMC manufactures about 90% of the world's most advanced chips, like those that we're talking about here. And I just want to pause for a moment as we all think about the geopolitical environment and people sometimes wonder about Taiwan and the US position vis-à-vis Taiwan and just pausing on the fact that TSMC, a Taiwanese company, manufactures 90% of the world's most advanced chips. So if you're ever wondering about one of the issues that makes Taiwan so important, think about that.
Scott Caravello: And I think one other, sort of, background note that might be helpful there actually is that because we're throwing this number 90% around and so folks might be thinking, how does that square with talking about Nvidia and the fact that, you know, they are such an important player in the market and you have all of these other huge players in the market. And that's because TSMC is actually the one that is manufacturing the chips for them. They have what's called a foundry, which makes chips for a lot of different customers. So anyway, moving beyond the hyperscalers, there's competition. I'm sorry, I won't say competition. Take a step back. And so moving beyond the hyperscalers, there's also a startup landscape that's specifically focused on chip manufacturing. And that's really worth covering too. And so those are companies like Cerebras, Groq, and that's Groq with a Q, not a K, so to be distinguished from xAI's Grok model, and SambaNova. And they're all taking novel approaches. Cerebras, I think, has a really interesting one to flag because it's focused on wafer-scale chips, which means single chips with four trillion transistors—
Katherine Forrest: Four trillion transistors, four trillion transistors at a wafer scale, right?
Scott Caravello: Exactly, so super, super thin, right? And so what does that actually mean? You have this thin wafer-scale chip, right? It's one thin chip that's dinner-plate sized. And again, by doing that, the idea is that it allows all of the electricity and signals to move faster across the chip.
Katherine Forrest: Right, and, as we know, let's just go back to Cerebras—with these wafer-scale chips with four trillion transistors—it just IPO'd this month.
Scott Caravello: Yeah, absolutely. And then on top of the Cerebras and its IPO, Groq has secured a 1.5 billion investment commitment from Saudi Arabia to expand delivery of its inference-focused chips. And then SambaNova unveiled its SN50 chip in February 2026. And they're claiming speeds that are five-times faster than competitive chips for agentic AI workloads. So there is a lot happening in this space outside of NVIDIA and the hyperscalers that's really worth paying attention to.
Katherine Forrest: Okay, so we've talked about all of the different companies making chips, some of whom are perhaps unexpected, the different types of chips, the various ways in which they're architected to move faster. And it's pretty incredible. But, you know, here is a fundamental fact. There is no single country that can fully manufacture and design a cutting-edge AI chip independently. And you mentioned just a moment ago that TSMC in Taiwan actually manufactures roughly 90 percent of, well I mentioned that it manufactures 90% of the world's most advanced semiconductor chips, but you also then mentioned its foundry. But the supply chain for a single chip, even some of the other ones that we've been talking about, can cross, you know, international borders back and forth over 70 times before reaching consumers.
Scott Caravello: Yeah, and even so, despite all that back and forth, right? When we talk about the concentration in manufacturing these chips in one place, it creates geopolitical risk for businesses that are relying on the chips for their operations, whether it be AI-specific or otherwise, because Taiwan sits 100 or 110 miles from mainland China. So, an attempt at reunification and a resulting disruption to Taiwan's chip exports could have a serious impact on global GDP and ripple through every advanced manufacturing sector for years.
Katherine Forrest: And through the AI development world. So, the United States has pursued this two-pronged strategy of one, export controls, which have gotten progressively tighter over the last couple of years. And these controls restrict the export of advanced AI chips and semiconductor manufacturing equipment to China and have a goal of limiting China's ability to develop or run AI models at scale.
Scott Caravello: And then the second prong of that strategy is on-shoring. So, through the CHIPS and Science Act, TSMC is building fabs, which is the semiconductor term for plants, and it's shorthand for fabrication facility, in Arizona. And Intel received major subsidies to expand its domestic production. And the idea behind all of that is to reduce the nation's strategic dependence on Taiwan.
Katherine Forrest: Right, but how effective have the export controls been in preventing China from accessing the most advanced chips? And actually, I want to pause because that actually has this embedded assumption, which is that China is getting the most advanced chips from the outside and not the inside. We actually don't really know the full picture of what China is able to make itself in terms of advanced chips. But according to a December 2025 analysis from the Council on Foreign Relations, the known, and I want to emphasize that again, the known performance gap between US and Chinese chips is significant and it is growing. But again, we may not have the full picture. The best US chips are currently about five times more powerful than Huawei's best offerings as one example. But that's their known best offerings. And by 2027, the gap is projected to widen to 17 times. But I really just want to suggest that we don't know the whole picture of what's going on in China right now.
Scott Caravello: And so then I can caveat the other example that I wanted to give by again saying that this is just based on what's known. But another one of China's advanced chip makers, SMIC, is making chips that are at seven nanometers because of the restrictions that have been put in place by the U.S. and its allies on the actual semiconductor manufacturing equipment used to make the chips. And again, by way of comparison, NVIDIA's chips are being manufactured at three and four nanometers, which like I had also mentioned before is cutting edge.
Katherine Forrest: Right, so we have to be careful not to overstate the impact because if we've talked about before, as recently as just two weeks ago, the Chinese models like DeepSeek, Alibaba, they remain very, very competitive. These are high-capability AI models despite what I'm going to call, again, known or expected hardware limitations. And the export controls that have been imposed may have constrained China's ability to produce chips, but they haven't totally stopped Chinese labs from building cutting-edge models. And as we know from Trump's visit to China last week and some of the press around that, there are reports of China making its own very fast chips. We just don't know that much about them.
Scott Caravello: Exactly. I mean, it's very clear that the Chinese chip-making industry is not standing still.
Katherine Forrest: Right. And, you know, for instance, Huawei's Ascend 910C chip, while less powerful than Nvidia's offerings, is being deployed domestically in China. And Chinese companies like Cambricon, Baidu, and Kunlun chips, and Biren are all developing alternatives to US chips. Now, by the way, I'm going to spell all those names because while I just, they just tripped off my tongue like I could actually pronounce them, but I don't really know if I did. So Cambricon is C-A-M-B-R-I-C-O-N. Baidu, I know I've gotten right. And Kunlun, which is K-U-N-L-U-N chips, Kunlun chips. And then Biren, B-I-R-E-N. So, they're all Chinese companies developing alternatives to U.S. chips. So, there is, we know, a nationwide effort within China to make their own chips and to become an independent, you know, have their own independent sources and bypass any limitations that may be, you know, being imposed on them. So what are all of our takeaways here, Scott, except that we now know you have scintillating conversation for your whatever parties you're going to tonight for this wedding.
Scott Caravello: Exactly, exactly. I cannot wait. But I can take the first takeaway, which is just that the AI chip market is expanding and the players are mixing up. Nvidia, rather, remains key, but its own customers are taking a role in the chip-making market as well.
Katherine Forrest: Right. And I would add as a second takeaway that chips are inseparable from geopolitics and they are, as a third, inseparable from AI policy, which is embedded with geopolitics. So, we've got export controls, geographic concentration, and related on-shoring initiatives. And all of these things are defining and are going to continue to define the landscape for AI for some time to come. So there you go. We've given you a lot to talk about tonight, Scott. I want to just, you know, when you're sitting there with that, you know, special designated cocktail in your hand, I want to hear you hold forth about chips.
Scott Caravello: I am gonna go find the bride and I am gonna take up all her time telling her all about semiconductors.
Katherine Forrest: That'll be really just exciting and thrilling. But that's all we've got time for today, folks. And I'm Katherine Forrest.
Scott Caravello: And I'm Scott Caravello. Don't forget to like and subscribe.
