Loading summary
Transcript77 lines
- [00:00]
A
Today I wrote about will AWS buy TPUs from Google? In the front page of the Wall Street Journal's Business and Finance section, they're singing the Trainium chips praises. Amazon's chips pose risk to Nvidia. The whole week we've been talking to people.
- [00:16]
B
Is that clickbait?
- [00:17]
A
I don't know what we're gonna find out. We'll see. It certainly doesn't seem, you know, good to have more competition in the, in the market. There's a lot of losers if Google winds up winning with TPU. The losers came out to to fight, apparently. Amazon.com is the latest big tech company to muscle in on Nvidia's turf. Give me a sound cue from the fallen.
- [00:40]
B
How about this? There we go.
- [00:42]
A
That's right. On Tuesday, Amazon Web Services announced the public launch of its Trainium 3 custom AI chip, which it says is four times as fast as its previous generation of artificial intelligence chips. 4x speed up. That's actually very significant. That's great. The company said Trainium 3, produced by AWS's Annapurna Labs. Fascinating company. Acquired a decade ago for around 350 billion or 350 million. So it's pretty small acquisition actually. 350 million in AI, you never know. But back then you start a custom silicon company, you could barely clear nine figures on the way out the door. Inner Partner Labs has been working on custom silicon for Amazon for a long time. They actually do have a custom CPU at AWS to accelerate silicon CPU based workloads. Then for the last few years they've been working on GPUs or ASICs for accelerated workloads. And so this custom chip design business, Annapurna Labs, can reduce the cost of training and operating AI models by up to 50% compared with systems that use equivalent GPUs. The chips are meant to provide a stronger backbone of computing power for software developers like Dean Leiters. Leiters, the co founder and co an executive chief executive officer of the startup Descartes who we had on the show and Descartes.
- [02:06]
Is valued now at $3.1 billion. Let's go. So if you don't remember, Descartes came on and Dean was doing live AI video generation while he was doing the interview with us. It was really crazy.
- [02:20]
B
Yeah, he basically, yeah, it was real time. He looked like he was in a video game, but it was happening with little to no delay. Really, really cool demo.
- [02:29]
A
He said his company had a breakthrough enabled by a Trainium 3 chip. By the Trainium 3 chip. After trying out several other competitor chips, including Nvidia's processors. Dozens of programmers and AI researchers from his San Francisco based company had been trying four months to train a version of Descartes flagship AI powered video generation application known as Lucy, that would be able to render footage in real time without bugs or hiccups. AWS gave Descartes early access to training 3. After meeting with the startup and being impressed with founders, the company was two weeks into a marathon coding session in a rented house in Silicon Valley, which I think he took us on a tour of while he was in wizard land, an AI generated sci fi world. It was very fun that a few of his employees were celebrating wildly behind him. The moment that I saw it worked, I saw four people just start jumping up and down, said Dean. The next question was how fast can we get it to market and start changing industries with it? The launch of Trainium 3 is the latest broadside against Nvidia, which dominates the GPU market. A flurry of deals in recent months have caught the attention of investors, indicating that more AI firms are seeking to diversify their suppliers by buying chips. Meta Platforms is in talk with Google to buy billions of dollars worth of advanced AI processors known as TPUs. And OpenAI has struck deals with rival Nvidia, rival AMD as well as Broadcom. Very exciting that Descartes got good results out of the Trainium chip. That's awesome. Obviously, I'm sure everyone over at Amazon has been working very hard on that. At the same time, we've heard that Anthropic maybe didn't have that great of an experience with Trainium and that's why maybe they're moving over to TPU a.
- [04:12]
B
Little bit more even though Amazon remains a major shareholder.
- [04:16]
A
Exactly, yeah.
- [04:16]
B
Anthropic.
- [04:17]
A
And so my question is, will AWS buy TPU from Google? I asked Matt Garmin that question.
- [04:23]
B
You didn't ask me that question. Yes, I said they will be mocked.
- [04:26]
A
They would be mocke, which would be mocked, which is ridiculous.
- [04:29]
B
Where I was like, please, please, my arch rival, can I please get some chips for my data center to compete with your data center?
- [04:38]
A
Okay, well let's actually go to what Matt Garman, the CEO of AWS, said on TBPN yesterday, because I asked him, Will you be buying TPUs? And he said, hey look, we are very excited about Trainium and we think it has enormous potential and we absolutely think there's a benefit to optimizing every layer of that stack. People were joking on the timeline, you know, oh, there's Trainium chip. And somebody was like, all five people using Trainium are ecstatic. You know that there's this new news. Amazon's so bad at hype. Trainium is used by 500 million people through Bedrock, but their marketing team just can't. AWS is undervalued, blah, blah, blah. And he's obviously a bull on the stock. But what's interesting is that like, it is, it is deployed.
- [05:20]
B
I met some of their GTM staff today. Let's just say you'll have years to accumulate stock at cheap prices.
- [05:26]
A
Yes, there obviously is value, even if Trainium winds up being a particular niche. Like maybe it's for real time video. Like maybe that's the. Maybe that's what it gets really good at.
- [05:35]
B
The thing with real time video, that's interesting. Something that Descartes is focused on is working with live streamers, specifically on Twitch.
- [05:43]
A
Yeah.
- [05:43]
B
Amazon owns Twitch.
- [05:44]
A
Oh, that'd be cool.
- [05:45]
B
That makes that kind of partnership more interesting.
- [05:49]
A
I like that. So obviously there is value to saying, hey, if you go to aws, you can get Bedrock and some services that have been fine tuned specifically for Trainium. You go all the way down, you're going to get very good performance because we have a stack from top to bottom that's efficient. But at the same time, if you're trying to do something that's sort of like not within the Trainium ecosystem, you might have a rough go, you might wind up on a different chip. But he did say something. He said, we are going to support choice for our customers as well. And so we'll continue to offer GPUs from Nvidia as an example. And we have a very tight partnership there. So this idea of customer choice I think is important. And if you go back to Jeff Bezos, he said, we're not competitor obsessed. This idea that Google is their arch rival, that's not in Amazon's DNA. Jeff Bezos said, we're not competitor obsessed, we're customer obsessed. We're customer obsessed. And so if the customer says, look, it's great that you acquired annapurna labs for $350 million, I'm really happy with what you've done with Trainium 3. It doesn't work for me. I'm the customer and I want you to give me an Nvidia GPU in your server or in your data center, or I want you to give me a TPU in your server, they might do that because that's actually in Amazon's DNA.
- [07:04]
B
Yeah. And then the follow up question Is, is there any world where Google sells TPU to Amazon?
- [07:11]
A
Already they are partnering. Like this was another partnership that came out separately. There was an announcement of an AWS partnership with Google Cloud. Now they aren't buying TPUs, but what they're doing is they're enabling customers to establish private high speed links, links between the two companies, computing platforms in minutes instead of weeks. And so the general here, the general idea here is that Google has some amazing AI capabilities that customers are just struggling to match on AWS at this point. And the same thing's happening on Microsoft as well, because on Azure you have access to OpenAI models that you might not have access to on on aws. And so even though your whole infrastructure might be on aws, you might be going back and forth to GCP constantly. Companies used to think about AI as a special piece of their application. So it would be fine to bounce around to another cloud to get the best possible results. But if the next generation of companies. I'm sure we'll talk to some of the AI focused YC demo day companies today about this.
- [08:11]
B
I hope there's at least one, I.
- [08:12]
A
Hope there's at least one company that's doing something with AI that would be a real treat. So it used to be fine to bounce around. Now the next generation companies, they're maybe making their entire infrastructure decision based on who has the best AI products. What are you laughing at?
- [08:27]
B
I'm laughing because I texted Simon. They have Turbo Puffer has a booth at aws. I said, how's it going at Re Invent? And he says, I'm not there. I just make it seem like I'm there as a joke because the VCs keep going to the booth. And then our growth intern is like, oh, Simon, I don't know, I think I saw him over there.
- [08:47]
Just continuing, continuing to MOG while ARR skyrockets.
- [08:53]
A
Amazon needs to fight back against this. And allowing high speed interconnect between AWS and GCP solves a piece of that. But will they go further? Nvidia has an insane amount of power right now. They've just ramped full year revenue from 27 billion in 2023 to 60 billion in 2024 to 130 billion in 2025. That's like one of the greatest revenue ramps at scale in history. And then also they grew their net profit margin from 16% to 56%. That's insane. Insane. Yes, Goat. That's why Jensen Huang is on Joe Rogan. And I'm sure it's going to be a fantastic episode. All the hyperscalers and OpenAI. But that creates problems, right, because all the hyperscalers and OpenAI are now sort of incentivized to form a bit of an anti Nvidia alliance to commoditize the accelerator market and drive down those margins a bit. So 56% net profit margins on 130 billion of revenue. They were just sitting there and they're like, there's $50 billion of profit over there. Like, that's a lot of acquisition.
- [10:00]
B
And that's our Perna Labs. That's our cost.
- [10:01]
A
Yeah, that's our cost. Like you're just, you're, you're just eating a lot off of these plates.
- [10:05]
B
How much do you think it hurts Amazon that they don't have a dedicated podcast guy? Like they don't have a Sholto, they don't have a Sam, they don't have a Satya.
- [10:14]
A
You know how much that hurts? Because they definitely have someone in that role. You just don't know them.
- [10:17]
B
That's what I'm saying.
- [10:17]
A
Yeah, they don't.
- [10:18]
B
They might have, they might have the tit, but they're not really in the driver's seat. Right?
- [10:22]
A
Yeah, they don't have a rune.
- [10:24]
B
They don't have the rune. Right. They don't have a Sholto.
- [10:26]
A
Yeah, they should step it up. They should definitely get someone. Fortunately, I mean the semianalysis crew was over there taking pictures, sharing photos in the timeline of the Trainium 3 Ultra server. Liquid cooled with a lot of hard eyes. That's a glowing endorsement from the semianalysis crew. And look at this. Very purple. I wonder if that's like intentional Google is having this kind of success with TPUs. What about Amazon's Trainium? Trainium is new and underpowered. Just 667 t flops. BF16. It has lots of HBM, but the bandwidth is lower than the H100. TPU v6e is competitive 8 100. Not on HBM or bandwidth. And Ironwood is competitive with Blackwell on FLOPS bandwidth and HBM capacity. I expect Ironwood to quickly gain market share as it ramps up. As you can see from throughput/tco Nvidia vs. Trainium, Rubin, Mogs, Trainium 3 harder than Blackwell vs. Trainium 2 on TCO training flops and reduces the gap by 5% on TCO MEM bandwidth. So the gap between Nvidia and Trainium is actually increasing rather than decreasing. By the way, this math was done before CPX was introduced. I won't be surprised If CPX plus Ruben is cheaper than Trainium for inference. So I do think that there's a world where.
- [11:40]
Where there's something specialized like what's going on with Descartes, some sort of special model that thrives in what Trainium is good at and they can further niche down. But we'll see. I mean maybe they come from behind and they just destroy tpu and we're all talking about Trainium next year. We gotta say a little rest in peace.
- [11:58]
B
Rest in peace to Claude.
- [12:00]
A
San Francisco's beloved albino alligator has passed away at age 3:30. That's a good age. I don't know how long alligators typically live, but I'm glad looking it up.
- [12:11]
B
30 to 50 years.
- [12:13]
A
Okay. A little bit short. But Claude was of course often supporting.
- [12:18]
B
Often reaching 70 years or more.
- [12:20]
A
Yes.
- [12:21]
B
You know, obviously people started speculating immediately. Anthropic of course was the sponsor Claude. And you know, people were wondering was, was there foul play involved? Was it possible this poor dinosaur, not dinosaur alligator passed the day that they, that it, that it got announced that they've hired IPO lawyers. Some people were speculating could. Is it possible Claude was sacrificed to the Capitol markets gods in some type of ritual? But anyways, he, look at the, look at this expression he has on his face. Can we zoom in a little bit? What, what a, what a cool guy. And he will be remembered.
- [13:08]
A
Yeah. Trump administration will invest $150 million into a lithography startup called XLite. Its first Chips act award. Chatted this morning with XLite CEO. There's a few lithography companies now. We've had some on the show. It's a, it's a very interesting tier of investment, like $150 million from the government. That feels like a Series B. They did raise a Series B this past summer led by Playground Global. Makes sense that the government's investing in Intel. Pat Gelsinger, of course, former Intel CEO, now he's getting involved in XLite. Marshaled 40 million of capital, went and got 150 from the government. There's also another AI startup that wants to remake the $800 billion chip industry. This one's in the Wall Street Journal. Founded by ex Google researchers. Recursive intelligence raised 35 million with backing from Sequoia to automate chip design. Obviously this is not lithography, this is the design process.
- [14:02]
B
But still this is AI for, for AI chip design.
- [14:05]
A
Oh, that's right, yes. On a quiet residential street a few blocks from Stanford University, two former Google researchers are launching A startup they hope will remake the $800 billion chip industry. Trying to build software that can automate the design of cutting edge chips. A prospect that would allow every company to build their own chips from scratch. Working from the top floor of a suburban home, the duo recently raised 35 million to kickstart Recursive Intelligence with funding from Sequoia Capital and Strike. We got it.
- [14:31]
B
The recursive. We got to add putting it in the name.
- [14:34]
A
We got to add that to the list of. Because there's standard capital, modern capital, standard intelligence, modern intelligence, raw intelligence, raw intelligence. Wow. The company 35 million for evaluation of 750 million. That's very low dilution. What 5% or something like that.
- [14:51]
B
VCs were mocked high.
- [14:53]
A
Yeah, I would have assumed this would be a very, very capital intensive business. But I suppose if it's just a software that they're develop, maybe, maybe they have more control here. Thoughts on AI progress? He says he's moderately bearish in the short term, but explosively bullish in the long term. Very interesting. So he says he's confused why some people have short timelines. They say AGI is coming soon, but at the same time they're bullish on rlvr which is reinforcement learning with verifiable rewards. He says if we're actually close to a human like learner, this whole approach is doomed. Currently the labs are trying to bake in a bunch of skills into these models through mid training. There's an entire supply chain of companies building RL environments which teach the model how to use Excel to Write financial models for example. I think we're actually talking to an AI XL analyst for Excel power users called Crunched at 1250 YC Co. In the context of when does AGI arrive? When does superintelligence arrive? I understand Dwarkesh's point. Either these models will soon learn on the job in a self directed way making all of this pre baking pointless, or they won't. Which means AGI is not imminent. Humans don't have to go through a special training phase where they need to rehearse every single piece of software we might ever use. When we see frontier models improving at various benchmarks, we should not, we should think not just of increased scale and clever ML research ideas, but billions of dollars spent paying PhDs, MDs and other experts. One counterargument I've heard from the takeoff within five years crew is that we have to do this clue GRL in service of building a superhuman AI researcher and then the million copies of Automated ILIA can go figure out how to solve robust and efficient learning from experience. This gives the vibes of, we're losing money on every sale, but we'll make it up in volume. This automated researcher is somehow going to figure out the algorithm for AGI, something humans have been banging their heads against for the better part of a century while not having the basic learning capabilities that children have. That seems super implausible to me. You've been asking about economic diffusion. What is the rate that we're diffusing? Let's see what Dwarkesh has to say about economic diffusion. He says that economic diffusion lag is cope for missing capabilities. I'm very sympathetic to this because when I go to the doctor's office and they hand me a piece of paper, I know that a web form is good enough. The capabilities of the digital form are complete. No, it's just a diffusion problem. There's just someone who runs that doctor's office is like, I like doing it the old way. Right. And that's the economic diffusion lag problem that I think is real in a.
- [17:42]
B
Lot of scenarios right now. AI is great at generating text. Right. It's great at kind of analyzing a piece of content and then generating text based on that.
- [17:52]
A
Yeah.
- [17:53]
B
And yet we still have multiple people on the team at TVPN whose job is to find interesting moments of the show and then create captions around that and share it to X and Instagram and YouTube and other platforms.
- [18:05]
A
And Dwarkesh did that too, where he was trying to find the most interesting pieces of a full podcast with one Big Gemini prompt. And he was trying all the different models and couldn't get it to actually find, like, the most salient and viral points.
- [18:19]
B
The other thing that stands out is, like, one of the seeming missing capabilities is, like, ability to identify humor or even something like, it's almost emotional. So Ilya and Dwarkesh talked about this, where I think Ilya was giving the example of scientists studied people who had had various brain injuries that limited their ability to experience emotion. And when they took out emotion, it took them. It can take somebody two hours to figure out which pair of socks to choose. And they were kind of, like, stunned. It's just a pair of socks. You know, what's going on in your day? Why do you need emotion in order to make that kind of decision? And so it seems like, at least in AI, a missing capability is like, okay, finding out, like, what's an interesting moment of a podcast. Right. Is it something that makes the audience member Feel something. Right. The other thing that's notable is like, on wap, one of the best, one of like the top jobs that people do on WAP or way they make their first dollar online is just like clipping for various content creators and media companies. And some of the clips that they make are so sloppy. Like it's literally just like a random segment of the show and they're blasting it out from like 20 different accounts. And the fact that we're still paying humans to do that, still, I mean, it just feels notable.
- [19:40]
A
Well, let's read Dwarkash's take on economic diffusion. Lab lag being COPE for missing capabilities. Sometimes people will say that the reason that AIs aren't more widely deployed across firms and already providing lots of value outside of coding is that technology takes a long time to diffuse. Dorkash thinks this is cope. He says people are using this COPE to gloss over the fact that these models just lack the capabilities necessary for broad economic value. New technologies take a long time to integrate into the economy. Well, ask yourself, how do highly skilled, experienced and entrepreneurial immigrant humans manage to integrate into the economy immediately? Once you've answered that question, note that AGI will be able to do those things too. If these models were actually like humans on a server, they'd diffuse incredibly quickly. In fact, they'd be so much easier to integrate and onboard than a normal human employee. They could read your entire slack and drive in minutes and immediately distill all the skills that your other AI employees have. Yeah, I agree with that. The one thing that I don't necessarily agree with here, he says, well, ask yourself this quote from Stephen Burns. How do highly skilled, experienced and entrepreneurial immigrant humans manage to integrate into the economy immediately? I mean, they do sort of integrate into, into the economy immediately, but like the, the immigration flow is like a slow process. Like, it doesn't just happen immediately. It's not just like, you know, the amount of immigration went from like zero to like, I don't know, a million people or something. Like, it's like people move around. There is like a, there is a bit of a drag. But I understand what he's saying here and it does make sense. Silicon Valley is rallying behind a guy who sucks.
- [21:25]
What does that mean?
- [21:27]
B
Like pure ad hominen.
- [21:29]
A
It's, it's, it's rage bait. It's going to go hard. It already got a thousand likes on a linked article. The Verge is not putting up a thousand likes per link. So this is outperformance and it's heavily paywalled. You cannot learn how David Sack sucks without subscribing to that thing. They did a good job. You got to pay. You got to pay. You want to know why he sucks?
- [21:49]
B
That'd be really funny. If behind the paywall is like, we're just kidding. He's actually, we think that. We think the New York Times missed on this one. The startup told me that one of their investors didn't like that they were selling to newly founded startups and wanted them to sell to bigger companies who have more money. If investors tell you this, write them off as idiots. Selling to startups is the best thing you can do. I'm sure many of the companies we're talking with today will be selling to other companies in the batch. A lot of people say that's bad. They try to say, like, YC is a circular economy, but you have to ignore the hundreds of very real businesses that have been created through YC and gone on to work with every kind of company in the world.
- [22:36]
A
Yeah, even if there's some sort of insular circular economy in the startup ecosystem, there's a pretty immense amount of pressure to actually deliver something that's valuable.
- [22:47]
B
Yeah, they're being rational. It's not like. I'm sure there's been small instances where companies were actually, you know, had had somewhat bad behavior. But in general, it's like, if I'm going to pay for your the SaaS tool or the beta that you're running, it has to be good.
- [23:03]
A
So there's a $1.5 billion judgment against Anthropic for including 480,000 books in training. Their AIs. Five of my books are among them. Word is there might be $1,500 payout per book. According to my agent, Max Brockman, I wrote to my agen the following. If any payment comes to me, please send it back to Anthropic with my thanks for including my books and their AIs. The Judgment website offers a way to opt out of the payment, but I found it cumbersome, so I didn't. I'm principled, but too lazy to be highly principled.
- [23:37]
B
The secondary markets are rife with fraud and bad actors, and it pains me to see these bottom feeders profiting off of Anduril's growth while fleecing retail investors through unreasonable or opaque fee structures. In this week's episode of Nonsense, Ignite vc, a fund we've never taken a meeting with or had any contact with whatsoever. Founded by Brian, who we've never met. Is soliciting investors via public Google Doc to invest in an SPV that will in turn invest in another SPV that will in turn potentially enter into a forward contract with a supposedly though unnamed early Enduril employee. A few problems here. First off, so called forward contracts are notoriously hard to settle in private companies and counterparty risk is extremely real. What about the many complicating corner cases like acquisitions where shares don't trade or marriages, divorces or deaths where ownership of the underlying shares is complicated, just generally a risky structure to close that I don't think most folks actually understand? Yeah, if you enter into a forward contract and you basically buy the right to the future value of some shares and then somebody gets again married or divorced or passes away or bankruptcy is another situation where you might not be actually able to collect even if the even if your investment should have generated some return. Matt says Second, this deal memo includes basically no details about Anduril's performance, no revenue figures whatsoever, no product specifics. I guess that's good, right? Like if they were if they were just floating around information that they had acquired, almost as if it's soliciting investors to invest on hype and momentum and not fundamentals. Generally I'd advise folks to be skeptical of any deal memo lacking basic details. Third, forward contracts are explicitly disallowed by Anduril's stock plan and bylaws, which means that Anduril will never consent to Team Ignite's SPV actually taking possession of these shares while we are privately held. Zero chance. And finally, the memo spends most of its time talking about the structure and fees, which are insane. A double layered SPV with all legal and admin costs pass through in addition to an 8% upfront fee, 3% annual fee for two years, 20% carried interest, and the craziest part, an implied price per share that is completely insane. In this case the implied pps is 115% higher than the most recent preferred raise from 9 months ago. Flattered I suppose, but also puts these investors in an almost absurd position by paying more than double the price per share of our most recent transaction as stated at the top. I don't know Brian or Team Ignite at all. Maybe they're kind of wholesome people and this is all a big misunderstanding. But if I were an investor looking at this quote, I'd run for the hills. And I believe the founder replied and said appreciate the heads up. The document reference was an internal draft prepared for discussion with an existing LP and was not intended for public circulation. It appears someone shared it without authorization. And we're looking into how that happened.
- [26:23]
A
But do you see what and then there's like seven people that share a screenshot of like a direct email we got with this exact number.
- [26:29]
B
Okay. And the other thing is they say not soliciting investment for any anduril related vehicle. Matt says, really? The draft was written by your founder and managing part, watched him edit the doc in real time, and he has a screenshot of like the founder's name in Google Doc.
- [26:48]
A
What a mess.
- [26:49]
B
Don't do this.
- [26:50]
A
Don't do it. Instead, why don't you start a company and apply to Y Combinator, build an actual business instead of going around hustling SPVs and companies that don't want to sell shares. We will talk to you later.
- [27:03]
B
Cheers.
- [27:03]
A
Goodbye.