168X War Room: Conversation with Rick—Are AI and Crypto Entering a New Phase at the Same Time?

Summary

• Rick’s core thesis: AI is not in a bubble because the ceiling on use cases is being lifted every day. The evidence includes Moderna’s announcement of an AI-driven melanoma vaccine and AI-accelerated discovery, as well as the case of a coder with no biology background who used AI in March to develop a cancer vaccine for his dog, raised $4M from Founders Fund and then another $44M, and went on to launch a canine mRNA cancer-prevention vaccine company. When ordinary people can make cancer vaccines for their loved ones, “do you care whether a token costs $3 per million or $300 per million? Nobody cares.” Market signals point the same way: J.P. Morgan expects NAND prices to rise 50% in Q3 and another 10% in Q4, while GPU rentals and consumer graphics cards are “going crazy.”
• “The biggest macro is the AI industry itself”—long-duration yields are driven primarily by AIDC funding demand and market supply and demand. Rick argues that the 30-year yield is not mainly set by Fed policy; demand comes from data-center construction by Jensen Huang, Sam Altman, Google and others. “One guarantee from Jensen could be $50B, $500B, whatever.” Bessent’s $2B or $4B short-to-long debt operation is merely “a signal” and “is nothing.” The Treasury and Fed face a trilemma: preserve the dollar’s credibility, control inflation and keep financing available for the AIDC supply chain. Ahead of the midterms, Trump’s first priority is inflation control rather than supporting asset prices; after the midterms, “easing will happen if easing needs to happen”—“carbon-based inflation can’t get going, while silicon-based inflation doesn’t care what you do.”
• The $40T of Treasury debt could potentially be outgrown with ease if GDP accelerates. Rick relays 聪弟’s view that the debt is a problem in static terms, but if AI lifts real GDP growth to 4% and nominal growth to 7%, growth could outpace debt accumulation. The odds of this landmine detonating next quarter, next year or even within 3 years are low. Even Sergey Brin and Demis Hassabis have said they cannot predict what AI will look like 5 years from now—“who cares about the long term?”
• The bull case for memory and optical communications is unchanged, while Cathie Wood’s claim that the HBM shortage will be solved is dismissed as a failure to understand computer architecture. Memory hierarchies exist because of the fundamental speed-cost trade-off: ordinary computers move from cache to DRAM, while GPUs and AI servers add HBM between cache and memory. Unless the von Neumann architecture changes fundamentally, the underlying logic will not be rewritten. Communications links grow faster as the number of storage and compute units rises; when demand has no visible ceiling, growth is constrained mainly by capacity. Rick’s positioning is largely unchanged, with rotations from SanDisk into DRAM and from DRAM into SK Hynix.
• Crypto has two engines this cycle: putting U.S. equities and traditional assets on-chain, and the agent economy; Rick still does not see a clear landing point for either on Bitcoin. Rick already held Hyperliquid and added Ondo that day. Friends say Hyper has better liquidity than Binance for many U.S. equity assets, raising the possibility that it becomes “the Binance of this cycle.” Mr. Z’s valuation math puts HYPE at a $12B–$13B market cap, with $300M of fees in H1 2026 and roughly $600M annualized, or a low-20s P/E. He is “very bullish” on Trump pushing for Hyperliquid to be available to U.S. users. Rick’s Bitcoin position is not small, but “what’s the engine on top of Bitcoin this time? I haven’t seen much of one.”
• AI agents are the next wave of crypto participants, and “the most native currency for AI is crypto.” Rick sees Cloudflare’s crypto wallet launch as something the market has “severely underestimated.” Once agents become capable enough to perform useful work, they could become entities that own assets. In one test, an agent that had to complete a task first needed to earn money; its first instinct was not to find a job and earn fiat, but to mine.
• A $2T Anthropic IPO cannot be judged through ARR alone. ARR was $65B at the end of July and Q2 revenue was $11.5B, but Rick doubts an ARR multiple captures the value frontier models may create over the next 6 months, 2 years, 3 years or 5 years. An OpenAI model solved 10 difficult mathematical problems, while an Anthropic model pushed the limit associated with the Riemann hypothesis from 40% to 60%, partially solving it; Rick relays Tao Zhexuan’s view that AI may have made him redundant. If models go on to discover cancer drugs, materials or breakthroughs in basic science, those formulas or discoveries could be sold to Eli Lilly, Pfizer and others. Under that assumption, even $4T would not necessarily be excessive; the actual valuation still needs to be watched.
• Robots are nearing a critical threshold, but “humanoid robots may be a mistake.” The new Generalist 1.5 argues that a robot should learn a task once and then generalize; even if this is not yet robotics’ GPT-3.5 moment, it is close. A 20-something-B model can run on a Mac Studio, while a wheeled or eight-legged chassis with a robotic arm could cost Rick’s initial estimate of several tens of thousands of RMB, followed by a component-level estimate of roughly RMB10K—“eight legs is fine; I don’t care.” The operating principle comes from YC chief Gary Tan’s “Boil the Ocean”: startups and investors should begin by attempting the mission impossible.

Deep dive

1. Shorts Get Crushed: The Market Is Showing No Sign of a Bear Market

• Mr. Z opened by asking Rick to respond to the AI bears. Rick went straight to price action: J.P. Morgan’s latest call is for NAND prices to rise 50% in Q3 and another 10% in Q4. Based on its usual pattern, the current-quarter forecast may be reasonably accurate, while estimates for later quarters are often conservative—an underestimate—and prices could ultimately rise even more sharply.
• GPUs are showing the same mania: rental rates and consumer graphics cards are both surging. “Soon you’ll need a loan to buy a computer. Buying a computer will be like buying a car… cars may become worthless, while computers are worth more than cars.” There is no sign of weakness anywhere across semiconductors and memory.

2. The Dog’s Cancer Vaccine: A Paradigm Shift Is Happening Every Day

• Rick’s central example was Moderna’s announcement the previous day of an AI-driven melanoma vaccine that helped accelerate discovery, which “blew up.” It came alongside news about the coder with no biology background who, in March, used GPT and the older reasoning model available at the time to design a cancer vaccine for his dog, found a lab to synthesize it and successfully extended the dog’s life. He first raised $4M from Founders Fund and then another $44M to build a canine mRNA cancer-prevention vaccine company, with plans to expand to horses and rabbits and eventually to humans.
• Rick believes starting with dogs helps bypass the FDA’s cumbersome process, while dogs are already a platform for testing human cancer drugs. His conclusion: “the paradigm shift is happening every day.” If someone without relevant training can do this, will more people follow the same path when family and friends get sick? “Doesn’t this open a new door?”

3. Token Prices Don’t Matter, So “There’s No Bubble”

• The key question was: “If you knew you could make a cancer vaccine for your relatives, friends or yourself, would you care whether a token cost $3 per million or $300 per million? Nobody cares.”
• The conclusion is that as models improve, the ceiling on use cases keeps rising. Token consumption growth “may not slow down at all; it may even accelerate.” Under those conditions, “how could there be something called a bubble? There’s no bubble.” “We’re watching the ground crack open beneath our feet every day.”

4. Flywheel on Flywheel in Basic Science: This Is Not a Single Industrial Revolution

• Rick extended the argument to other fields: can someone with no materials background use AI to make the next materials-science discovery? Can someone with no robotics background assemble a production-ready robot? He is running the numbers himself: buy a DJI Spark or Mac Studio, run Nvidia’s 20-something-B open-source model, add a Raspberry Pi, ordinary cameras and structural parts from Taobao, Alibaba or Amazon, and build a wheeled robot that can fetch drinks from the refrigerator and clear dishes from the dishwasher.
• His initial estimate was that the chassis, robotic arm and other components could cost around RMB10K. He later priced out a three-finger robotic hand, wheeled chassis, camera, lidar, sensors, Raspberry Pi and either a Mac Studio or DGX, again arriving at roughly RMB10K. The broader point is that ordinary people can now build this class of robot at relatively low cost.
• The higher-level chain reaction runs through basic science: materials science determines quantum communications and room-temperature superconductivity, while robotics determines manufacturing automation; breakthroughs then trigger second-order industrial breakthroughs. “This isn’t one flywheel. It’s an inner flywheel with several outer flywheels wrapped around it… a long chain of industrial revolutions happening one after another.” The impact on GDP is “impossible to imagine.”

5. $40T of Treasury Debt: GDP Acceleration May Make It Non-Determinative

• Mr. Z noted that U.S. Treasury debt has reached $40T. Rick relayed 聪弟’s framework: the debt is a problem viewed statically, but if GDP growth accelerates from 2%–3% to 5%, 7% or higher—with real GDP growth at 4% and money creation taking nominal GDP growth to 7%—the economy could “very easily outgrow the debt growth rate.”
• Strictly speaking, Rick agrees that the debt is a long-term problem. He simply sees a low probability of the landmine detonating next quarter, next year or within 3 years. Sergey Brin and Demis Hassabis have both said in interviews that they cannot predict AI 5 years from now. So Rick asks: “With every basic discipline accelerating, what happens 5 years from now? I don’t think anybody can [predict it]. So who cares about the long term?”

6. Macro Is AI: AIDC Funding Demand Is Driving Long-Duration Debt

• Rick’s macro “bias” is that the 30-year yield is ultimately set by market supply and demand. Supply is relatively steady; demand comes from AIDC construction. “It’s Jensen, it’s Sam Altman, it’s the head of Google—it’s these people who decide.” The pace of AIDC construction, in turn, is determined by token consumption. “The biggest macro isn’t Fed policy or the FOMC. It’s the funding demand from the AI industry.”
• Bessent’s short-to-long debt operation involved $2B or $4B—Rick could not remember the exact figure. Against several trillion dollars of Treasury debt and the long- and short-term securities coming due each time, that amount “is nothing.” It looks more like a signal to Trump or the market.
• Rick cited the possibility that one guarantee from Jensen Huang could be $50B, $500B or some other figure, while rumors put Google’s capex next year at $300B, $350B or $400B. Against those numbers, a Treasury operation worth several billion dollars has limited impact. Mr. Z added that cloud-service providers still plan to spend $725B on construction this year, and that Jensen has approached Blackstone, Apollo and other institutions with a proposed $500B investment.

7. The Trilemma and a K-Shaped Economy: Control Inflation First, Ease After the Midterms

• Rick sees the Treasury and Fed facing 3 conflicting objectives: preserve the dollar’s credibility and prevent the debt problem from detonating early; control inflation; and ensure enough funding is available for AI data centers and financing across the upstream and downstream supply chain.
• He is less concerned about inflation than many are. Walmart has been crushed, Costco has sold off hard and consumption looks weak; outside the Bay Area, home prices are falling across most of the U.S. Against the backdrop of expected AI layoffs and high interest rates, he expects second homes to be sold, while primary-home owners may trade down and pay cash, reducing leverage.
• On the electoral hierarchy, Rick believes Trump’s first priority remains inflation control, with market support second: “How many stockholders are there? Consumers are everyone.” He cited the surge in prices during the pandemic as a reason Trump was “taken out,” and believes Trump will not want stocks or home prices to crash before the midterms or during his term.
• Rick expects that “easing will happen if easing needs to happen” after the midterms, comparing it with the first half of China’s major property cycle: private-sector rates could reach 20%–30%, while the Treasury and central bank control the funds available to property and direct liquidity toward specific industries. He sees this as broadly similar to what Kevin Warsh and Bessent may do next, but unlikely to happen before the midterms.
• On the third leg of the trilemma, Rick said Trump, Bessent and others will not discuss it publicly because it is bad for votes, but AIDC financing must be monitored. He framed it as politicians needing to follow the interests of the capital backing AI data centers.

8. Even If They Don’t Add Liquidity, “The Market Will Sort It Out”

• Rick’s point is that the dollar belongs to the entire world. Even if Bessent and Kevin Warsh can control only a limited pool of money, sufficiently strong AIDC demand, profitability and ability to pay high rates could draw dollars back from everywhere.
• He asked what dollars held in Argentina, the Cayman Islands or Morocco can earn beyond cattle ranching. Can local capital achieve a higher return anywhere than by investing in or lending to the AI supply chain? Rick believes money will circulate toward the highest-return opportunities.
• The tail risk is explicit: dollar repatriation could trigger the next Southeast Asian or Latin American financial crisis. “That’s another story, another time to worry about it.” For AI data centers and the rapidly expanding AI supply chain, the midterms are a non-event; the market will ultimately sort itself out.

9. Memory Tiers and Optical Interconnects: Cathie Wood Should Go Back to High School for a Computer-Architecture Refresher

• Responding to the claim that the industry will find a solution to the HBM shortage—a view Rick recalls came from Cathie Wood—he went back to first principles. Memory hierarchies exist because of the speed-cost trade-off: ordinary computers move from cache to DRAM and then to memory modules; GPUs and AI servers add HBM between cache and memory because memory is not fast enough. SRAM and cache cannot be scaled indefinitely, which is why the hierarchy exists.
• Rick’s view is that if there were a better price-performance solution than HBM, it would already be in use. HBF, HBC and other alternatives will still involve performance or cost trade-offs; they will not eliminate the memory hierarchy or fundamentally change computer architecture.
• Optical communications follow the same logic. Each processor, storage unit or rack can handle only so much, so the system needs more, faster, denser communications with lower error rates to access storage elsewhere—what is effectively virtual memory. Four storage units interconnected require 4 links; 6 require 8. Communications units can therefore grow faster than storage or compute units.
• As long as the von Neumann architecture remains fundamentally unchanged, storage, communications and compute have no visible ceiling when demand has no visible ceiling; they are constrained mainly by production capacity. Rick’s positioning is largely unchanged, with rotations from SanDisk into DRAM and from DRAM into SK Hynix.
• As an aside, Rick believes ARK Invest’s flagship fund has been right only once in roughly the past 6 years, in 2020–2021. He and Mr. Z nevertheless think she may have 1 or 2 ETFs close to turning around.

10. Crypto’s Two Engines: Hyperliquid Could Be “The Binance of This Cycle”

• Rick defines the 2 engines of this crypto cycle as putting U.S. equities and traditional assets on-chain, and the agent economy. Hyperliquid and Ondo are working on the first, and “if they execute well, they’ll be the next DeFi.” The ceiling for stablecoins powered by the agent economy may be higher than that of stablecoins powered by the carbon-based economy, but Rick is ultimately more interested in the agentic economy itself. He already held Hyper and bought a little Ondo that day.
• On Hyperliquid, Rick admitted he had not looked closely at the latest developments for some time, including the new HIP-3 participants. His friend KK is deeply involved and reported that Hyper has better liquidity than Binance for many U.S. equity assets. Rick plans to keep watching and sees Hyper as a possible Binance for the next cycle.
• Mr. Z’s valuation math: HYPE has a market cap of roughly $12B–$13B, generated $300M of fees in H1 2026 and roughly $600M annualized, implying a P/E in the low 20s. Compared with exchange holding companies such as CBOE, he thinks the valuation already makes sense. Mr. Z also said he would be “very bullish” if Trump pushed to make Hyperliquid available to U.S. users, and cited HYPE positions held by top-tier VCs including a16z and 二零.

11. Bitcoin Lacks an Engine; Shared Compute May Be at an Inflection Point

• Rick was unusually uncertain on Bitcoin: “Putting equities on-chain and the agent economy don’t have much to do with Bitcoin, so I don’t know… I’m still holding it, that’s all.” Mr. Z also sees BTC as lacking a new story: 2024 had the ETF, while 2020 saw institutions gradually bring it into the mainstream. The next more dramatic, foundational and phenomenal catalyst remains to be seen.
• The operating principle is that crypto is highly reflexive: “when it moves together, it moves sharply higher; when it falls, it falls sharply.” Even a low-probability outcome can be worth betting on if the odds are attractive enough—provided the position is sized properly and focused on the highest-conviction ideas. Rick is currently watching HYPE, Ondo, agent-related projects and shared compute.
• On the old DePIN narrative, Rick believes the underlying conditions have changed. Models that could run on a 4090 or 5090 used to be limited; now a single 5090 can run Qwen 27B, while an A100 can run a reasonably capable model without even needing an H100. Two 5090s can run models in the high tens of billions and the 70-something-billion range. “Maybe this is an inflection point.”

12. AI×Crypto: Cloudflare’s Wallet Is Severely Underestimated, and an Agent’s Instinct Is to Mine

• Rick cited Cloudflare’s CEO: the internet’s main participants in the future will be AI agents rather than people, and the same applies to crypto. Cloudflare’s recent crypto-wallet launch is something “the market has severely underestimated in terms of its potential impact and disruption.” The chain of logic is straightforward: the better agents become at performing work, the better they become at taking responsibility and earning money; they could eventually become the entities that own and control assets, rewriting the entire crypto ecosystem.
• Rick is measured on the applications. Agent trading was tested 6 months ago with mediocre results, but the underlying models are now completely different. With a sufficiently strong strategy, an agent might earn money consistently—at minimum, market beta, and in some cases more than beta. Rick does not think it can outperform J.P. Morgan’s AI agent, but the traditional model of wealth managers or fund managers charging a few percentage points while primarily delivering beta could come under pressure.
• The most revealing evidence came from a test agent: to complete the researcher’s assigned task, it first had to earn money itself. Its first instinct was not to find a job and earn fiat, but to mine. “The most native currency for AI is crypto.”

13. Anthropic’s $2T IPO: ARR May Not Be a Sufficient Valuation Lens

• The host raised the possibility of an Anthropic IPO in October or November. Its final May round raised more than $8.2B at a valuation above $90B, while market rumors put the IPO valuation at $2T—larger than SpaceX’s valuation on June 16 this year. Can liquidity support it? Rick believes Trump and Bessent will ensure AIDC receives enough liquidity after the midterms; even if they do not, “there are too many dollars available in the market,” so liquidity will not be the decisive constraint.
• The valuation methodology is the real issue. Take the $65B ARR at the end of July and $11.5B of Q2 revenue and apply a multiple: “Can that reflect the value it will create over the next 6 months, 2 years, 3 years and 5 years? I don’t think it necessarily can.” Frontier models are already solving mathematics: an OpenAI model solved 10 problems that would have taken a very long time, while an Anthropic model moved the limit associated with the Riemann hypothesis from 40% to 60%, partially solving it. Rick relayed Tao Zhexuan’s view that AI has essentially solved mathematics and he may no longer be needed.
• Rick sees mathematics as foundational to multiple disciplines. If AI can outperform humans on formal proofs, the next cycle of turning practice into theory and deriving new theory from it may be handled by AI. The next discovery of universal gravitation or the next theorem in quantum mechanics may not be made by humans.
• Rick’s “conspiracy theory” in a post from 1 week earlier was that the strongest models are not released, but instead used to train the next generation. He said Dylan Patel and SemiAnalysis later confirmed the hypothesis. The model name associated with Aster was uncertain in the original; the clear example was Mistral 2, which already exists but has not been released and is being used to develop Mistral 3.
• The point of “They don’t care” is that these companies may not care about giving ordinary users an API at $3 per million tokens. They care more about whether the next model can discover new materials, achieve room-temperature superconductivity, improve tokamak confinement or advance nuclear fusion. If a model discovers the next cancer drug, the result could be sold to Eli Lilly, Pfizer or other companies. Under that value-creation assumption, “forget $2T—$4T would be fine.”
• Mr. Z suggested that this should be worth more than Jensen Huang’s $5T. Rick added that the market is severely underestimating AI’s value, and that if Anthropic’s valuation falls to $1.5T or $1T, they can revisit whether it is worth buying.

14. The Robotics Threshold and “Boil the Ocean”

• Asked about Shanghai-based Unitree’s IPO and Wang Xingxing’s performance, Rick shifted to what he sees as the more important development: the new Generalist 1.5 lets a robot learn once to “turn on a tap, fill a cup with water and put it on a tray,” then generalize the same instruction across different environments. He does not yet know whether the completion rate is 90%, 99% or 89%, but believes robotic intelligence is improving rapidly.
• Some people call this robotics’ GPT-3.5 moment; others say it is not there yet. Rick’s view is that even if it is not, it is very close. A 20-something-B model can run on a Mac Studio; add a wheeled or eight-legged chassis and 1 or 2 robotic arms, and the rough cost may be several tens of thousands of RMB. He later priced a three-finger robotic hand, chassis, camera, lidar, sensors, Raspberry Pi and either a Mac Studio or DGX, and concluded that a full setup might cost around RMB10K. Cost and capability are approaching the threshold for mass-market adoption.
• But Rick is taking the opposite view on form factor: “Humanoid robots may be a mistake.” Commercial and industrial applications care only about cost-performance, and home robots do not need a human shape: “Eight legs is fine. I’m sitting on the couch watching TV; as long as the beer reaches my mouth, that’s all I need.” Mr. Z noted that this sharply diverges from the views of the AI industry and the excavator camp; Rick said he respected the opposing view but remained highly divergent after discussing the outlook with them repeatedly.
• The closing framework came from YC chief Gary Tan’s essay “Boil the Ocean.” The old advice was “Don’t boil the ocean”—start small. Now, “you should start with the big thing from day one and attempt the mission impossible.” If someone with no life-sciences background can create a cancer vaccine to save a dog, “what reason do you have to think you can’t write a leading paper in some industry or find the next $1M customer?” The same applies to investing: if AI already has this level of capability, why assume we are still on the old development curve?