David Friedberg says Google's AI stars are walking out because the company is funding data centers, not their models. "So now if you're one of the great computer scientists, you're Demis, you're Jeff Dean, you're this whole crew, and you're inside at Google and they're allocating capital not to your models, not to the things that you're most interested in." "But they're allocating capital to infrastructure and data centers and supporting the broad ecosystem of models." "You start to say, well, given the fact that I can go down the road and visit Brad Gerstner and a couple other people and raise a couple billion dollars at a multi-billion dollar pre-money with a PowerPoint deck because I'm the greatest in the world at doing this, that might be a better path for me." "And I think that that's the moment. So the way I would frame it is capex is high alpha, low beta in data center infrastructure. That capital and model development theoretically could be high alpha, but it's very high beta. It's a very risky way to deploy capital."

Palmer Luckey reveals Anduril is building something that makes 50 years of firearms development irrelevant, but he can't tell you what that is today. "I'm building things today that we could have built 100 years ago and it would have changed everything." "We're building things, for example, that would have made the last 50 years of firearms development completely irrelevant." "Like, let's make every gun that has been designed since the 50s or so, since Eugene Stoner invented the AR platform at ArmaLite Rifle, we're going to make everything since that irrelevant. And you could have actually done this in the 20s." "I can't tell you exactly what it is."

$AMZN CEO Andy Jassy reveals customers want Trainium chips outside AWS and says there's a real chance Amazon sells them that way. "On the question about selling Trainium, we're quite excited about what's happening in our chips business." "As I mentioned earlier, it's over $25 billion in annual revenue at this point. We think we have the leading price performance chip in both the AI space with Trainium and in the CPU space with Graviton." "The fact that we have multi-year, multi-gigawatt commitments from the two largest AI labs, Anthropic and OpenAI, and more and more companies, as I mentioned in my opening comments, using Trainium is exciting and promising." "And we just have an incredible amount of demand for Trainium. So there are a lot of customers who are very excited about using it in the form that we're providing right now." "We do have an increasing number of customers who are interested in us providing the Trainium chips to them separate from our cloud. And we're actively having those conversations and exploring. And I expect there's a real chance we'll do that in the future."

Stripe's Patrick Collison says we should no more mourn coding than we mourned hand-written assembly, and admits he misses writing it himself. "Obviously on the one hand, it used to be really fun to write all this assembly and machine code and to optimize your instructions and layout and memory and everything." "And now we don't have to do that anymore. Compilers do it for us. We don't mourn it too much." "And so maybe in the same way we shouldn't mourn source code. We should just transcend the plane of instructions to Claude et al." "But emotionally, I miss it." ______ More from his fireside here:

Jordi Visser reveals he's playing the AI buildout through its bottlenecks that will still grow earnings, naming Micron and GE Vernova as scarcity right next to Bitcoin. "But right now, I'm trying to invest in scarcity because I think we're at the phase of innovation where we can't actually produce the energy necessary." "So we're converting, as David Friedberg said in the podcast this past weekend, we're converting molecules into bits. And I want to be involved in the molecule side right now." "And that's why I focus my attention on scarcity. So scarcity is not just Bitcoin." "Scarcity is Micron. Scarcity is GE Vernova. Scarcity is all of these things that are necessary to build out the compute that are still going to have the earnings growing."

Patrick Collison reveals 25% of all Delaware corporations are now started through Stripe Atlas. "And in the case of Stripe, I really love it because we're working with the world's most interesting and innovative companies." "Like we're 25% of all Delaware corporations are started with Stripe via Atlas." "And then we get to partner with them and work with them and hear from them and get their feedback and get the request and everything through the entirety of the journey, up to being the Shopifys and the OpenAIs and all the standout successes." ______ Takeaways from Patrick's fireside:

Simile's Joon Sung Park reveals Simile won its first customers by rerunning a large consulting firm's three-to-six-month study on the opening call and answering in two minutes, and says that's how it won early customers who had been running on gut instinct. "There are so many questions that they are truly relying on their gut decision today. That if they can get some form of evidence to at least directionally guide them in the right path, then they're ready to try it." "And then they very quickly realize that, oh, this is actually an amazing way to interact with a lot of data. This is amazing way to gain evidence that is actually quite accurate." "And one of the ways we actually got some of our first customers was in the first call, they actually had a finding from large consulting companies and they basically queried our system. Hey, if we were to rerun this, what would the system say?" "And we predicted the outcome of studies that took three to six months, but just within two minutes."

Meta's AI chief Alexandr Wang says building a startup used to be David versus Goliath when he started Scale ten years ago, but agents have now enabled it to be Goliath versus Goliath. "When I started Scale 10 years ago, if you started a company, you had to be, it was like David versus Goliath." "And you had to be clever and you had to find an angle into the market and you had to figure out a way to compete even though you had much fewer resources." "And now I actually think with the power of agents and AI, broadly speaking, it's much closer to Goliath versus Goliath."

SemiAnalysis's Doug O'Laughlin predicts H100s go worthless once inferencing one model takes 100 of them saying it'll be better to "just let the old girl go" and get a B300, and says a B200 to B300 price gap would confirm it. "We talk about these new models. It's very clear to me that the old chips will become worthless." "You know, everyone's like, oh, the H100 is an appreciating asset. But at some point in time, it's going to take like 100 H100s to inference one of these models, and you're just like, dude, just let the old girl go." "Just get a B300 or, you know, a VR200 instead, right? That's my speculation, and we'll see, we'll obviously be writing notes about it." "But it's like the true confirmation of this trend is if there is a pricing divergence between B200 and B300."

Benchmark's Chetan Puttagunta says Legora is differentiating itself on the last mile tooling and expertise of what lawyers need rather than the next marginal line of code. "There's a lot of innovation here where you'll see some companies building custom models as an example for their application. And that's a mode of differentiation, because those custom models are very different than the foundation models." "So whether it's like custom image models, custom video models, wherever it is, those are differentiated against the large language models." "And there's a business being built around the differentiation of the model itself." "And then what we've talked about with companies like Legora is that they've built so much tooling and expertise on the last mile of what lawyers need that it's not the next marginal line of code that people are paying for. It's really that expertise of helping lawyers do their job better."

The CEO of America's HBM champion Micron, who also founded SanDisk, was almost never let into the country, denied multiple times. A persistent father paved the way. There was fateful timing too, but you make your own luck and sometimes someone else makes it for you.

$LITE CEO Michael Hurlston says replacing a circuit board's copper traces with optics means a 1000x more optical lanes, though he doesn't expect it before 2029, and that disaggregating HBM from the GPU will need optical interconnect too. "Yeah, this is literally a thousand X. So we go one, ten, hundred, thousand. So it's a massive, massive market. And this is exactly what you're describing." "So on a printed circuit board, on a PCB itself, you have the need now for the same reasons, the resistance of copper on the printed circuit board, the distance over which high bandwidth data needs to travel." "You have a need to replace the copper traces that sit on a printed circuit board with optics." "And so what we've seen from our customers, again, it's not next year, it's not the year after, it's probably 2029 before we start to see this, but even in the high bandwidth memory, where you've covered one of the memory bottlenecks right now, and one of the problems is you have memory stacked in stack die configuration with the compute, with an NVIDIA GPU or a TPU, what have you, and people are looking to disaggregate that for cost reasons." "And when you do that, now the bandwidth, right, between the memory and the compute block is so high that you want to use optical interconnect rather than copper."

people want to dunk on this by saying their weekend harness does better on some dumb one shot benchmark but at scale this is true. these tools are complicated and must factor in user behavior. you do not even understand what this means unless you have a huge user base sending you data all of that gets sent as feedback to the model which gets trained into specific behavior i think the part that's misrepresented though is it's extremely easy to recreate the exact same environment there nothing hidden in the harness, you can observe it all and recreate it. bit tedious to keep up with it but not hard so ultimately even if they're right it's not defensible

This is accurate. That said, we are currently working through a rigorous experiment testing closed and open models with and without Software Factory. Early data is super fascinating and backs up the experiment below.

Gavin Baker: "I profoundly believe the future is composable models" This is him laying out the stack every enterprise ends up running and putting 85% of the queries on a model you trained yourself. His words: "You're going to have what Andrej Karpathy called the Council of LLMs. You're going to have Grok, you're going to have Anthropic, you're going to have OpenAI, Google. You're going to have at least two of those." "But you're also going to have your own open weights model that you RL'd on your data. And you're going to put those two together, the frontier models and your own model, and you are going to get real Pareto dominant outcomes." "Only the hardest ones are then checked by the frontier models." "A misconception that a lot of people have is that open source models are somehow bad for AI. They're awesome for the AI infrastructure providers. They just shift economic value from the margins of the frontier labs to the infrastructure." "Frontier tokens are capturing 90% of the economic value and open source tokens are probably 80% plus of tokens processed, and those ratios may be here to stay." "So I think this is the future. It's coming." Bookmark & watch today ↓

"I'm not perfect but I'm very good at predicting the future. My batting average is very high for a human." - Elon Musk His slugging is a magnitude greater. Legit 1 billion OPS.

Microsoft CFO Amy Hood says that if the AI build turns into an overbuild, Microsoft can throttle back quickly, because the bulk of the spend is short-lived chips it can simply stop buying, not the land and the buildings. "But when you start to think about over the duration, I try to remind people a lot of the expense, especially you see it in CapEx, you see our CapEx really pivot toward what I would call and do call short-lived assets, which really, right, that's CPUs and GPUs that have relatively shorter lead times." "And so if the demand environment changes, you just slow down what is, in fact, the largest component and the driver of COGS." "The investment into land and data center builds is actually quite flexible, right? It's a smaller percentage of the overall cost structure, and timing can be changed on much of that, especially on the builds." "Or you can stagger the timing of the build-out of some of the GPUs and CPUs that you plan to put in." "And so when you think about being able to manage through that, hyperscalers have been doing that for quite a long time in terms of having the flexibility and the understanding to manage those changes in demand."

RecSys is the crux of the bull/bear case on Meta, I wouldn't get too distracted by much else: 1. Meta CFO Susan Li saying on the follow-up call that the Meta RecSys still has gaps and can't fully reason user interests, which is interesting. - "Our current recommendation systems are very, very good at leveraging user interaction histories, but unlike LLMs, they don’t have the ability to reason about content or user interest from first principles." 2. IG Head Adam Mosseri had a similar point on Lenny's pod earlier this quarter that the RecSys is not as sophisticated as people assume but critically he said it's "only now getting as sophisticated as people assumed for many years" ie some form of inflection. - "I think a misconception historically is, until recently, we don't really know as much about you as you think. We were just like, oh, you liked these photos, these people also liked those same photos, and they like these other photos, so you might like those other photos. That's kind of how -- I'm oversimplifying, but that's kind of how it worked. Only now are we actually getting as sophisticated as I think people have assumed we've been for many years." 3. They also talked about the RecSys upgrade at their @Scale conference in a session. - "Recommendation systems are still very limited as it... relies on something called collaborative filtering which means it recommends content based on what others could engage in our platform. So it doesn't address dynamic and personalized motivations and needs for the users." - Basically, they log what you're clicking or watching but they never understood the "why" which is where they're headed next. All in, how fast or slow Meta's RecSys improves or not is the real determinant of whether the infra spend has legit ROI or not. Plus, moving over to a new system probably causes stop gaps on topline too. Talk to any advertiser and they'll tell you every system upgrade is always a cluster.

Microsoft CFO Amy Hood says the AI compute crunch is now showing up in the spot market, and because demand keeps outrunning supply, every efficiency gain Microsoft wrings out of its fleet gets sold almost the moment it's freed up. "There are still constraints in the system. I think we've continued to say, I think now for a number of quarters, that demand continues to exceed available supply, and that certainly remains true." "You can even see it, I think, in some of the pricing that's occurring in the spot market for assets." "So when you think about being able to deliver better, the first thing we focus on, and I tried to talk a little bit about it in my prepared remarks, is efficiency, being able to get more out of everything that we've got in the fleet. That applies to efficiency gains in the CPU fleet. It's going to be efficiency gains in the GPU fleet." "We saw a good work this quarter, in particular, from our engineering teams to make as much of that available as we could. And because of the supply-demand imbalance we've been talking about, when we can make efficiency gains, they are quickly monetized in quarter. And I think that dynamic certainly impacted the quarter positively."