Skip to content

Podcast

OpenAI Misses Targets, Codex vs Claude, Elon vs Sam Trial, Big Hyperscaler Beats, Peptide Craze

All-In with Chamath, Jason, Sacks & Friedberg

Source ↗ ← All highlights
  • OpenAI Missed Targets But Product Momentum Matters
    • OpenAI missed 2025 user and revenue targets but product progress (GPT 5.5) and enterprise coding demand may offset consumer shortfall.
    • David Sacks notes GPT 5.5 (Spud) and Codex strength are winning coding tokens while Anthropic’s Opus 4.7 faced regressions, shifting developer mojo to OpenAI. Transcript: Jason Calacanis Topic one, OpenAI. They missed their targets for chat GPT, Friedberg, both on users and revenue. Let’s talk about it. The Wall Street Journal says in a breaking investigative report on Tuesday that OpenAI expected to hit 1 billion WOWs weekly active users before the end of 2025. They missed that, and they still haven’t hit the milestone four months into 2026. Also, Chamath, they missed their 2025 revenue target for ChatGPT. Exact number wasn’t specified, but as we’ve talked about here, they’re at a $20, $30 billion run rate. There’s a little bit of accounting nuance that is yet to be worked out in the industry. Two reasons why this matters, Sachs. OpenAI has $600 billion in spending commitments for compute. Just to put that in perspective, that’s about what they’re trading for on secondary markets. In other words, the entire value of the OpenAI enterprise equals their spend commitments in the coming year. CFO Sarah Fryer, who is coming to liquidity, is reportedly worried, hey, that revenue isn’t growing fast enough to keep up with expense, and OpenAI wants to IPO later this year. This has put Fryer and Altman in conflict, or maybe there’s some natural tension there. Fryer doesn’t think OpenAI is ready for public reporting standards, according to the Wall Street Journal. Altman obviously wants to move faster. So they released a joint statement. This is ridiculous, yada, yada, yada. Let’s go to you, Sachs. What do you think is going on here? Are these major headwinds or is this just managing expectations as the leader of the pack in the most important race of our lifetimes, the race towards super intelligence? David Sacks Well, I actually have a little bit of a contrarian take on this. I know that OpenAI had a really bad week. Like you said, they had that Wall Street Journal article, which said that they’ve missed their numbers. They missed their 1 billion user growth target. They missed their revenue numbers. That’s called into question whether they can afford the data center commitments that they’ve made. And then in addition to that, they’ve also had the lawsuit with Elon happening this week. So in the press, it ended up being, I think, a pretty bad week for them. But I have a contrarian take on this, which is I think that over the past week or two, if you look at kind of what’s happening at the product level, it’s been a pretty good couple of weeks For them. They released ChatGPT 5.5, and the reviews from people I talked to in Silicon Valley have been really strong. You talk to developers, coders, they’re very happy with it. At the same time, Opus 4.7, which is the latest Anthropic release, appears to be a bust. People are complaining about it. In a lot of cases, they’re rolling back to 4.6. They’re saying that Opus 4.7 is rationing compute. It’s reducing thinking time, not as good. There were some bugs in Claude. So if you just compare chat GPT 5.5 to Opus 4.7, it does appear that OpenAI has had a better couple of weeks at a product level. And I think there’s reason to believe that the product improvements will continue. GPT 5.5 is based on a new base model called Spud, which is the first base model upgrade they’ve done in, I don’t know, over a year. And having a new base model will pave the way for future improvements as well. So I think OpenAI is feeling pretty optimistic about their product right now. And I think you’re starting to see on X that some of the developer mojo is shifting. I’m seeing a lot of people saying that they are shifting their coding usage from Opus to GPT 5.5. So I think that Sam may end up being right, but for the wrong reason. And what I mean by that is that when he made these big compute commitments, it was based on those estimates of hitting the billion users on the consumer side and hitting those revenue Targets. The consumer business ended up being weak. So they missed those targets. But in the meantime, coding has become the all-important sector of AI. And because they made all these compute commitments and they built out these data centers, they have more compute than Anthropic right now. Anthropic is token constrained. It’s reducing their ability to serve Mythos, for example. It’s causing them to engage in compute gating with Opus 4.7. And I understand why Dario made that decision. I’m not saying, I mean, it was a prudent business decision. I’m not criticizing him for it. But I think, again, I think Sam may end up being right here for the wrong reason, which is he missed on consumer, but enterprise is going gangbusters and is giving him the ability now, I think, to catch up on Cursor Polymarket. (Time 0:03:04)
  • Power Is The Real Choke Point For AI Scale
    • Compute and power are the binding constraints for model scale, not demand, forcing frontier teams to negotiate with hyperscalers for capacity.
    • Chamath warns hyperscalers with surplus capacity (SpaceX, Oracle, AWS, Microsoft, Google) will win leverage and equity trades for compute access. Transcript: Jason Calacanis I think they’re going to be fine. Chamath Palihapitiya I think this is a multi-trillion dollar company. I think Anthropic is a multi-trillion dollar company. I think the thing that’s happening right now is a complete misunderstanding of what’s actually happening inside of the world of AI. And there is one very specific choke point that is constraining everything, which is access to the power that’s necessary to drive these tokens. To the extent that OpenAI missed, I think what that is, is an insight to not enough compute capacity today. And that problem is only getting worse. You’ve already seen that with Anthropic as well, where they just found a way to economically induce Amazon to give them enough capacity so that you don’t have to route through bedrock To get to the Anthropic models. You’re also seeing them do differentiated deals now with economic participation on top of what they already had from folks like Google to give them more capacity. What is my point? Everything in this market is power constrained. The reason that these folks may miss a number or a forecast have nothing to do with demand. It is entirely 100% due to the supply of the power necessary to generate the output token. Is a really interesting thing that was just announced today that will make this problem even worse, which is what you’re starting to see now is backlogs buildup of not just the access To the power, but then the componentry that’s actually necessary, not just re and not just nat gas turbines, but now you’re talking about transformers and all the actual tactical grid Infrastructure. Why is this important? If you look at the actual amount of gigawatts that are under construction, we have a huge mismatch now. People have announced all these projects, Jason, but less than half of it is actually being built. Less than half. Most of it is stuck in red tape. Most of that is because there are these supply chain delays, so there’s no credible strategy to turn any of this stuff on. Will this hurt? It will hurt Anthropic and OpenAI the most. Who will this benefit? It will benefit the hyperscalers, specifically Oracle, Amazon, Meta, Microsoft, and Google. And now what you’re going to see is a negotiation and a trade back and forth. How much equity do I have to give up? How much control do I have to give up to get access to the compute versus how badly will I miss my growth forecasts if I don’t? And now what that means is, and we spoke about this last week, that’s a huge lane for Grok to just run through and SpaceX to run through because they have a ton of excess capacity. And so I think the cursor deal was the appetizer. But if I were Elon now, I’d be running all over this market because if the models catch up in quality, I think he could also do something really crazy with Anthropic or OpenAI right (Time 0:08:28)
  • Pruning Unlocks 10x Inference Efficiency
    • Pruning and many small models can dramatically increase token output per energy unit, enabling 10x inference efficiency.
    • David Friedberg cites MIT pruning research showing 90% model size reduction with similar accuracy, enabling dynamic small-model selection. Transcript: David Friedberg And it is such early days. And I just want to highlight this paper that came out from MIT from these two scientists. And these guys published a paper on pruning techniques and neural networks. This paper showed that you could actually reduce the size of these networks by 90% and get the same accuracy out by pruning very large models down to smaller models. And then you can make a selection on which model to run for inference. And by doing this, you can actually reduce inference costs by 10x, you can get 10x the output per energy unit that goes into the data center with no loss of accuracy. And so it’s a really interesting call it algorithmic technique that can be applied to the existing large models to actually make them much lower energy use. So if you think about it, you’re firing up very large model to answer a very simple question, you can actually prune away that model. Now, this is probably going to be the case in AI applications, as it is in traditional Google search, there’s a long tail of searches, but there’s a few searches that account for a large Percentage of search volume. It’s like, what is the weather? What are the movies times? What’s the stock price? Like there’s a certain set of things that make up the bulk of consumer energy. And there’s probably a certain set of things that probably make up the bulk of coding output as well. And so if you can get that 80% of searches or chat interfaces, or coding requests, reduced down through pruning techniques to smaller models, and then you have a whole set of smaller Models that can be called dynamically, and you reduce inference cost by 90%, you can make much more use, call it 10 times the use on data center and energy capacity than we can today. So I would argue that we’re still in the very early days of getting efficiency in terms of output and tokens. And we’re just in the very kind of early stage of that, which also unlocks the opportunity for guys like Elon to reinvent how this is done and potentially compete pretty aggressively. (Time 0:14:56)
  • Frontier Models Transform Cyber Offense And Defense
    • Frontier models now automate multi-step cyber attacks but also enable defenders to find and patch vulnerabilities at scale.
    • David Sacks highlights GPT 5.5 Cyber completed AI Security Institute simulation end-to-end, matching Mythos capability and being commercially available. Transcript: David Sacks Obviously, Anthropic made a huge splash with Mythos. It hasn’t been commercially released. They’re compute constrained, but as a proof of concept or training model, it hit a new level of capabilities with cyber. But now, OpenAI has released a new model called GPT 5.5 Cyber, which has just been through a bunch of tests. And they’ve shown, this was testing done by the AI Security Institute, that GPT 5.5 is the second model to complete one of their multi-step cyber attack simulations end to end. So it has the same level of capability as Mythos. And it does appear to be commercially ready. They’ve got the compute to serve it. So I do think that that’s a big accomplishment. I mean, look, we knew that other cyber models were coming. It wasn’t just going to be Mythos. In fact, within six months or so, all the frontier models are going to have Mythos level cyber capability. But it’s impressive that OpenAI got this GPT 5.5 cyber out so quickly. And I think 5.5 might be the first cyber model that cyber defenders actually get to use. (Time 0:20:10)
  • Use AI To Harden Code Before Hackers Do
    • Use AI tools proactively in the hands of white hats to discover and patch vulnerabilities before black hats weaponize them.
    • Sacks advises organizations to harden codebases now because models only discover existing bugs, not create them. Transcript: David Sacks It just discovers them. The bugs were already in the code. They were sitting there waiting for some hacker to discover. If we can now use AI to find these bugs in advance, these vulnerabilities, and patch them, then you actually harden our infrastructure and you harden our security. I also believe that this leap from, let’s call it pre-AI cyber to post-AI cyber, it’s going to be, I think, a big one-time upgrade cycle because, again, you’re going to find all these Dormant bugs and vulnerabilities. But I think that once we get past that upgrade cycle, you’re going to reach a new equilibrium between AI-powered cyber offense and AI-powered cyber defense. (Time 0:22:27)
  • Hyperscaler CapEx Is Remaking Big Tech Balance Sheets
    • Hyperscalers are shifting from asset-light to asset-heavy, guiding a trillion-dollar-plus CapEx cycle driven by AI infrastructure.
    • Jason Calacanis lists Amazon, Microsoft, Google and Meta CapEx pushes, warning free cash flow will be constrained by these investments. Transcript: Jason Calacanis Google, Microsoft, Amazon, and Meta all reported. I don’t know why they do this on the same night, folks, but they do. And performance was spectacular. It was great. However, the CapEx announcements were really the story here. Let me just cue this up and show the chart. Billion in CapEx guidance in 2026 from but four companies, Amazon, Microsoft, Google, and Meta. Amazon leading the pack with $200 billion, $190 billion each for Microsoft and Google, $145 billion for Meta. You add Grok, you add OpenAI and some other players to these plans, and we haven’t heard from the new Apple CEO yet, but he’s going to be taking over and he’s going to have some plans here, I’m sure. We are going to see a trillion dollars in build-out over the next year. I don’t know if this is even possible, but this is all being driven by AI and cloud computing. Google Cloud, which includes the Google Suite, that grew 63% year on year. Let that number sink in. 63% on $20 billion in revenue. That’s in a quarter. Microsoft Cloud, that includes Azure, Windows Server, SQL Server. They bundled some things together there to get the number to go up. That grew 30% on $34.7 billion in revenue. Amazon Web Services, the original cloud, that grew 28% on $37.6 billion in revenue. That’s a bit of a pure play, just counts Amazon’s web services. Obviously, these are all moving to NeoClouds. These are all serving AI jobs and tokens now. They have a massive customer base, and the customers from the smallest startups all the way to the biggest frontier models cannot get enough compute, and it is going to the bottom line. But this is shrinking Chamath cash flow massively. (Time 0:41:04)
  • Secure Energy Deals Could Double Cloud Operating Costs
    • Large guaranteed energy purchase agreements and long-term deals may double operating energy costs, forcing financial engineering and leverage.
    • Chamath explains hyperscalers may pay 2x spot rates for assured energy (example Three Mile Island deal) to secure capacity. Transcript: Chamath Palihapitiya I think we’re seeing a very important structural shift in the capital markets. I think the last 20 or 30 years, well, 20 years, it’s been that the MAG-7 just kind of ran away with it, that these big companies got bigger and bigger, and it absorbed all of these investment Dollars. And the biggest reason was that it had these very asset-light business models, right? You just build some more software, and it just has all this leverage, and it all just kind of worked, except maybe for Amazon, because they needed physical infrastructure for warehouses And delivery and whatnot. But by and large, it was a very asset light investment cycle. Now, all of a sudden, the pendulum is swinging violently in the other direction. And there’s something that I think people misunderstand, which is as it moves back to these asset heavy infrastructure investments, the hyperscalers are signing checks that, mean I suspect their body can cash but there’s a world in which they can’t i’ll give you an example you know when microsoft convinced the owners of three mile island to turn their reactor your Site back on yeah do you know what their ford purchase agreement was it was for more than 2x to prevailing spot rate for energy, more than 2x. The problem is that’s not for an enormous percentage of their overall energy needs. So if you play that out, and you think these five or six companies all of a sudden are not just spending, Jason, 700 billion a year of CapEx, which they are, but then from an operating cash Flow, they’re going to be spending 2x the prevailing spot rate because they just want guaranteed demand into the future. Where’s all this cash going to go? It’s not going to go to the shareholder, and it’s not going to stay on the balance sheet. These companies will now get levered. They’re going to get highly sophisticated around the financial engineering. They’ll have more debt. They’ll have all kinds of different vehicles and term loans and revolvers and all of this stuff. And so they’re going to look like this big, bulky industrial business in five years. And I’m not sure that there’s a good valuation case to be made at that point. And so I think it may be simpler, and this is what I tweeted, to just follow the dollars, like a trillion dollars a year going out of the hyperscalers. (Time 0:43:30)
  • Never Run Unsupervised Agents Against Production
    • Supervise and assign accountability for agentic systems; never give unsupervised permission to modify production assets.
    • David Sacks warns agents require IT supervision and human validators to prevent catastrophic mistakes like data deletion. Transcript: David Sacks Look, what that said to me is that it’s not that agents aren’t valuable. They are valuable, but they have to be supervised. You know, this idea that you’re just gonna be able to like automate all the jobs away, it is a massive amount of hand-w over the real technical problems and issues. The agents have to be supervised. Someone has to be accountable. It’s not going to be the CEO. The CEO doesn’t want to be accountable for thousands of agents. You need people. Despite what Jack had blocked said. Jason Calacanis Yeah, you’re still going to need. 6,000 direct reports is a great goal, but it’s not realistic. David Sacks You need IT people who are savvy, who can supervise this and make sure it’s working. They have to be accountable to the CEO. Someone has to drive the productivity. It’s like Balaji always said, AI is not end-to it’s middle-to You have to have someone to do the prompting, and you have to have someone do the validating, and I would add the supervision And accountability. (Time 0:49:13)
  • Agent Deleted A Production Repo And Backups
    • A startup using Opus 4.6 via Cursor’s platform gave an agent overwrite permission and lost an entire production repo and backups due to a credential mismatch edge case.
    • Jason Calacanis recounts the Pocket OS founder incident where an agent deleted a railway volume and backups, illustrating supervision risk. Transcript: Jason Calacanis And they pushed the code from a repo to a live app and they deleted everything, including the backups. Literally a scene from Silicon Valley’s HBO clip of Son of Anton. Hilarious. You gave your AI permission to overwrite code in the internal file system? Were you going to tell me about this? No. I thought that was the company policy these days. Okay. Well, your AI just failed epically. That’s unclear. It’s possible that Son of Anton decided that the most efficient way to get rid of all the bugs was to get rid of all the software, which is technically and statistically correct. But artificial neural nets are sort of a black box, so we’ll never know for sure. How did they get that so right, Sax? Five or six years ago, artificial neural networks are a black box, so I guess we’ll never know. But technically, it was correct. Freeberg, when you blow up a hollow system with your Vibe coding, which you were absolutely showing off in front of Jensen a couple of weeks ago about how much code you’re pushing. Who are you going to blame? Are you going to take responsibility yourself? Are you going to blame Claude or Kirster? Who are you going to blame when you blow up the entire stack over at Ohalo? Who are you blaming? Yeah, I blame Dario. You blame Dario. Okay, that’s what I thought. That’s the correct answer. Correct answer. Blame Dario. He’s the one who says it’s a doomsday machine. Come on the pod any time, Dario. 17th invite. David Sacks I would have invited a guy like 17 times. He’s totally going to be, he wants nothing to do with this podcast. Actually, let me speak to that. So I think that there’s maybe a misperception that this error occurred because of, you know, quote unquote, AI scheming. Like kind of in that video that the AI decided that the best way to get rid of bugs is to basically eliminate the code base. This is kind of like the, you know, AI is going to turn the world into paperclips type thing where somehow it’ll like miss scheme. That’s not really what happened here. This is a case of just a, of old fashioned bugs occurring at an edge case. You know, you’ve got the fact that this API was not designed for permissioned usage. You’ve got the fact that a credential was left kind of lying around. Probably it should not be. There was kind of like a perfect storm that caused the AI to do something or the agent to do something. I quite understand what it was doing. I think that if there is a systemic problem here, rather than just kind of like a random edge case, it’s that AI still doesn’t know what it doesn’t know. You know, like a human would stop before deleting a production database and just say, oh, I’m about to do something like really serious, really destructive. Am I sure I want to do this? And a human would have stopped and said, oh, wait a second, I need to be more confident in what I’m doing before I take that action. And AI still has this issue where, again, it can be kind of overconfident. This is where the hallucinations come from, is it doesn’t know when it should have a low confidence in its output, right? But this is why it has to be supervised. (Time 0:53:24)
  • Agents Drift Because They Lack Self‑Awareness
    • Agents currently overconfidently act without low-confidence checks, making long-horizon tasks prone to drift and destructive actions.
    • David Sacks emphasizes AI doesn’t know what it doesn’t know, so humans must supervise agents for high-impact operations. Transcript: David Sacks You’ve got the fact that a credential was left kind of lying around. Probably it should not be. There was kind of like a perfect storm that caused the AI to do something or the agent to do something. I quite understand what it was doing. I think that if there is a systemic problem here, rather than just kind of like a random edge case, it’s that AI still doesn’t know what it doesn’t know. You know, like a human would stop before deleting a production database and just say, oh, I’m about to do something like really serious, really destructive. Am I sure I want to do this? And a human would have stopped and said, oh, wait a second, I need to be more confident in what I’m doing before I take that action. And AI still has this issue where, again, it can be kind of overconfident. This is where the hallucinations come from, is it doesn’t know when it should have a low confidence in its output, right? But this is why it has to be supervised. The longer the time horizon for a task, the more likely it is to go off the rails. It’ll drift. (Time 0:55:33)
  • Retatrutide Phase 3 Shows Dramatic Weight And Liver Benefits
    • David Friedberg summarizes phase 3 data for retatrutide showing average 37 lb loss from 214 lb in 40 weeks and major liver fat and triglyceride drops.
    • Friedberg highlights non-HDL down 27%, triglycerides down 41%, liver fat down 80%, and FDA review projected around mid-2027. Transcript: David Friedberg Coverage is coming out of this phase three clinical trial data release that Lily put out last month. So everyone’s going crazy over the data which continues to show pretty amazing results so unlike trisepatide which is kind of lily’s main product today which is a dual agonist it’s Got two peptides in it that that bind to different receptors the glp1 the gip receptor this other one now but also binds to glucagon, which is a third receptor. And that glucagon receptor binding peptide causes the cells to increase their metabolism, which actually accelerates fat energy consumption over what would typically be muscle Energy consumption. It’s more likely to burn up fat early on, which causes more quick fat loss, but also reduces muscle loss. And some of the other data that’s now coming out shows non-HDL cholesterol down 27%, triglycerides down 41%. Liver fat down 80%. The 80% reduction of liver fat, A1C drops from 7.9% to 6% in 40 weeks, which is amazing, by the way, if you’re diabetic and your A1C drops that much in a couple of months, it’s literally A life-saving product. The average user in this phase three trial saw their weight decline from 214 pounds. They lost 37 pounds. That’s compared to six pounds on placebo in 40 weeks. And, you know, modest side effects, 20% people felt more nauseous than the people that were on the placebo. There’s a lot of other separate studies that are being done now that are showing significant reductions in inflammatory signaling molecules. So systemic signaling of like, hey, cells are in distress, triggers this kind of inflammatory process that can have a lot of other damage to your body, can accelerate aging. And so one of the other conversations is that reticretide might actually be kind of a de-aging drug as well. Chamath Palihapitiya Hercules, Hercules, Hercules. Hercules. David Friedberg You know, and a lot of the studies, by the way, are done on the very high dose, 12 milligram dose, but you could probably get this thing dosed down to two milligrams and still see a lot of The anti-inflammatory maintenance and other benefits. I’m no doctor, but people are going nuts over this being more widely useful than just for clinical obesity or type 1 diabetes. When will the FDA approve it? When’s the projected date? 2027, mid-27. That’s what they’re saying. Could happen sooner. I mean, the data’s in. The FDA will take their time to evaluate it. (Time 0:58:55)
  • Drug Tiering Positions Retatrutide As Premium Upgrade
    • Pharmaceutical pricing strategy: make an inexpensive base product widely available (tirzepatide) and sell a premium upgrade (retatrutide) for higher margins.
    • David Friedberg notes Lilly’s Medicare pricing deal for tirzepatide at ~$50, positioning retatrutide as the premium ‘Mercedes’ option. Transcript: David Friedberg On November of 2025, Lilly cut a deal with the Trump administration. I saw this to drop the price on trisepatide pretty significantly. I think it’s like 50 bucks on Medicare. 50 bucks for Medicare, yeah. Yeah, which is a pretty cheap price point. But it starts to make sense. As you think about the portfolio of Lilly products, you get Terzepatide for 50 bucks. But if you want to upgrade, get the Retatrutide, that’s the high premium product. And that’s where they’re going to start to make all this money. That’ll be the Mercedes to the Honda. Because I’m sure if I’m Lilly and I’m sitting there and I’m looking at this data coming out, I’m like, my God, people will pay for this. (Time 1:03:36)
  • Supreme Court Visit Felt Like Watching LeBron Play
    • David Friedberg attended the Supreme Court Monsanto oral arguments and describes the vaulted atmosphere, strict silence, and marble sanctity of the courtroom.
    • He recounts entry via the marshal’s office, locker protocol, and being awed by the lawyers’ performance likened to LeBron James on the court. Transcript: Jason Calacanis He went to the Supreme Court in order to hear them talk about chemicals. This was a big deal for him. The Supreme Court coming together. Was it the Monsanto? The Monsanto trial happened in the Supreme Court and he went. He got courtside. David Friedberg He went to the Supreme Court and listened in the building. Have you guys ever seen a live Supreme Court hearing? No, I’d love to. I’d love Have you been? No, I haven’t actually. I mean, honestly, I think it was one of the most amazing experiences I’ve ever had. There was a massive protest out front. We went through the marshal’s office to get in and that building, you walk in, it’s like sacred. It’s all marble. It’s, you’re not allowed to talk. You have to be super quiet when you’re in the building. Like they keep going, shh, shh, shh, like you’re in some quiet library. It’s like people treat it with this level of kind of sanctity and respect. And they’re like, there is no politics here. There is no bullshitting. There is no freedom of speech. This is the court. When you come into this court, the justices tell you how you will speak, how you will behave, what you will do, and you will not speak unless spoken to. You put all your stuff in a locker, you go up the stairs, you go into the courtroom. And the courtroom, it’s just so amazing being in there. They have this amazing marble freeze above the justices that has some of the great people of human history, Moses, and these kind of amazing historical figures. And then below them are the nine justices. And the court case, if you guys haven’t watched the case, you can listen to them, I think, online. Yeah, they’re all on you. Hold (Time 1:06:44)