Anthropic’s AI Slowdown, OpenAI’s $1.2T Valuation & Salesforce’s AI Bet

The handpicked topics for this week are:

1. Anthropic’s AI Slowdown: Safety or a Market Power Play? Anthropic asks for a slowdown, regulation, and antitrust waivers so labs can coordinate; Jensen Huang, Mark Zuckerberg, David Sacks, and the U.S. and Chinese governments push back, reframing the issue as ordinary product safety. Daniel argues the slowdown request reads as regulatory capture by two private labs that want to lock in a duopoly before open models and cheaper compute erode model-layer pricing power. Patrick lands on the product-safety frame: test before you ship, keep the testing procedures contained, and avoid an FDA-style regime that would stall deployment. (The Decode)

2. Trilateral safety coordination without the waiver. OpenAI, Anthropic, and Google DeepMind confirm they are coordinating on safety in the same week they disagree on whether they need antitrust protection to do it. Patrick treats the coordination as plan B after the regulatory ask failed to land, and flags the competitive risk to smaller labs and open-source developers if three frontier companies set standards together. Daniel argues self-regulation plus regulators "along for the ride" beats leading with regulation, pointing to Europe as the cautionary case for growth. (The Decode)

3. The frontier sets the pace for open models. Both hosts agree on one point last week’s debate largely missed. Open-weight models from China and the UAE match frontier output at a fraction of the cost, but many are distilled from frontier models, so a frontier slowdown slows open source too. Patrick adds that these model makers are also shipping real architectural gains (an 85% KV cache reduction, on-device DeepSeek V4.1 Flash running comparably to a frontier model on Spark hardware), which drives inference cost down regardless of who leads on capability. (The Decode)

4. Dreamforce and the end of the "SaaSpocalypse" narrative. Sam Altman, Dario Amodei, and Jensen Huang all appear at Salesforce's Dreamforce, which Daniel reads as confirmation that systems of record remain the control point for enterprise AI workflows. Salesforce's Agentforce and Data Cloud lines are growing at triple-digit rates. The company is pairing a Claude-based Clawed Force with a multi-model harness, and training its own custom model, Koa, on NVIDIA Nemotron and 30 years of CRM data. Patrick sees Koa as the week's underappreciated signal: enterprise software vendors building on open weights instead of renting frontier intelligence. (The Decode)

5. OpenAI pulls even in developer traffic. OpenRouter data cited by Gavin Baker shows OpenAI climbing from roughly 20% to 50% of API share versus Anthropic since June, and Signal65's Pinnacle benchmark now places Gemini and Grok above Anthropic's latest model on several enterprise personas. Patrick credits OpenAI's early compute lock-in for the swing. Daniel's takeaway is that leadership rotates every few months, which means the model is a shrinking moat and compute, energy, and distribution are the durable ones. (The Decode)

6. AI Infra Summit: CPUs re-enter the conversation. NVIDIA presents Signal65 data on its main stage showing Vera running 1.64x faster per core than the leading shipping x86 processor, even as NVIDIA hedges with Intel on head nodes. AWS commits to more NVIDIA capacity despite owning its own silicon, and co-packaged optics, networking, and memory optimization dominate the floor. Patrick notes a bipartisan ratepayer protection measure on data center energy costs arriving the same week, an early sign that power politics will shape build-out timing. (The Decode)

7. The Flip: Is OpenAI worth $1.2 trillion pre-IPO? Daniel argues yes: a $40 billion run rate, a 41% step-up in six months, investors initiating the round, and a safety regime only the largest labs can afford all favor incumbents. Patrick argues no: historic cash burn, second place in revenue behind Anthropic, a higher cost of capital as rates rise, an IPO slipping to 2027 that puts part of Amazon's committed capital at risk, and open models cutting inference costs by 90%. The disagreement narrows to one question: whether a slowdown at the top protects or erodes the economics of a company priced for acceleration. (The Flip)

8. Doomer sell-off flips into a chip rally. Semiconductors reverse a safety-headline sell-off into a sharp mid-week rally, shrugging off oil and rate pressure along the way. Daniel frames every "end of AI" narrative as a buying window and adds a Jevons-style twist: safer models require more training runs, which means more compute demand, not less. (Bulls & Bears)

9. Salesforce clears the model and still falls. Salesforce reaffirms $63 billion of revenue by fiscal 2030, beats Street models by $3.8 billion, and shows current AI ARR above $1.5 billion growing more than 240% year over year, yet the stock declines. Patrick reads the drop as a show-me story on target history rather than on AI itself. Daniel, in the room, points to a same-day platform outage, dependence on Anthropic, and pressure to prove cost-per-task economics across the model landscape. (Bulls & Bears)

10. Nebius raises prices into scarcity. Nebius lifts pricing more than 20% across H100, H200, and Blackwell capacity, with Vera to follow, a move Daniel treats as direct evidence that compute constraint, not demand, is the binding variable for neoclouds. CoreWeave trades in the same range on the same concern: adjusted EBITDA is not cash, and reinvestment intensity stays high through 2030. Patrick favors Nebius's vertical integration, since building its own infrastructure sets it apart from peers renting theirs. (Bulls & Bears)

Watch the full video at sixfivemedia.com, and subscribe to our YouTube channel so you never miss an episode.

Disclaimer: Six Five Media is for information and entertainment purposes only. Over the course of this video, we may discuss companies that are publicly traded, and we may reference their equity share prices. Nothing discussed during this webcast should be considered investment advice or a recommendation to buy or sell any security. We are not investment advisors, and you should not rely on this content as financial advice. Six Five Media collaborates with technology companies and industry leaders to produce research-driven interviews and multimedia programming for enterprise technology audiences.

Patrick Moorhead:
Welcome back to Six Five Pod, everybody. We made it to episode 320. 320 episodes. I cannot believe it. Daniel, look at that number. And it's just mind boggling. How are you doing this morning?

Daniel Newman: 

Yeah, we were young men when we started.

Patrick Moorhead:

 No, we're not. I don't know cry, you know Physiologically, I think I might be even younger. So I mean I was I was a fat pig, you know You can see the I I think the the lead-in videos were after I've lost 55 pounds and they edited you down Yeah, they did you down, but don't worry you know, they take clearly the producers have a you know more love for you than me I

Daniel Newman: 

because I'm pretty sure they actually edited me up. They actually made me even fatter in some of those videos than I even was, which is hard to believe.

Patrick Moorhead: 

It's kind of like Chamath's sweater, right?

Daniel Newman: 

It's Chamath's sweater. It's how big Benioff was next to Simmons. You got to love Jensen's million years of evolution between me and you. We realized that we don't need him that big anymore. We don't need him that big anymore.

Patrick Moorhead: 

Oh, that stuff just makes me laugh.

Daniel Newman: 

You know, I'd like to be six five though. I gotta be honest. I think I think the world will look good up there. I can't tell because I've never been, but when I stand on a mox, it's kind of cool.

Patrick Moorhead: 

You know, I always used to think I was six feet and then I was like, well, OK, I was five eleven and three quarters and like, and I'm like five eleven, like barely, you know, it's crazy.

Daniel Newman: 

You're shrinking. By the way, you could probably get like half an inch back. I've been dead hanging every day. Yes. Twice a day for like, I try. I know you can do two minutes. I got like 100 seconds before my arms feel numb and they're like, they're going to fall off, but I'm a heavy boy, you know, 220 pounds. But like I did notice like my spine and my like shoulder, like rotation has gotten better. So anyways, but did you know you're taller when you lay down than when you stand up? I had no idea. Yeah, this is actually true. So if you measure yourself laying down, that's your true height. Not joking, but when you stand up, so even when you wake up in the morning, you'll be a little bit taller than you are throughout the day. Cause as the day goes on, you, you get some compression and you'll usually lose like half an inch. Yeah. Yep. Useless information for everybody out there, but I'm sure you all wanted a little lesson to know that when you're measuring yourself, lay down, then you can lie to everybody about how tall you are when you stand up.

Patrick Moorhead: 

I know. No, it's great. I mean, if you don't like what the current state of measurements are, then you just change the measurement. You change the measure of merit.

Daniel Newman: As founders of a benchmarking company, I'd like the world to believe that our measures of merit are legit. But yeah, I mean, you can definitely play with all the inputs to change the output. Let's just put it that way.

Patrick Moorhead: 

Benchmarking at its finest. So, hey, we had a great show today. You know, we film on a Friday. And therefore, this first topic made its way over the weekend. But let me tell you what, it was the discussion point over the weekend. That's pacing the frontier essay that Dario brought out. I mean, everybody came out of the woodwork. Sam, Musk, Sabis, Zuckerberg, Jensen, Sachs. Uh, and even fricking China, uh, came out on that. And, you know, it's funny, uh, who would have thought that, uh, it would have culminated in, in, in the event that you attended, which was Salesforce's dream force. Uh, but it did cause Sam and Dario were there and so was Jensen. So it just kept this, uh, this whole thing, uh, moving on. So. Yeah, we're going to talk about that. We're going to talk about OpenAI, looking at a $1.2 trillion pre-IPO round. We had AI Infra Summit, just when you think there wasn't enough events that you can attend. We had analysts from Moore and Futurum on the ground there doing videos with the 6.5. And maybe or maybe not, Signal 65 benchmarks were shown on the NVIDIA stage. We're also going to talk a little bit about, it's funny, what started off as a AI chip route turned into absolutely a crushing bull week. I looked at the semi-picks from Wednesday to hopefully today, and they were absolutely you know off the chain. Did you buy the top?

Patrick Moorhead: 

Daniel you know for once I didn’t. In fact I took um the market went down, and I went against the Pat Moorhead stock selection strategy and I actually bought the in the troph

Daniel Newman: 

You bought the dip?

Patrick Moorhead: 

I bought but I haven't sold at the peak.

Daniel Newman: 

Did you buy the dip-dippity-dip-dip? All right. Sell some. Dude, sell some. Just sell some because you need to make a dollar and maybe you'll love how that feels. The buy high, sell low thing is exhausting and I want you to sell the top on one of these.

Patrick Moorhead: 

Yeah, it's funny. I think of every time I sell something that I get 60 cents on the dollar. So. Well, you get the tax break on the loss. Yeah. Hey, let's dive into this show. Let's start off with the decode. All right, Dario kicked it off with his Pace the Frontier essay, which essentially said, slow down. We need to be regulated. We need to have antitrust waivers so we can all work together and just slow down. And oh, by the way, keep an eye on those open models and all the infrastructure. that's being able to China. Everybody came out. Sam, Elon, Hasabis, Zuck, Jensen, Xi. Right. It was it was it was it was pretty awesome. Help us break this down, Daniel.

Daniel Newman: 

Well, it ended up becoming like, you know, the battle of the the accelerationists versus the the doomers. And it was very interesting. I mean, look, last week, our show was all about that kid, Jacob Coxson. That came out, became like a news sensation. 160 million people viewed his tweet, which is great. You and I probably in our 20 years on Twitter haven't gotten that many views. I mean, joking, but not by much. And the bottom line is then it built up to, you and I thinking that was like this big narrative and we woke up on the weekend to this pacing thing, which I don't know about you, but it became like my entire weekend and into Monday. And then we had to, you know, like I was at Dreamforce and so Zuck was there, not Zuck, Jensen was there, Sam was there, Dario was there. I don't know if they were allowed to have Dario and Jensen in the room at the same time because I wasn't sure if they were going to box because Jensen basically said, you're full of shit. I mean, in so many, so many words. And by the way, the All In Summit was going on during this. So you had comments coming in from Elon and Jensen was there. China called it fear-mongering. But it's really interesting in breaking this down and being a little less hyperbolic about this and just being more intellectually honest. The bottom line is, is that you and I have been through some technology revolutions, industrial shifts. We've been, you know, you've been through more than me because you're old, but like, you know, you've been through Internet, we've been through social media, we've been through mobile, we've been through, you know, blockchains, AI now and all these different things that have come in and come out. And we've been through the hyperbole, the headlines. You know, you remember the there was there was all the ads against Edison and electricity. Remember the ads people electric electrocuting themselves. I mean, there's always this kind of doomerism that comes with all new technology about end of jobs and of certain industries. And this is why I think those that have studied history kind of tend to come back and be like, each revolution has brought more, not less, because that's what history has suggested. Now, again, every revolution is different. But here's the bottom thing. With each new technology that comes out, the optionality to use it for good or use it for bad exists, right? I mean, basically, you can When the internet came out or when when YouTube came out, like, you know, the things that they're saying, like, you know, they talk about Middle Eastern countries using Claude to try to build weapons. Right. And the fact is, is like people have been doing that as long as the Internet's been out. There's the dark web. I mean, there's been videos published on, you know, on YouTube that can teach you how to do things that you wouldn't want most people to know how to do. And what's happened with AI is it's all faster. It's more accessible. It's faster. It makes easier connections. It gets you to the right answers sooner. So certainly, we can use AI, whether it's for hacking, whether it's for weapons, whether it's for biological risks, whether it's for grid penetration of a grid. It's certainly there. At the same time, the same technologies we're building can also be used for good. It can be used to develop drugs cure cancer, potentially improve economic well being and impoverishment like these are all things that are possible with the same technology. And so you can't halt the evolution of good for bad. Now, in fairness. Nobody's coming out and saying stop. But the bottom line is what they are saying is slow down. And then there's all these questions, Pat, about how genuine is this? Are they genuinely wanting to stop because they're worried about these agents creating a bot swarm that's going to plant replications of itself so that it can make sure that you could never turn it off under any circumstances? And I'm pretty sure unless we have physical robots overseeing our data center, there has to be an ability to pull the plug on this stuff. But having said that, these are the worst possible scenarios. You talked about it last week. They get the football. They're able to launch nuclear weapons. They're able to go through all the different physical security. Because it's not just digital security. These things have digital and physical and all the things. But anyways, I could ramble about this for a long time. The bottom line here is that one, there is no real option to slow down. I just do not believe this. I believe the slowdown request is regulatory capture. The idea is that every major technological wave, whether it's been operating systems, whether it's been social media, has been won by a couple of companies in a combination of Best product and regulation and barriers to entry has made it where there's only a couple of social businesses. There's only a couple of search businesses. There's only a couple of OS companies. Anthropic and OpenAI want to be a couple here. They do not want to have open source models, disruptive capabilities. And I think a lot of this comes back to something I've said for a long time, Pat, is there is no moat in the model. The model is less and less of a moat. And the real moat is infrastructure. And you're seeing the race to being able to have enough compute to be able to keep up demand for the models. And so this is just this cycle. But now what we have is trillion dollar valuations that are having to be justified, that are going to try to go public. And all of a sudden, someone realizes like, crap, I can do 99% on GLM 5.3 or 98% on Muse or 97% on Brock that I can do on Frontier Model. And then the questions around the business's long-term value and sustainability start to come up. Regulatory capture can fix this. I think that's what these guys were aiming for. But at the same time, I'll just summate this with, we do need to pay attention to safety. We do need to pay attention to security. It is the responsibility of the industry to make sure they ship safe products, secure products. Jensen was right about that. I agree more with Jensen and more with Zuck than I do with Dario and Sam. But again, this is probably a two hour long conversation. I said a lot and I've probably not even gotten close to really putting my opinion out there.

Patrick Moorhead: 

That was a great breakdown, Daniel. And at times like this where you have these multivariate issues that are very difficult, I try to simplify, or I glom on and admire what someone else has done to simplify. And Jensen really hit it first. And I think it was the short form of what David Sachs had brought out, which was essentially, it's about product safety. If you do not think that your product is safe, then don't bring it out and do the testing on it as much as you can. By the way, we know that it's the testing procedure that has had the circus animals leaving the tent. and keep your freaking animals in your tent. Keep the lions and the tigers and the monkeys in the cages and just do that. And it just that really hits my Midwestern sensibility. It just makes sense, Daniel. And people will say, oh, but this is different. This is so much more complex. No, this is product safety. This is what it is. It's self-driving cars, product safety. It is drug safety. And by the way, when I say drug and highway, I don't want NHTSA or an FDA type of thing that will slow everything down to an absolute crawl. And then Jensen simplified it even more, right? And Jensen is just so, so good at this. I have so much respect for his ability to simplify the problem. Essentially, he said it's all about omelets. If you can't bring out a good omelet and this is, you know, going back to his Denny's where NVIDIA was created, then don't don't serve it. OK, like it's not a good omelet. Don't serve it. I am in super common sense. Right. So it just makes sense. So, Daniel, great take. There's nothing else that I have to add.

Daniel Newman: 

We could spend a whole show kind of thinking through what we should and can do here and what's important and what the intent underneath all this stuff is for everyone out there. We don't usually do like PSAs on this show, but like just my recommendation is read everything with a bit of skepticism right now. The media narrative requires critical thinking, be a little bit curious and realize that it's really popular. The hyperbole is popular. People love these AI will blah, blah, blah. AI will end humanity. AI will displace all jobs. AI will replace all software. AI will end all security. Like if you think about the number of times we've heard this, this is just another deep-seek moment.

Patrick Moorhead: 

Yeah, yeah, exactly. So the same week, let's move to the next topic, which was OpenAI, Anthropic, and Google DeepMind confirmed that they are working together, trilateral coordination on AI safety. Almost in response, even though OpenAI Global Chief Policy Director said that they've been in active talks for several weeks, I viewed this as a rebound. They got the message and said, listen, if regulatory capture doesn't work, then there's got to be a plan B, and hey, we can all work together. So I think that this, quite frankly, makes a whole lot of sense than a waiver. And I do find it ironic that the three companies decided to get together. They first said they needed some air cover on regulatory. And by the way, for those who don't fully understand it, You can't have three big companies working together on something that might look like a monopoly or a triopoly or oligopoly. It's not even an oligopoly, but it's three parties working together to squeeze out the small people, namely the smaller labs and the open source. Companies just there's not a whole lot of meat together right we haven't seen a coordinated paper yet so i'm kind of keeping my. keep my options open. There's not a whole lot to debate and pull around it. But I do find it very ironic that the three labs are confirming that they're coordinating on safety. And in the same week, they disagree on whether they need an antitrust waiver to do it.

Daniel Newman: 

Yeah. I mean, look, the the. Idea of the three three coming together and trying to solve all this is probably the definition of sort of an antitrust coalition. I mean, the three companies that have the advanced frontier, and I know some people argue that Google is behind, but I actually think it's funny because Google and Meta kind of more aligned on this topic of like, they're behind because as public companies, they've got certain beliefs in what safety and and governance looks like, and they're not releasing things that they don't know if are safe, whereas these private frontier lab VC companies are kind of full steam ahead. But I do think those are the, you know, that those three companies could certainly drive alignment. But it's also interesting to me that Zuck's staying outside of this all. Look, I think this is a more to come. I think to your point, there's not a lot of meat here. I think the industry, what I will agree is, is like the industry needs to be self-regulating. That's how capitalism goes fast. I mean, I do agree that like with the FAA and the NTSB, with the autonomy of vehicles, like we need, you know, when there's physical harm and safety, and I think in this case, when there's potential digital risk, I mean, you know, the hacking risk, the risk to our grids and infrastructure, it's real. And, you know, what I put out there is this, is like, I think the regulatory needs to come along for the ride. I don't think regulatory doesn't get involved, but I think if we are trying to lead by regulatory and regulation, we become Europe. And, you know, when, when, when, you know, laughingly when Europe says, you know, we need to slow down or anything, it's like, did you ever start? I mean, over there, like, you know, like, like Europe's. The exact reason why we have to be really thoughtful about how going too hard on regulation could be damaging because Europe's economy is basically for half a century now because it refuses to actually allow its technology to to to innovate at pace. It's a capture. It's a capture economy. It's it funds itself by, you know, antitrust and anti-competitive suits against mostly American companies. It's the, it's the wrong approach. We need balance here. And that's why I said, like, I went on TV, Pat, and I said, common freaking sense, like use common sense. I don't know why we have so little of it.

Patrick Moorhead: 

Yeah, totally. Those are, those are good ads here. You know, one of the things we, we didn't discuss, I think enough on this is the impact to open source. Uh, you know, I think we're going to hit, I think we're going to hit some of that when we hit the, uh, the flip. But if you look at, at, at what could have happened if we went along with what, uh, uh, Dario, uh, and, and it's what Dario wants. And it's so funny. We say, Oh, you know, Sam agreed. Well, Sam said he agreed and they comes out with a completely different proposal or a modified version of it. So. Anyways, this is this is what makes tech analysts and influencers worth something, I guess, you know, and just like it's worth taking a moment here.

Daniel Newman: 

But like, Open source, probably the one thing we missed on this whole multi-topic here in the decode is the frontier is the pace. Let's be honest, right? I mean, I think that was one of the smartest things that was said that we didn't comment on. The frontier actually does dictate their own pace. If they want to slow down, slow down. There's only a couple of them. I mean, Google did slow down. Meta did slow down. They actually slowed down. Anthropic and OpenAI have not slowed down. And the net of it is that everyone can talk about how great open source and open weight models are, but you and I both know that the open source and open weight models would not be as good as they are if the frontier wasn't moving as fast as it is. They are doing some interesting things architecturally, and they do interesting things for token generation, optimization, for improvement, all the way back to test time compute and how they did that with DeepSeq 1. But even DeepSeq, even these GLM models, these QEM models, They think they're anthropic and there's a reason that stuff happens is because they are using those foundational models to build their own models and so if the frontier slows down open source will to the real debate is is. Dario said 5% to 10% of the capabilities are being used. So for the average activity that an enterprise needs to do, one of the five skills that we measure on our pinnacle benchmark for an executive skill, an analytical skill, a customer skill, you don't need Astra. To do it, you can probably use Sol. You could probably use a model even older than that to do it. And that's the point where people are worried where the economics are, because at this point, like 90% of the economics still flow to two companies. Sorry, I just want to back in and talk about that, because I think that's a real thing people need to be aware of. The open doesn't move if the frontier doesn't move.

Patrick Moorhead: 

Yeah, I agree. And check out our Signal 65 Pinnacle results that puts that up there. One thing Pinnacle doesn't test yet are client-based devices. Pico and I, we have four Sparks at this point that's sitting on his sitting on his kitchen table. And we've got DeepSeek V4.1 Flash, which is very similar in output quality to 5.6 sol. Right? Like think about that. It is crazy. And everybody says, oh, you know, there's like a three month difference. I agree with the distillation point. I did a victory lap on that last week, but I also want to add that that these Chinese model makers and even the IFM out of the UAE are bringing some very novel architectural capabilities, like an 85% reduction in KV cash. I mean, that is unbelievable. What that means is it's just freaking cheaper, right? So anyways, check out the Signal 65 Pinnacle stuff, do your own comparison there. Hey, let's get into the next topic. Daniel, you spent an inordinate amount of time at Dreamforce. $36, 36 hours. Well, if you're Daniel Newman, that's that's a lot. 

Daniel Newman: 

For me that's a long time.

Patrick Moorhead: 

Yeah.

Daniel Newman: 

I mean, you and I have, you know, jetted in and we are we are we are notorious for our short stays. But look, I mean, given everything that was going on, we were at the center. We were the amazing. technological center of the universe. I mean, like I said, literally within an hour of time, I heard Mark Benioff talk to Sam, to Dario, and to Jensen. And Amidst all this, I mean, there's a lot going on and probably, you know, well, a lot of that wasn't directly related to Dreamforce. What it does kind of circle back to is when you think about security, you think about safety, you think about governance, was that whole AI eating software narrative, you know, which was that, you know, AI, SaaSpocalypse, the end of enterprise software. And, you know, Salesforce rolled out AI force built on top of cloud force. And you have these, these three showing up. There's a reason for this. And the reason here is that the narrative The hyperbole of AI ending enterprise software maybe met the moment for a lot of people who have a beef with enterprise software systems of record, don't like their applications, maybe have some data entry misery in their past, but it's just not playing out in reality. And so what the smart software money are doing now, and this is what Salesforce is doing with its AI force, is it's understanding that the future is different. The future, you know, Mark was on our Six Five Summit, talked about the future is headless. I know we have some debates on whether headless is the word we like, but the net of it is, is that, you know, a model plus a system of record with your, uh, critical deterministic business data, plus your infrastructure is going to be the way that we interact with the business's workflows. And so, you know, I believe, and I went to the investor day, like the growth rate is intact. The business is seeing strength. They're building agent personas across AI force. They're partnering deeply with Claude, planning to partner with OpenAI. building an infrastructure that can work across different models and harnesses to basically do enterprise workflows on AI force. And what I think we're seeing is the fact that people aren't going to move off their systems of record, not their critical ones. And so that's why this group showed up. That's what is basically happening here. This is why their agent and data cloud businesses are growing at triple digit percentages. It's a new company. It's got a new cadence of growth. You heard Mark talk about this. He hasn't had a launch like this ever. But at the same time, agents and AI force are certainly traversing a narrative that had long made it look like these companies may be in trouble. Market hasn't fully caught up yet, but seeing what they're building and making available, looked really encouraging. And like I said, the people that showed up to me are indicators that the AI pivot was a great headline, but it isn't reality.

Patrick Moorhead: 

Yeah, from the start, and I think this came out in our discussion on all the cesspocalypse that brought these tickers down like 40%, 30, 40%, some even more, was that if you are quickly integrating valuable features that leverage AI, you're going to be okay if you're an enterprise play. And it's based upon two premises. First of all, it's more than just doing a POC, right? I can go on perplexity or name your end-to-end service lovable or something like that, and I can build a POC. Now, can it hit every piece of data, protect that? Can it give a known good output across 500 different situations? And can it be updated and secure in infinite item down the road that's that's a very different proposition so. The SaaS providers like Salesforce didn't have to be the first with these capabilities, but they had to very quickly put the integration in so their customers could derive value. I think the rub is always, in the end, Dreamforce, sorry, Salesforce has to grow. It has to grow profit dollars. And then you ask yourself, Well, where is that money going to come from? Right? Where? What budget is this? Did the marketing budget go up? Did the sales budget go up? Or is it more about the effectiveness? Can I drive more sales? Can I increase my marketing spend, effectiveness, and efficiency? And I think Salesforce, although in addition to ServiceNow, should kind of be put up on a pedestal as this is how it is done. Some of the notables that I noticed, you know, we've talked a lot about open models. We've talked a lot about Um, frontier models, but what we haven't talked a lot about on this show lately, uh, are custom models and, uh, Salesforce took, uh, Nemo Tron. And they created their own custom model called Koa. And I think that was the big deal. And what they did is, so first of all, Nemotron is not distilled from any other model. And two, it scores some of the highest openness ratings on artificial analysis. By the way, I did a diatribe on what open actually means on a Forbes article, if you want to check that out. Open is a generic, but there's open weights, there's open data. There is open architecture. There's five different classifications of open and hats off to artificial analysis, but Nemotron, Nemotron Ultra scores, I think the max score on that in addition to the IFM models. And they took that and they modified it and they trained against 30 years of data from CRM deployments. They called it synthetic data sets to not freak out their customers. But this just makes sense. By the way, it's the same model that Alex Karp talked about that he rolled out inside of his service offering. So that was a big takeaway for me. But if I dial back, this is all about multimodality and utility. It's it's hey, if you want an AI type of interface to get to the same result, if I'm Salesforce, I got that. If you want classic Salesforce UI, but adding the power of AI, we can we can offer that as well. And some people are like, oh, my gosh, like you giving away the farm by doing this with people like Anthropic. I think that's missing the reality of where the value is. You know, Daniel, you and I have debated or discussed where does the true value of AI come from, right? Technology is going to be homogenous. So to me, it's timing and it's people and it's data, right? Like how quickly can you get there and hoover in the customers? That's the timing. What unique data do you have? And what people do you have to be the human in the loop? And from a Salesforce perspective, that's their Salesforce, right? You remember when they laid off a bunch of their salespeople because their AI was so good? And they're like, you know what? Uh, this actually isn't working. And then they went out and they hire thousands of human, uh, sales salespeople, uh, out there. So what a big show. Uh, I'm sorry to have missed it, but, uh, I was unfortunately, uh, tied up at a, I covered it for you. Don't worry. No, I appreciate that. I was at a client advisory that had been baked in quite frankly, uh, before I got invited to, uh, Dreamforce client advisory, he says, prove it. and All right, let's go into the next piece here. So OpenRouter says OpenAI is pulling even with Anthropic on its developers. So I want to put a caveat out there, and that is that OpenRouter monitors maybe 10 or 20% of the entire AI traffic out there. The other modalities are going through a CSP, going directly to the Frontier Lab, but, and this came out from a Gavin Baker, a tradies management tweet out there, that quite frankly, And during the Anthropic Claude 4.6 moment, you had OpenAI basically not flatlining, but losing a lot of market share to Anthropic. And what this new September 16th data showed that OpenAI captured 50% of the API traffic share versus Anthropic, up from 20% in June, which is just absolutely fascinating. And I think it just shows that this ping pong effect of who's in the lead and who are the laggards. And just when you think that somebody has checkmate, something else comes along the frontier and says, ah, hold on a second, that's too fast. So, you've got Google, you have Anthropic, and you have OpenAI, and then you've got open source, and then you've got custom, like we saw I think this is good for the market to have all these different options and this back and forth. One thing for sure, this puts an exclamation point on, is that Sam's upfront investment and locking in the infrastructure is part of why this happened. God, I just love the free resets from Tebow, right? Tebow, I don't know his exact title at OpenAI, but this guy is like, you know, AI Jesus for free token resets. And he takes feedback and he's out there shitposting and responding to the critics and asking people kind of what they want. So congrats to OpenAI on this.

Daniel Newman: 

Look, this is where I'm just going to double, triple click is the race ain't over, folks. It's just not. You know, these are new caveats. It's it's, by the way, it's not even just limited to the two. You know, it's. We thought OpenAI had won, and then we thought that OpenAI had had lost and that Claude had built a, you know, an unachievable gap for any other company. And then Sol looked better. Astra blew it out of the water on our benchmarks. And then, you know, we saw what Muse did, came back with something that was not quite but close, out of nowhere, right? They were like nowhere. Everyone had said meta was done. And a lot of people have ruled Google out and say Google can't come back. And I think that's a fool's errand. You mentioned some of what NVIDIA is doing with Nemotron and to think a company with that type of capital resources can't build something. So that's the first kind of thing I'm just thinking about as I watch this happen. Because clearly there's less of a moat itself in the model, meaning that if someone finds another model that can do it, and of course, the economics of the model becomes more efficient, where it can be deployed, the size of the model, the amount of compute required. And you and I have talked a lot about this, but like, you know, the aggressiveness on buying compute has a meaning. People like kind of thought it was crazy, but look at how much compute Zuck's been buying, look at how much compute Sam went and bought when people thought it was crazy. There's no way you can get a return on that. And I believe Jensen this week came out and talked about how, you know, 50 billion in investment will return about 50 in a year on Vera. Rubin on a gigawatt of compute. And that means in year two and three, it starts to return pretty handsomely. Well, these folks are buying many. I think Anthropic's note today came at they're buying, they're trying to have five up by the end of the year. And so if you look at that revenue ramp, there's a lot going on. But here's the thing, like I said, it's fun because the race ain't over. It's fun because there's more companies to enter the competition. It's fun because all the things you mentioned about open source creating a back end pressure for the frontiers to keep building. But again, we are still seeing that this is a race that's going to be won on having enough resources. And the resources here are energy, compute, all the constraints. And by the way, that's why the markets keep ripping each time that we keep thinking it's going to pull back. Because, you know, look at Michael Dell, by the way. Congratulations. I just want to make a quick shout, but I think he's now the second Wealthiest human being in the world, you know, and someone who, you know, by the way, just one of the absolute best humans that I've met. I know you know him very well. So, but like, I'm just so happy for him, man. I mean, God, what an incredible run. I mean, it's not like he was hurting before.

Patrick Moorhead: 

Isn't it amazing that, you know, the number one is Elon and he has like more haters. than ever. And Michael, pretty much everybody universally respects and likes Michael. So it just shows that you can be mega rich and not be a D-bag and offensive and show value to everybody. And now him unlocking value with the children of tomorrow is just so great. It really is. One thing that one thing to add on, you know, something that we didn't even talk about that just came out of nowhere. Talk about competition. When's the last time we talked about Google? I know. And by the way, Google is number two on the pinnacle benchmark. Gemini 3.8 flash thinking high gets better. Oh, and Grok six thinking gets higher scores than the best Claude Fable 5.1. on what on which were the scores on the overall and a knowledge worker. It's number two on data analyst. It's number two on IT professional had to refresh.

Daniel Newman: 

Sorry, I refreshed. Yeah, yeah, you're right. The Astra thinking max is one and now you got Gemini. Grok, which I didn't even mention in the competition. And so in the top 10, it's basically GPT, then Google, then Grok, then Muse. I mean, Claude's in there too, sorry. Claude, then Muse, then Quinn, which as we know, and then I think Deep Seek rounds out the 10. So it's, the race is on, man.

Patrick Moorhead: 

Daniel, you had ended your segment talking about infrastructure. There was a huge summit last week, AI Infrastructure Summit, where a lot of things were announced. I wouldn't say new product announced, but more new information came out, doubling down on positions and hey, maybe even Signal Six Five results.

Daniel Newman: 

We got some main stage action, right, when Ian Buck presented from NVIDIA and we put out some new benchmarks around CPUs. Again, remember that was a hot topic for a while. You know, it still is important by the way, we're just not talking about it as much. And this is a little bit of a controversial one. I think people are going to have some debates. And obviously, NVIDIA is hedging. They're partnering deeply with Intel on future head nodes, CPU head nodes. But they're also believe that Vera is a 20 plus billion dollar annual business. But it came out with a because I wasn't at the event, but the high level data point that came out was we ran 1.64x faster per core than the leading x86 processor. Now that's the leading what's out now. Of course, some companies might debate what's about to come out, but we can't test what we can't test. So that was kind of the big thing for us there. I mean, that's basically ARM versus x86, just a note. And right now, the ARM core, which, obviously, NVIDIA has optimized theirs for Vera Rubin, and that'll continue to be debated. But there was a bunch of other stuff, too, that just kind of high level, since we weren't there. But we had AWS made some big commits to a ton more compute from NVIDIA, again, even though they have their own. We had the big announcement last week. with Qualcomm AWS, these are kind of around the summit news. But I mean, you had a ton of focus there on CPO. There was a ton of on network and memory optimization technologies. I don't know about you, I, Brendan Burke from my team was there. I know we had an interesting six, five session with Ambarella that should come out this week. But like, it seemed like there was a lot more focus just, you know, on continuing to optimize the stack there. I would wait for our team to put out their nerdy notes because I got to be candid with all the stuff I was covering. I didn't get into the weeds at this one.

Patrick Moorhead: 

Yeah. Yeah, me either. It was interesting and ironic, I guess, maybe the same week, this ratepayer protection act. on data center energy costs that came out. And essentially they passed something that said that consumer prices for energy will not go up based on any data center that's built. And I think that's a good thing. And I think that is a fair thing. It does need to be metered against all the other benefits that are put out there in the community, right? Um, so, uh, anyways, I thought that was a, uh, an adder here. So great decode here, uh, Daniel. Um, but it's now time to get a little bit spicy and jump into the flip here. And, uh, this topic is essentially that the open AI is $1.2 trillion pre IPO private valuation. remains justified after all of this AI doomer news. And also their IPO was pushed out to 2027. So let's dive in. All right, Daniel, it looked like your final head on there. So you're going to argue that OpenAI is worth a $1.2 trillion.

Daniel Newman: 

Yeah, listen, I mean, last week was a safety headline, not a demand headline. And the market already told you which one matters. And you're seeing that already as the stocks continue to rise. Saturday, Dario says, pace the frontier. Sam and Elon agree. Sam pulls his 26 IPO. Monday, chip stocks dropped. Tuesday, smartest money in the world, knocking at OpenAI's door, wanting to come in at $1.2 trillion. A $1.5 trillion number is being thrown out there. $1.2 is a 41% step up in only six months. And investors are starting the conversation. Sam's not asking. The money's trying to come in because they see the potential here. And look, the numbers support it. I mean, the run rate crossed $40 billion. It went up 20% in a single model release. We just talked about re-acceleration that nobody had priced it in March when it was already valuing at close to a trillion dollars. The burn against the $40 billion base is growing, but it's not growing the way people were once concerned. And we've already said, Sam has gobbled up the compute that he needs to continue to make these investments, continue training. And let's be candid, no one's really slowing down. So if they were to slow down though, I mean, Pacing doesn't hurt anything in terms of an already deployed revenue. Enterprises are paying for what works today. And as Dario said, 5% to 10% of what's available is basically being used, meaning there's still so much more that can be done, even with the models that already exist. So even if it slows down, that only helps them even further. And if the capability at the top were to slow, the race becomes, as we said, distribution, contracts, brand, cash. OpenAI owns all four. Third-party evaluators inside every lab is a compliance regime that only the giants can afford. Sam is playing offense. That's why he's focused on safety. He's basically rolling off the field behind him like every great company has done when they've been monopolizing an industry, which is why it's going to happen again. Sam's got what he needs. And frankly, if Anthropic is marketing at $2 trillion and still planning its IPO, 1.2 feels very conservative given the fact that the open router data shows that the newest open AI models are better. They're being used faster. The revenue is growing more quickly and the demand is being created. So not a bubble in open AI. And maybe the Doomer Week changed the story a little bit, but it didn't change the demand and it certainly hasn't hurt the valuation of open AI.

Patrick Moorhead: 

Yeah, that's a great story. Unfortunately, it's wrong. I mean, open AI is absolutely not worth it. And I'm going to tell you, I'm going to tell you why here. So it's burning, it's burning cash at a historic rate. And it obviously needs this money, even though they did massive payments and commitments on all this infrastructure, and they're giving away tokens left and right. At some point, this company actually has to make money like Anthropic. It's second in revenue too, and it's asking for a first place treatment. Anthropic's run rate is 60% larger. And its investors will see $100 billion to $120 billion by year end. Let's not even talk about the discount rate, the amount of the Fed hiked 3.75% to 4%, which increases the cost of capital pretty much everywhere. 10-year touched 5%. So long-duration private marks are the most rate-sensitive asset in the AI trade. And it's so funny. Part of the capital that OpenAI is expected to get is conditioned on the event that just slipped. March round, 35 billion of Amazon's $50 billion depends on an IPO or AGI. And IPO is now 2027. And I don't think we can consider the Astra AGI at this point. And the other thing is, even Sam, he's endorsing slowing down while raising at a higher price. If the pacing is real, then the growth slows. If not, there's a credibility cost. And I already asked what happens. which, by the way, I totally agree. Final, and I think the most important element here is open source. Pinnacle's 90% cost reduction with open models that we just saw on September 11th. Salesforce just built COA on open weights instead of renting Frontier Intelligence. So I rest my case here, Daniel.

Daniel Newman: 

Yeah, and they also built Clawed Force, run on a Frontier model. They also realize that the open weight models depend on the frontiers to be any good. But sure, you can win.

Patrick Moorhead: 

Not true. Not true. Not true. NVIDIA is on the record saying that their models are not distilled.

Daniel Newman: 

Fair. Now, let's be very clear. Cause I, you know, I like NVIDIA. Who do you know that's running Nemotron to do all their AI work right now? All their AI work? Like who's actually centralizing or building their loops and their skills and their tools around all their enterprise software right now using Nemotron? Give me one.

Patrick Moorhead: I mean, Alex and Benioff.

Daniel Newman: 

But like they have Cloud Force, like literally all their customers are doing Cloud Force. Like the entire, I was in there yesterday. The entire thing was demonstrated on Cloud. I get it, like I do get it. By the way, I think personally, I do think if you have a Nemotron and you have Palantir level ontology, you can probably freaking do some amazing stuff on a completely open source model. I'm just saying, by the way, where is Nemotron on our benchmark? Good question. Because I would love to see how it does, because I think that would be a real interesting proof point. We should do that. We should run Nemotron against our five personas and see how it does against, because here's what I mean, as an enterprise, like not as a stock picker or a person that talks about the market path, like I would love, love, love, love, love. that if my cost of compute in my consumption and API costs went down by 90%, now I'm not spending what some are, but we have probably in a multi six figure annual run rate in token costs now, I would like to spend less. Well, did you see that we put out a survey today? We got 50 enterprise- I saw that. To tell us what they think about this pacing the frontier, so.

Patrick Moorhead: 

Yeah, tell us about your Doomer audience.

Daniel Newman: 

Well, no, it's just interesting because like we've spent a ton of time focusing on what the politicians say, what do the pundits say, what do the labs say, and then what do the largest providers of infrastructure say. But the bottom line is it kind of goes back to enterprise deployment is freaking hard. And I think, I think Dario talking about that five to 10% of actual, uh, the power models is being used is probably right. And most of the enterprises, they're trying to build a harness to be able to keep up with all the model diversity, but they having a hard time deploying, um, you know, compliance governed, uh, AI that, that meets all their requirements to, to act semi or fully autonomously. And so they're kind of saying like, look, we don't need more models. We need help to actually deploy this crap, which is why. We have forward deployed analysts and this is why Palantir has forward deployed engineers is because building what's possible to your point earlier about POCs very possible building something and deploying it safely and securely with real customer data. hard. And so the enterprises are screaming, we're okay if you slow down. But by the way, none of them believe that the labs are doing it altruistically.

Patrick Moorhead: 

Great, great flip here. Let's now tune into bulls and bears. So week off started pretty shaky and then it just started to absolutely rip. A lot of that was the AI Doomer, Doomer sell off at the end of the week. And then starting on, I believe, Wednesday, everything just started to absolutely boom.

Daniel Newman: 

Oh, I tweeted. on Coxon's news and then on Dario's news that every one of these FUD slash Doomer slash end of moments has presented a buying opportunity and this one will be no different. I stand by like, look, there's a lot of macro things going on. Like we do need to figure out how to exit this war. This oil thing is getting scary. Like we got to figure out how to either finish it or exit it. But like we, this, he does not want his legacy to be that. I don't think, cause he's criticized that a lot over time. The interest rate sensitivity, but obviously the big tech company, there is no price sensitivity. So there's no sensitivity to interest rates because if these guys have higher input costs, we've seen it with memory, they can just charge more. So, but what we've seen, Pat, is every single one of these Doomer claims has led to a Jevons paradox, which we like to use that. But like, we're actually where we think there'll be less, there needs to be more. And the biggest and the most analogous thing I can say here is, guess what happens if we want to run a safer AI? We got to spend more money on compute to do more training to actually build the safety into the models. There's more training work, which guess what? Exponentially more compute. So it's like, even the thing we think would slow us down, actually the labs to pace the frontier are going to need to invest more in safety. And much of that investment is more in training, which means they're going to need more compute. So we get it wrong every time. I just need bigger cojones to make bigger bets because I've tweeted correctly but I haven't always placed my financial bets correctly because there's always this overwhelming fear because the narratives get so strong that maybe this time will be wrong. We just got to believe but FUD equals opportunities.

Patrick Moorhead: 

Yeah, it was. And I think right now we're going to see blocks move together after a certain period of time. And if you look at the net increase over kind of this week versus last week in terms of capitalization, it's way up. And one marker that you know, hasn't been on a percentage basis moving a lot this year has been Nvidia. And, you know, even though they doubled down on literally doubling revenue for next year, it's just kind of been sitting there. But the market is finally reacting to that. I'm really amazed at how the market, you know, didn't didn't respond to oil, it really didn't respond to interest rates, and it rebounded so quickly on the dumerism, but it was good to see. All right, at Salesforce, and I don't know if you attended this, but they had their financial analyst day. They reaffirmed over $63 billion of revenue for 2020, 2030, but the stock went down anyways, which I think we have to imply. that this is what a kind of a show me story, show me story looks like. So, yeah, I mean, it cleared the streets model by $3.8 billion, which was, you know, I think even JPM said, quote unquote, incredibly positive for the rewriting thesis, you know, and, you know, he came out and counted six distinct monetization paths, including premium SKU. upgrades. The AI metrics in the outline though, they were over a year old. Current ARR was above 1.5 billion, up more than 240% year over year. So I think the bear case here is the company's target history, not AI in itself. This one was a little bit of a head-scratcher, but all I can think of is that the street wants to actually see even more evidence. The sell-side guys were or primarily positive, right? Stiefel, Canaccord, Freedom Capital, you know, equal weighted average of those guys that moved went from $258 to $298. So maybe, Daniel, you can shed some light on why this stopped.

Daniel Newman: 

I mean, there was a there was a weird irony. They had a massive outage the day of. which is never great on your investor day when your whole system went down. They did have it fixed by midday, but that was noteworthy. I think there was some concern. I mean, listen, it was a packed house. Like I was one of only a couple of analysts that are non-financial that was even allowed in the room because the room was packed to the gill. So people were definitely there and I did like they were actually showing demos, meaning they were doing things with CloudForce. building agents, building artifacts, showing how a business might perform certain tasks through pipeline or an agent to do a certain customer service workflow. So that was great. Having said that, I think there is some concern about the dependence on Anthropic. Your mention of COA and building on NVIDIA, I think that's going to be a big test is how much that grows. And then again, I think we're in the middle of the ROI era, the hard wire, as I like to call it, and I think they need to be able to show, you know, they have, they built kind of a cost, like we built for our benchmark, they built like a cost per task type pricing. And I think getting that cost per task optimized so that the most tasks at volume are being run across the Salesforce ecosystem on the right models is going to be key. They've got personas for all parts of the sales stack that they're building and they're continuing to expand them. And I think in the end, the math and the governance are going to be the things. Salesforce can deliver the governance by being inside that system of record and running it all through them. But then again, people are going to want to optimize for cost. And so I thought it was a good number. They're talking about data AI remaining 100 plus percent growth for the next two fiscal years. Again, the building across the model landscape, very, very strong. Uh, you know, it was a good, it was a good showing, but again, having your entire system go down on the day of your investor day, probably, uh, you know, I don't know if that's why it went down, but it's just not, it's just, it just probably wasn't the look they were going for. Yeah.

Patrick Moorhead: 

And let's move to the next, the last topic. I'm going to skip interest rates. And and I want you would like you to talk about Nebbius future equities. Had some discussion on this. So, Daniel, what what's going on here? I mean, literally no free cash flow until after 2020, 2030 here. Yeah. I mean, look.

Daniel Newman: 

First of all, I can't believe you skipped interest rates. I love interest rates. Don't you all love interest rates? Okay. Look, I think the real news around nebbiest this week is that, you know, They showed some big pricing moves. We had some interesting conversations about the fact that in 28, we're expecting them to build around 15 billion of EBITDA, and they're already sitting at a $60 billion market cap. If you actually look at that as a multiple of two years out, they're actually priced okay. I think the real question is about what Nebbias can charge. And I think this week they came out, I believe they raised their prices across the board by more than 20%. That's across their H100, H200, and Blackwell, and soon Vera products. And when you start seeing the fact that they have this kind of price sensitivity and the ability to raise prices at this rate, you start to realize that The constraint is very real. The demand still remains elevated. The market's pricing these things, probably undervaluing them because the ability to generate growth and EBITDA remains high. CoreWeave took a bit of a hit this week. Nebius is in the same ballpark. And I think the reason that's happening has largely to do with the fact that while they are creating EBITDA, it's adjusted earnings. It's not cash. These businesses are going to have to keep investing in incredibly intense rates for a long time. But it looks like with Nebius's pricing power and this demand curve that I think they're going to remain a strong play. And the future equity side likes them a lot. I continue to like their token factory approach and the profitability that they have. And, and, you know, so that BS looks good.

Patrick Moorhead: 

Yeah. I like their vertical integration, meaning they're not buying, they're creating their own infrastructure, which is, is unique across the, um, Neoclouds. So, hey, Daniel, great show. I want to thank everybody for tuning in. Hope you have a great week. You had a great weekend. We really appreciate you. Hit that subscribe button. Tell your friends, your family, your pets, pretty much everybody who parks themselves in front of some form of video and is looking for insights on tech. Take us for a walk. Get out there, be healthy, touch grass, right? Yeah, you can touch grass too, and listen to the audio portion of it as well. I'm not that good looking, at least not me. So, anyways, thanks everybody, take care.