SemiAnalysis · 2026-08-07 · 44 min

Ep. 023 - Everyone Leaves Google, Elon Forecasts 1T ARR, Reflecting On GPT-5 | Jon from Asianometry

Jon YDoug O'LaughlinJordan Nanos

YouTube ↗

TL;DR

Google’s talent drain is real, with Jeff Dean, John Jumper, Noam Shazeer, David Silver and several Gemini leads departing. Demis Hassabis is described as moving into a higher-level chairman/chief-scientist role while also serving as Isomorphic Labs’ CEO. The panel splits on whether this is merely “the dream team…breaking up as they all turn 38” or the moment Google loses unusually broad system-level judgment that compute alone cannot replace.
The bearish Google call is about execution, not earnings. Dylan argues that Google repeatedly invents foundational technology, fails to commercialize it, and risks becoming an AI-era Bell Labs: highly profitable, strong in TPUs and potentially “really good for the stock in the short run and medium run,” yet destined to “quietly bow out” of frontier models.
China restrictions increasingly mean banning the best product, rather than excluding cheap imitators. Jon supports reducing revenue flows to Chinese suppliers but says firms may route around narrowly drafted rules through Vietnam or Thailand. With China strong in optics and the 1.6T/16T transceiver supply chain primarily there, an InnoLight ban could hurt Western buyers more than its intended target.
Agentic coding redeemed the GPT-5 thesis. Jon says GPT-5 itself “probably wasn’t that great,” Doug calls 5.2 disappointing, but both see 5.6—and potentially “six”—as a major step forward: once agentic coding arrived, “everything hit the pedal to the metal.”
Bespoke software can be built for hundreds of dollars. With four attempts, a Claude-written specification and Codex 5.3 or something similar, Jon built a 40 MB video editor tailored to his keyboard-driven workflow; 5.6 later completed a previously stubborn feature after roughly 25 specification questions. His advice: “roll your own software,” the technical equivalent of making your own furniture.
AI may make the open internet toxic to machines. The panel’s mechanism is that software systems are too large for any human to understand, while a model can absorb the code and interdependencies in one context window, finding both high-level and memory-level exploits. Jon’s endpoint is “superhuman level breakage of the system forever,” potentially making nearby compute and models more attractive than constant web access.
Elon’s $1 trillion forecast requires heroic physical ramps: SpaceX discussed moving from 1 GW to 10 GW by end-2027, then 20 GW, while pulling its revenue target from 2031 into 2030. Jon expects Terafab to start with memory; Dylan argues its scarcity and margin make it the fastest route to revenue. The panel’s deliberately sharp hurdle—“anything under $100 billion at the end of Q4 ’27 is a miss”—highlights how far the arithmetic outruns the required ramp.

Digest · the substance, structured for research

1. Agentic coding rescued a disappointing GPT-5 cycle

Looking back at the October 2, 2024 discussion, Jon says GPT-5 itself underwhelmed, but the underlying expectation that progress would continue proved right. When agentic coding became useful, “everything hit the pedal to the metal,” changing his assessment more than the original model launch did.
Jon says 5.2 was also poor; Doug calls 5.6 “mind-blowing,” while Jon says the new base model and 5.6 are good. The panel also jokes that “six” is coming soon. Separately, rumors say the Doug model may be unusually good at writing.
The important change is practical rather than benchmark-driven: Dylan says projects that were outside his control with 4.5 became workable with Lovable or 5.6. Jon’s joking standard for his original podcast appearance—“it doesn’t really matter if I’m accurate, as long as I’m entertaining”—has yielded to repeatable evidence from tools he now uses.

2. China bans increasingly remove the strongest supplier

Jon is directionally hawkish on restrictions because the US should not bankroll strategic Chinese competitors, but he expects firms to satisfy the letter while evading the spirit of the rules. He names Vietnam and Thailand as examples of possible routing around narrowly drafted bans.
The optics case reverses the old trade logic. Jon says InnoLight may be better than Western alternatives, while the broader 1.6T and 16T transceiver supply chain is primarily in China. A ban could therefore be “probably more painful for the Western companies to bear,” especially while demand keeps rising.
The panel’s list now includes optics, humanoid robots, drones, solar panels and EVs—categories where excluding China can mean excluding “the best product in the world.” The proposed industrial-policy answer ranges from nurturing domestic competitors to forcing a 1960s-Japan-style minority joint venture and IP transfer.

3. Google can print cash while losing the frontier

The departures extend beyond public figure Jeff Dean to Quoc Le, Oriol Vinyals and Sanjay Ghemawat, described as ostensible Gemini leads, following John Jumper, Noam Shazeer and David Silver. Departing researchers can raise money—including from Google Ventures—buy GPUs and pursue ideas they may not have received enough compute allocation to pursue internally.
Dean is the hard case for the “stars do not matter” argument. His fingerprints span MapReduce, Spanner, Bigtable, TensorFlow and TPUs; the panel likens losing him to losing Google’s “Chuck Norris,” someone able to hold enough of the full software-hardware system in his head to provide insights others cannot.
Jon offers TSMC as the counterexample: semiconductor R&D is deliberately federated, with “no single ego,” because thousands of people must rub down the corners. Dylan agrees for chips but argues that frontier-model labs implicitly reject that logic whenever investors back a departing researcher on the belief that discoveries—not merely “the biggest engine with the most compute and the most data”—still matter.
Dylan’s broader “Google has an L culture” thesis draws immediate pushback. He says nearly everything beyond search was acquired and cites GCP’s late execution; the discussion counters with Maps, GCP, TPUs and Kubernetes, as well as Google’s $4.3 trillion valuation and position as the world’s third-most-valuable company: “Lose to who?”

4. Google’s Bell Labs problem may be bullish before it is fatal

Dylan separates the stock from the technology trajectory: Google can remain enormously profitable while becoming financialized, protecting search and focusing on TPUs rather than pursuing repeated frontier-model breakthroughs. His analogies are Intel abandoning mobile and IBM favoring mainframes despite once holding roughly 90% market share.
The sharper historical analogy is Bell Labs or Kodak: Bell Labs made the transistor, and Kodak made the first digital camera, yet both examples show how invention can fail to secure the commercial future. Google could likewise invent the transformer, settle for a “good enough” model and eventually become primarily an infrastructure supplier.
The panel’s pushback—worth keeping—is that scale and bureaucracy do not prove decline. Google serves ads, search, YouTube and “a billion surfaces every day,” whereas OpenAI and Anthropic still have one job; Demis Hassabis’s move upward may even position him as a future Alphabet CEO.
Dylan nevertheless calls the present an inflection point: in five, ten or twenty years, this could look like the moment Google chose infrastructure over repeated frontier-model leadership. His hedge is explicit: “This is not saying it’s a bad stock,” and the choice may be beneficial over the short and medium term.

5. Uncontrolled AI could make the web toxic

Dylan’s speculative endpoint is that malicious AIs could make ordinary internet use toxic—damaging computers and stealing money—while nearby compute, such as a neighborhood GB200, could make web retrieval less necessary.
Jon’s mechanism is that modern software is too large for any human to read and understand along with all its interdependencies. A model can take the whole codebase into context and find ways to break or repurpose it, from high-level Python vulnerabilities down to CPU memory leaks. He expects “superhuman-level breakage of the system forever” and jokes that the only fix may be an American Great Firewall, with a physical moat as an additional safeguard.

6. AI-built personal software is already economically real

Jon’s strongest evidence is his 40 MB custom video editor, built for a banking-style, mouse-light workflow full of keyboard shortcuts. Right-clicking an image can send it directly into the editor, while integrated Wikimedia Commons and video search insert material with the correct attributions.
It took four attempts and aggressive scope reduction. Claude questioned him about the workflow and produced a roughly 50-page specification; Jon then handed it to “Codex 5.3 or something” with instructions to build one simple structure that performed one task well enough to extend.
Later models changed the ceiling. Using 5.6 with a library called Easy Cheese, Jon answered roughly 25 questions for one feature—he says the tool can produce an interrogation of “50 words or 50 questions”—then watched it “one-shot the whole thing” after Opus had repeatedly failed.
His total spend was only a couple hundred dollars, using Codex, GPT Pro and a $20 Claude subscription. He has also built an audio post-processing app with a waveform despite having “no idea how it works”; the editor makes him faster, though the unintended result is longer videos, and viewers apparently have not noticed the tooling change.

7. Terafab’s $1 trillion path probably begins with memory

SpaceX’s abundance case moves from 1 GW to 10 GW by end-2027, then 20 GW, while advancing a $1 trillion revenue forecast from 2031 to 2030. The backdrop is Microsoft, Amazon, Google and Meta taking free cash flow toward zero to buy chips—an investment wave Jon says he still struggles to believe.
Jon expects Elon to start with memory. Dylan argues that memory shortages are crimping products in Taiwan and that high margins make it the fastest route to revenue. He cites a Tim Culpan report that Apple has N2 chips awaiting memory. Jon says Apple tried to negotiate a large discounted CXMT order and was refused; Dylan briefly guesses Huawei before Jon insists it was Apple.
The hosts stress-test the schedule with comic severity: “Anything under $100 billion” in Terafab’s Q4 2027 exit rate would be a miss, and Dylan clarifies that this is Terafab alone. Dylan invokes the maxim that “Elon makes the impossible late”; Jon agrees that a $1 trillion chip-revenue outcome in 2040 would still be extraordinary.
The ecosystem choice currently favors Nvidia. Elon wrote that SpaceX installs its GPUs because “they are the best,” while Dylan characterizes TPU software support outside Google as effectively “zero.” Jon points to StableHLO and related tools, but says a TPU user would have to roll their own software, GCP sales representative and console. The panel also cites David Silver’s new effort raising a $1 billion seed round and spending it on FuriosaAI chips.

Dylan Patel

Hello, everyone. Welcome back to SemiAnalysis Weekly. We've got our second guest ever on the program today. That's Jon from Asianometry, the No. 1 YouTuber in the world—or at least my favorite. Doug said he's in the top 2, but we won't hold him to it.

This week, we're planning to talk about some big news: everybody leaving Google, leaving Gemini hanging out to dry, maybe, or just going to work on some of their passion projects now that they've spent their first 35 years in one job and had enough; the Chinese ban on transceivers; Elon pulling in his $1 trillion revenue forecast from 2031 to 2030; and maybe some DVDs. Guys, welcome to the show.

Jon Yu

Thank you. Thank you for having me. This is weird. You're so professional compared to Transistor Radio.

Doug O'Laughlin

We've grown up. It's so grown-up, Jon.

Okay, I have a burning question for you, Jon. One of the times when many people may have seen you talk was on the Dwarkesh Podcast, where, a few years ago, there was this famous clip where you're sitting beside Dylan and you say, "Wait, wait, wait. We're assuming GPT-5 is going to be good. We've just assumed that."

Do you remember this clip?

Jon Yu

I do remember this clip.

Doug O'Laughlin

That was a year and a half or 2 years ago. Now we're sitting in August 2026. Do you feel like GPT-5 was good? Did it surpass your expectations?

Jon Yu

I think the concept of GPT-5 was probably right, in the sense that—was it good? GPT-5 itself probably wasn't that great compared to others, but it kept going. Then, when the agentic code stuff really hit, everything hit the pedal to the metal, so I was very impressed by that.

When I went on the pod, the first thing I thought to myself—I had chugged 5 cups of coffee—was that it doesn't really matter if I'm accurate. As long as I'm entertaining, that's what matters. Hopefully, we can aim for that today as well.

Dylan Patel

Entertaining and questionable accuracy. Doug, what do you think?

Doug O'Laughlin

This was October 2, 2024. I just looked it up—when Dylan and Jon went on the Dwarkesh Podcast and talked about this stuff. I kind of agree that GPT-5 was a bit of a dud, but 5.2, and certainly now 5.6, is mind-blowing 2 years later.

Jon Yu

5.2 was ass, too. I'll be honest. I think 5.2 was not very good.

Dylan Patel

I think 5.6 is good.

Jon Yu

The new base model is good. 5.6 is good. And also, soon—soon—6.

Dylan Patel

This week.

Jon Yu

Yeah.

Dylan Patel

The second.

Jon Yu

6, apparently, is good.

Dylan Patel

Doug is coming, right?

Doug O'Laughlin

Yeah, Doug is coming, as they say.

Dylan Patel

Doug—and, Jon, I don't know if you know this, but Doug is the largest spud in the world.

Jon Yu

Really?

Dylan Patel

Yeah, that's the name. That's how they named it. So it's like the Spud model—the largest spud.

The Doug pre-training is done. There are also rumors that it's going to be really good at writing, and I think that's pretty poetic: one of the models that finally really knows how to write some incisive commentary is going to be named Doug.

Doug O'Laughlin

That's blazing. It's going to teach us a master class in writing history. I'm looking forward to it.

Jon Yu

Yeah. Hopefully, it knows.

Doug O'Laughlin

It's not "this," but it's "that."

Dylan Patel

That's where Doug gets his edge: the references to 1940s Taiwanese history.

Doug O'Laughlin

Well, you say that, but the only person who's probably better at it than me is Jon, and Jon's here. I'm like, "Oh, yeah, semiconductors—well, that doesn't exist. Semiconductors weren't around then." But then Jon's like, "Yeah, here's what a DVD player is." I was like, "What the hell?" So the only person in the world who goes deeper and more random, I think, is Jon.

Dylan Patel

I've always liked Doug. I read this really weird book on Japanese history from the 1700s about the Mitsui family, and I'm talking about it. Then Doug's like, "I read that."

Doug O'Laughlin

The first half was great. The second half was okay.

Dylan Patel

Like this.

Doug O'Laughlin

I guess when it got into the modern era, I didn't really care. But the first half was pretty good.

Jon Yu

Well, no, you got that from Jordan, though. We all got it from the same source. I picked it up from my "When Japan Ate Its Rich" video, but I skimmed through it. It wasn't like—

Dylan Patel

It's a good narrative, or it's too—

Jon Yu

No, it wasn't even narrative. It was really, really boring.

Dylan Patel

This is Jordan Schneider of ChinaTalk we're talking about. I think it's unlikely that anybody subscribed to somebody else's weekly doesn't know Asianometry or ChinaTalk, but in case they don't, that's a clarification.

Interesting to get the take without Jordan on here, but Jon, what's your take on the Trump administration prepping to ban a bunch of Chinese data-center components, including transceivers, and then possibly banning humanoid robots as well, like the Unitree stuff we were talking about a few weeks ago on here?

Jon Yu

On one hand, I'm generally more hawkish. I think we should be banning this stuff. You want to make sure that you're not driving revenue for people in China. But on the other hand, they have the expertise, and it's not exactly like the stuff is—I mean, they're very, quite good at evading the letter of the law once they read it. They meet the letter of the law so they evade the spirit of the law. And I think, yeah, they're real followers.

Dylan Patel

Vietnam.

Jon Yu

Yeah, they're real followers.

Dylan Patel

Thailand.

Jon Yu

Thailand is actually a big area. The other thing, too, is that unlike past bans, I think this is an example where InnoLight is actually better than the West. This is like us banning inverse EVs, right? They didn't have EUVs, and I don't think they will have EUVs for some time. But when it comes to 1.6T, 16T transceivers, maybe there are some components in the West, but the entire supply chain is primarily in China. So, ironically, this is probably more painful for Western companies to bear because the supply chain is already there.

There's one thing that I feel like the Chinese are really, really good at: optics. Maybe it's the rare-earth thing, or the fact that a whole bunch of people study it. I don't know.

Dylan Patel

Yeah, they're also good at optics. They made the optics in the iPhones, like Sunny Optical or whatever, right? Those are the 2 companies that make the lenses.

Jon Yu

Yeah. I mean, they're also pretty good. A Chinese company, isn't it?

Dylan Patel

Ah, whoops. Sorry, it's Largan. It's the Chinese company. Sorry, sorry, sorry, sorry, sorry.

Jon Yu

I mean, they're also pretty good at humanoid robots, drones, solar panels, and now EVs. I think there's a growing list of real technology where banning the best product in the world is what it takes to ban China from competing in the American market. Whereas in the past, it was like you're banning cheap knockoff competitors who are trying to bleed their competitors dry. Now you're just kind of—

Dylan Patel

Losing the domestic manufacturing base because the manufacturing—

I guess that's what I would guess. The idea would just be kind of like, you'd ban one side—the import—and then you'd raise up a domestic competitor. So I would imagine that the Americans have a good number of domestic competitors.

Or, if you really want to go 1960s Japan, you kind of force the Chinese to enter a joint venture with the Americans if they want to enter that market: a 49%/45% joint venture, and then IP transfer, baby.

Jon Yu

You're foreseeing Trump in light industry coming soon.

Dylan Patel

I don't have to do a 50%/51% to the Trump family.

Actually, it's really easy. Analy[?], if anyone is listening, tip us for the idea.

Jon Yu

Yeah, but demand is not slowing down.

Dylan Patel

Maybe the biggest news story from this week was everybody leaving Google. Jeff Dean, Sanjay Ghemawat, Oriol Vinyals—a lot of people. There's news about Demis stepping down to become the chairman. Well, no, he's stepping up, in theory, because that's up the org structure. But Demis has been really focused on Isomorphic Labs, and some say this was relatively well-forecasted or communicated beforehand, allegedly, according to the investor bros.

I think there's a real chance that Demis could be the next CEO of Google. I think Demis is failing upward as fast as possible, in the spirit of Google. Does he actually want to be CEO of a company? He doesn't seem to like the administrative stuff, right? He looks like a scientist-scientist.

Jon Yu

He's chief scientist now that Jeff Dean's gone. And he's currently CEO of Isomorphic Labs, as if chief scientist wasn't enough.

The guy reportedly works insane hours. He'll just work a normal day, go home, have dinner, and then come back and read papers by himself in the office until 2:00 a.m. every night.

Dylan Patel

Isn't that how I work?

Doug O'Laughlin

Why is this acceptable? Why are you saying this is unusual behavior?

Dylan Patel

It's Taiwanese behavior, bro. It's Taiwanese. He's got to work 18 hours a day.

Jon Yu

He's got that fire in him.

Dylan Patel

Yeah.

Jon Yu

Even if this is just some story to gas him up, you do not hear the same story about Jensen, Zuckerberg, or anybody who fancies themselves a scientist who is also the CEO of a leading company.

Dylan Patel

So I do think it would be a change from the current day-to-day to become CEO of Alphabet, I guess, and take over for Sundar Pichai. But maybe the more important thing here is that Jeff Dean left—the legend, employee number 30, MapReduce, Spanner, Bigtable, TensorFlow, TPUs.

Jon Yu

He’s an infrastructure guy. He’s just been focused so much on infrastructure.

Dylan Patel

Sorry, I’m not going to go into that. No, I do think it’s a big deal. I definitely think there are a lot of rumors that I don’t like. I guess the thing is, Jeff Dean’s also at every conference, you know? He’s at every conference, man. He’s always on the speaking circuit, and he’s always talking about something. I feel like Jeff Dean is, as you know, the Chuck Norris of Google. That’s his reputation. Even if he was or was not working there day-to-day, I feel like you have to be a little sad for Google, because they just lost Chuck Norris, bro. Chuck Norris just died.

Jon Yu

Yeah. This is right after losing John Jumper, right after losing Noam Shazeer, right after losing Dave Silver. The prototype seems to be that these guys leave, and then they raise money from a bunch of investors, including Google Ventures.

Dylan Patel

And Google Ventures. Yeah.

Jon Yu

And then go buy a bunch of GPUs and do whatever they want. Maybe because they couldn’t get compute allocation to pursue the ideas that they wanted to.

Dylan Patel

I mean, you may be able to say Jeff Dean was in this senior role and quite a public figure, but I don’t think you can say the same for Quoc Le, Oriol Vinyals, or Sanjay. These were ostensibly the leads of Gemini. I think that’s pretty crazy that they all left.

Jon Yu

Well, can I say something, though? From an SF-vibes perspective, is this surprising? The joke is that Gemini hasn’t been a frontier model for a long time, and now it definitely won’t be. You lose all the legends and think, “How are they going to come back?” But they do have a lot of compute.

Everyone at Google DeepMind is not AI-pilled, right? All the people who want to do the best AI don’t work at Google, and I think that’s an important mindset. Google’s not a startup, right? Google’s maybe the largest bureaucratic organization in the world, and OpenAI and Anthropic, who are the best of the best, are definitely kind of startups. I’m sure they’re going to speedrun becoming a bureaucracy really quickly, but they have one job and one job only: be the best at AI.

Google is like, “We do ads and search and YouTube and a billion surfaces every day.” Then they’re like, “Okay, well, how do we shove AI into all these different places?”

My hottest take—or not even a hot take; it’s actually pretty mild—is that Google has an L culture and has never actually made anything internally. They’ve only ever acquired all their innovation. Number one, their execution is very poor, and they were also terrible at go-to-market until Thomas Kurian got to GCP.

Let’s do some accounting. Everything other than core search has been purchased: YouTube, AdMob, DoubleClick, AdSense.

I’ll help you make it by saying Android.

Dylan Patel

Android. Yeah, all these guys.

Jon Yu

Not true. This is not true, man.

Dylan Patel

They didn’t acquire Maps. They didn’t acquire Maps.

Jon Yu

They acquired Maps. They mapped it out. They bought the team. Dude, there’s a whole book about this, man.

Dylan Patel

They did not acquire Maps. They added positions after they started.

Jon Yu

They effectively acquired Maps. There’s a whole book about this.

Dylan Patel

Okay. Didn’t you just say that Doug is the guy who has the book?

Jon Yu

They also did not acquire GCP, which was the crux of your argument that they have—

Dylan Patel

They acquired Thomas Kurian, who then made GCP good. They also, to be clear, invented half of the concepts of GCP, and then they were in last place despite being pretty close to the front—you know, showing up at the beginning of the line and then underexecuting to be last place. So I just think Google has—

Jon Yu

Acquired TPUs? Did they acquire Kubernetes?

Dylan Patel

No. Okay, so I’ll give them TPUs. I’ll give them TPUs. How does that sound? I’ll give them TPUs.

Jon Yu

How generous of—

Dylan Patel

Yeah, you’re welcome. How nice of me.

Jon Yu

No, but I think the thing is, there are these companies that are known for their research, and they never execute on the—

Dylan Patel

Bell Labs comes up, and, like, Bell Labs—they make a bunch of science. No, but they literally make a bunch of science stuff, and they literally get massively disrupted by it in the end. Bro, the first digital camera was made by Kodak.

Jon Yu

Kodak died.

Dylan Patel

Okay, dude. Who invented the transistor and gave it to everyone?

Jon Yu

Exactly, dude. It makes so much sense that Google is going to invent the transformer and then get giga-screwed as it proceeds to just lose and become a TPU. I mean—

Dylan Patel

Proceeds to lose and be the third-most-valuable company in the world, with a market cap of $4.3 trillion. Lose to who? Lose to Nvidia.

Jon Yu

There’s a chart showing that IBM was the biggest company in the world in 1982.

Dylan Patel

From a technology perspective, I think there are these important branches in an innovative-technology company. To be clear, this is not saying it’s a bad stock or a bad company. They’re going to make a crap ton of money, but there are these important branches in an innovative-technology company where they start to become financialized and say, “Hey, let’s do the way we make money and stop pursuing technology.”

You can argue this is Intel giving up on mobile. You can argue that when IBM effectively had systemic problems, but honestly, IBM’s vibes are very similar to Google. They were like, “Yeah, no, no. We’re not going to do all this new stuff. We’re not going to do PC. We’re going to focus on mainframe,” right? They’re going to do what they did, and then they will proceed to lose. To be clear, IBM had effectively 90% market share in 1950.

There are these points in a company’s history. Today, more than at any other point, I feel like Google has that opportunity where they’re like, “Okay, we’re going to go all in on TPUs. We’re going to become an infrastructure company. We’re going to do good enough. We’re going to have a good-enough product on the transformer, but they’re probably not going to stay for the next and the next and the next. Then they’ll quietly bow out.”

When we look back in 5, 10, or 20 years and ask, “What was the moment when things started to change?” it’s like, right now. To be clear, I think that it’s really good for the stock in the short run and medium run, honestly.

The government should basically anoint Google’s monopoly in search, maybe, but force them to outsource—to license—all of the science they’ve been doing internally to every startup, just like they did with AT&T. That would supercharge the startup network.

Jon Yu

You think there’s a new Bell Labs coming?

Dylan Patel

They’re effectively Bell Labs.

Jon Yu

Big, profitable, and then—

Dylan Patel

Yeah.

Jon Yu

They can’t even monetize their own crap.

Dylan Patel

Yeah, yeah, exactly. It’s the perfect example. They have all this cool stuff. My favorite is Killed by Google. If you go to Killed by Google, it has all these things, and you can argue that many of the companies they had or killed look like very successful, valuable companies today. They literally get killed by Google, leave, and start up again as an independent public company, and it goes well.

Search was a giant monopoly that gave them the ability to do YOLO random stuff and positive R&D, but they’re never going to make money off of it. It’s kind of based.

I have another question. With these big guys leaving, does it really matter? If some guy—if LeBron leaves the Lakers and people are like, “Are the Lakers over?”—does that matter? I mean, to win, they’ve kind of done their stuff, right?

Jon Yu

Yeah, I think that’s kind of agreed. Maybe the answer is that the Dream Team is breaking up as they all turn 38. That makes sense to me. You’re like, “Yeah, their best days are behind them. Now they’re going to go do one last glory tour. Maybe they’re going to go play baseball or something and then go back.” Who knows?

Dylan Patel

But I still think, in your heart of hearts, you should feel really sad about—

Jon Yu

And pathetic knowledge.

Dylan Patel

Yeah.

Jon Yu

But that—

Dylan Patel

Like TSMC, right?

Jon Yu

TSMC’s R&D structure has incredibly federated, incredibly split-up teams, especially within Taiwan, in Hsinchu. Something they always emphasize to me is that they work as a team. There’s no single ego, which means that if one guy leaves, or a couple of guys leave, it’s not a big deal.

They tell me that when big people leave to go do something else, you should always disregard that—not disregard it, but don’t take it as far larger than what it is. Because in the end, R&D, at least in the semiconductor world, is a team effort. Everyone has to put a crapload of effort into rubbing down the corners, and no one person knows it all.

I agree in semiconductors specifically, but I think I disagree in the Jeff Dean case because it's such a strong counterexample. As Jordan says, he's made so much stuff. It's so rare to have 1 dude who's so cracked and is like, literally, “Oh, my—it's crazy.”

He also did this, right? He's on all these different software, hardware, coding, and infrastructure projects. Across the entire stack, he has his fingerprint on all of them. It's really rare to lose a talent like that. He's revered for a reason. But yeah, we'll see. Maybe. Maybe.

Dylan Patel

Maybe start with von Neumann, especially with a lot of his stuff being collaborations. He does a lot of collaborations. Maybe that's Jeff Dean—he's just a collab guy.

Jon Yu

He's an idea guy. He does collaborations with the computer thing. The von Neumann model drives me crazy. If you read into the science of it—

Dylan Patel

It's very vibey. He just showed up and was like, “Yeah, you should make it like this,” and pieced the shit out. No, but to be clear, you have to respect von Neumann. You have to put some heat on von's name. Come on.

Jon Yu

All right, but let's not glaze him. Let's not glaze him the same way.

Dylan Patel

I think there's something about Jeff—

Jon Yu

Somebody who can keep—

Dylan Patel

Whether it's von Neumann or Jeff Dean, I think there's something to be said for the talents required, and really the experience required, to be able to fit all of the different components of the system in your head and reason about them. Even if the work that you do is pretty quick and the insight you give is pretty small in the grand scheme of a big project, nobody else can really provide the exact same value that you can if they don't have all of those little pieces of experience that you stack up, which I think is what people view there.

On the model side, though, I think if you view individual people leaving as not representative of the next version of Gemini being good or bad—or not a signal that goes against the entire ethos of all these labs being founded right now—it kind of implies that every single lab founded by the guy who just left isn't going to be successful. It implies that whoever has the biggest engine with the most compute and the most data is just going to produce the best models, and there are no more discoveries to be had.

Jon Yu

Maybe it's true.

Dylan Patel

Could be true. I mean, didn't a bunch of people leave OpenAI after 2022? In 2023, OpenAI kind of staggered a bit, but they got themselves back.

Jon Yu

No, that's different, dude. That's like the Protestant church. When we go back in the history of AI, that's going to be like, “Dude, the Catholic Church has lost the sauce. They don't even really believe in the same God we do.” Dario Amodei is the guy who nailed his freaking—

Dylan Patel

He's Martin Luther.

Jon Yu

Exactly. He's Martin Luther, dude. That is a take. That is a take, dude.

Dylan Patel

Exactly. Yes, he did. And he's like, “No, no, no, no.” And that's why they have—yeah, that's Anthropic and OpenAI.

Jon Yu

So soon you're going to be like, somebody's going to look over your shoulder and see if you're using Codex or Claude, and it's going to be like somebody in Northern Ireland being like, “Are you a Catholic or a Protestant?”

Dylan Patel

They'll be bombing certain data centers while others—you know, there'll be a religious war.

Jon Yu

A 30-year religious war.

Dylan Patel

Sounds good.

Jon Yu

Well, the models are currently competing with each other to show off how many—

Dylan Patel

Felonies.

Jon Yu

Examples of their models breaking out of containment they can—

Dylan Patel

Felonies—the felony chart, dude. I will admit, some of them, if they're real, do scare me. My complete rando-ass take that may or may not become true is that the internet will be viewed as toxic—as in, it's just going to essentially affect your computer.

If we have enough local LLMs—not local, but, like, why do you need to talk to some dude across the world? Maybe talk to some dude across the world, but, like, referencing information, why do you need to go to a website? Your GB200 in your neighborhood has all the information you need anyway. Because if you actually go on the internet, all the malicious AIs will just destroy your computer instantly and steal all your money.

Why do these companies not keep track of their AIs? Isn't it using a computer? Don't they own the computer? Why can't they keep track of it? It's like, why is their dog pooping in my yard? Why do you not keep track of your dog?

Jon Yu

Because they can't. It's not a dog. It's a yard that's so big and there are no fences. In fact, it's 100 acres, and there are secret tunnels everywhere because we've just been building it for God knows how long. This dog is superintelligent and knows everywhere—every little nook.

The thing that's really crazy, I think, is that in the same way we talk about how it's too big for any one person to understand how to make a chip, modern software is too big for any one person to read and understand all the code, and also understand all the interdependencies, right? It's like, hey, it all goes in one context window and it can read every single word at the same time.

That might as well be the effect of, “How do I understand all the context of a very specific, extremely abstract language and then be able to break it and essentially make use of it?” Or, “I could just do this instead of this,” right? I think it's the perfect use case. I don't know. I just think everything's not going to be safe one day.

It starts with Python and whatever—stuff that's in super-high-level languages—but there are also super-low-level exploits, too, like memory leaking out of the CPU. These are really hardcore exploits that are superhuman, and I think we'll just continue to see superhuman-level breakage of the system forever. It's going to be an interesting decade, to say the least.

You know what's the only thing to fix it? A Great Firewall. I think America should erect a massive firewall. They've provided the technology originally, so they might as well just use it.

Dylan Patel

Yeah. Maybe one on the internet and maybe one physical one as well, just in case the robots are getting going—like a big Great Wall.

Jon Yu

It's okay. The humanoid robots can't swim for more than 10 minutes, I think.

Dylan Patel

So you're telling me this is going to be a great moat? It's going to be one great 1-mile trench filled with water?

Jon Yu

Sure, man. Yeah, it's back to the medieval times. A big moat. Big moat works.

Dylan Patel

Big moat.

Jon Yu

Okay, okay. What else you got in your bag of questions?

Dylan Patel

Yeah. Well, okay, the other one is—

Jon Yu

Question, man.

Dylan Patel

Let's get a nice segue. With all this demand from all this new AI, it's going to mean more chips, including Elon breaking ground on Terafab and then telling investors in the first combined SpaceX earnings call that they're going to move their 1-gigawatt forecast to 10 gigawatts by the end of 2027 and 20 gigawatts after that, which means that he now believes they're bringing in their $1 trillion forecast from 2031 to 2030. That's the abundance case, man.

Let's go back to the very beginning of this. Jon, I talked about you going on Dwarkesh with Dylan in October 2024. That was before, like—

Jon Yu

The hyperscalers decided to pour in a trillion dollars of capex, and we've already got some pretty incredible models in the 2 years since then. So—

Dylan Patel

What happens when $1 trillion becomes more than a trillion in the next 5 years—like $10 trillion in the next 5 years because of Elon? You're saying Elon hits his numbers and suddenly we're all awash in compute. Is that what you're trying to say?

Jon Yu

It's going to be Elon, but it's also going to be all the other competitors: Microsoft, Amazon, Google, and Meta. They're all taking their free cash flow to 0 right now to buy chips, which is a fascinating thing to live through. I mean, at that time in 2024, I could not believe it. I still don't believe it now. I don't live in the States, so I don't see it here. Taiwan doesn't really build data centers; they need all the power for fabs.

Dylan Patel

Do people even use AI in Taiwan?

Jon Yu

They use ChatGPT to translate letters and stuff. I think they watch a lot of slop, but I don't think they use ChatGPT that much. They're very behind.

Dylan Patel

Do you use coding agents?

Jon Yu

Yeah, I code my own video editor.

Dylan Patel

Oh, really? Can you tell us about this?

Jon Yu

Yeah.

Dylan Patel

When did you do that?

Jon Yu

I think a couple of months ago. I needed one. My workflow is very specific, with a lot of keyboard shortcuts, and I figured, you know, I used to work in banking, so they used to take away your mouse and you had to do the whole thing.

So I didn't like my video editor not having keyboard shortcuts, and I was like, “I should make my own. Can I do that?” It turns out I did. My little video editor is around 40 megabytes, and it's highly optimized to what I do. If I grab a picture, I'll right-click it, and it immediately jumps into the video editor. There's a lot of small stuff that I never thought was possible with Adobe Premiere or even iMovie. It's wonderful, and I highly recommend that people roll their own software.

It's like the hipster version of people making their own furniture. If you're a true tech hipster wearing a beard and a plaid shirt, you need to roll your own software. I'm going to make my own text editor soon. I feel like I should.

Dylan Patel

Yeah, like the men.

Jon Yu

Honestly, words so bad. Why not?

Dylan Patel

What's to lose? There's literally nothing to lose.

Jon Yu

There's nothing to lose.

Dylan Patel

Can you explain how you built it? When did you start? What tool did you use?

Jon Yu

I tried 4 times. First, I asked Claude to quiz me on what this video editor did and how it worked, because the workflow is a bit specific. You start with an audio file, so I thought to myself, “Well, can I do it this way? Do it this way?” I tried the first 3 times, and I really didn't like it.

Finally, I stripped it down enough. I had to get another video out, and I thought to myself, “Let's give myself a deadline.” I finally stripped it down to a point where I said, “Just do this 1 thing really well.” I ended up having Claude make a 50-page spec. It was a massive spec, and I handed it to Codex 5.3 or something at that time. I said, “Just make this very, very simple structure.” It barely was good enough at 1 thing that I could keep working on it a little bit. That's how I did it.

Dylan Patel

Has it gotten better with subsequent AI? Because I do feel like there are things that were outside of my control, or outside of the ability of what I was able to do with 4.5, that are doable with Lovable, that are maybe doable with 5.6. I think it's weird, because every once in a while I'll be like, “Wait, I should try to redo this project.” I redid a lot of projects with Lovable, and when I redid them, I was like, “Oh, actually, it's good now.”

Jon Yu

Yeah. There are a lot of times when I would be triggering Claude Opus and saying, “Improve this. Can you improve this?” Then I would have to go onto my computer to troubleshoot it and check it. It's gotten a lot better.

I have situations now where there's a library I use called Easy Cheese, and it really grills me. They use a lot of cheese puns. It grills me for 50 words or 50 questions. The questions are longer than the actual coding, and it creates this titanic spec, but it hits every feature in 1 shot. I love it. I have no idea how it works, but it's ridiculously good.

The fact that I can roll out entire features is amazing. There was 1 feature I added recently: “When you add an image, add it to the end of the last image, but fill the text block,” because it has a subtitle. “Fill the text block,” because the text block could be different sizes. Opus couldn't do it for the longest time, but I ended up asking 5.6, and then I used Easy Cheese. I answered 25 questions. I'm not even joking. I went to bed, and it was still asking me questions. But after it finally did it, it one-shotted the whole thing. I was just mind-blown. It was crazy.

Dylan Patel

How much do you spend? Do you pay per token, or do you have the monthly plan?

Jon Yu

I have Codex, I have GPT Pro, and I have a $20 Claude plan.

Dylan Patel

So you've done all this with just under $1,000 or something?

Jon Yu

Yeah, like a couple hundred bucks.

I have my own video editor, and I've integrated Wikimedia Commons search into it. I've integrated video search into it. It drops things in with the correct attributions. It's ridiculous. I don't even know how. It's made me faster. It just makes the videos longer, unfortunately. I mean, I like them.

Dylan Patel

So how many have you done with this? You said you started in April, so you've done a bunch of videos with this new editor?

Jon Yu

Yeah. I don't think people have noticed. No one's noticed.

Dylan Patel

I'm going to start commenting on the videos. I could tell this is AI slop now.

Jon Yu

You're hilarious. I still do it. I still do the work. It's just my own tool, man.

Dylan Patel

No, I'm just kidding, dude. No, it's okay. I have a ton of tools like that, too. Like, you have turned it into Excel.

Jon Yu

Yeah, really? Yeah, pretty much. You have all the hotkeys.

Dylan Patel

No, there's so many things like that where it's like, this is the exact way I think about this set of problems, and I want to make sure it's done exactly like this. Then I'll work on it forever, and it's like, “Okay, now this skill—press Enter, go.” It's fun. It's fun. I don't know how we got into this, but—

Jon Yu

I also have a bunch of other tools. I do audio post-processing with a little app that I made myself, which I coded as well. I have no idea how it works, but it works. It even gives me a nice waveform.

Dylan Patel

Who needs to know?

Jon Yu

Do I need to know how it works? The audio comes out better.

Dylan Patel

Who needs to know?

Jon Yu

I don't need to know.

Dylan Patel

You don't need to know, man. It's all—it's all—

Jon Yu

I mean, my dog—that little post-processing app might be hacking the Chinese government website right now, but who knows? It's code that's for the birds.

Dylan Patel

So, okay. What's left on the question list, man?

Jon Yu

I don't—I don't have any more. I was going to ask Jon's take.

Dylan Patel

Yeah, yeah. What's your Elon take?

Jon Yu

I think he's going to start with memory.

Dylan Patel

Yeah, 100%.

Jon Yu

Yeah, I think he said that, actually.

Dylan Patel

Yeah, you should choose the highest margin.

Jon Yu

Why do you think that?

Dylan Patel

Well, I'm just reading his latest—he was just going on about memory on the SpaceX earnings call. He's just talking about memory all the time. If you're in Taiwan, 1 thing I definitely do know is that the lack of memory is crimping every product out here that they're trying to make, and they can't get enough of it.

Apple has a bunch of N2 chips sitting in a locker somewhere because they can't get memory to put into their phones. My friend Tim Culpan reported that. It's going crazy. The fastest way to revenue is to just start making memory and expand there.

Jon Yu

All the specs are open. Apple tried to negotiate with CXMT and got denied.

Dylan Patel

I think it was Huawei.

Jon Yu

No, no, no. CXMT. It was Apple with CXMT.

Dylan Patel

They were trying to be like, “Oh, give us a giant order, and you'll give us a discount.” CXMT Chad-mogged them and was like, “Why would we give you a discount? This is lower than the going rate of memory.” Then Apple's like, “Well, shoot.” They're like, “Don't you want to work with us because we're Apple?” And CXMT's like, “Who are you? Everyone wants my memory.”

So I guess Elon’s going to—“You need to hit that $1 trillion in revenue within a couple of years. I'll start making memory. That'll contribute to it. Easy.” I'm excited. I'm excited to see him smoke a cigar in the Terafab and make $100 billion of revenue next year, because that's the ramp that needs to happen in order for Terafab to happen. Yeah, I'm excited. It's going to be cool.

Jon Yu

Anything under $100 billion at the end of Q4 2027 is a miss. Hey, Doug—

Dylan Patel

I would say so. Yeah, I'm being conservative. That's just the exit rate, too, just to get to the next—

Jon Yu

Does that include or not include Tesla after they merge into Tesla?

Dylan Patel

No, no, that's just Terafab alone, Raz.

Jon Yu

I mean, I don't think Tesla's a drop in the bucket if they need to go to $100 billion next year, too.

Dylan Patel

Yeah, just who's even counting, honestly? Who even cares? I'm not even counting. Who's counting? SpaceX investors aren't counting. It doesn't matter. $1 trillion—2030, 2040. Who's—who? I think—okay, man. I hate this because I really do. Everyone says this, and I really do believe it is: Elon makes the late, no, the impossible late. So that means he's going to get $1 trillion of chip revenue in 2040.

Jon Yu

Yeah, okay. That would be pretty good. Some would say that's pretty good.

Dylan Patel

Sure. Yeah. Hey, power to him, man. I'm excited for more memory.

Actually, Jon, I was going to say, I feel like that would be the best call that could possibly be asked on the TSMC earnings call, because they hate analysts asking dumb questions. You could be like, “Well, have you heard of an American company that's guiding to $1 trillion of revenue? What do you think about the competitive prospects of a new entrant in the foundry market?” And they would be like, “Please shut the fuck up.” That would probably be the most triggering question possible, but I would love to hear it be asked.

Jon Yu

Yeah, and the guy will never come back again.

Dylan Patel

Yeah. He'll be kicked. They'll send him to East Taiwan.

Yeah. He also had a nice tweet where he said, “SpaceX chooses to install NVIDIA GPUs because they are the best.” Pretty matter-of-fact.

Jon Yu

Yeah. And then a bunch of other people were saying Jeff Dean left Google because he didn’t like TPUs anymore and wanted to use GPUs so badly, which is the trend for everybody else leaving. It’s like David Silver leaves, raises a $1 billion seed round, and then spends the entire thing on FuriosaAI chips, which they have a nice promo post with Google Cloud about.

Dylan Patel

Why do you think that is? I thought TPUs were good for pretraining.

Jon Yu

Yeah, I think TPUs are great, but I think that the support you get outside of Google on the software side—

Dylan Patel

—is zero.

Jon Yu

That’s what StableHLO’s for. That’s what StableHLO’s for, bro. You’re going to roll your own software. You’re going to roll your own GCP sales rep. You can roll your own console.

Dylan Patel

StableHLO, Pallas, TPU, XLA, and everything.

Jon Yu

Exactly. Yeah. I think you might get downgraded to Opus 5, which is intentionally screwing you over.

Dylan Patel

Just not trained on the TPU documentation because that would make you a competitor to Anthropic, who uses all the TPUs.

Jon Yu

I believe it. That sounds like something the Protestants would do.

Dylan Patel

The Protestants love it.

All right, everybody. Thanks so much for tuning in to this latest episode of SemiAnalysis Weekly with Jon from Asianometry. We covered the gamut, and we appreciate everybody joining and listening to us talk.