@firesidealpha
Account based inUnited States
About this account
- Account based in
- United States
- Connected via
- United States App Store
Account-level information from X, not a live location or the device used for a specific post.
Learn what the best in business and tech are seeing, then go find your edge @firesidetapes for business legends
United States
Joined June 2026
- Tweets2.2K
- Following371
- Followers35K
- Likes3.4K
Pinned Tweet
Gavin Baker says the physics PhDs calling orbital data centers impossible on X have not out-thought the 10,000 SpaceX engineers who treat it as solved
"It's very hard for me to engage. There are all these people on X and they're like, I am a physics PhD and this is impossible."
"There's a friend, another investor, who actually is a physics PhD, who I had many arguments about it with him, and he's like, I am a PhD and this is impossible. And then he goes to the SpaceX day and talks to the SpaceX engineers, and he's like, well, I was wrong."
"So let's say you're an astrophysics PhD. You are brilliant, you're hanging 100 IQ points on me. Have you thought about this for an hour? Have you thought about it for 10 hours?"
"Because you have 10,000 of the world's smartest engineers at SpaceX who've thought about this each for hundreds if not thousands of hours. And the sum of that, working with very sophisticated engineering tools, is it's a solved problem."
_______
For more on Gavin Baker: firesidealpha.substack.com/p…
Gavin Baker re $SPCX: Investors should start to consider Grok when they think about SpaceX, very few of them talk about Grok at all
"I talk to a lot of investors about SpaceX. And what's just amazing to me is everyone will have a nuanced view on Starlink."
"Maybe they're a skeptic on Starlink direct-to-cell. I think the plan to use these distributed Starlink effectively almost as, we'll call them, small cells, femtocells, is interesting."
"Everybody will have a nuanced view on orbital compute -- whether they're bullish or bearish, they will have a view. They'll have a view on terrestrial compute."
"Very few of these investors talk about Grok at all."
"And it is Pareto dominant on a lot of measures."
"And we have an existence proof that Anthropic went from, whatever it is, $1 billion to $50 billion really fast."
"And if Anthropic is worth, with the $2 trillion per the FT, $3 or $4 trillion, maybe the numbers settle out -- I said I thought that was a market-clearing price, I didn't say that's where I thought it was worth."
"Maybe people should start to consider Grok when they think about SpaceX."
DEVELOPING: Southwest CEO announces Starlink will be on Southwest’s entire fleet by this time next year, making it the first airline to get there
United Airlines CEO says it has more Starlink flights than the rest of the US airline industry combined, with about half of its departures already on Starlink and the whole fleet done by the end of 2027
Starlink is now a “motivating factor” in which airline people fly, so United launched a new tracker to show which flights have it.
“You’ve heard me say in the past that hope is not a strategy. Instead, we’re focused on things we can control, and those are the big investments we’ve made over the past several years in products, service, and technology."
"And one of those things is Starlink, a new and truly differentiating RASM tailwind in 2027."
"Today, about half of United departures are on Starlink-equipped aircraft, and we offer more Starlink flights than the entire U.S. airline industry combined."
"We currently expect to reach 61% of departures by the end of 2026, 87% by June 1st, and available fleet-wide by the end of 2027."
"Importantly, we’re also seeing evidence that Starlink is a motivating factor for customers to choose which airline they fly. That’s why we launched a new tracker so people can check if their flight has it and can follow along as we outfit our fleet.”
Meta CFO Susan Li says over 20% of U.S. Muse users are already $1,000-plus annual ARPU users, skewing to Meta’s highest-value base
“Muse right now skews to our highest value users. I think over 20% of U.S. Muse users today are $1,000-plus annual ARPU users. So, it gives you a sense of kind of what the early adopters look like, and so they are not totally representative of the overall U.S. user base.”
Mark Zuckerberg says Muse users already bring Meta “several hundreds of dollars” a year each across its family of apps, and Muse itself should eventually earn even more per user, mostly through fees on the shopping it handles
“From a business perspective, the average revenue per American user for our existing Muse users is, sorry, for our existing family of apps users is several hundreds of dollars annually."
"And we expect that Muse will reach an even higher level over time due to the amount of commerce that we already see taking place. This will take some time, but we’re starting to work on this now."
"Our business model will have a few components. For most people, the product should be free and we will build commerce tools. So we get paid a fee based on the transactions that we help facilitate. We expect this will be paid for by businesses."
"For people who want the most advanced models or to use extra tokens, they subscribe directly and pay. For business owners who want additional features to use Muse to grow their businesses or for companies that want people to use Muse for work, they will be able to subscribe directly and pay as well."
"And for those who want Muse to help improve the personalization and quality of their content and ads across Meta, they will be able to connect to those experiences as well."
"Just to clarify, the average revenue that I was talking about before is the family of apps revenue for our existing Muse users.”
Sam Altman says OpenAI cannot responsibly open-source its current best internal model until internet cyber defenses are fundamentally rebuilt, and admits similarly capable open models may not be far behind
“I don’t think we could responsibly open-source our current best internal model anytime soon.
"I'm trying to think of what would have to happen to the world. The total cyber-defense model of the way the internet works would have to change. We’d have to build up this bio layer and a long list of other stuff too.
“The reason that this is an interesting and live question is: How far ahead does anyone really believe that model is from someone who will open-source a model of that level? Not super.”
Elon Musk says Sam Altman is “so deceptive, it's crazy” and explains why AI labs should test each other’s models before release
“In fact, I’ve recommended that we do that exact thing. Because there’s a lot of talk of having a regulatory agency. But the thing is that regulatory agencies, it takes a long time to get one established. And their rulemaking is very slow."
"And I said, guys, why don’t we just have a call once a week, just like you’re saying. And that maybe a week before we release a model, we give access to our competitors to say, hey guys, why don’t you try out the new model, see if you see any problems with it."
"I think competitors would keep each other honest in that situation. You know, if they say, oh, I was able to make a nuclear bomb with your AI, it’s like, oh, we better fix that before we release it, you know, type of thing.”
“He’s so deceptive, it’s crazy.”
Fireside Alpha retweeted
Sundar Pichai on Google's new space data center effort, Project Suncatcher:
"We will have a nice 'hello world' moment coming up... We'll get a chance to say hello to the TPUs, which I'm really excited about.
I think we are taking the early steps, but we are deeply committed to the journey. The team is world-class, and in a 10-to-20-year time frame, they're going to have amazing milestones. That's what I'm excited about.
If you think through it from a first-principles standpoint, with energy availability and sun-synchronous orbits, getting access to 24/7 energy is a pretty good reason. There'll be more reasons, I think.
We have so many opportunities to use the physical space, the constrained physical space on planet Earth. The data centers today look big to us, but they are not... It is tough for us to exponentially extrapolate our compute needs 20 years out, so I don't think Earth would look nice if that's how you plan to do it."
Sundar Pichai on Gemini 4 and the AI compute crunch
Even Google can’t get enough compute. I sit down with Google and Alphabet CEO @sundarpichai shortly after the wider release of Gemini 4 Argon to discuss its early reception, where he still wants the model to improve, and what it takes to keep Google at the frontier of AI.
Pichai explains the recent leadership shakeup at Google DeepMind and why building frontier models has become an “all-consuming” job. We also get into Google’s answer to the personal agent race, how he predicts Gemini will do more work on your behalf, and whether agents will finally bring Search, Gmail, and Gemini together.
I ask him about Google turning free cash flow negative, the risks in the AI infrastructure buildout, and how he plans for demand years before knowing what the models will be capable of. He explains the signals giving him confidence to keep spending and how he uses AI agents to follow model development and compute usage inside Google.
Timestamps:
00:00 Gemini Argon launch reception
01:49 Model access and safety
06:20 The AI model race
08:35 Agentic breakthroughs and 2027
11:34 Recursive self-improvement
15:41 DeepMind leadership changes
19:52 Google's AI compute investments
25:20 Is AI spending sustainable?
33:18 How Sundar uses AI agents
35:35 Gemini's personal agent plans
38:37 How AI agents change the web
43:52 Google's space data centers
47:38 Waymo's expansion strategy
50:26 Wing's drone delivery plans
51:36 AI safety and regulation
55:58 Google's quantum computing bet
YouTube: youtu.be/fJBSNEFdm8c
Spotify: open.spotify.com/episode/0cJ…
Substack: sources.news/p/sundar-pichai…
Apple Podcasts: podcasts.apple.com/us/podcas…
This episode's presenting sponsor is @hyperagentapp. Thanks to the show’s premiere sponsors: @Atlassian, @meetgranola, and @mercury.
SemiAnalysis says GPUs are already oversupplied relative to data-center capacity, and warns the imbalance could flip into data-center oversupply in 2028
Jeremie Ontiveros: “The overall forecast for data centers is slightly easier in the sense that it’s based on infrastructure signals you can track ahead of time. Chips can be slightly more volatile, but we also have a lot of very strong indicators pointing to what’s going to happen in 2026 and 2027.
“Generally speaking, we’re a bit more worried about the risk of an oversupply of data centers relative to chips than the other way around. We’re already in a sort of oversupply of chips today, as we speak.
“One always has to be a slightly bigger constraint than the other. If you want to deploy something right now, it’s easier to get GPUs than data-center megawatts.
“But when we compare how these different supply chains are expanding into next year—and especially in 2028, we do see a risk of it flipping in 2028 and potential data-center oversupply.
“That’s going to be the thing we watch most closely: How fast can the chip supply chain expand?
“That is obviously a strong positive for capex—investing in new semiconductor fabs, buying new equipment and that type of stuff. We’ve been flagging that for our clients for several months already.”
Elon Musk says there's no AI demand issue and explains why SpaceX rents its compute on short-term contracts: AI compute is worth more per watt every year, renewals are coming in at higher prices, and the market is a “cheese shop” with no chips left
“Our belief is that the value of compute per watt will increase over time, not decrease. And the reason for that, I think it should be obvious to people really, is that as you improve the AI software, as the models get better and better, they’re able to, you get more useful intelligence per watt. And the value of that compute is proportionate to the useful intelligence per watt. And that’s why you see things like even old computers like H100s renting for more than they rented when they were new.”
“And I think that’s likely to continue. I think we’re going to face a shortage of superintelligence compute. And that shortage is going to remain for a long time. So, I think the smart move is to do short-term contracts and to essentially place the bet on useful intelligence per watt increasing rapidly, which obviously is done, so that’s what’s happened so far.”
“And also if you look at, say, how are our short-term contracts renewing? Are they renewing at higher dollars per watt or lower? And we see them renewing at higher dollars per watt, not lower dollars per watt. We are also frankly hounded relentlessly by every hyperscaler and by dozens of smaller companies to rent more of our compute. So we could rent out our compute several times over if we said yes to everyone. That’s just the situation here.”
“That’s why I posted that sort of video meme of the Monty Python cheese shop sketch where the guy goes into a cheese shop to get some cheese and they actually have no cheese. And so he goes like, what about this kind of cheese? What about that kind of cheese? And they really just don’t have any cheese.”
“So we’ve actually tried to offload some of our inference compute to other cloud providers so that we can use more compute for training. And we get the cheese shop sketch. You know, they’ll say, like, well, we’ll ask them about what about some, you know, GB300s? Nope, don’t have any of those. What about, you know, some GB200s? Nope, don’t have any of those. What about some TPUs? Nope. What about some Trainiums? Nope. What about some, okay, how about some H100s? And like, well, we might have some overseas somewhere and then they’ll check and come back and say, nope, we don’t have that either. So it’s like ridiculous.”
“Yeah. So I don’t think we’re going to have a demand issue here. And like, it’s not impossible. I’m not saying this will happen. It’s not impossible that we will have a six-month payback period for compute next year.”
$LRCX Lam Research CEO Tim Archer pushes back after Lam stock fell when a memory maker said it will prioritize 3D DRAM over 4F²: if the industry leaps straight to 3D DRAM, etch and deposition intensity “is going to rise even faster”
“Obviously, we don’t know exactly what the reaction was or why people would have reacted in that way, because we’ve said for a long time that as all device architectures go 3D, etch and deposition intensity rises. Now, I would say that in some ways, interestingly, the longer it takes to get there, the more we’re seeing momentum for some of our other scaling activities.”
“And so we focused on this, even in this call, advanced patterning. If you want to extend the current architecture, in which patterns are getting even smaller, that’s why we’re starting to see this uptick in processes like Aether, processes like Akara for conductor etch, the processes that are needed to etch and create those smaller structures.”
“So I would say in the short term, etch and dep intensity are rising, and if really we’re making a leap right from 6F squared to 3D DRAM, then etch and dep intensity is going to rise even faster, even sooner than we would have thought, and that’s a great thing.”
$MU Micron CEO Sanjay Mehrotra pushes back on peak-cycle fears, saying memory gets tighter in 2027 and 2028, not looser
"In calendar '27 as well as '2028, we see demand exceeding supply. And in fact, we see greater tightness in the industry in '27 and in '28 versus '26."
"So overall, supply-demand environment is only getting tighter."
"And of course, even as we work hard to bring up capacity, even with any new clean room space coming up in 2028, we see continuing tight supply conditions because, first of all, clean rooms take a long while to build."
Elon Musk says Tesla needs to build 10,000 to 20,000 Optimus bots and train them through self-play at an Optimus Academy, the way a toddler goes from “wobbly” to learning “to wipe their own ass”
“And really, in order for Optimus to get the training data we need, we actually need to make a lot of Optimi. So we need to make, I think, on the order of 10 or 20,000 bots and then have them do self-play. And we’re building up an Optimus Academy where Optimi can practice doing a very wide range of tasks. And then we can use that data in sort of a flywheel to make it better and better at these various tasks.”
“A big difference with Optimus versus the cars is we don’t have this vast fleet of cars from which we can gather training data. So with Optimus, we have to bootstrap it by making the robots and having them do self-play, which is really how a human learns to navigate the world.”
“You can see this when you have a toddler and they grow up, and they initially are sort of wobbly and they don’t know how to do things, and eventually they learn to wipe their own ass, which is a big milestone. Anyone who’s had kids knows. And then they learn to read and do all sorts of things and be able to do these things. And they learn this by playing. And we actually need Optimus to play at scale in order to have it be really good at tasks in the real world."
Elon Musk says Optimus’s remaining AI, design and durability problems should be solved “around the end of next year,” when Tesla starts delivering robots to outside companies, and he remains “convinced that Optimus will be the biggest product on Earth”
“But like I said, I would expect that the AI, the electromechanical design, and the reliability and usage duration, like the usage life of the robot, I do expect these issues to be solved around the end of next year. So that’s the time to expect when we’re delivering Optimus to companies outside of Tesla.”
“And I remain convinced that Optimus will be the biggest product on Earth. And humanoid robots in general will be the biggest product on Earth. So it’s a very big deal, but I don’t want to temper expectations, we still have a lot to get through, and it’s about a year or so that it will take us to get through all that stuff. But to the best of my knowledge, there is no one that is even close to Tesla with respect to a general purpose humanoid robot.”
Elon Musk says Optimus’s remaining AI, design and durability problems should be solved “around the end of next year,” when Tesla starts delivering robots to outside companies, and he remains “convinced that Optimus will be the biggest product on Earth”
“But like I said, I would expect that the AI, the electromechanical design, and the reliability and usage duration, like the usage life of the robot, I do expect these issues to be solved around the end of next year. So that’s the time to expect when we’re delivering Optimus to companies outside of Tesla.”
“And I remain convinced that Optimus will be the biggest product on Earth. And humanoid robots in general will be the biggest product on Earth. So it’s a very big deal, but I don’t want to temper expectations, we still have a lot to get through, and it’s about a year or so that it will take us to get through all that stuff. But to the best of my knowledge, there is no one that is even close to Tesla with respect to a general purpose humanoid robot.”
Elon Musk says Tesla’s Cybercab fleet in Austin has grown 10x since launching about a month ago, and fully autonomous miles are growing so fast the weekly chart “looks like a rocket going to orbit”
“Regarding Robotaxi, we’re seeing the Robotaxi miles driven, fully autonomous miles driven, scale exponentially, meaning scaling very rapidly. And it would be hard to overstate this point. It really is an extremely fast week-over-week growth in autonomous miles. And I don’t think there’s any company in the world that can match our scale with autonomous vehicles as we continue to ramp.”
“The Cybercab fleet in Austin, for example, has grown by a factor of 10 since launch about a month ago. And the cumulative mileage is really, when I look at the weekly chart, it’s wow. I mean, it looks like a rocket going to orbit, basically. It’s growing very fast.”
Elon Musk says Tesla’s Cybercab fleet in Austin has grown 10x since launching about a month ago, and fully autonomous miles are growing so fast the weekly chart “looks like a rocket going to orbit”
“Regarding Robotaxi, we’re seeing the Robotaxi miles driven, fully autonomous miles driven, scale exponentially, meaning scaling very rapidly. And it would be hard to overstate this point. It really is an extremely fast week-over-week growth in autonomous miles. And I don’t think there’s any company in the world that can match our scale with autonomous vehicles as we continue to ramp.”
“The Cybercab fleet in Austin, for example, has grown by a factor of 10 since launch about a month ago. And the cumulative mileage is really, when I look at the weekly chart, it’s wow. I mean, it looks like a rocket going to orbit, basically. It’s growing very fast.”
Elon Musk reveals Tesla demand is “exceeding our production by a significant margin,” with FSD the biggest driver, and he’d “be surprised” if EU-wide FSD approval takes longer than Q1
“So the third quarter was very strong, and I’d like to give a big congratulations to the whole Tesla team for working hard to make it happen. Demand is growing rapidly for our vehicles. It’s exceeding our production by a significant margin, and obviously we’re going to be ramping, and we are ramping production meaningfully. The biggest driver for demand is Full Self-Driving. And today there’s really no competitor for FSD in any car that you can buy.”
“And to be clear, we are producing and delivering as many cars as we can, and we’re pushing very hard for FSD approval in additional markets. We’re now approved in many countries in Europe and we’re hoping for an EU-wide approval. It’s hard to predict, but maybe within a few months, but I would be surprised if it takes longer than Q1. But obviously that’s not a timeframe we control, but we do anticipate being fully approved throughout Europe and many other countries.”
Elon Musk reveals Tesla demand is “exceeding our production by a significant margin,” with FSD the biggest driver, and he’d “be surprised” if EU-wide FSD approval takes longer than Q1
“So the third quarter was very strong, and I’d like to give a big congratulations to the whole Tesla team for working hard to make it happen. Demand is growing rapidly for our vehicles. It’s exceeding our production by a significant margin, and obviously we’re going to be ramping, and we are ramping production meaningfully. The biggest driver for demand is Full Self-Driving. And today there’s really no competitor for FSD in any car that you can buy.”
“And to be clear, we are producing and delivering as many cars as we can, and we’re pushing very hard for FSD approval in additional markets. We’re now approved in many countries in Europe and we’re hoping for an EU-wide approval. It’s hard to predict, but maybe within a few months, but I would be surprised if it takes longer than Q1. But obviously that’s not a timeframe we control, but we do anticipate being fully approved throughout Europe and many other countries.”
Elon Musk says there's no AI demand issue and explains why SpaceX rents its compute on short-term contracts: AI compute is worth more per watt every year, renewals are coming in at higher prices, and the market is a “cheese shop” with no chips left
“Our belief is that the value of compute per watt will increase over time, not decrease. And the reason for that, I think it should be obvious to people really, is that as you improve the AI software, as the models get better and better, they’re able to, you get more useful intelligence per watt. And the value of that compute is proportionate to the useful intelligence per watt. And that’s why you see things like even old computers like H100s renting for more than they rented when they were new.”
“And I think that’s likely to continue. I think we’re going to face a shortage of superintelligence compute. And that shortage is going to remain for a long time. So, I think the smart move is to do short-term contracts and to essentially place the bet on useful intelligence per watt increasing rapidly, which obviously is done, so that’s what’s happened so far.”
“And also if you look at, say, how are our short-term contracts renewing? Are they renewing at higher dollars per watt or lower? And we see them renewing at higher dollars per watt, not lower dollars per watt. We are also frankly hounded relentlessly by every hyperscaler and by dozens of smaller companies to rent more of our compute. So we could rent out our compute several times over if we said yes to everyone. That’s just the situation here.”
“That’s why I posted that sort of video meme of the Monty Python cheese shop sketch where the guy goes into a cheese shop to get some cheese and they actually have no cheese. And so he goes like, what about this kind of cheese? What about that kind of cheese? And they really just don’t have any cheese.”
“So we’ve actually tried to offload some of our inference compute to other cloud providers so that we can use more compute for training. And we get the cheese shop sketch. You know, they’ll say, like, well, we’ll ask them about what about some, you know, GB300s? Nope, don’t have any of those. What about, you know, some GB200s? Nope, don’t have any of those. What about some TPUs? Nope. What about some Trainiums? Nope. What about some, okay, how about some H100s? And like, well, we might have some overseas somewhere and then they’ll check and come back and say, nope, we don’t have that either. So it’s like ridiculous.”
“Yeah. So I don’t think we’re going to have a demand issue here. And like, it’s not impossible. I’m not saying this will happen. It’s not impossible that we will have a six-month payback period for compute next year.”
Meta CFO Susan Li: Meta’s ads engine is moving to LLM-based infrastructure that can “more deeply reason about people’s preferences,” and generative AI tools will let advertisers turn an image into video or sound for each viewer
“The second theme I would say is that we are excited about some of the new research bets that we have as we move into sort of a more LLM-oriented AI infrastructure to help us more deeply reason about people’s preferences, what they’re interested in, both across organic content and recommendations. And we expect that we’ll be able to unlock benefits to our ranking and recommendations engine from that research too.”
“And then the third thing is that we expect that generative AI tools, which have already been very valuable for advertisers, will continue to grow in terms of the features that we make available to advertisers, primarily through the Advantage+ suite, but features that make it easier for advertisers to quickly and seamlessly generate, whether it is creative or take an image and generate video or generate sound, all, I think, in service of making the ads increasingly personalized for the person that they’re being served to. So I would say we feel like there is meaningful runway across all three of those things.”
$META Devansh Tandon reveals the tokens-in, engagement-out flywheel behind Meta's recommendation system, where better models lift Reels watch time 30% and the monetization pays for the next training run
"These scaling curves aren't just academic research. They're driving real product impact at scale for some of the biggest consumer businesses in the world."
"Here's a couple of examples I have from Meta's recent earnings reports. Instagram Reels had a strong quarter, 30% year on year watch time."
"And the optimizations we made to improve the quality of recommendations included simplifying our ranking architecture to enable efficient model scaling and longer interaction histories to identify a person's interests."
"We doubled the length of user interaction sequences used for training Instagram and increased the richness of each user interaction."
"So these are direct parallels to the power law scaling curves for LLMs."
"And I want to introduce this idea of a flywheel of tokens in engagement out, which is what's powering all of these recs models scaling."
"You train a model, you then run inference on it, which is the tokens in, that recommendation model results in better content recommendations, it drives consumer engagement, daily active users' time spent, it translates to monetization and ads or subscription, which pays for the next model training run."
"And so every step on the scaling curve is one loop around this flywheel, and a lot of consumer apps are spinning this core flywheel at the heart of their business. So we have a long way to scale these recommender systems."
_______
For the full breakdown of the lecture on Meta's ads recommendation system: firesidealpha.substack.com/p…
Mark Zuckerberg says compute is "one of the biggest constraints" on Muse: it’s expensive enough to run that Meta hasn’t launched it globally, even with 10 million sign-ups in the US and Canada alone
“Right now for Muse, one of the things I pointed out is we have more than 10 million people in the US and Canada who have set it up."
"But I think one of the biggest constraints so far is that because this is a compute intensive service to run, we have not launched it globally yet. That’s partially because of wanting to optimize it for each place, and there’s some work around connectors for that, but it’s partially that we want to make sure that we do this efficiently.”
“So there’s pros and cons there. The pro of course is that that just makes it even more impressive that you have 10 million people using it and signing up to use it in the first six weeks, because those are all Americans and Canadians in a small number of markets. So if we actually did roll it out more broadly, then you would expect that number to be significantly larger than what we’re seeing today. But then on the flip side, we actually have to go do the work to bring the efficiency in line to make that make sense.”
Mark Zuckerberg says Muse users already bring Meta “several hundreds of dollars” a year each across its family of apps, and Muse itself should eventually earn even more per user, mostly through fees on the shopping it handles
“From a business perspective, the average revenue per American user for our existing Muse users is, sorry, for our existing family of apps users is several hundreds of dollars annually."
"And we expect that Muse will reach an even higher level over time due to the amount of commerce that we already see taking place. This will take some time, but we’re starting to work on this now."
"Our business model will have a few components. For most people, the product should be free and we will build commerce tools. So we get paid a fee based on the transactions that we help facilitate. We expect this will be paid for by businesses."
"For people who want the most advanced models or to use extra tokens, they subscribe directly and pay. For business owners who want additional features to use Muse to grow their businesses or for companies that want people to use Muse for work, they will be able to subscribe directly and pay as well."
"And for those who want Muse to help improve the personalization and quality of their content and ads across Meta, they will be able to connect to those experiences as well."
"Just to clarify, the average revenue that I was talking about before is the family of apps revenue for our existing Muse users.”
Meta CFO Susan Li says Meta isn’t giving a 2027 revenue outlook yet, but expects AI gains in the core ads business to “sustain strong overall revenue growth throughout 2027,” while Reality Labs losses narrow to around $15 billion next year
“Our AI investments are already delivering substantial incremental advertising revenue through stronger engagement, better recommendations, and more efficient monetization. Building on this progress, our MSL investments are creating the technology and infrastructure needed to unlock significant new revenue opportunities and strengthen capabilities across the business."
"While these investments will take longer to generate returns than those in our core business, the breadth of monetization opportunities and internal efficiencies gives us confidence in their long-term potential.”
“Looking ahead, although we are not providing a full year 2027 revenue outlook, the benefits our AI investments are delivering to the core business, combined with emerging opportunities from model advances and product launches, position us to sustain strong overall revenue growth throughout 2027.”
“Separately, we remain focused on improving Reality Labs profitability. As Mark mentioned, we expect losses to narrow next year to around $15 billion as we reallocate investment towards other priority areas, primarily AI."
Meta CFO Susan Li reveals LLM-based ad systems are lifting conversions: a new “Meta Generative Recommender” added nearly 2% on Facebook feed, and the GEM ranking model added 9.6% for app ads
“Within our ad systems, we are continuing to drive performance gains as we deploy more predictive models that understand users more deeply and deliver ad experiences better aligned with their unique preferences while also making our infrastructure more efficient.”
“Last quarter, we introduced Meta Generative Recommender, which uses LLMs to reason across ad content and user preferences to predict the best ad for each person. In Q3, we launched this approach to generative retrieval on Facebook feed, driving a nearly 2% increase in conversions across targeted segments.”
“In addition, we expanded our GEM model for ads ranking to additional surfaces, driving a 1.7% increase in conversions for off-site conversion campaigns and a 9.6% increase in app ad conversions. And we continue to scale our retrieval systems efficiently. In Q3, improvements to Andromeda, such as scaling model complexity and throughput, increased conversions by 1.4% on Facebook.”
Meta CFO Susan Li reveals LLM-based ad systems are lifting conversions: a new “Meta Generative Recommender” added nearly 2% on Facebook feed, and the GEM ranking model added 9.6% for app ads
“Within our ad systems, we are continuing to drive performance gains as we deploy more predictive models that understand users more deeply and deliver ad experiences better aligned with their unique preferences while also making our infrastructure more efficient.”
“Last quarter, we introduced Meta Generative Recommender, which uses LLMs to reason across ad content and user preferences to predict the best ad for each person. In Q3, we launched this approach to generative retrieval on Facebook feed, driving a nearly 2% increase in conversions across targeted segments.”
“In addition, we expanded our GEM model for ads ranking to additional surfaces, driving a 1.7% increase in conversions for off-site conversion campaigns and a 9.6% increase in app ad conversions. And we continue to scale our retrieval systems efficiently. In Q3, improvements to Andromeda, such as scaling model complexity and throughput, increased conversions by 1.4% on Facebook.”
$META Devansh Tandon reveals the tokens-in, engagement-out flywheel behind Meta's recommendation system, where better models lift Reels watch time 30% and the monetization pays for the next training run
"These scaling curves aren't just academic research. They're driving real product impact at scale for some of the biggest consumer businesses in the world."
"Here's a couple of examples I have from Meta's recent earnings reports. Instagram Reels had a strong quarter, 30% year on year watch time."
"And the optimizations we made to improve the quality of recommendations included simplifying our ranking architecture to enable efficient model scaling and longer interaction histories to identify a person's interests."
"We doubled the length of user interaction sequences used for training Instagram and increased the richness of each user interaction."
"So these are direct parallels to the power law scaling curves for LLMs."
"And I want to introduce this idea of a flywheel of tokens in engagement out, which is what's powering all of these recs models scaling."
"You train a model, you then run inference on it, which is the tokens in, that recommendation model results in better content recommendations, it drives consumer engagement, daily active users' time spent, it translates to monetization and ads or subscription, which pays for the next model training run."
"And so every step on the scaling curve is one loop around this flywheel, and a lot of consumer apps are spinning this core flywheel at the heart of their business. So we have a long way to scale these recommender systems."
_______
For the full breakdown of the lecture on Meta's ads recommendation system: firesidealpha.substack.com/p…
Mark Zuckerberg says Muse users already bring Meta “several hundreds of dollars” a year each across its family of apps, and Muse itself should eventually earn even more per user, mostly through fees on the shopping it handles
“From a business perspective, the average revenue per American user for our existing Muse users is, sorry, for our existing family of apps users is several hundreds of dollars annually."
"And we expect that Muse will reach an even higher level over time due to the amount of commerce that we already see taking place. This will take some time, but we’re starting to work on this now."
"Our business model will have a few components. For most people, the product should be free and we will build commerce tools. So we get paid a fee based on the transactions that we help facilitate. We expect this will be paid for by businesses."
"For people who want the most advanced models or to use extra tokens, they subscribe directly and pay. For business owners who want additional features to use Muse to grow their businesses or for companies that want people to use Muse for work, they will be able to subscribe directly and pay as well."
"And for those who want Muse to help improve the personalization and quality of their content and ads across Meta, they will be able to connect to those experiences as well."
"Just to clarify, the average revenue that I was talking about before is the family of apps revenue for our existing Muse users.”
Mark Zuckerberg says 10 million people set up Muse agents in six weeks, a quarter use it daily, and shopping volume was so far above expectations that Meta and Stripe had to raise credit limits in week one
“In just a month and a half, more than 10 million people in the U.S. and Canada have set up Muse personal agents. Growth continues to be strong, and we have been number one in the app store almost every day."
"Engagement and retention are also off to a strong start. About 25% of people who have set up Muse use it on any given day. For our earliest adopters, the longer that they use Muse, the better that retention seems to get."
"So to put this in perspective, this is more than twice the level that Threads started at, or where we believe that other AI apps started at as well."
"The most common use case for Muse is shopping, with more than a quarter of people who are using Muse engaging in this, with an average purchase size of over $200. Plaid told us that they hadn’t seen activity like they’re seeing with their connector on any other service before."
"And in the first week, we had to work with Stripe to raise credit limits since the volumes were above what any of us expected.”
Mark Zuckerberg says 10 million people set up Muse agents in six weeks, a quarter use it daily, and shopping volume was so far above expectations that Meta and Stripe had to raise credit limits in week one
“In just a month and a half, more than 10 million people in the U.S. and Canada have set up Muse personal agents. Growth continues to be strong, and we have been number one in the app store almost every day."
"Engagement and retention are also off to a strong start. About 25% of people who have set up Muse use it on any given day. For our earliest adopters, the longer that they use Muse, the better that retention seems to get."
"So to put this in perspective, this is more than twice the level that Threads started at, or where we believe that other AI apps started at as well."
"The most common use case for Muse is shopping, with more than a quarter of people who are using Muse engaging in this, with an average purchase size of over $200. Plaid told us that they hadn’t seen activity like they’re seeing with their connector on any other service before."
"And in the first week, we had to work with Stripe to raise credit limits since the volumes were above what any of us expected.”
Alexandr Wang reveals Meta CFO Susan Li pulled him aside "many times" at an exec offsite in Hawaii to warn that if the model didn't work, it would be very, very bad in financial terms
ZUCKERBERG: "So I think this is part of the excitement of booting up a frontier lab now, where it is basically like the ante is extremely high."
WANG: "And I think this is, and I kind of viscerally remember this. We were in Hawaii at the end of last year. It was kind of like the whole exec team got together."
"And yeah, Susan pulled me aside many times and was like, you know, I just want to tell you if this doesn't work, this is going to be really, really bad. It's going to be very, very bad in financial terms."
"It's like, it has to be good. If it's good, it's all good. We're great. But if it's bad, it's really, really, really bad."
$SPCX Mississippi bishop says SpaceX renting his church lot was no problem, with hotels and restaurants full and sees the shuttle lot as "a way to bring new members to God"
"I've noticed that hotels are full. Restaurants have been full and the traffic has increased."
"We were given a call to ask and we helped alleviate a parking issue that a company had. We didn't know who the company was. We were willing to do it. And once we found out it was SpaceX, it was no problem."
WREG's reporter on the Bishop: "Bishop Matthews did not go into details, but he says in his opinion, the data center is a done deal. And whether folks agree or disagree on the project, he sees the shuttle lot there on his church as a way to bring church members, new church members to God."
Elon Musk says there's no AI demand issue and explains why SpaceX rents its compute on short-term contracts: AI compute is worth more per watt every year, renewals are coming in at higher prices, and the market is a “cheese shop” with no chips left
“Our belief is that the value of compute per watt will increase over time, not decrease. And the reason for that, I think it should be obvious to people really, is that as you improve the AI software, as the models get better and better, they’re able to, you get more useful intelligence per watt. And the value of that compute is proportionate to the useful intelligence per watt. And that’s why you see things like even old computers like H100s renting for more than they rented when they were new.”
“And I think that’s likely to continue. I think we’re going to face a shortage of superintelligence compute. And that shortage is going to remain for a long time. So, I think the smart move is to do short-term contracts and to essentially place the bet on useful intelligence per watt increasing rapidly, which obviously is done, so that’s what’s happened so far.”
“And also if you look at, say, how are our short-term contracts renewing? Are they renewing at higher dollars per watt or lower? And we see them renewing at higher dollars per watt, not lower dollars per watt. We are also frankly hounded relentlessly by every hyperscaler and by dozens of smaller companies to rent more of our compute. So we could rent out our compute several times over if we said yes to everyone. That’s just the situation here.”
“That’s why I posted that sort of video meme of the Monty Python cheese shop sketch where the guy goes into a cheese shop to get some cheese and they actually have no cheese. And so he goes like, what about this kind of cheese? What about that kind of cheese? And they really just don’t have any cheese.”
“So we’ve actually tried to offload some of our inference compute to other cloud providers so that we can use more compute for training. And we get the cheese shop sketch. You know, they’ll say, like, well, we’ll ask them about what about some, you know, GB300s? Nope, don’t have any of those. What about, you know, some GB200s? Nope, don’t have any of those. What about some TPUs? Nope. What about some Trainiums? Nope. What about some, okay, how about some H100s? And like, well, we might have some overseas somewhere and then they’ll check and come back and say, nope, we don’t have that either. So it’s like ridiculous.”
“Yeah. So I don’t think we’re going to have a demand issue here. And like, it’s not impossible. I’m not saying this will happen. It’s not impossible that we will have a six-month payback period for compute next year.”
Sundar Pichai predicts Google Cloud will be negotiating quantum-computing deployments at some scale within five years
“First of all, welcome to the club. When you say you don’t understand quantum, you’re joining each and every one of us.
“It’s a technology where, the more time you spend with it, in some ways, the less you understand it.
“Quantum is an area where we set internal milestones, and the team is tracking or exceeding those milestones over almost a 10-year time frame in terms of robustly scaling quantum systems with a view toward commercial applications.
“The future possibilities of quantum are both vast and, in some ways, uncertain. But I genuinely see practical commercial applications in a five-year time frame that we would get excited by.
“If I were sitting in this chair five years out, our Google Cloud team would be negotiating deals on quantum deployments at some scale.”
Sundar Pichai predicts 2027 will be a “golden age” of productivity inside Google after Argon began solving tasks across millions of lines of Chromium code
“With Argon, we have seen a step change in internal productivity ever since a version of Argon was rolled out internally.
“There are large codebases on which we've pointed it to where we’ve had problems we know we need to fix. It’s millions of lines of code. In Chromium, if you want to create an experience where you can drag out a tab, move it around and put it in an AI experience, how do you keep it live without reloading?
“Historically for teams, it cuts through all layers of Chromium. But you can point Argon at it, and it can persistently go and work through it and get it done. We’re seeing examples like that everywhere.
“In some ways, I view 2027 as a golden age of that productivity frontier for us because of the new class of models and how well we are integrating them with our internal environments.”
