Is The AI Bubble About to Pop?
CeterisJasonTommy ShaughnessyJoseKevin
- Tommy Shaughnessy's “token panic” thesis: enterprises are blowing through AI budgets and may shift most low-budget workloads to open source. Salesforce reportedly spent $300M on Anthropic in Q1 alone, others burned full-year budgets in a quarter, and Microsoft—despite owning 27% of OpenAI—is releasing DeepSeek models for Copilot. DeepSeek is “20 to 40x cheaper,” GLM-5.2 is roughly a fifth the cost, and privacy/data-retention boxes can now be checked through offerings such as Venice and OpenRouter. Tommy is bullish on AGI but concerned about near-term disruption from enterprises shifting off expensive cloud models.
- The frontier-side counter is that frontier intelligence is orders of magnitude more valuable, and there is no obvious way to bet on open source. The $5M startup analogy contrasts hiring top-tier engineers with spending it all on Karpathy; the arms-race dynamic means frontier models stay expensive, “justifiably so.” Open source may have more users and inferences, like Android, but the value of those tasks may be lower and the models may remain four months to a year behind.
- The proposed convergence is a rough 10/90 frontier/open-source split with the whole pie growing—and frontier labs may be undercharging. Nobody on the call thinks Anthropic or OpenAI are at peak revenue. Jose spends up to roughly $1,500/month on frontier models for a 2–3× efficiency gain and would pay 10× more, while penetration remains low: many people do not know how to use these systems, and many corporations still use Copilot.
- On the bubble, Jose says bubbles can last a long time, while the leaders look different from 2000. Shiller CAPE is around 42 versus roughly 44 during the dot-com boom; Cisco traded around 120× earnings, Oracle around 150×, and Sun Microsystems around 10× sales, while Nvidia is around 20× forward earnings with 47% free-cash-flow margins, Meta around 18×, and Google a little over 20×. The memory trade is defensible if supply shortages persist.
- The pop risks are named specifically: circular financing, leverage, regulation, and the 2026 midterms. Kevin flags open-source substitution as a possible local-top trigger and is also concerned about Google raising money and investing in Anthropic while Anthropic uses that growth financing. Another speaker says debt and off-balance-sheet financing are the historical bubble-poppers. Jose notes that AI is extremely unpopular in the U.S. and that roughly 25% of data-center projects are blocked. The host frames the midterms as a binary risk, while Ceteris compares it to the 2024 Gensler trade; Tommy pushes back that trading around political control has not historically been profitable.
- Jose’s Carlota Pérez read: this bubble may inflate much further before it breaks, and his own FOMO is a local-top signal. Roughly $700B of annual capex eventually needs trillions in revenue, but chips are worth more three years later than they were when a two-year depreciation period was expected, rental prices are up, and AI still looks underestimated. He expects 10%–15% drawdowns but thinks the direction of travel is up.
- Strategy is described as a major unforced error. STRC is around $92; an 11.5% dividend on roughly $10B outstanding implies about $1.1B–$1.15B annually against a reserve near $1.1B. A speaker says mNAV could be driven toward 0.5 and that selling $3B–$5B in the market would have been cleaner. Bitcoin is described as heavily tied to Saylor, with the market needing resolution there.
- In crypto, HYPE and Zcash carry the discussion. Jason calls HYPE compelling on pure flows—smaller-than-expected unlocks, Assistance Fund and ETF buying, a $1B credit facility, and Circle USDC revenue not yet reaching buybacks—and says hacks are a bigger risk than regulation. Jose made Zcash his biggest crypto position, viewing the Orchard bug as a JELLY moment to buy; Ironwood is targeted for the end of July as a formally verified replacement pool.
1. Jason’s macro: markets are primed to run into midterms
- Jason’s non-quant view is that “there’s a lot of incentive for the markets to go up between now and midterms,” anchored by new Fed chair Warsh, who has described inflation as a choice by the Fed. If forced to choose, Jason thinks the Fed would choose higher inflation over raising rates in the immediate term. Add “the 70th Iran deal we’ve had this year,” which he thinks has more legs than prior attempts, and he is staying risk-on, reassessing around the end of Q3.
- His IPO tell: SpaceX initially had a “liquidity-sucking black-hole effect,” but the Anthropic and OpenAI IPOs expected in the next couple of months may be more telling. His refrain remains: “as long as you don’t really own Bitcoin or most of crypto, you’re probably feeling really good.”
2. Token panic: enterprises are blowing through budgets, and open source just checked its boxes
- Tommy’s viral tweet started with a Delphi colleague repeatedly hitting the ceiling on a Claude business account. That led Tommy to question whether enterprise AI spending is sustainable. Benioff was reported to have spent $300M on Anthropic in Q1, more than 5% of Salesforce’s roughly $5B–$5.5B annual salary bill, while other companies reportedly burned their full annual AI budgets in one quarter. Tommy’s conclusion is that businesses may not have the capital or expense budget to keep paying for this.
- Microsoft, despite owning 27% of OpenAI, is releasing DeepSeek models for Copilot, potentially reducing revenue flowing to OpenAI. Tommy remains bullish on AGI but is concerned about near-term disruption as enterprises shift away from expensive cloud models.
- His nuance is that open source has improved along two difficult dimensions. Privacy and data-retention concerns can now be addressed through private models and private data centers, including Venice and OpenRouter offerings. Cost is also far lower: DeepSeek is “20 to 40 times cheaper,” while GLM-5.2 is roughly one-fifth the cost. Tommy says GLM-5.2 is below Fable and 4.8, roughly in line with 4.8, and above 5.5 and ChatGPT 5.5 on some metrics and Arena scores. The substitution offers similar or near-similar intelligence for 90% less, or 90 times more usage.
3. The frontier-side counter: the frontier commands the value
- The counter is that frontier intelligence can be orders of magnitude more valuable. The startup analogy asks whether, with a $5M budget, one would hire top-tier engineers for roughly $500,000 or spend the full amount on Karpathy. The point is not that the choice is literally binary, but that the value of the top 1% or 0.1% varies enormously by field.
- Programming is a clear example because compute is the substrate of the modern economy, arguably the equivalent of oil. Legal work, investment, and scientific discovery are other fields where the value of better intelligence can be enormous or uncapped.
- Frontier models also face an arms race: competitors will use the most expensive models, creating pressure to do the same. Open source may resemble Android—more users, more tasks, perhaps more inferences—but the tasks may be less valuable. There is no obvious way to bet on the trend because open-source inference may be served by low-margin neoclouds or run locally.
- The speaker is surprised by how well open source has performed despite China’s chip constraints. The prevailing explanation is distillation through thousands of calls to frontier models, though he says he does not know enough to judge it. His base case is that open source remains four months to a year behind and therefore is not competitive for most frontier applications.
4. The convergence: 10/90, undercharging, and low penetration
- The proposed mental model is that roughly 10% of requests should go to frontier models and 90% to open source. This is not a bearish view on frontier models: the overall pie can grow dramatically while open source takes a larger share of low-budget, non-frontier workloads. The frontier models may be “undercharging by a huge degree” for the highest-value use cases.
- Jose’s personal experience is evidence of the consumer surplus. He runs essentially all his work through frontier systems, spends up to roughly $1,500 per month between subscriptions and API usage, and sees a 2–3× efficiency gain. He would readily pay 10× more.
- Penetration remains the larger story. Many people still do not know how to use these systems, and many corporations are still using Copilot, which Jose considers poor. As capabilities continue to improve, users psychologically want the newest frontier model for everything. Once capability growth slows, cost optimization, routing, and orchestration will matter more.
- Perplexity-style orchestration could become a major business, routing requests to frontier models, open-source models, or specialized fine-tunes depending on the task. The labs will also compete in this market by offering cheaper earlier versions of their models.
5. CAPE at 42, but the leaders look different from 2000
- Ceteris says he has no idea whether there is a bubble and continues to own the indexes. He does not want to time the macro top because the upside risk is that this is a new paradigm and investors never get a chance to buy back in.
- Jose says the market is definitely in a bubble, but bubbles can last a very long time and this one is supported by real fundamentals. Shiller CAPE is around 42, versus roughly 44 during the dot-com boom and around 27 in 2007.
- The leaders look different from 2000. Cisco traded around 120× earnings, Oracle around 150×, and Sun Microsystems around 10× sales. Nvidia is now around 20× forward earnings with 47% free-cash-flow margins, Meta around 18×, and Google a little over 20×. Jose considers these companies extremely cheap by comparison.
- The memory trade is also discussed. SK Hynix is around 6× forward earnings, Micron around 10×–11×, and SanDisk around 25× after rising roughly 5× this year. The discussion notes that the companies trade below 7× forward revenue while supply shortages may persist for at least another one to two years.
- SanDisk’s RSI reached 99.13 after a roughly 3,000% twelve-month run. The technical-analysis point is that RSI is often a confirmation of momentum rather than a topping signal: SanDisk’s weekly RSI breached 70 in September 2025, when the stock was around $50–$60, and momentum continued.
6. What is genuinely different: issuance, hated data centers, circular money, and debt
- The issuance discussion identifies a major difference from prior bubbles: not only the SpaceX, Anthropic, and OpenAI IPOs, but also large equity raises by companies such as Google. Google’s reported $80B raise was described as larger than SpaceX’s. The comparison is to crypto’s low-float, high-FDV projects, where some assets can still perform even while broader supply weighs on the market.
- Kevin mentions a tongue-in-cheek post comparing SpaceX to the Trump memecoin and Anthropic and OpenAI to the Melania memecoin, implying that successive offerings could drain liquidity from existing assets.
- Jose’s main risk is regulation. He cites the view that AI is extremely unpopular in the U.S. and says roughly 25% of data-center projects are currently blocked. The room contrasts wanting more data centers and nuclear plants with not wanting to live next to one. Building in remote locations is difficult because power and grid infrastructure must also be brought there.
- Kevin identifies two concerns. The “low-IQ” concern is that cheap, capable open-source models could trigger a local top by making substitution easier than expected. The other is circular financing: Google raises money and invests in Anthropic, while Anthropic uses that financing to subsidize growth and complete another round.
- Another speaker identifies leverage as the historical bubble-popper. The build-out is shifting from hyperscaler cash flow toward off-balance-sheet and debt financing, including private-credit structures such as the Apollo and Blackstone deal for Anthropic’s compute. This is not necessarily an immediate problem, but if projects are delayed, costs rise, or compute produces less ROI than expected, debt obligations could become a major headwind.
7. The 2026 midterms as a political risk
- The host argues that every pullback while Trump is in office “almost has to be a buy” because the administration has said it will buy equity in AI companies and backstop dips. A Democratic takeover amid anti-AI and anti-data-center sentiment would pull risks associated with 2028 and afterward into the immediate term; a Republican sweep or continued control would push those risks out roughly two years.
- Tommy pushes back that buying or selling based on which political party is in charge has never been a reliably profitable strategy. He cites crypto’s post-Gensler experience as a warning: the market did well for roughly three months and topped when Trump entered office.
- Tommy nevertheless says AI cannot be separated from politics or national security. The U.S. government cannot afford to lose the AI race, which is why he finds it difficult not to buy dips. He says he is holding his longs. Ceteris similarly says he prefers to stay long for the next few years.
8. Jose’s Carlota Pérez read: the bubble may inflate much further
- Asked for a historical analogy, Jose first flags his own emotions: he wishes he had bought more and is trying to talk himself into deploying cash. He considers that internal conversation a possible local-top signal.
- Even so, Jose thinks this bubble will inflate much further. The overlapping technological waves include AI, data centers, robotics, drones, and biotech. The fact that many people are already calling it a bubble may itself be the “wall of worry” needed for a continued advance.
- His warning signs would be a sharp inflection higher in hyperscaler and large-cap multiples combined with slowing revenue. Annual capex is discussed at roughly $700B, which would eventually require trillions of dollars in revenue to justify, while current revenue is not yet close to that scale.
- The operational evidence remains strong. Chips are worth more three years later despite earlier expectations of two-year depreciation, and rental prices are up. Jose sees the main froth in real-estate REITs and some data-center and infrastructure companies.
- His broader point is that current capabilities would have seemed astonishing in 2022: entire COBOL codebases rewritten in weeks, gene-editing work involving a dog, and Karpathy no longer writing code manually. Software costs may compress toward zero, enabling much more software creation. Jose still thinks AI is underestimated. He expects 10%–15% drawdowns and greater volatility but believes the direction of travel is up.
9. Strategy is a “really big unforced error”
- A speaker says STRC is around $92 and questions whether it can return to $100. With an 11.5% dividend yield and roughly $10B outstanding, annual dividends are about $1.1B–$1.15B against a reserve near $1.1B—roughly one year of coverage, while investors would prefer at least two years. Raising only about $100M per ATM offering would take many offerings to rebuild the reserve.
- The speaker says mNAV could be driven down as far as 0.5, compared with roughly 1.17 on enterprise value and below 1 on circulating shares. A cleaner strategy, in his view, would have been to sell $3B–$5B in the market and continue buying programmatically.
- The situation is described as messy and a major unforced error. Earlier, Saylor could be defended to non-crypto audiences; now, the growing liability and repeated fundraising make that defense harder.
- Another participant says Bitcoin is difficult because it is currently so tied to Saylor and that the market needs resolution there. He is more bullish on some other crypto assets.
10. HYPE: flows over fundamentals, with hacks as the bigger risk
- Jason deliberately sets fundamental valuation aside, saying HYPE is clearly overvalued relative to current fees, and focuses instead on price drivers: aggressive buying, selling pressure, and available supply.
- He sees limited supply because unlocks are much smaller than previously expected. The Assistance Fund buys millions of dollars of HYPE each week, ETFs buy millions or tens of millions weekly, and he says the market-cap-adjusted buying impact of the HYPE ETF has exceeded that of the Bitcoin ETFs since inception.
- Hyperliquid DATs are raising cash while also acquiring HYPE. There is a $1B credit facility, and the Circle USDC deal and associated revenue have not yet reached buybacks. Jason calls HYPE the most compelling crypto-specific investment he has seen in years.
- Pre-IPO perps are not a major direct fee driver, but Jason views them as a “Trojan horse” that gets TradFi aware of Hyperliquid. SpaceX trading has prompted TradFi contacts to ask about HYPE, creating attention that can later monetize through perps, HIP-3 markets, ETFs, or other products.
- Tommy argues that regulation should not permanently block access to the venue with the most open interest, volume, and usage. Jason says the larger risk is a hack or a North Korean attack rather than regulation. Ceteris cautions that no-KYC perp platforms have historically been allowed to scale for a period and then faced regulatory pressure.
- Another participant expects meaningful clarity within six months, whether through KYC-enabled U.S. access or a decision not to pursue it. The potential trade-off is losing retail traders while gaining substantial institutional flow. The discussion compares HYPE with Binance and notes Binance’s user losses and the reported rejection of its Greece MiCA application.
- One speaker says HYPE has clear product-market fit but struggles to see it reaching $1,000 because the no-KYC market may be capped and the narrative is already hot. The speaker also calls Jeff’s Colossus profile overdramatic. In response, another participant jokes, “Corporations are the equivalent of curing cancer. What are you talking about?”
11. Zcash: the Orchard bug was the JELLY moment
- Jose says Zcash has become his biggest crypto position. He viewed the sell-off after the Orchard bug as a buying opportunity, comparable to the JELLY moment for Hyperliquid. The thesis does not require an immaculate, never-hacked system; it requires prompt, competent handling of bugs and evidence that the issue is unlikely to recur.
- The immediate evidence suggested the bug had not been meaningfully exploited, and the team acted quickly. Mythos audited the system, and another speaker says the audit came back clean. Jose sees the private-Bitcoin store-of-value narrative as small relative to its potential market, with additional developer talent moving into Zcash.
- Ironwood is targeted for the end of July. It is the new Orchard pool with the bug fixed, formal verification, and audits by Mythos and Open AI. Orchard will become withdraw-only. The expectation is that once several million Zcash move from Orchard to Ironwood, confidence will improve.
- The speaker discussing the mechanics says Orchard should never reach exactly zero because some users will inevitably lose keys. If it does reach zero, that would be evidence that someone likely exploited the pool. Ironwood could become one of the most battle-tested shielded pools, though formal verification and audits do not reduce bug risk to zero. Running two proof systems is mentioned as a further, potentially costly defense.
- Sentiment recovered with price, but the sell-off showed that many holders did not understand the privacy trade-offs. The shielded pool had not grown much since October, suggesting that the marginal buyers were momentum traders rather than long-term privacy users.
- The Zcash holder says he retained the position without adding. A move back above pre-exploit levels could quickly revive the narrative; the one-year chart was described as a possible inverse head-and-shoulders pattern. The speaker recalls waking to prices around 350 and then seeing them fall toward 250, calling the move extraordinary.
Full transcript
All right, welcome back to the Delphi Hive Mind podcast. The show about markets, crypto, AI, and unfiltered opinions. We have some of the usual suspects today joining me. We've got Jose, Sattarous, and Jason returning. We also have a very special guest, our other co-founder of Delphi, one of our partners on the Delphi venture side, Tommy Shaughnessy. Tommy, great to have you today. There's a lot of AI talk we're going to get into, and that's why I wanted to bring you into this conversation, because I know you've got some great thoughts. You've been deep in this world.
Before we get there, let's start with the usual market check. How are we feeling? What's everybody looking at? Bitcoin's kind of hovering around the $65K range. It seems like we got a little glimpse of potentially some bull-market action, and that obviously got stomped out. It hasn't really broken out of a range. We've got a new Fed chair and an FOMC meeting later today. By the time this comes out, we'll have a decision, so we can talk a little bit about that. But there's also the SpaceX IPO and a ton of stuff going on. Jason, maybe I'll kick it to you to start. How are we feeling? What are you looking at? What's exciting?
1. Open Source vs Frontier Models
I mean, as long as you don't really own Bitcoin or most of crypto, you're probably feeling really good, right? That's the same thing we've said every Hive Mind for the last 4 or 5 months at this point. My very non-quant view of the market is that there's a lot of incentive for the markets to go up between now and midterms and the end of the year.
When you look at Warsh, his view on inflation—what he's written in past essays or talked about in interviews—is that he views inflation as a choice by the Fed. So I think if push comes to shove, they will choose higher inflation over other things, over raising rates, at least in the immediate term. Then, obviously, with the 70th Iran deal that we've had this year, this one seems to have maybe a little more legs than the prior ones. When you add all of those things together, to me it seems like markets are being primed to run it up into midterms.
What happens after is up for debate, obviously, depending on which party wins, how big they win, how much control they have of the different branches, et cetera. When I look at things, I'm just like, I don't know. I've got a bunch of risk, and I'm not planning on taking any risk off anytime soon. Probably as we get closer to the end of Q3 is when I start to reassess that. Obviously, if anything else happens in the interim, you've got to reassess, but I'm very comfortable with the risk I have.
Most people know that risk is mostly HYPE and HYPE alternatives or HYPE derivatives, but equities look great. I mean, most of them look great. The SpaceX IPO is obviously something to watch. Clearly, you had a little bit of a liquidity-sucking black-hole effect initially. I'm curious to see how the Anthropic and OpenAI IPOs go in the next couple of months. I think that will be way more telling than this one.
I think the conditions are good enough for things to continue going higher now, and I'm bullish into essentially Q4. Depending on what inflation prints come in at, depending on whether there's any derailment of this deal they're working on and oil spikes—stuff like that, continued oil shocks—I think that could start to weigh. But I don't think it's going to weigh in the short term. I think the market will look through it until midterms. So, yeah, I'm just staying risk-on. I'm not really eager to add a ton of risk, but I'm very happy to hold everything I have.
How's everybody else feeling? This can be a broader market or stock-market conversation, too. There are arguments that equities have absolutely ripped and that some of these AI sectors have had massive outperformance. There are also arguments that this run still has a lot of legs and that we're only at the beginning of this cycle. I'm curious to hear what you guys think.
2. OpenAI, Anthropic & Enterprise AI Spending
Yeah, I think it'd be nice to get Tommy's take, maybe. He had this viral tweet a bit ago fearmongering about the end of token maxing and how this could lead to a crash in the AI trade. Maybe you can introduce that, and then I can retort.
My bad, sir.
I guess the genesis for the tweet was that we have an excellent colleague of ours, Rossin, who works on the venture side, and we got him a Claude business account. He kept hitting the ceiling on that business account, and it was starting to annoy me because I just wanted to get him a seat subscription. But we needed the protections on the data-privacy side to appease our operations team and get the users that he wanted.
It got me thinking about how much enterprises are spending with AI and whether it's sustainable. When you switch from a subscription that you know to an enterprise API, generally the seat of the subscription side—
Salesforce spent like $300 million, right, in the first quarter? I think I heard that recently, like Benioff said he spent $300 million on Anthropic in the first quarter, which is kind of wild. There were a bunch of others saying they spent their entire annual budget in the first quarter.
I mean, Uber could pull that off, right? Then everyone just came out of the woodwork and raised their hand: “By the way, us too. We're also blowing through this.”
Yeah. So everyone's been blowing through their subscriptions or API spend because I think it's just too expensive. Businesses are realizing that they just don't have the capital or the expense budget to pay for this stuff, and they're going to look for alternatives. People are already doing that.
You guys mentioned Salesforce. There's Uber, there's Microsoft. Actually, Microsoft, despite owning 27% of OpenAI, is releasing DeepSeek models to use for Copilot, which obviously reduces the revenue flow to OpenAI itself. So, yeah, I'm happy to take this any way you guys want. I'm bullish on AGI, but I think the near-term disruption from enterprises shifting off expensive cloud models is my concern.
Yeah, but you also have a separate take about open source, right? Basically, you think this is going to shift to open source. Open-source models are getting really good. I think GLM-5.2 is probably the best open-source model so far, and arguably—I haven't seen the latest Epoch benchmarks—it would be the closest one to state of the art in a while because the gap has actually been widening.
So your thesis is also that a lot of the spend will shift to people using open-source models, right?
Well, yeah. It's hard. My take's a little nuanced. I think the open-source models are good along 2 spectrums that have previously been really hard.
The first one is that if you're a business or enterprise, you can't use open source because you have data-retention and privacy issues, and your CCO, CLO, or operations person just won't let you. Now that's being checked. Venice has private models on private data centers. They have end-to-end encrypted models. OpenRouter offers private models. You can check that box.
The other side is that it's ridiculously cheap. DeepSeek is 20 to 40 times cheaper. GLM-5.2, I actually don't think, is that much cheaper. I think it's a fifth the cost, when usually we talk about things being 10 to 99 times cheaper. So generally, I think the cost is way less and the privacy guarantees are there. It sort of makes sense for them to switch over.
Yeah.
Sorry, go ahead, Kevin.
No, I was just going to say I think the framing—because I'm glad we just got right into this conversation, because I think it's going to be one of the big ones today—and the framing I had, and the question I wanted to pose to all of you, especially Tommy, since you brought him on, is that it seems like there's been this back-and-forth with open source and closed source for years now, right?
But it kind of seems like, at this point, earlier on—because this was pre-agents really taking off—there was this idea that there was a huge consumer surplus with some of these big AI model companies and inference providers, right? Essentially, for $20 a month or even eventually the $200-a-month plans, you were getting so much power from that. With the early use cases, there was a lot of value the consumer was getting, but the model companies themselves and the big labs weren't necessarily capturing that.
Now that's starting to transition and shift to more API-focused pricing and more usage-based pricing. These companies are now some of the largest companies on the planet, and both of the 2 big major labs are now going toward an IPO. So the timing makes a lot of sense, because they need to be generating hundreds of billions in revenue eventually to justify the valuations that they're going to go to market with.
Now, finally, that cost aspect is starting to hit, where you go from token maxing to token panic, in a sense. My question—and the overall framing I want to have the conversation around—is: Is this kind of open-source AI's moment? If it is, how long or how big is that window?
It seems like this is the best opportunity to put open source front and center, or the best opportunity I've seen over the last few years to make the argument for open source. But I'm curious to hear your guys' take.
Tell me, do you want to make that argument first?
He’s just teeing up to knock me out of the argument.
Because he’s like, “I’m going to go ahead and take the other side.”
Yeah, I’ll take the other side. I’m taking the other side, I think.
My thesis is pretty simple. I think these APIs for frontier intelligence are extremely expensive. It’s hard to know what expensive means because we’re not at these companies, but what we’re seeing is companies switching off. You’re seeing Uber switch off, you’re seeing Microsoft switch off, you’re seeing Salesforce switch off.
People want frontier intelligence; there’s no doubt. They like the closed-source models because they have the best models, but at the end of the day, it’s just too expensive. When you have a substitution—open-source models that are really good—it makes sense to switch over.
I mean, GLM-5.2, which they mentioned, is below Fable, obviously. It’s below 4.8. I think it’s in line with 4.8. It’s above 5.5 and ChatGPT 5.5 on a couple of different metrics and Arena scores.
When you have a substitution that’s so cheap and offers you similar privacy guarantees, or data-privacy guarantees, that allow your company to switch over, it makes sense to do that because you’re getting the same intelligence, or close to the same intelligence, for 90% less. Or, looking at it the other way, you can get 90 times more usage.
Are you saying they’re expensive relative to the value they create or relative to alternatives? Are you saying the spend that people are having on them isn’t justifying the value they’re creating, or that they could be even more efficient if they used open-source models?
I’m looking at it strictly from a spend perspective—the second one.
I think the frontier is always going to be way more expensive and create way more economic value. Let’s say you have a $5 million budget for your startup. Do you want to spend, let’s say, $500,000 and get some top-tier engineers, or do you want to spend $5 million on Karpathy? Which one would you prefer for your startup? Which one do you think leads to more success?
Obviously, $500,000 is a very good salary for an engineer. You’re going to hire some pretty good people. I just don’t know if it’s that black and white, though. It’s definitely not black and white; it’s an analogy.
But I think, in general, frontier intelligence is orders of magnitude more valuable, especially as the intelligence goes up more and more. Obviously, it’s not like that for all professions. The delta of value that the top 1% creates varies based on what you’re doing, and the same is true of the top 0.1%.
I think programming is the obvious one, where compute is the substrate of everything in our world. It’s arguably the equivalent of oil, so writing better code is direct leverage on that. I think that’s a very clear one.
I think legal is an obvious one, too. The best lawyers are paid tens of thousands of dollars an hour, and it’s for a reason. Investment is going to be one of those, obviously. Science and scientific discoveries are another one, where the value is arguably uncapped. If you find a cure for cancer, reverse aging, or do all these kinds of things, the value is enormous.
I just think the frontier is always going to be really expensive, and justifiably so. I don’t think it’s expensive in terms of the value it’s generating, at least from personal experience and from seeing some of the startups we work with—what they’re able to do with way less money.
Even in the Salesforce case, which is an extreme one, I’m sure a lot of that was wasted spend, with people running stupid loops and stuff. But they spend $5 billion or $5.5 billion a year on salaries. $300 million is 5%, a little over 5%, of that. Is that the right number?
I feel like, in the future, you’re going to be spending more on compute—inference—than you are on software engineers. I think it’s pretty inevitable, in my mind, that you’ll end up spending multiples more.
The other dynamic that I think is important is that it’s a competitive market. Most of the value is always at the frontier in terms of technology. All the value is created at the frontier, whether it’s frontier models or anything else.
Obviously, the frontier is always moving, but I do think it’s hypercompetitive. Your competitors are going to be using the most expensive models, so you have to as well. It’s kind of like an arms race. I think that will always drive up the price for frontier intelligence, and I think justifiably so.
I do think open source is going to be very widespread. It might be like Android: it has more users, it’s doing more tasks, and maybe even more inferences. I just think they’ll be way less valuable.
I also don’t think there’s an obvious way to bet on that trend right now, because a lot of open source is just going to be served either by one of these neoclouds at relatively low margins, or people will run it on local machines.
I’ve been surprised by how well open source has done, especially GLM-5.2, because it doesn’t seem like this should be the case, especially with China not getting access to these chips and so on. The narrative is that they’re just very good at distillation—that they’re very good at effectively copying these models through thousands of calls and doing that.
I don’t know enough about it. It seems like it should be easy to stop, but clearly it isn’t, unless people are just repeating a narrative and there’s actually something else happening.
I think open source is going to continue to be somewhere between 4 months and a year behind, and therefore not really competitive in the majority of frontier applications that require frontier intelligence.
That’s a good take. I agree that the value of frontier models is always going to be greater in terms of new science, new medicines, and new companies that they can create. If you’re dealing with frontier intelligence, that’s what you’re paying for.
My mental model is that 10% of requests should go to a frontier model and 90% should go to open source. I’m not bearish on frontier models. The part I’m nuanced-bearish about is that, at this point in time, I feel there’s a big shift where the majority of workloads could move to very cheap substitute open-source models.
You don’t need to run every request through Fable 5, ChatGPT 5.5, or 4.8, because, to your point, it’s mostly loops and things like that. I think the majority goes to open source, while the single-digit percentage to 10% of usage goes to frontier models. I honestly think the frontier models are undercharging for that by a huge degree.
So, do you think we’re at peak—
[Snorts]
—revenue for Anthropic and OpenAI over the next year or something like that? Is it down from here?
No, I don’t think so.
You think they’ll continue to grow revenue?
I think the whole pie continues to grow.
Okay.
I think we agree.
Well, again, I don’t think we’re trying to make it too binary. I agree that frontier intelligence grows. My point is that, for the short to medium term, I think it’s a very strong argument to shift the majority of low-budget, non-frontier workloads to the open-source providers that they’re currently running on the clouds.
Mhm.
The pie gets bigger—massively bigger. The market share that open source takes also gets bigger compared to where it is today, but that still leaves a huge TAM for frontier models. Both can win in this sense.
For your Salesforce example, what percentage do you think should be frontier versus run on open source? What do you think, dollar-wise?
I don’t know. It’s hard to say. It’s really hard to say at this point. I think no one has the answers to this.
Dollar-wise, if you’re developing or architecting a new system and trying to solve hard problems, it should always be the frontier. Once you’re implementing and running it, you can run it on open-source models.
I also think there’s going to be a massive business in being the orchestrator, kind of what Perplexity does really well, and to some extent what the harnesses do as well. It’s about being able to route your request in the most cost-efficient way to answer it.
In some cases, that will be open-source models. In some cases, it might be specific fine-tunes of open-source models. In some cases, it’ll be the frontier.
So I do think that’ll be the case, and I think there are a lot of people working on that, even in their own personal stacks. But I was curious: how much do you spend monthly on inference for the value you get out of it?
Oh God. I don’t know—a lot. Whatever Ventures approves. I don’t spend a ton on API usage. I think I just have subscriptions to everything. I have a subscription to Venice and all the apps.
Me too. I do all my work through AI. There’s not a single thing I don’t do through AI. I have the subscriptions, and between subscriptions and API usage, I’m probably spending $1,500 a month at most, which is way less than I would pay.
I’m not using any open-source models; I’m just using the frontier models. The value I get out of it has definitely been a 2–3× efficiency gain, especially on most of my mundane tasks. It’s harder to quantify on the investment side, but that just feels cheap to me.
It feels like if I could spend more to get more efficiency gains, I would. In fact, I’m trying to find ways to do that. That’s more of my problem right now than the opposite. I just think the penetration of these models is so low.
I’ve been trying to show as many of my friends as possible, and I’ve gotten a few of them into it. I think most people still have no idea how to use this stuff. Most corporations are still using Copilot, which isn’t even on this ranking. It’s terrible.
I think there is some overspending by people like Meta and Salesforce, for sure, but the penetration is also incredibly low. No one is using this stuff yet. That’s why I think we’re saying the same thing: I’m bullish that the pie continues to grow, which is also why I don’t think we’re in a stock market bubble.
I mean, we are in a bubble, but not in my bags.
What would you pay per month if it wasn’t subsidized? If you could only use APIs and had a maximum budget, what would you pay for what you’re getting today?
I’m not 100% sure that my spending is subsidized, or that I’m spending enough to qualify for a subsidy. I have to do the math on how many tokens I’m actually spending.
But I think a bigger question is the implicit subsidy: how much work you’ve been able to offload and how much more work you’ve been able to do by using these systems. That’s where the consumer surplus number has been really big.
Let’s say, to Tommy’s point, you were paying $10,000 a month for the current setup.
I’ll gladly pay that. You’d still pay it.
Yeah. And that’s where I think frontier models still have an advantage in terms of pricing power, and why I think revenues will increase. For that top 1%, or maybe even 5%, of use cases where frontier intelligence actually moves the needle in a massive way, they can charge more because of the value that power users and people in more technical fields and industries get from it.
There’s still a pretty big consumer surplus.
For sure, even though people are on APIs. Most people just don’t use that many tokens. Even I think most people have way less imagination for what to build than they think they do, including me. I’m spending $1,500 a month, and I’d easily pay 10× more.
And I do think that while capabilities are inflecting this much, it’s all about the frontier. Psychologically, you feel that, too. A new model comes out, and you only want to use the new model for everything.
3. Are We In An AI Bubble?
I think once capabilities start slowing down, people will start worrying a lot more about costs, optimization, orchestration, and what to shift to open source. That will definitely be interesting, but at that point the labs will play there, too. They’re not going to just drop that market.
If you want to use earlier versions of Claude or whatever, they’re going to be much cheaper. So it’s going to be interesting. I don’t think it’s an easy road for open source.
Maybe we can talk about markets generally—the AI bubble and how everyone feels about that. Things have obviously been going up for a long time, and I think a lot of people feel like there’s a bubble. I’m curious where Ceteris stands, since he hasn’t said much, and maybe Kevin and Jason. What do you guys think?
I have no idea if there’s a bubble or not. I still own the indexes. Obviously, there are a lot of differences versus past bubbles, given the speed at which earnings are going up and everything.
I’m not going to try to time the macro top. I’ve talked about this before: the upside risk is that this really is a new paradigm, and you’ll never get a chance to buy back in. I’m hoping some of that money starts flowing to crypto at some point, though.
I think it will.
Which bags? Which crypto bags?
Select assets.
Post that Bitcoin channel, Kevin.
I know how much you love that channel. You love those channels.
Bitcoin is tough because it’s just Saylor right now. He’s in such a tough spot, and I feel like the market needs some resolution there or something.
I honestly don’t know about Bitcoin now. There are definitely other assets I’m more bullish on than Bitcoin right now. STRC is down to $92 today. Is that thing going to get back to $100? I don’t know. It doesn’t seem like it is.
He’s going to keep using the ATM and doing half-Bitcoin buys and half-cash-balance buys. I don’t know what Michael Saylor’s strategy is now, but I wouldn’t be surprised if he runs mNAV down as much as he can—down to 0.5 or something. They have it at 1.17 based on enterprise value. I think it’s below 1 if you do it on circulating shares.
I wouldn’t be surprised if he just runs that thing down as much as he can, because it’s not like mNAV needs to be above 1. You see all these other treasury companies where they’re below 1. But if you’re holding Strategy right now, you’re essentially—
You’re a dick.
You are a dick.
Over and over.
I don’t know. Maybe there are enough people who hold it that he’ll be able to do this for a while, or maybe he has some new crazy thing up his sleeve. It’s just hard to tell. There are too many products.
Now it seems like the STRC game is potentially over, and it’s just a perpetual liability from the dividends. If he gets to a point where he has enough cash to fund the dividends—if he slams the ATM for a while and has enough cash to fund them for multiple years—then I guess there’s less of a spiral risk.
How many times would he have to raise cash at his current rate of about $100 million per offering?
This is why I’m long Zcash in crypto.
I’m still quite bullish on Zcash.
I’m still quite bullish on Zcash.
I think, to Jason’s point, though—
He has to do a lot of these things.
He has to do a lot of these things. Isn’t the annual dividend way—
The current dividend yield is 11.5%, and he’s got what, $10 billion outstanding? So that’s $1.1–$1.15 billion, I guess. His reserve is $1.1 billion now, so I think he has around a year. People would want to see that at least be 2 years.
Yeah, but if he’s only raising $100 million in cash with the ATM sales, that’s going to take quite a while.
Yeah, I still think it would have made more sense to just market-sell $3–$5 billion and then continue buying on schedule, programmatically. Now it’s just messy.
It’s messy.
It’s messy because a year or 2 ago, if you were talking to people in normal, non-crypto land about Bitcoin and they brought up Saylor as an issue, there were some things you could say. You could be like, “Yeah, I know. It’s fine.”
It was pretty much all equity, so—
Those conversations today—you can’t, in good conscience, be like, “Yeah, it’s all fine.” The conversations are totally different.
It's just a super messy situation and, in my opinion, a really big unforced error.
It definitely was an unforced error. He did not need to do all this.
In terms of the bubble question, we're definitely in a bubble, but bubbles can last forever, right? They can last a long, long time. If you read any financial history book about markets going back any period of time, bubbles last a while. They can last a long, long time, and this one just happens to be buoyed by actual fundamentals, too, for a lot of it. Yeah, I think—what are you going to say?
No, the Shiller CAPE is around 42 now, which I think is the highest it's been in 150 years other than the dot-com boom, which was around 44. For reference, I think 2007 was around 27, so it's definitely pretty high.
Most of that is not in my bags, to TL;DR. I think the Magnificent 7 and some of the core AI players are extremely cheap. When you think about 2000, the Shiller CAPE was the same, but when you think about the leaders, Cisco was trading at around 120× price-to-earnings, Oracle was at 150×, and Sun Microsystems was trading at around 10× sales. Nowadays, you've literally got Nvidia, which has 47% free-cash-flow margins, trading at around 20× forward earnings. Meta is trading at around 18×, and Google is trading at a little over 20×. These things all look extremely cheap to me.
Even the memory trade, which has run up around 5× this year and that people are losing their minds about, doesn't look that expensive, either. Most of them are trading at between, I think, 6× and 25× forward earnings. SK Hynix is around 6× forward price-to-earnings, Micron is 10× to 11×, and SanDisk is around 25×. Obviously, these are extremely high for these names cyclically, but I do think this time is definitely different for me in terms of the capex cycle and where we are with it.
Yeah.
I don't know.
I'm not trying to time—
This shows that, too. You look at revenue growth. I've got Micron, SanDisk, and SK Hynix as proxies, right? Revenue growth has obviously been exploding. There are huge supply shortages that realistically probably aren't going to be resolved because the capacity isn't there for at least another 1–2 years, at a minimum.
What's interesting is you've seen it on a forward basis. You're basically paying less than 7× forward revenues on these things. The implication is that if you think supply is going to come online, or you think prices for memory are just going to collapse after this recent rise, you could still make the argument that these are fairly valued in that environment. If you think supply isn't going to be solved and growth is going to continue, these wind up looking cheap, to Jose's point.
I mean, if you look at SanDisk, it is crazy. It hit an RSI of—what is it?—99.13, almost an overbought level for any stock in history. It's up around 3,000% or something crazy in the last 12 months.
Yeah. I don't know. I don't really feel like this time is different, but I'm not deep on the memory trade, like you say.
I don't know.
Here's what I will say quickly about the RSI stuff. I used to do a lot of this, especially technical analysis, in a prior life. Oftentimes, I think RSI gets misconstrued as a topping signal, whereas in reality it's a confirmation of momentum.
A good example, to Tommy's point, is SanDisk. If I put this on a log chart, it still looks exponential. Your RSI down here breached 70 on a weekly basis in September 2025, when this thing was trading at $50 to $60. It's just ridden momentum, momentum, momentum.
4. SpaceX IPO & Equity Issuance
All right, this time is different. I think one thing that's different with this period versus prior bubbles, though, is the amount of equity issuance that's coming. It's not just because of the 3 mega-IPOs with SpaceX, Anthropic, and ChatGPT. With SpaceX, we had the IPO, but there's still massive issuance when the float unlocks. You also have the Googles of the world doing these equity issuances now, which haven't been done in the past.
Yeah.
Google did an $80 billion one. It's more than SpaceX.
Yeah, so you're getting essentially more mega-IPOs with the Magnificent 7 doing their own equity issuances, right? This doesn't mean that the smaller fundamental things can't keep doing well. They certainly could. You've seen it in crypto over the past couple of years, where the low-float, high-FDV projects have gone down a lot and you've seen some things do well. Obviously, in crypto there's a smaller universe of actual things that can do well than in the stock market, but it's similar in that sense.
Even if you think this time is different, Jose, which I think there are a lot of reasons to say, what are some things that you would see that would shake your confidence that the bubble isn't inflating anymore? I think we can probably all agree that there is a bubble, and you just want to ride it as long as you can until you start to see some warning signs.
What are those warning signs to you? I was thinking about the SpaceX and Google issuances and Anthropic. Somebody made a funny post on Twitter: “Okay, so SpaceX is the Trump meme coin. Then Anthropic and OpenAI are just going to suck liquidity from everything after the fact, like the Melania meme coin did.” Obviously, that's tongue-in-cheek, but I do think those dynamics are still at play for the most part.
5. AI Regulation & Political Risk
Then, say you get a midterm election cycle and the Democrats win control. There's a huge amount of rhetoric around anti-AI and anti-data centers, and it's only going to pick up. A combination of tons of supply hitting the market and the market not being able to absorb it, possibly having to increase interest rates in early 2027 if the oil shock persists, on top of an unfavorable policy coming down from the administration—kind of like we saw with crypto during the Gensler era, to an extent. Do you think something like that could prick the bubble?
For sure. I think Dalio did a good job talking about some of the risks. Someone said he just shows up every 6 months to remind you he's sidelined and tell you China is going to win, but I think his post actually did a good job talking about some of the risks.
I do think one of the biggest ones is regulation. A lot of it has to do with how the CEOs have handled public relations, but AI is extremely hated in the US. No one wants data centers near them. I think something like 25% of projects are blocked right now.
I mean, it's hard to blame them. I wouldn't want a data center near me, either.
I agree. I want more data centers.
I want more data centers. I want more nuclear plants, too, but I don't want to live by a nuclear plant.
Right. That's the argument, right? If you agree that data centers can be good—and there's a lot of rhetoric that data centers are environmentally bad, along with a lot of misunderstanding around them—you want this, but you don't want it in your backyard.
The hard part is that everyone's saying, “Why don't you just build these in remote places in the US where there is no population? No one's around.” But the cost to do that, actually shuttle in power and energy, and get the infrastructure connected to some type of grid where there isn't a grid is a massive cost challenge. No wonder a bunch of these are popping up in places like Northern Virginia, where there's real infrastructure to be able to slide in and build. It's a hard problem.
We've got to designate one state that we all don't like to build all these things—the power plants and the data centers. Just one.
Let's alienate an entire listener base right now and just put up a state.
Yeah, I think that is a big risk.
Jose, I think one thing is that you're a huge Carlota Pérez fan. I know you read that book, and I think you interviewed her. You've looked at the history of bubbles intimately. There obviously is a really good place to add fuel to the fire and make a lot of money. When you're in it, earnings and everything look great. The metrics are there and they're reinforcing.
In hindsight, obviously, we realize there was a sell-off. Is there any historical analogy from past bubbles that you think relates to where we are right now with the asset?
Yeah. That’s a great question. First of all, I’d say that the way I feel emotionally right now about the market is a bubble signal to me: I wish I’d bought way more, right? I’m trying to talk myself into deploying my cash, which generally is a pretty top—at least local-top—signal. Then there’s that channel chart I sent you, Kevin.
I do think, in terms of a bubble, that we are in a bubble, and it’s inevitable that this will bubble up. Every new technology does, and I think this one has a lot of legs too. For me, the signals I’d be looking for are my view that the bubble will inflate much bigger than this. I think there are so many overlapping technological trends happening right now between AI, the data center build-out, robotics, the drone stuff, and biotech.
The things I’d be looking for are the multiples really inflecting on some of the hyperscalers and large caps, and revenue slowing down. Right now, I think we’re spending something like $700 billion a year in CapEx. I think this year or next year is set to be $700 billion. You do need trillions of dollars of revenue to justify that, at least over the long term, and we’re not even an order of magnitude close to that right now.
But what I’m seeing using these models—and that’s what I’ve had for the last few years—makes me so bullish on the ability for people to do way more with them than they’re doing right now. I don’t see that they’re overrated yet. I still think they’re underestimated. Even the fact that there are so many people calling for bubbles is kind of the wall of worry that you want to climb on the way up.
I do think there will be local maxima. Maybe this is one of them, and people will panic. I think it’ll be much more volatile. I think we’ll see 10%–15% drops over the next few months for sure. But in general, I think the direction of travel is up because it’s just hard to convey how bullish AI is. We’re all in it, so we don’t get the time to zoom out.
If you showed us in 2022 what Claude can do now, we would lose our minds, right? There are people rewriting entire codebases that were written in COBOL in a new language in a few weeks. There are people gene-editing their dog using Claude. Andrej Karpathy isn’t writing a single line of code anymore. There are just all these things that, when you’re in it, look and feel normal, but when you step back, they’re so meaningful.
I think the cost of software is going to compress to zero, which means you’re going to be able to write a lot more software. I think the whole CapEx cycle for data centers—even what you’re seeing with the price of chips—is still healthy. There was supposed to be 2-year depreciation, according to Michael Burry, but 3 years later, those chips are worth more than they were back then. The rental prices are up. All the signals to me are still pretty healthy, other than the froth in the REITs—the real estate REITs—and some of these other data-center-related and infrastructure companies.
That’s a good answer. I know how much time you spent with her and reading through that. My take is that I’m just as bullish as you are on AI. I think we get AGI, and I think people are vastly underusing the capabilities. If you gave someone Fable, they don’t even know how to use it the right way to get everything out of it. It’s extremely powerful.
The 2 things that scare me are a low-IQ take and maybe a higher-IQ take. I haven’t figured out the second one; I’m not smart enough to understand it. But the low-IQ one is that I think the substitution with open-source models could actually be a local-top trigger. I think we’re underestimating how easy it is to switch, how cheap these models are, and how good they are.
GLM-5.2 is really good. On any code arena, design arena, or anything, it’s second or third on the list, which is pretty nuts for a cheap price. The second thing that scares me is the circular financing: Google raising money and investing in Anthropic, with Anthropic using that to subsidize growth and do a round. That stuff scares me, and I’m not smart enough to dig into what it means.
I’m bullish that we’ll get AGI, but I think the sentiment and bubble indicators definitely feel local-toppy. I’m in your camp. I do think the direction of travel is up and to the right.
Yeah, I’ll build off that quickly because I think those are great points. The second one specifically is, to me, where the biggest potential risk—where we get a real, true topping pattern—comes from. It’s not just the concept of circular financing.
Up until, you know, whatever it was, 6 months ago, a lot of the AI infrastructure build-out was funded by big hyperscalers, or largely funded by big hyperscalers, that were the most profitable, asset-light companies we’d ever seen, using cash flow. Now that the CapEx is so big, and the infrastructure investment to bring future compute online—compared with what we have today versus future compute needs—is astronomical, companies have to find a way to finance that.
That’s why you have companies like Google going out and raising $80 billion in an equity sale, because you can no longer do that straight off the balance sheet. That can all be problematic, but the biggest problem, and the thing that historically pops major bubbles, is leverage and the overuse of debt financing. We’ve started to see more debt financing, whether it’s private credit or this Apollo and Blackstone deal for funding Anthropic’s compute.
Debt is what historically has been the major coolant and bubble-popper because that’s where you’re actually taking on debt to hopefully invest in much more productive capital or CapEx that generates an ROI. If we see any type of dip or inability for some of these big data center projects—based on delays, higher costs, or eventually not generating as much ROI on the compute they have, for whatever reason—if they’re unable to start meeting those obligations, or the risk of that starts to be perceived as higher, that’s where debt financing can wind up being a big bubble-popper.
Why I’m still in the camp that this can continue to run is, first, that the debt isn’t going to come due for a little while. Some of these deals are just being inked, and we’re seeing the massive data center build-out now. So it’s not a tomorrow problem, but it’s one of those things lingering in the background that you want to keep track of, because eventually it can be a pretty massive headwind.
The numbers we’re talking about here are so large that there’s no way companies can fund that off their balance sheets. They’re going to go to off-balance-sheet financing for it. That’s great for building it out and generating some ROI on it, but that intersection is something we definitely have to keep tabs on.
No, these are all great thoughts. For me, I think—and I’ve been saying this more and more—that the 2026 midterms are a really, really important moment for AI and its continued build-out. If it goes poorly, in terms of people taking control of different branches who are opposed to AI for various reasons—whether constituents don’t want data centers, there’s some environmental bullshit, or whatever it is—I think, depending on how the 2026 midterms go in November, it either pulls forward a lot of the risk or pushes it out a little bit more.
If things shift a lot and the current administration and Republicans lose a lot of their seats and power, I think a lot of the risks for 2028 and what happens after that get pulled forward into the immediate term. The opposite is also true. If Republicans sweep or keep massive control, I think a lot of the risk people perceive gets shifted out 2 more years.
The reason is that Trump is already— they’ve already said that every pullback while Trump is in office is a buy. It almost has to be a buy for me because they’ve literally said they will buy AI and take equity in AI companies. They will backstop any dips that come. The issue will arise if they’re no longer the ones calling the shots.
My counter to that would be that it has never been a profitable strategy to buy or sell based on what political party is in charge. It’s always made sense to stay invested.
Well, it’s more of a simple thing, right? I view it as almost a very similar trade to the 2024 Gensler trade. It’s almost as binary to me as that: if they win, the next 2 years are not as bullish as they would be if they lost. So now you need to start thinking about what other people are going to do—de-risk. Again, interest rates going up.
If Trump is in office and they have full control for the next 2 years, I don’t know what Warsh is going to do, but if they lose a lot of political capital and Trump is facing 2 years of impeachment trials, the outlook can deteriorate.
For sure. But look at crypto. Everybody thought that once Gensler was out and Trump was in, it would be up only.
It did well for a year.
It did well for 3 months until he went into office. It literally topped when he went into office.
Enough.
But I would say we were kind of front-running a lot of that for a while. But that’s essentially how I’m trying to think about just general perception around the industry and risk.
I don’t think you can discount it because, historically, I think you’re right that buying or selling based on who’s in office or who wins elections is probably not a dominant strategy. But clearly, markets and politics are significantly more intertwined than they’ve ever been and will continue to be.
I don’t think you can separate AI from politics. It’s a national security thing, which is also why I think it’s hard not to buy dips because the U.S. government can’t afford to lose the AI race, really.
You can’t untie AI from the U.S. economy.
So, yeah, I don’t know. It’s tough. I’m just holding on to my longs.
Yeah, that’s kind of my feeling. I feel like you could probably, if you’re a good trader, trade that cycle, but for me I just want to be long. For the next few years, I think.
You mentioned you were thinking about deploying cash. You said that almost as a mistake, like usually when you feel like that—when you’re talking yourself into it—it’s generally a local top. That’s kind of how I felt. I’ve been trying to talk myself into buying some [bleep], and I just can’t.
Yeah, me too.
But I want to so badly.
6. Hyperliquid & HYPE
It does make it harder when you start that internal conversation and then a month later you look at it and you’re like, “Yep, that was the right decision in the moment,” but yeah.
Do you guys own HYPE? I know Jason is max bullish.
Yeah.
Great segue, Tommy, because that was what I was feeling. I wanted to bring it back to a quick position—almost like a mini portfolio update. No, this is perfect because I know Jason’s got thoughts.
Yes, I do still own HYPE, for those wondering.
Yeah. It is interesting. This kind of ties full circle. The SpaceX IPO is obviously massive. There was a ton of concern that the IPO and the supply coming to market were going to suck liquidity out of just about everything to go and buy it. There was some of that—you saw some other space-adjacent stocks, for example, sell off.
But I think what’s interesting, and it’s not necessarily the biggest fee driver, like we’ve talked about, Jason, is that there’s a lot more press coverage around pre-IPO pricing and SpaceX. Obviously, with pre-IPO perps on Hyperliquid, it wound up being a really good proxy for where this thing was going to wind up pricing. You don’t have to rehash your whole HYPE thesis, but I do think that’s another—
Another shot on goal. It’s just like every shot is on goal, and they’re all hitting.
I’ve never looked at pre-IPO markets as a needle mover for underlying revenue or anything, because they’re not. But I’ve mentioned this in the pro chat, maybe on previous podcasts or whatever, and the pre-IPO markets are more of a Trojan horse to get TradFi aware of Hyperliquid and talking about it, and eventually allocated to it, either through the ETFs or through perps or whatever it is.
It’s their best marketing, for sure. That has tons of value.
It’s a fantastic tool for that, right? All of my TradFi friends—I’ve been talking to them about HYPE for a while. Most of them weren’t aware of it until I brought it up. I’ve actually had a couple of people reach out to me in the last couple of days about the SpaceX IPO and talking about the SpaceX perps, as well as the pre-IPO trading on SpaceX.
So I view pre-IPO stuff as just a marketing thing. It’s not a marketing scam or anything, but it’s the narrative that people can talk about, and that essentially drives attention to Hyperliquid, which gets monetized in other ways via perps, HIP-3 markets, or whatever it is.
For HYPE, I kind of remove the fundamental valuation from the conversation for a bit because it’s clearly overvalued right now based on what fees it makes and what its current valuation is. I set that to the side, and I just talk about what moves prices generally over the short to medium term. Valuation can play a part in that, but that’s not what drives prices.
What drives prices higher or lower is more aggressive buyers or sellers in one direction over a period of time, and how much supply exists, right?
When you look at it from a simple perspective like that, there really isn’t much supply for HYPE. The supply comes mainly from unlocks, which are significantly less than previously expected. A huge portion of our bull case heading into 2026 was that the unlocks would be significantly less than everybody was expecting because the team is not dumb.
When you look at the buying pressure, you have the Assistance Fund buying millions of dollars a week. You have ETFs buying millions of dollars a week—tens of millions of dollars a week. When you do a market-cap-adjusted comparison to Bitcoin and the Bitcoin ETFs, there’s been more impactful buying for the HYPE ETF since inception than you saw with the Bitcoin ETFs.
You have the Hyperliquid DATs, right? Again, I think DATs are pretty lazy vehicles in general, but if your incentives are properly aligned as a DAT operator, they can be beneficial in a way. They’re actually increasing cash balances and HYPE balances at the same time, right? They’re not just doing ATM and buying; they’re raising cash and acquiring HYPE through execution skill over the last couple of months.
There’s also a billion-dollar credit facility. The Circle USDC deal and revenue from that haven’t taken effect and haven’t hit the buybacks yet. There are just so many reasons to own HYPE relative to anything else in crypto from a pure flows perspective, and then you have all the narrative and everything else.
It’s the most compelling thing I’ve seen in years in crypto from a crypto-specific investment. I think this is going to be trading much higher in the coming weeks and months once all those things start to take effect. There’s too much buying pressure, and there aren’t enough sellers. There’s just not enough supply right now, I think.
I’m curious. I have a bit of a different view on the risk for HYPE. A lot of people say the biggest risk to HYPE is regulatory. I’m curious what you guys think, because I disagree.
I think that if you have the venue with the most open interest, volume, and usage, as an American, you should have access to that. You’re going to take a ton of slippage and pay fees on CoW Swap to do perps, or Coinbase, because you can’t—it just doesn’t make sense to me that Hyperliquid won’t be allowed in the U.S. What do you guys think?
I think the biggest risk is them getting hacked, or North Korea or some shit like that. I don’t think it’s regulation. Regulation may have been the biggest risk maybe a year and a half ago.
I think Hyperliquid will be allowed. They’ll just KYC, and whether or not you think that’s a risk to their model, I don’t know if it is anymore because—
Historically, you can run a no-KYC perp platform in crypto for 2 years, right? Then you reach scale and they clamp down on you. Every centralized exchange has run this playbook over time.
I don’t know. Maybe it’s like Polymarket and Kalshi, right? You’re supposed to be licensed to do gambling, but at some point the regulators are like, “Fuck it, we ball. We just let them do gambling here.” Obviously, all the gambling companies are pissed. Maybe that happens.
It just seems like here the interest is—it’s just a no-KYC derivatives exchange, right? And it’s not particularly decentralized. So, I don’t know. I do think it’s a risk. I still own HYPE, Zcash, Ethena, and Lyra, basically. Those are my bags.
But I'm bullish on HYPE. I don't know how much upside there is from here, though. I struggle to see this at $1,000.
Oh, no.
HYPE is a trillion-dollar asset—I mean, a $1,000 token.
Pretty sick.
Personally.
I think the market for no-KYC is big, but it's capped, right? All the big pools of money sit in KYC institutions that aren't going to use HYPE. Or maybe not. Maybe I'm wrong. Maybe they will, or they have the obligation to, once the liquidity is good enough.
But, yeah, it almost felt to me like that market's tapped, and the narrative is very hot as well. I don't know. I thought Jeff's Colossus profile was slightly ridiculous. He was positioning himself like he was curing cancer or something, doing this really important thing for the world. He's crushed it, and he's obviously an amazing founder.
Corporations are the equivalent of curing cancer. What are you talking about?
And, yeah, all the hype was—I don't know. But how can you not be long HYPE if you're in crypto, right? It's the one thing that has clear product-market fit.
I think a good comp is Binance. Just compare it to Binance. I think they'll be pretty equally valued at some point in the not-so-distant future. And again—
Binance is losing users, so maybe HYPE can take that.
The MiCA thing, yeah. I think their application in Greece got rejected, or is getting rejected. So, yeah, I just—
That sucks.
I think it's hard for me to envision a world in which Hyperliquid doesn't somehow—especially with HIP-3 and Jake Chervinsky—work on figuring out this KYC-regulated-markets aspect, right? The easy thing to say would be, "Okay, if they get KYC stuff and can service U.S. people, they're going to lose a bunch of retail traders on Hyperliquid," which may or may not be true, probably mostly true.
But in terms of trading activity and volume, you're losing retail and gaining tons and tons and tons of institutional flow, right? So it's probably more bullish in that sense for Hyperliquid for them to go after that. It just—
[Snorts]
7. Zcash & The Private Bitcoin Thesis
I would be super surprised if we were talking about this a year from today and there wasn't meaningful progress on this front. Whether that progress is KYC-enabled and allowed to trade in the U.S. for U.S. people, or they're foregoing that altogether, one or the other, I think you'll have a lot of clarity on that within the next 6 months. So—
Agreed. Sorry, guys, I have to jump for something else. I know we want to end on this private-BTC/Zcash discussion, so I want to say my 2 cents and then jump off.
Zcash has become my biggest crypto position, at least for now. I do think the question people were discussing was whether the private-BTC thesis was damaged by this, right? I saw the news, read about the hack, and posted in our chat that this seemed like a really good buying opportunity for Zcash. It really felt to me like the JELLY moment for Hyperliquid, which was when I managed to size Hyperliquid. There was this JELLY exploit, and people were like, "Oh, this proves the model doesn't work," and all this kind of stuff.
I think it's the exact same thing here. I don't think the mistake people are looking for is this immaculate thing—never being hacked—breaking. I think everyone understands that's impossible to guarantee. What people really want is to see hacks dealt with promptly and well, and that the team has it handled, basically, and that these things are unlikely to happen again.
All the actions the Zcash team took—I mean, obviously, it could have been a lot worse, don't get me wrong. If this had been exploited properly, it could have been a lot worse. The immediate evidence was that it hadn't really been exploited, and I think it's pretty clear that it hasn't. They handled it really well. They've got access to Mythos now, right? They're using it to protect the system.
I just think every crypto is exposed to hacks, and what really matters is how damaging the hack is. If you mint 20% of the supply and dump it, that's going to be hard to recover from. Then it's how you deal with it, and I think the Zcash team dealt with it pretty well, honestly. So I do think it was a buy-the-dip opportunity.
With Bitcoin being challenged by the Saylor overhang and stuff like this, I think Zcash has a narrative of being that private Bitcoin, right? The store-of-value narrative is still pretty small relative to what that TAM is. So to me, it's one of the most—
There's just a lot of really good developer talent behind it. Dave, who used to be at Osmosis and was a co-founder alongside Sunny, and one of the smartest cryptographers in crypto, recently moved over there. There have been a lot of talented people I know who've moved over there in various ways. It just seems to have momentum and a very strong narrative.
Jose, they were also audited by Mythos, and they came back clean, which I think is pretty unique.
Basically, yeah. For Zcash, I think the sell-off was just typical proof that a lot of people didn't really understand what they owned.
Yeah, I know. I would push back on Jose's point that it was common knowledge—
That this is something that all privacy coins face.
I saw the post that there was probably an exploit—not an exploit, but a bug had been found—and my reaction was, "Oh, damn. It's good that they caught this bug, right?" But I didn't think to sell because I didn't fully think about how everybody else would take this news, right?
In hindsight, I guess this was a pretty big thing. Zcash has been a big momentum trade, with a lot of borrowed conviction on it. A lot of people were hopping in on the quantum narrative and all this without really understanding privacy and the trade-offs at a fundamental level.
But I do think that once Ironwood goes live, which is aimed for the end of July and is the new Orchard pool, things will look different. Orchard is the pool where the bug was found. They're going to release Ironwood, which is the same pool but with the bug fixed. It will be formally verified, and Orchard will go into withdraw-only mode.
Once a few million Zcash from Orchard move to Ironwood, people will feel more confident. Orchard will never hit a zero balance, in my opinion. If Orchard hit a zero balance, I would actually expect that there wasn't an exploit. How many people lose their keys in crypto? If literally 1 person lost the keys to their Zcash in the Orchard pool, Orchard should never hit zero balance.
So if it goes to zero, it's like, "Oh, probably somebody exploited this," because hitting exactly zero is highly unlikely. Then Ironwood could be one of the most battle-tested shielded pools. They found the bug, they're doing the formal verification, and they had Mythos audit it. They had Open AI audit it as well.
That's a few layers of review. It doesn't mean the risk of a bug is zero, but it's pretty close. There's also a fundamental trade-off that you're going to have with privacy. One thing they could do is run 2 proof systems in the shielded pool. If there's a bug in 1 circuit, you'd have to exploit both proof systems. That impacts scalability a bit, but it's another thing you could possibly do. I brought it up to Zooko, who said it was a good idea. I don't know how feasible it is, though.
Do you feel like, generally, sentiment—or common understanding and consensus—is more aligned with what the realities are?
No, I don't think so.
Do you think sentiment has bounced back, or do you think it's still—? That's part of the question, too. Some people might—or you could argue a lot of people might—look at this and not fully understand what the so-called exploit or bug really was, or the fact that this technically plagues all privacy coins. They could look at it and go, "Yeah, this is a thesis-invalidating event," or, "The narrative has been violated," and just move on to the next thing. Do you feel like sentiment has rebalanced at all, or is it still—
No, I think sentiment has obviously come back because price does that.
But I think that there are a lot of people who probably will never touch it again because they didn't understand the risk they were taking to begin with, and now they do. And so, yeah, I don't know if Zcash is around 500 now. I don't know if it's going to rocket over the next couple months. Maybe Ironwood needs to happen and all of that, and people need more time for this to simmer a bit.
Because, obviously, a lot of the marginal buyers—you see that dump in the chart. Everybody I know who was a Zcash bull from earlier—none of them were really fazed by this at all. So it's kind of just proof that that entire dump is just all the momentum trade, right? And it was the marginal buyer of this coin. People weren't buying Zcash from October until June because they wanted to put it in the shielded pool. The shielded pool actually hasn't gone up too much since October, right?
So, I mean, if it's going to take some time for the marginal buyers to come back online, then, yeah. At the same time, crypto's super reflexive. So if Zcash breaks out and gets back above pre-exploit levels, and then if you zoom out to the 1-year chart, it's a really nice inverse head-and-shoulders. You can see the narrative kind of change on this pretty quickly if you wanted to.
So, yeah. I don't know. I own it still. I didn't sell it. I didn't buy more. I didn't think it was going to drop as much as it did—definitely not. I remember when I woke up in the middle of the night and it was like 350, and I was like, “What the fuck?” Then I saw it hit like 250, and I was like, “That is insane.” But, yeah.
8. Final Thoughts
I was in a very similar camp. But there's a lot to keep tabs on, a lot to keep updated on. We've covered a bunch of ground today. I think that's a great place to wrap it. I really appreciate you guys all joining. Tommy, thanks for showing up and chatting some open-source AI; that was really great.
Sick mouse.
It's a news station.
News. Can you grab it?
Bring it a little closer. Your name is covering it. It's kind of sick.
It is sick, and I love it.
Oh, it's massive.
Oh, it lights up, too. I could plug this thing in.
Next time we bring Tommy on, maybe we'll do a night session.
Dude, it looks a lot smaller in the background. It looks kind of like a picture, but it's as big as your torso, maybe bigger.
The meme is Nous Research, or Nous Hermes' boyfriend, Hermes' bad girlfriend. That's the current meme.
Okay.
It's a meme that five people will understand. I think more people need to understand the memes.
For the people that do understand, that's going to be hard. It's a good niche. In fairness, you have to be extremely high-IQ to understand Tommy's jokes.
It's undervalued as a meme, for sure. Thank you, senators.
All right. Good episode, guys.
Yeah.
Guys, thanks so much for having me. This was a blast.
Yeah. We'll see everybody in a couple weeks.
See you boys.