CNBC Halftime Report with Brad Gerstner - October 15th, 2026
- Brad Gerstner is tactically cautious after AI stocks’ April–May surge, yet more bullish on 2027–28. He cut exposure to “medium-small,” expecting consolidation amid higher rates, inflation and political risk, but may begin buying over the next couple of weeks as prices become enticing. His distinction is crucial: hedge-fund capital must manage drawdowns, while growth and venture funds can underwrite five to ten years—“two simultaneous but different worlds.”
- For Gerstner, the AI trade still turns on whether lab revenue can justify the capex buildout. He recalls asking Sam Altman how companies could make $1 trillion of spending commitments against $13 billion of GAAP revenue; this year’s AI revenues then “exploded higher,” with Anthropic expectations moving from roughly $30 billion at the year’s start to reports of at least $100 billion. Consolidation reflects questions about sustainability, open-source commoditization and financing costs, but he welcomes the churn because “it avoids bubbles.”
- Gerstner left Anthropic’s investor day more bullish on the company, its prospective IPO and the road ahead. His “vibe check” among major investors was “a nine or a 10 out of 10,” despite a negative public-market reaction to reported Q3 revenue just under $17 billion, up 45% sequentially versus roughly 140% earlier in the year. He expects an IPO in “weeks, not months” and suspects demand could exceed $500 billion.
- His central thesis is that revenue follows model capability, and recursive AI development should accelerate the next wave of products. Opus 4.5 enabled Claude Code, which he credits with unlocking this year’s revenue; machines are now writing “80 or 90% of the code” for subsequent releases. Gerstner expects analogous “Claude Code moments” in cyber, knowledge work, biology and materials science, with completed but unreleased models providing potential catalysts.
- Gerstner’s non-consensus view is that frontier labs are widening, not losing, their lead over open models. He argues that recursive intelligence compounds, “intelligence is the log of compute,” and the leading labs will add the most compute next year; public comparisons therefore match Chinese or open models against older released systems, not internal generations under safety testing. He still wants vigorous open-source competition and sees price cuts as rational sharing of increased “intelligence density,” not proof that frontier economics are collapsing.
- CoreWeave illustrates Gerstner’s distinction between a violent valuation reset and a broken business. He owns no shares today after reducing AI exposure and reallocating 10% of his book into Meta and another 10% into Tesla, but calls much of the CoreWeave debt narrative misinformation: the 15% public mark concerns about 20% of borrowings, while 80% is GPU-secured and relatively cheap. He sees potential opportunity in both debt and equity, while accepting that a GPU-rental business may not deserve an Anthropic-style multiple.
- On AI safety, Gerstner supports engineered containment and a human-controlled emergency brake, but rejects an “FDA for AI” as a static answer to an evolving technology. He describes Commerce-led frontier-model review, audit and board oversight under the White House superintelligence accord, and an AI task force overseeing cyber and critical industries. He allows that further formalization may eventually be needed, while arguing that safeguards must evolve without letting China reach the frontier first; his approach to China is “trust but verify.”
1. The AI correction changed Gerstner’s entry points, not his long-term thesis
Gerstner’s framing begins with the market’s central question: “Do the AI revenues justify the capex buildout?” He recalled asking Sam Altman how companies could make $1 trillion of spending commitments while producing only $13 billion in GAAP revenue. AI revenue then “exploded higher,” but softer momentum, open-source fears, higher rates and political uncertainty have forced the market to test the thesis again.
His positioning became “medium-small” after semiconductor stocks made what he called their biggest move in history during April and May. He believed they had moved ahead of fundamentals, then inflation and rates added further pressure; the resulting chop and consolidation neither surprised him nor invalidated the secular case.
Wapner pressed the apparent contradiction between Gerstner’s bullish language and lower exposure. Gerstner’s answer was mandate-specific: hedge-fund capital must avoid drawdowns, while growth and venture portfolios invest across five to ten years. With prices falling, he was “leaning in” and considering purchases over the following weeks as valuations became enticing.
2. Anthropic’s reported slowdown did not weaken his IPO conviction
Unable to disclose investor-day details, Gerstner offered a “vibe check”: major investors in the room registered “a nine or a 10 out of 10.” Over a one-to-three-year horizon, the information he had gathered made Altimeter more bullish on Anthropic, its IPO and “the road ahead,” not less.
The on-air counterweight was stark: Anthropic reportedly generated just under $17 billion in Q3, up 45% sequentially after roughly 140% quarterly growth earlier in the year. Reported adjusted profitability and margins above 80% supported bulls, but the market treated the deceleration as “not good enough,” amid a floated $2 trillion valuation.
Gerstner’s response rested on forward capability. Jensen Huang, he said, was “pretty sure”—or words to that effect—that lab revenue could at least double next year. If revenue exits this year around the reported $100 billion, that would imply roughly $200 billion exiting next year. Gerstner said investors closest to the model pipeline are seeing completed systems, not merely speculating about future breakthroughs.
His final trade was the prospective Anthropic IPO, which he expected in “weeks, not months.” Gerstner called Anthropic a potential enduring company of the Google-or-Amazon class and suspected more than $500 billion of demand. He compared that potential imbalance with the SpaceX IPO, where he recalled roughly $100 billion of stock being offered against about $350 billion of demand. Whether Anthropic priced at “$150,” “$175” or “$200” mattered less to a one-to-two-year holder.
3. Model capability is the load-bearing variable behind AI revenue
Gerstner’s strongest causal claim was categorical: “All of the revenue is derivative of model capability.” Without the breakthrough represented by Claude Opus 4.5 and the resulting Claude Code, he argued, this year’s revenue would not have appeared; Claude Code became better than any software engineer at writing software, and revenue consequently skyrocketed.
The next acceleration comes from a recursive loop: machines now write “80 or 90% of the code” behind subsequent product releases. Some call that the singularity; Gerstner’s narrower claim was that this machine-built development cycle is what he expects to see in 2027.
That mechanism also explains his rejection of the consensus that frontier leads are shrinking. “Intelligence is the log of compute,” the leading labs enter the recursive loop from ahead, and they will add the most compute next year. Public benchmarks compare open or Chinese models with released frontier products rather than internal models undergoing safety work.
As examples of the hidden gap, he argued that Astra had not solved Navier–Stokes and that an unreleased OpenAI model had clearly solved Millennium Prize problems. His investment implication was broader: cyber, knowledge work, biology and materials science could each experience their own “Claude Code moments.”
4. Open-source pressure is real, but Gerstner sees a much larger market
Gerstner resisted blaming Anthropic’s sequential slowdown on competition that had existed throughout the year. One possible explanation, carefully hedged, was reduced model-distillation traffic: if reports that such activity represented roughly 5–7% of revenue were right, billions of dollars of revenue might have appeared in Q2 but not Q3 as Anthropic restricted it.
Millions or billions of queries from Chinese models can generate revenue while helping competitors catch up, which Gerstner called a national-security threat. He said Anthropic was devoting substantial compute and intellectual resources to stopping it, and normalization could make the quarter-on-quarter picture look different.
He nevertheless wants “a very vibrant ecosystem of competition,” including open source, and called AI perhaps technology’s most competitive market. His disagreement is about the frontier: he expects closed-source models such as Haiku and Sonnet to become as cheap as or cheaper than open-source models, while customers will continue paying more for frontier intelligence.
Price cuts, in his account, reflect higher capability from the same compute rather than automatic share loss. Like AWS repeatedly cutting prices, labs can retain the productivity gain or share it to expand their moat. He prefers customer additions, spend and Fortune 500 or Fortune 2000 penetration to volatile monthly run-rate disclosures.
His market sizing was deliberately expansive: every person and company will use AI. Even under his illustrative assumptions that open source captures 20–30% of usage and frontier systems receive about 50% of usage over time, demand from major enterprises—including companies such as Lilly and Bristol—would still leave “the largest TAM in the history of the world.”
5. CoreWeave is being repriced as infrastructure, not written off
Gerstner owns no CoreWeave today, despite Altimeter having invested before its IPO. The exit was not a company-specific judgment: his team reduced overall AI exposure and funded diversification by shifting 10% of its book into Meta and another 10% into Tesla.
He called “90%” of commentary about CoreWeave misinformation. The quoted 15% yield was the public mark on roughly 20% of its debt, while the company was still paying about 9%; the remaining 80%, secured by GPUs, was “rock solid and relatively cheap.” He therefore saw possible opportunity in both its equity and debt.
Gerstner expects some weak hands to wash out, including businesses whose IPOs were pulled, but regards CoreWeave’s team, capital stack and Nvidia partnership as strong enough to survive as one of the neoclouds that makes it to the other side.
On valuation, he said CoreWeave went public, “I think,” at about $30 per share, reached “maybe” $180, and was around $60 when he spoke—still roughly double its IPO price, which occurred sometime within the prior two years. Investors had first given it an AI multiple, then reconsidered whether a GPU-rental business deserved the same multiple as Anthropic or SpaceX. Gerstner viewed that separation and application of more appropriate multiples as the kind of market behavior he wanted to see.
6. Safety must advance with the frontier, not freeze it
Gerstner endorsed Satya Nadella’s call for a human-controlled emergency brake and described Nvidia’s OpenShell as another approach to safe deployment. Enterprises will not rely solely on model makers; they will use software providers and controls across their own data estates to contain unexpected or compromised behavior.
His regulatory case rested on a “three-legged stool”: Commerce-led review of every frontier model before release; a White House superintelligence accord using internal audits, external audits and board oversight; and an AI task force led by Jay Clayton, Emil Michael, Scott Kupor and others to oversee cyber and critical industries.
Gerstner conceded that parts of this framework may eventually require codification, but rejected claims that nothing is being done. He also rejected the premise that AI is a new species: “This is a piece of software” trained by humans, so training frameworks should be disclosed, reviewed and audited as capabilities evolve.
His harder safety conclusion was that “intelligence is going to have to be the one to oversee intelligence,” because advanced models’ cyber capabilities may become the best monitors. An “FDA for AI” could set America back, while losing the frontier race would also destroy safety: “If China gets there first, I can guarantee we’re not safe.”
On China itself, Gerstner adopted “trust but verify.” He found the second Trump–Xi meeting that year, plans for further meetings, and AI discussions between both governments encouraging. He allowed that shared interests could be real, but would not outsource verification or assume those interests were sufficient safeguards.
Full transcript
Today: AI in America. I am Scott Wapner at One Market in San Francisco. Over the next hour, we’ll look deeply into the AI trade, assess the risks and rewards of this societal transformation, and examine what the recent action in the stock market suggests about the road ahead for investors.
Brad Gerstner is with me today for the hour. He’s been in nearly every major meeting at the White House on AI. Snowflake’s CEO is going to join me as well today, as the future of software in this AI age remains a hotly debated subject. The investment committee is with me as well: Josh Brown, Stephanie Link, and Malcolm Ethridge.
You know what the market picture looks like, but let’s show you, because the Dow is up big, but AI stocks are not. I want to start with you. As we think about semiconductors, what’s happening with CoreWeave—and we’ll get to CoreWeave more specifically in a moment—is this about AI slowdown concerns? Is this about debt levels and interest rates going up? Is this choice C: all of the above? What do you think?
1. AI Revenues Meet Capex
Yeah, of course it’s all of the above. First, it’s great to be back here. We talked about this maybe a month ago, right? The number-one question in the market today is: do the AI revenues justify the capex buildout?
Frankly, it goes back to the conversation I had with Sam Altman on my podcast last October, when I asked him the question: how can you make a trillion dollars of spending commitments when you have $13 billion in GAAP revenue?
What happened this year is that revenues exploded higher. Remember the forecast for Anthropic at the start of this year: $30 billion. Now, a lot of the reports have them doing at least $100 billion. We’ll get into that a little bit today.
But the fact of the matter is that AI revenues exploded higher this year, and now we’re in this phase of consolidation. People are starting to worry: is open source commoditizing the frontier labs? Are the revenues sustainable? At the same time, they’re hearing in the marketplace that things may have slowed down.
That, plus higher rates and some of the political stuff going on in the background, has people scared. I will tell you, I’m on the more positive side of this, but I think it’s good that we have this churn in the market. It avoids bubbles.
Okay, you’re on the more positive side, but your actions suggest that you are getting more cautious, don’t they? You’ve taken your exposure down.
Yes, I did. I gave a presentation at the All-In Summit about a month ago and said I had taken my exposure down to medium-small. I think I had shared that with you the week before.
The reason for that is that I saw a lot of these AI stocks in April and May. We had the biggest move in semiconductors in history, and I just thought they got ahead of themselves, that we were going to consolidate a little bit. It turns out there were even more fears that came along—inflation, rates, et cetera.
How can you then profess how positive you are?
Remember, I live in two simultaneous but different worlds. One world is that I manage money for people in a hedge fund, so my job is to make sure I avoid some of the drawdown in my long-only fund. In my growth funds and venture funds, we’re investing for the next 5 to 10 years.
I’m telling you, when I look at 2027 and 2028, I’m even more positive, and we can get into why. But you have to acknowledge that there’s going to be chop along the way, so it doesn’t surprise me at all that we’ve had some of that chop and consolidation.
Why am I getting more positive? Why am I leaning in and maybe starting to buy over the course of the next couple of weeks? Because I’m seeing prices come in and start to get enticing again.
Well, Josh, that’s the trillion-dollar question: whether there’s more pain to be had in this market, whether we’re going to get to some levels that are attractive not only to a Brad Gerstner, but to a Josh Brown and any investor who is now watching this program, thinking about where we are and where they think we’re going.
2. Rates Reveal AI Crowding
Yeah, look, I think something really interesting happened this week. We finally got the answer, or part of the answer, to the riddle that everybody’s been wondering about.
The question was: why have rates been marching higher really since June, but accelerating into the end of September and early October? It was this big debate. Is it oil prices feeding into inflation expectations? Is it reverse crowding out, where all this debt coming from the AI capex buildout is making it so the government has to raise the rate in order to entice buyers?
I think what we learned yesterday was that the latter is a really big driver of rates. Think about what happened. We had hot PPI and hot CPI. It doesn’t really matter, because in the end, bonds fell when it looked like there would be fewer bonds coming from CoreWeave and the neoclouds, and maybe even some of the established incumbents in data centers.
All of a sudden, we had this crazy move in rates—what looks like a blowoff top in the 5-year and the 10-year. That tells you how much of the rise in rates is coming from just too many bonds being offered by corporate America.
If we think there’s going to be a slowdown there, and that’s what the stock prices of Nebius, CoreWeave, and the rest tell you, all of a sudden we get rates coming in. Look what’s working in the market today: everything other than AI caught a big bid, from health care to even staples. We’re seeing housing-related plays move.
That’s a really interesting situation for people trying to figure out where to go from here. The answer for stock market investors is: do yourself a favor. Instead of adding the fifth semiconductor name to your portfolio, look elsewhere. That’s what the market is telling us today.
Well, Steph, there’s been a debate about looking elsewhere, right? The market became much more concentrated at the expense of the elsewhere, which you’ve been advocating for repeatedly. Now maybe the pendulum is starting to swing back the other way as we ask more serious questions about where we are in the AI trade.
Well, I don’t think it’s one or the other. I think it’s both. I have been adding to technology on any of the declines, but I also have been very much exposed to the 493 other stocks and other sectors.
The reason is because the economy continues to grow in the face of higher rates and higher inflation. We’re still growing north of 3%. Like today’s data—I know we were all focused on PPI and yesterday’s CPI—but did you look at retail sales and how great that was? Weekly jobless claims had a 4-week moving average of 198,000, the lowest level since 1969.
You look at some of the manufacturing regional series, like the Philly Fed and the Empire Fed. There are a lot of good things happening beyond technology and beyond AI. I think we should see a catch-up, because I think you are going to see earnings.
I thought the bank earnings this week were a little sloppy, but outside of capital markets and M&A, the rest of the businesses are humming. You had these stocks that had been up significantly at one point. Morgan Stanley was up 25% this year and is now flat on the year, so the valuations are coming down.
At the same time, Scott, in the semiconductor space, I do not think that AI and the food chain are dead. We haven’t seen any kind of change in the $1 trillion spend that we’re seeing from the hyperscalers.
Rates are certainly going to hurt some of the bad players that have bad balance sheets. CoreWeave has $50 billion in debt. Oracle has $167 billion in debt. So sure, higher rates are going to hurt them, but you’ve got cash generators that are actually down just as much, and those are opportunities.
NVIDIA, Broadcom, Micron—we’ll get to all of them later—but you have very healthy companies getting caught up in this tailwind of concern. I think you’re going to see opportunities in tech as well as beyond tech.
Malcolm, what do you make of what’s been happening with the AI-centric trade?
Yeah, I think what we’re seeing is potentially a changing of the guard. I heard Brad mention that we’re probably going to see a tremendous amount of chop between now and when we finally get there.
I’m listening to Josh and Stephanie talk about looking outside of the Mag 7, let’s say, or the hyperscalers, for the opportunity to stick with the AI trade or to diversify. What I think we’re looking at is not necessarily a changing of the guard in the sense that AI is no longer going to be the thing powering the economy as well as the stock market.
Maybe what we’re looking at is just some other companies beneath the surface that we haven’t really paid attention to because they’re not the chips. They’re not the CoreWeaves or the Nebiuses. They’re not the hyperscalers or the Mag 7 doing all the building.
They’re the companies that are really going to weaponize the AI technology as it’s being constructed right now. They’re going to figure out ways to do a lot more with a lot less, and that is what’s going to drive the second leg of the AI revolution.
I think that’s probably the question we should be looking to answer. Is it one where we’ve got to actually take the market down entirely to make that rotation, or is it one where we can successfully make the leap between where we are right now and those companies that will come out as the new leaders at the forefront of the AI revolution?
I’m thinking software. Obviously, that’s a very direct link. But I’m looking also at the companies beyond that.
You heard Brian Moynihan, for example, talking about how they've had to add 1,000 people to their 3,000-person workforce to help build out their cybersecurity outfit. But they're also expected to see something like $700 million in savings starting next year just from what they've done to deploy AI across the bank. And that's one company out of however many thousands of global enterprises. So I think that's where the real opportunity is that investors should be paying attention to and figuring out how to get to where the puck is going.
You, Brad, have really said—and I think you said it as much already—that all roads lead back to OpenAI and Anthropic, right? The hope is that their revenues are going to grow fast enough to meet the commitments that the market expects them to make. It's kind of the whole story at this point. You were at the investor day yesterday.
I was.
You've been an early investor in Anthropic. What do you make of what you heard?
3. Anthropic Faces the Market
I can't get into anything that was said in the room yesterday, but what I will share is this: If I had to do a vibe check coming out of the room, I found it fascinating. You had a lot of the most important investors in the world in the room, and I would say the vibe check was a 9 or a 10 out of 10, right? Very bullish in terms of how they feel about this company.
Now, they're probably looking at it over a 1-, 2-, or 3-year time horizon, not over the next 30 days. If you look at what was reported as having been said in the room and the reaction in the market, that was quite negative, or at least my interpretation of it was quite negative. So you had this asymmetry between how people feel about it.
Certainly from an Altimeter perspective, the information I've gathered over the last few weeks, including yesterday, makes me more bullish on Anthropic, their IPO, and the road ahead, not less bullish. But they're going to have to wait until they flip the S-1 and can share some of that information with the world, do their roadshow, and I can't wait for this company to get public. Retail investors are going to be excited about this, as will long-only funds, mutual funds, sovereigns, and lots of folks. They should get the chance to invest in this business, and then I'll be able to share a lot more of what we want to say.
The fact of the matter is, I think AI is accelerating. I just did this podcast with Jensen, and I think it's accelerating because we're entering a recursive loop where the machine is building more of what comes next, right? The machines are building 80% or 90% of the code that drives the next product release. That is a recursive loop that accelerates. Some people call this the singularity, right? But I think that's what we're going to see in 2027.
Remember the breakthrough we saw with Claude Opus 4.5 last December that led to Claude Code? Without that, we wouldn't have had any of these revenues. That was one product. Now imagine having a variety of those breakthroughs for cyber, for knowledge work, for bio, and for material sciences. I think we're on the verge of having Claude Code moments for a lot of these industries, and it's going to drive a lot of excitement.
I know you're limited, as you said, to what you're allowed to say, but I appreciate you confirming that you're in the room as an early investor. Our Kate Rooney is not limited in what she can say about this, and you do have the numbers that were reported and how investors are reacting to what we know. What can you tell us?
Exactly. Well, Brad, we'll let you off the hook. You don't need to share the numbers. We've talked to some folks who were in the room. We do have those new numbers that were shared at Anthropic's investor day here in San Francisco yesterday.
I talked to folks who were in there—not Brad—but Anthropic did tell investors that revenue was just shy of $17 billion for the third quarter. That was up 45% sequentially. Objectively, that is strong growth for any tech name, but for comparison, revenue was up about 140% quarter over quarter earlier this year. So the bar is just sky-high for Anthropic and a lot of these AI names, and the Street is reacting to that this morning.
DA Davidson put it well this morning. They said the market does appear to be reading the Anthropic update as a negative, which is understandable considering how on edge this market has been about anything around the AI trade. The bulls, though, are absorbing the news and pointing to profitability.
This was a topic in the room as well yesterday, from what we're hearing. Anthropic did tell investors it was profitable on an adjusted basis in the quarter, with 80%-plus margins. So the reaction to those results speaks again to expectations overall, plus some jitters about the spending behind the AI boom and whether all of it is justified based on some of this revenue growth.
Also, if Anthropic can support what has been reported to be a $2 trillion valuation—that is the number that we have heard being floated—Anthropic is targeting an IPO before Thanksgiving. No date quite yet, guys, but we're waiting on that S-1 flip so Brad can talk about it.
The “not good enough” seems to be the trending outcome, whether it's Piper Sandler today, and that's what they're saying. The numbers are obviously huge. You said you dropped the podcast that you just did with Jensen, in which he's as bullish as he has ever sounded. It sounds to me like he's bullish on where he thinks revenue growth is going to go for the labs, right?
Right. I asked Jensen in the podcast, “Are you confident that they're going to 2x revenue next year?” If they exited this year at what is reported to be $100 billion or something like that, that would be like $200 billion exiting next year. I wanted to know what his opinion was on that.
He's a big investor in Anthropic. He's a big investor in OpenAI, and he said—I think pretty sure, something to the effect of—at least 2x next year. So I think that a lot of the investors who are closest to the pipeline of what is happening on the models, remember this: All of the revenue is derivative of model capability. None of the revenue would have happened this year without Claude Code.
Claude Code became better than any software engineer at writing software, so it replaced them and that revenue skyrocketed. They're looking into the models that are already completed, Scott, not speculating on what they might do. They're looking at the models that are already completed in the labs and saying, with high confidence, that they're going to unlock other products like Claude Code.
I happen to agree with that, but that's what makes a market. I also like the consolidation in the chop. I think the debate you're seeing in the market is the right debate. The questions are the right questions.
We'll see when this thing comes public. I suspect if they choose to come public, there's going to be tremendous demand based on everything I know about the company. Remember the SpaceX IPO? I think it was a $100 billion offering, and you had about $350 billion of demand for that offering. The stock has done incredibly well coming out of that.
I suspect here you would have over $500 billion of demand for this offering, so I think the stock would do fine. Where it prices, none of us like—I don't really care, and I don't think the company really cares. Is it at $150? Is it at $175? Is it at $200? Those things are irrelevant if you're going to own the stock over the course of the next year or two.
What do you think about that? It's a great point. I wonder at what point competition starts to bite into the revenue growth, given what OpenAI has done. We know you're also an investor in OpenAI. At what point does competition start to bite into the revenue growth? There was a slowdown sequentially, so you look at the expectation, but that's slower than what they saw in the first quarter. What do you think is going on there? Is it more competition? Is it open source? Or are people just getting burnt out from spending big on these expensive models?
4. Open Source Tests the Moat
First, I would say the open-source competition and the OpenAI competition have existed all year, right? Some people look at their Q2 numbers and say, “Oh, those were good,” but they also had that competition there. So I might look for something else as an explanation.
I've seen a lot of reports out there that maybe they turned the knob down on distillation. There's been a lot of conversation in Washington and other places about the need for the labs themselves to cut back on distillation. These are Chinese models mostly running millions and billions of dollars of queries against these models, which generate revenue, right? But then they allow those competitors to catch up. It's a national security threat.
The company, as has been reported, suggested they were going to do something about that.
They were, and they were close to, if not solving it, making significant steps against it.
I think it's fair to say that Anthropic is devoting significant compute resources and significant intellectual resources to stopping this. Dario is very, very committed to that.
If they were to do that, and if that is what has been happening—you know, again, I've seen all these reports out there about 5% to 7% of revenue and all this stuff—that would be billions of dollars of revenue that would have been in Q2 but would have come out of Q3. So if you normalize those two things, it might look a little different.
Well, the irony is that these open-source models—the cheaper ones that are almost as good as Anthropic—you're saying have been essentially ripped off from Anthropic and OpenAI.
So I think for our audience, that is the full picture here: they make money on these, but what's happening in the background is that they are copying those models and offering basically a discount version.
Let me be very clear about two things. Number 1 is I want there to be a very vibrant ecosystem of competition, including open source, and there will be. You saw the results out of MUSE this week. Thinking Machines is on the verge of some big breakthroughs. Reflection just launched a new open-source model. We see what Google is doing.
So we have very, very healthy competition. I might argue this is the most competitive market in all of technology. And they're posting these results with that in mind. With that said, if you said to me, Brad, are they extending their lead over everybody else, or is the lead shrinking? I think the consensus view in the world—
—is that it's shrinking. I think I am on the exact opposite side of that. When you think about recursive loops of intelligence, they increase at an accelerating rate, and intelligence is the log of compute. These guys are adding the most compute next year. So you're ahead, you're entering a recursive loop, and you have the most compute.
Now you're comparing a Chinese model to their released models. Each of these labs has models complete inside the labs that they're doing safety testing on and lots of other things. So they may be generations ahead of the benchmarks that you're looking at and comparing to. For example, Navier–Stokes was not solved by Astra. So you have to ask, what solved it? What solved the Millennium Prize problems at OpenAI? Clearly, an unreleased model solved those things that is much further along than the open-source model.
So you disagree, then, with Gavin Baker, who posted, quote, “Why are the frontier models cutting prices? Might it possibly, just possibly, be the result of the fact that they're losing share to open-weight models?” Do you agree with that?
So look at what in fact happened. Number 1, every generation of model becomes much, much more capable. With the same amount of compute, you're producing a lot more intelligence density, and you get to choose what to do with that gain. You can either harvest it, in which case you keep the margin, or you share that margin back and you get more competitive.
What did AWS do over the last 10 years when they had those improvements? A hundred price cuts over the course of 10 years to expand their competitive moat. It was beneficial for the consumer and also beneficial for the company. What I would say is, look at their lower-level models, Haiku and Sonnet. They're going to have closed-source models that are as cheap as or cheaper than the open-source models.
But on the frontier, with frontier intelligence, people are still going to pay a lot of money for frontier intelligence.
You think they will? Well, one of the other things I thought was interesting that Gavin brought up was this run-rate revenue that everybody's been so focused on. I had heard that Anthropic is trying to move away from that, of course, as they go public, but it's a snapshot in time that people then extrapolate over the full year. Do you think we should be moving away from that? That has sort of been how all the AI labs in private markets have been disclosing their revenue.
Let me also add, before you answer that, on another post of his: no one—literally no one—in the supply chain takes the ridiculous $1 trillion in unfunded spending commitments from OpenAI and Anthropic even vaguely seriously. They're functionally irrelevant and that should be obvious to everyone. So put that with what he's saying. These are all questions that are being raised on the other side of the bull debate.
I talk to Gavin all the time.
And, by the way, he's still positive on OpenAI and Anthropic. I want to make sure that that's clear.
I talk to him all the time. He's a great analyst. He's on the right side of this. I'm just saying there's not a lot of daylight. Gavin's view is that the frontier labs are going to capture a huge portion of the economic value, and the open-source labs are, because we're dealing with the largest TAM in the history of the world.
Go back to run rate here for a second. I agree with you. Run rate is valuable because it tells us something about the pulse and health of the business at a given moment in time. I don't particularly love it when you release that as a public market company. It's like Roblox releasing DAUs every month—how many daily active users—because we habituate people to that monthly cadence, and the stock's all over the place. It's not great for investors.
But I think another way to think about the health of the business is: How many customers did you add last month? How much did they spend? How much of the Fortune 500 do you have? How much of the Fortune 2000 do you have? If you peeled back the onion on those things at Anthropic and OpenAI, they tell incredible stories.
I will stipulate this: this is the largest TAM. I've been out here for 25 years. There is not a single person or a single company, small, medium, or large, that's not going to use AI. One hundred percent of them are going to use it. Let's say 20% to 30% of their usage is open source, and let's say 50% of their usage over time is going to be on the frontier for companies like Lilly and Bristol and the most important companies in the world. It is a massive market for all these guys, and we are just getting started.
Stocks are down again. It seems to be the poster stock right now, at the heart of all of the concerns that we see. Steph was talking about $50 billion in debt. Yields are like 15%. CDS are blowing out, et cetera. Where does your position currently stand in CoreWeave?
I don't own CoreWeave today.
You don't own it. But, listen, we were pre-IPO, and I know you were pre-IPO, and you've been a big investor in this name.
5. CoreWeave Faces Debt Reality
We have. But I told you that over the course of the last 3 months, we took our exposures down—not because of anything specific to CoreWeave, but just because I said AI had overshot. We're going to have a period of consolidation. We wanted to make sure that we bought—we moved 10% of our book 2 months ago into Meta. We moved 10% into Tesla. We were diversifying into other names that we were excited about. We had to go somewhere to get that capital, so we diversified out of AI and took our overall exposures down.
Let me say a few things, though, about CoreWeave. Number 1, the team that's running CoreWeave is one of the best in the business. Ninety percent of the stuff that I read out there is misinformation with respect to CoreWeave. The idea that they have to pay 15% on their debt? No. Fifteen percent is the public mark on 20% of their debt. They're still paying 9% on that. All the stuff that is secured by the GPUs, which is 80% of everything they borrow, is rock-solid and relatively cheap.
So I think there's probably opportunity today in CoreWeave in both the equity and the debt. I'm sure my team is back at the shop working away on that, and somebody will make the pitch to me. But this is what happened, Scott, in these moments: everybody freaks out. “Oh my God, Anthropic's revenue was a little less than we thought it was going to be. Oh my God, we're not going to build out all this compute.” So they just sell everything in the compute stack because it had gone up in a parabolic fashion in April and May.
Jensen was speaking in Sweden today and said Nvidia doesn't need to backstop CoreWeave. He says CoreWeave's cost of capital is actually decreasing and that CoreWeave's strength is improving every day. But the market seems like it's going to be jittery around this name for a while, simply because of the amount of debt that it's taken out. You've got the data-center debate, which still feels like it's escalating.
I think there were probably some weak hands that got funded by venture capitalists, maybe even some in the public market, some of them that had their IPOs pulled. So I think you'll see some washing out.
But the fact of the matter is, I think CoreWeave, like I said, Mike Intrator, who runs that business, is one of the best. I think their capital stack is good. I think their partnerships, like the one they have with Nvidia, are fantastic. They're going to be one of the new AI hyperscalers, these neoclouds that make it to the other side.
But on a mark-to-market basis, let me just say one thing: they went public, I think, at $30 a share. I don't know. The stock's at $60 a share today. It's still a double off the IPO, which occurred sometime in the last 2 years. Not terrible. It's just that the stock went up to maybe $180 a share when people said, “I'm going to give it an AI multiple.” And now they're saying, “Oh, maybe it's a GPU rental business, and maybe a GPU rental business should not have the same multiple that I'm giving to Anthropic or SpaceX.”
That is the type of market behavior that we want. It's separating winners from losers. It's applying appropriate multiples. So I wouldn't panic in this moment. We'll come back.
AI safety and security has been the conversation ever since Anthropic's Dario Amodei published his “Pace the Frontier” letter. Satya Nadella of Microsoft had his own post over the weekend, in which he said, quote, “We need a human-controlled emergency brake in case models behave unexpectedly or get compromised, and can't simply rely on assurances from the model makers.”
What was your take on that?
6. AI Safety Needs Guardrails
I think it’s a good idea. We had this meeting at the White House two weeks ago, and he shared good ideas like that, as did Jensen, Elon, Jeff Bezos, Dario, Greg Brockman, and others. That’s what was so terrific about that meeting: everybody got to throw these ideas on the table.
I actually think you’d be surprised how little daylight there is. When I read Twitter, it sounds like there’s huge daylight between where Dario is and where Satya is, as an example. I think most people agree that we need some level of engineering, software-led containment. There are other things we need to do, and perhaps we ought to talk about what those other things are, but I like Satya’s idea. I think other people like Satya’s idea.
OpenShell, which was announced by NVIDIA, is another version of this, right? It’s the ability to deploy these models safely. Nobody’s going to rely just on the model companies. Of course, they’re going to look at their entire data estate. We’re going to have Sridhar Ramaswamy from Snowflake here in a little while, and they’re going to think about their current software providers and the things they can do to make sure that they contain or put a safety brake on these systems.
What do you mean when you say engineering software containment? Do you think we need a larger regulatory look at all of this? Because that seems to be where the heart of the debate is now: whether we need some sort of agency-level oversight into what’s happening here, because we truly don’t know what we don’t know about what these models are capable of doing.
I asked this question to Jensen on the podcast: Do you worry about the world that you’re bringing your kids into? I can tell you that everybody in that room had kids. We all care deeply about the future of the country and deeply about the future we create for our kids. I think we all have a lot of confidence we can do it safely.
Let’s first talk about what will happen.
Who’s the “we”?
You guys—everybody in that room, including a lot of regulators, like Jay Clayton, who was in that room.
Let me just talk about what we have, because I think it’s very mischaracterized. We have a frontier model review framework, right? That is led by the Commerce Department. Every frontier model is going to get reviewed before it gets released. We are the only country on the planet that’s actually sanctioned a frontier model and pulled it out of production, like we did earlier this year with Manus.
So this idea that we’re not reviewing the frontier models is a farce. Second, what we signed at the White House—the White House Accord on Superintelligence—is a Sarbanes-Oxley-like approach: internal audits, external audits, and board oversight. It creates massive liability for these companies.
And that was a law. Let me just give you the third.
Let me give you the third—the third leg on the stool. The third leg on the stool is the AI task force led by Jay Clayton, Emil Michael, Scott Kupor, and others. That’s going to oversee critical industries and cyber.
So we have a three-legged-stool approach to regulation in this country. I know that Bernie and Elizabeth and a lot of other people in the Senate are saying, “Oh my God, we’re not doing anything. This is crazy. We need to have the FDA for AI.” We need to take a breath and realize there is a massive amount of work going on to keep America safe.
We’re four years into this, okay? We have very powerful models. We just talked about all the anti-distillation work that’s getting done. You can imagine that Commerce is currently reviewing the next wave of models coming out of these two companies. Why haven’t we released the models, Scott? Well, maybe because we’re spending a lot of time reviewing the models.
But if you listen to senators, they would have you believe that no work is being done on this. So I think we first need to acknowledge that there’s a lot more being done than is being characterized.
Whether or not we need more—remember, the fourth bullet point in the White House Accord on Superintelligence was, “This is just a start. We will reconvene this group routinely. We will have to evolve the framework as the technology evolves.”
I feel good about that. I’m not saying that we never need to formalize or codify any of this stuff into law, but we did the right thing. We moved quickly, and we got a good regulatory framework in place.
What if we don’t know exactly what we’re supposed to be preparing for, or essentially putting guardrails around? It goes to what Bill Gates was talking about: that AI is going to be like an evolutionary event, that it’s going to grow insanely smarter at an exponential rate in the years ahead, and we’re not going to be able, as humans, to keep up with it.
So if we can’t keep up with it as humans, how do we know what its power ultimately is going to be and how to put those guardrails around it?
I think it’s really important that we have a pro-human approach to how we bring it into existence, right? We train the model. This is a piece of software at the end of the day. It’s not a new species that we’re bringing into being; it’s a piece of software that we are training.
So how we choose to train it—the frameworks—we want those disclosed. We want those audited. We want those reviewed. Those are all the things that these three legs of the stool are meant to address.
But I will also say this: It’s going to have to be the case that intelligence is going to have to be the one to oversee intelligence, right? The people who are in the best position to monitor this are the cyber capabilities out of the models themselves, because they’re going to be so capable.
To me, of course we have to be very thoughtful about this. We all have to get together and evolve this. But I don’t think there’s a group of humans that has the Platonic right truth as to how we set this up. Do I think there’s an FDA for AI that can give us the Ten Commandments on how we run it? No. I think that’ll set us back.
Remember, there is no safety off the frontier. If we end up in second place—if China gets there first—I can guarantee we’re not safe.
Okay. I want to go to that, because of what Alexandr Wang said, the head of Meta AI and the designer of Muse, who said recently, quote:
“I personally changed my own opinion on this matter,” talking about China. “In the past, I was very worried about China. I had read these old speeches that were given by PRC officials in the past that talked about how they viewed AI as this opportunity to leapfrog the West.”
That’s what this whole discussion is about. “But as time’s gone on, I think we’re painting them to be too much of a boogeyman when they’re going to see the same things that we see in terms of this being powerful technology. There are going to be a lot of shared goals in making sure that AI goes well.”
I think it’s well said by Alex, but I would say this: Trust but verify. Am I just going to trust that they’re going to take care of all of this for us? No, I’m not going to trust that.
But it is encouraging. Xi had his second meeting—I was at the Xi state dinner a few weeks ago—his second face-to-face meeting with the president this year. We’ve never had the presidents of the two countries meet more than once a year. Now they’ve met twice, and they’re committed to meeting another two times.
There was a lot of conversation about AI, not only between President Xi and President Trump, but also between Bessent and his counterpart in China. So I think there’s a lot of important conversation and coordination going on. I’m hopeful. I think some of the things Alex says may in fact be right, but I think we have to verify as well as trust.
All right. Three years ago, our next guest joined Brad and me as the CEO of Neeva, a company that was bought by Snowflake. Sridhar Ramaswamy was then named Snowflake’s CEO, and he is back with us now. I want you to listen to what you told us then, and then we can react to it after we listen. Pay attention to this.
Everything, as I said, from email to how you buy products, to how you approve things, to how you analyze what’s going on with your business, is going to be changed in a big way. So there’s a huge market for brand-new applications that are going to be built on top of the things that already exist.
Okay. You were obviously correct. We’ve taken leaps further, haven’t we?
7. The Agentic Enterprise Arrives
That’s right. I think pretty much everyone underestimated what we could do with AI. AI went from, “Here’s a cute chatbot,” to, “Oh, cool, I can attach it to the real world and get information about it,” to, “Oh, wait, this can actually think through complicated things and take actions on our behalf.”
That’s what we call the agentic enterprise in the world of enterprises. That’s the pull that you’re seeing with Snowflake’s success: how we did last quarter, what we guided to for this quarter, and how much more work we all have to do. It’s combining intelligence and data to drive how we work.
Do you still think in some ways we're underestimating the power of what artificial intelligence—superintelligence, as some are calling it now—presents?
Absolutely. I think it's very early. All of us, over the past 50 years, have navigated this world of information painfully. I joke to people that most of my life, for the past 50 years, I've pretty much been copying text from one screen into another screen. That's my life. That's kind of all our lives.
I think AI can distill all of that away and focus us on what's the thing that really matters. What's the judgment that you need to bring to the table? I think we are very, very early in that journey.
The software trade has been so volatile, as you know. The IGV is way off of its lows, and your stock has had an amazing year. Do you feel like we're getting better at understanding the winners from the losers and how the winners are actually going to win?
I think it's early. There's going to be a lot of impact on all kinds of things. But all of us also need to internalize that the act of software creation is just fundamentally different.
The example that I give to people is that software has become industrialized, and we are still trying to figure out exactly what that means. Software has gone from one dude in a garage, or five people on a team doing this slowly and painfully, crafting it, to being something you can mass-produce.
We think areas like data platforms become more powerful in a world like this, because storing, accessing, and acting on high-quality information ends up having such a big impact. But it's also a little bit like what happened to media. There's going to be lots and lots of software, and there are going to be fewer players that are going to be very effective at delivering it, both to consumers as well as enterprises.
You no longer own it through your hedge fund, but you have a lot of exposure personally to the stock still.
It's one of my biggest personal holdings.
Great job this year. The stock's up 50%.
Yeah. But to your point, Scott, part of the reason that I continued to own it in a big way was this difference. You talk about IGV as though all software is created equal. The fact of the matter is, talk to us a little bit about what makes the data platform so different from application software, perhaps, and what you see happening with your data platform that gives you so much excitement about the future.
First of all, what needed reams and reams of software purchases, installation, and integration has disappeared because you can now reason with data in English. In fact, one of my data teams that doesn't write a lot of software—they always just worked on data and set up systems—has had a huge impact on my company, the sales team, and the finance team, because they don't know any better.
That's what I mean by industrialization. You can write a sequence of English and have the underlying systems take actions on your behalf. That's the power of data platforms. The distance between what humans want to do and data has collapsed, and it's also collapsed the distance between how people implement software—the system integrators—and the data platforms, because you can do that a whole lot faster.
It's just made the aperture of what's possible with Snowflake so much bigger.
Josh, I'd love to get you and Steph involved, as you both own the stock and I believe are still bullish on it. Josh, you first.
Yes, I am bullish, and I love this awakening that's taken place this year, probably mostly in the last 3 months, where enterprises can run AI workflows directly in the data. We're not talking about layers on top of it anymore, and I think that's really helped the share price.
I'd love to hear your take on this. If a CIO were making a decision right now about where to put their agents, why would they choose Snowflake over Databricks? I know there's been a convergence between the lake, the warehouse, what you guys do, and what they do. Talk about how that decision gets made and why that should be a catalyst for Snowflake going forward.
We've always been the enterprise-grade platform where people have put their most valuable data. The big differentiator for us has been that our product is super tightly integrated. It's not a bunch of do-it-yourself pieces that you have to assemble.
If you create an agent, for example, it automatically obeys all the access restrictions on the data that's underneath. We make it super easy for you to get value from these agents. We're also enterprise-grade ready, whether it is permissions or things like disaster recovery, which a lot of our customers really care about.
It's the time to value and simplicity where we stand out. Customers like Craft went from perhaps getting competitive intelligence every couple of weeks to getting it in real time. There are tons and tons of examples like this of how quickly people get value from the data that they have brought into Snowflake.
How about you, Steph?
Sridhar, thanks for joining us today. I'm a long-term shareholder, very happy, so thank you.
You have done some M&A—Neeva and Observe—to increase your technology talent but also to expand your platform capabilities. How do you balance organic R&D against targeted acquisitions as you fill in the gaps in the agentic stack?
This is a great question. Brad and I were just talking. AI has made it possible for us to create software a whole lot faster. I would say the bar for big acquisitions has gone up in a big way because the messiness of things like actually getting the merger done and getting the teams to work together is extraordinarily high.
We continue to make strategic acquisitions. We acquired a company called Neeva, which brings this technology called MCP, which basically connects our agents to all the applications that are sitting within an enterprise.
We will be thoughtful about these acquisitions. They're more likely to be small rather than big, so that they're not distracting. But it's not an either-or. We are investing heavily in how we create new aspects of data platforms faster in-house, but we also acquire companies. I want young, hungry CEOs who want to go make a big difference in the world.
You have news today. Before I let you go, this partnership expansion with Lumen Technologies will build a private, secure lane that lets enterprises move data into Snowflake's platform without touching the public internet. That sounds to me like a safety and security conversation.
It's also an ease-of-use conversation because Lumen works with many, many companies on the planet. They power their infrastructure—their basic tech infrastructure of all kinds.
But getting highly capable agents and observability on top of the data is a pain. We want to make it super easy to bring that information over into Snowflake. We are very excited for this partnership.
All right. We appreciate your time. We're excited that you're back with us after more than 3 years as we take a pulse check on where we are in AI in America. Sridhar, thanks.
Thank you, Scott.
All right, halftime live from One Market in San Francisco. Back right after this.
At 3:00 Eastern this afternoon, Closing Bell, our special continues on AI in America. Nikesh Aurora, the Palo Alto CEO, joins us. Misha Asin, the Reflection AI CEO, will join us as well. Their new open-source model—that's a huge story. The Nobel economist on the future of jobs will join us today, and tech investor Jeff Richards. So, we've got a stacked hand later on. We look forward to that as well.
You have a final trade. I'll go to you first. What do you got?
Yeah, I like it. I'm going to go unorthodox.
There's nothing unorthodox here.
If we have an Anthropic IPO this year, that's my trade. It's non-consensus. There's a lot of chatter, a lot of negative chatter, coming out of yesterday.
I think when we look back 10 years from now, the Google and the Amazon of this era have now become more clear to me. I think Anthropic will be one of the enduring and great companies in the country.
I know you keep saying “if we do.” Do you expect that we will?
Yes, very much so.
When do you think—what's your best guess in terms of when they're going to go public?
I think we're talking weeks, not months.
Oh, okay. It's been great having you here.
Yeah. Thanks for having me.
Thanks. It's been awesome.