168X in Conversation with Jason Kam: Crypto & AI—How to Read This Cycle?

Summary

• Jason sees the next 12–18 months as a window to buy Crypto on weakness, but he is not advocating going all in now. Since June, he has been building a basket weighted toward BTC in tranches: add on dips, wait when prices rise. The key external variable is AI’s wealth effect. Assuming AI company employees make up 15%–20% of the workforce and allocate 2%–15% of their cash-outs to Crypto over the next 2–3 years, he estimates potential inflows of $10B–$60B, versus roughly $55B in cumulative BTC ETF inflows. “If AI does exceptionally well, Crypto will benefit too.”
• He sees Crypto’s proven use cases narrowing to onchain asset trading and bringing dollar access to the world, and thinks investors can now buy most of the sector’s beta directly. The primary-market “token factory” model is over. On the trading side, he points to Hyperliquid and Robinhood; on payments, to existing winners such as Stripe and Adyen. In his view, this kind of portfolio can capture about 80% of sector beta without paying 2/20. “Just buy the winners”—but still compare market cap with assets under custody, users and trading volume.
• BTC and other Crypto assets are becoming two different investment cases, not just assets with different sensitivity to the same rally. BTC is digital gold. Other projects that return real-world cash flow to token holders, with growth and credible management, are “a stock that doesn’t trade on Nasdaq.” BTC may not be “the fastest horse” either: if AI productivity can ease the dollar’s predicament, the boat that shelters investors from the flood will have to compete with “AI’s rocket” for capital.
• The hardest AI inflection point to call is not token pricing, but how much revenue each GPU ultimately earns. Lower token prices could be offset by model efficiency, chip optimization and higher output. He admits that “nobody knows” whether demand for compute from open source and post-training will make up for slower incremental demand from OpenAI and Anthropic. Waiting for GPU rental yields to actually fall could be too late, so he calls this the period when he has the least certainty and the least edge.
• For now, he is more comfortable buying cloud providers and hyperscalers than making another indiscriminate bet on semiconductors. He mentions Microsoft, Amazon, Google and cloud ETFs, while saying he would leave Oracle alone for now. That is not a call to hold regardless of the cycle, indefinitely. At the time of recording, if the broader market holds up and performs well, he thinks Crypto could see another major rally over the next 1–2 weeks as capital rotates back in. The trade remains concentrated in a handful of consensus coins.
• The China AI investment case is not a blanket bet on “cheap Chinese internet stocks”; it depends on DeepSeek’s cost strategy, Tencent’s monetization channels and ByteDance’s product flywheel. Tencent trades at about 13x PE; AI could improve ad efficiency and lower game development costs. At a $600B valuation and $50B in net profit, ByteDance trades at about 12–13x PE. He is most excited by its video generation and its ability to expand as an “app factory.” DeepSeek follows a Costco- and PDD-like playbook: first push open source, cost and performance to the limit. He would be interested in buying if it listed.
• PDD is in his portfolio more as protection against an AI bubble bursting or a K-shaped recession than as a cheap stock waiting to rebound on its own. He owns PDD, values it at about 3.5x PE and thinks it could still grow 10%–15% a year. Its substantial cash balance also has a role in competition and social responsibility. If a downturn deepens and PDD connects its low-cost supply chain to agents, he thinks its position would strengthen. The host summed that up as “downside protection.”
• He has high conviction that robotics is the next narrative, but no clear answer yet on what to buy. “Robotics is definitely the next thing to get hot” does not mean related assets are attractively priced today. He has invested in Robot Strategy and confirmed there were exits in May and June; he also worries that related assets could fall below NAV if AI breaks down. Who can monetize the gap created by cheaper intelligence, and where should capital go when GPU returns weaken? Those remain the research questions he most wants to solve.

Deep dive

1. Folius separates its investment vehicles, while AI takes the bulk of Jason’s attention

• Folius is now split between its legacy fund and a new holding company. The latter is like Jason’s personal family office, with outside capital as well, and can invest across assets such as equities. Once the old fund unlocked, it allowed LPs to redeem quarterly, making it a poor fit for new, long-duration primary investments.
• Since late January or early February, he has devoted 80%–90% of his time to AI, especially the public markets. His remaining Crypto work is mostly managing existing fund positions, not continuing to seed early-stage projects.
• Since June, the new vehicle has been buying a basket weighted toward Bitcoin, without accelerating the pace: buy slowly when prices fall each month, sit tight when they rise, and potentially sell calls if they run too far. “Buy whenever the market drops” is a cycle view implemented in tranches.

2. The token factory is broken; Crypto VC has to prove its edge again

• Jason says the fund stopped making new Crypto primary investments in September 2024, first and foremost because the “token factory” had run its course: the old model of a few people forming a project, issuing a token and locking it up no longer makes sense.
• Mainstreaming has shifted the industry’s center of gravity to the US, weakening Folius’s former edge in backing Asia-Pacific founders, especially in China. It also means competing head-on with Founders Fund, Sequoia and USV. The question of why a great founder would not take their money is unavoidable.
• The host adds that Trump’s meme coin damaged the marginal buyer outside the industry, while the high-FDV, low-float VC token game has been seen through. He recalls Haseeb asking, “What’s your edge?” Jason thinks leading funds are still active, but a firm like Dragonfly investing in Raincard looks more like a Ribbit-style Fintech investment.
• He expects the standalone category of “Crypto VC” to disappear, though liquid funds and prop funds will remain. Onchain analysis of token holdings, accumulated industry experience and information, and judgment of people’s character and backgrounds can all be sources of edge. A move into AI or Fintech, though, requires the relevant software, financial or investment network.

3. Proven demand is concentrated in trading and dollar payments

• In Jason’s framework, Crypto has two main use cases left: “putting everything onchain” and “the dollar covering every aspect of our lives.” That leads to trading, payments and neobanks—not rebuilding every internet application from scratch.
• His reflection on past bets is blunt: “I invested in all of them,” and they “all went to zero.” Users faced too much friction, did not care about financialization, or the technology was “a hammer looking for a nail.” In his view, the industry has become a branch of Fintech.
• He estimates the global financial product market at about 120 trillion, with Crypto penetration of roughly 2%–3%, and believes it will “definitely get to” 10% or 20%. Global payments are around 45 trillion–50 trillion; with agents in the mix, they “could” double over the next 10–20 years. Either market is large enough on its own.

4. New projects need a different wedge, not a copy of the leader

• The host challenges Jason with HIP-3: in his observation, TradeXYZ has taken about 95%–96% of trading volume, and founder Shoku was involved in the Hyperliquid ecosystem early. Is the lead due to the mechanism, or to first-mover advantage and capital stickiness? He doubts later entrants can catch up.
• On the host’s proposed A-share and Hong Kong stock perps wedge, Jason does not think adding a single new asset class is enough. The category would have to be large, demand would have to keep appearing, and each new market would have to launch faster. Otherwise, “you can do it, they can do it too”—and the incumbents might do it better.
• He would rather look at different entry points: Catalyst is trying to aggregate Twitter information, have AI interpret intent and place trades; Fomo is trying to turn social activity into a way to trade meme coins. He does not know if either has PMF, but “you need a different approach when you enter,” rather than building another Hyperliquid.

5. Sector beta is accessible; valuation and opportunity cost now matter more

• Jason thinks ordinary investors can now buy trading and payment winners directly and capture about 80% of sector beta, without necessarily paying 2/20 for an early-stage fund. That is one reason he returned a lot of LP capital near the top.
• On trading, he names Hyperliquid, Robinhood, IBKR and Kalshi. On payments, there are Stripe and Adyen; Circle is not the only option. He does not assume an early-stage stablecoin company will naturally beat incumbents with strong engineering teams and fast execution: “Just buy the winners.”
• But winners still have to be bought at the right price. He describes market cap as “Brownian motion” around adoption, with assets under custody, users and trading volume as possible anchors. For example, he thinks Hyperliquid could be a buy when market cap relative to TDR and volume gets low enough, around 25 to 30—not a reason to chase the leader at any valuation.
• He draws a parallel with the internet in 2002–2003: buying private shares in Amazon, Google or Facebook and holding them might have delivered annualized returns of 20–30%/yr, outperforming a VC fund. His personal return expectation for the Crypto theme is about 30% annualized, but he still asks whether it will outperform AI or robotics.

6. BTC is digital gold; other tokens increasingly look like Fintech stocks

• Jason insists on separating BTC from the rest: the former is a digital gold trade, the latter frontier Fintech. He sees Hyperliquid’s correlation looking more like US equities now, rather than being driven primarily by Bitcoin.
• A token willing to return real-world cash flow to holders, with credible future growth and management, is already “a stock that doesn’t trade on Nasdaq.” Investors should study how the company will develop, not just guess whether BTC will rise.
• The boat that helps people weather the flood is now up against “AI’s rocket.” If advanced productivity can resolve the dollar’s predicament, he thinks investors should buy productive assets; BTC may even be somewhat “anti-productivity.” People creating AI wealth may also prioritize putting it back into new productivity.
• He expects BTC to behave more like gold, rising slowly with the dollar cycle. He thinks BTC’s annual inflation rate will be below 0.5% going forward. Against the scale of the macro markets for ETFs and dollar assets, changes in new supply can no longer move BTC’s volatility, so investors should stop mechanically applying the four-year halving cycle.

7. Naive capital has left; long-term consensus is concentrated in a few assets

• Jason calls the early market a game for “young and naive capital”: sentiment could outperform fundamentals for a long time, and fund managers had to chase liquidity even when they did not believe in it. Now some players have been wiped out, gone bankrupt and exited; others made money and moved on. The capital left is smarter and more concentrated, and the list of coins people can hold long term is probably only 5–10.
• The host asks whether the Robinhood ecosystem can take the baton. Jason admits he has not looked closely, and Robinhood itself is not actively supporting it. The ecosystem’s potential remains to be seen.
• Outside trading and payments, he remains interested in Zcash-like demand for privacy, Venice’s private inference, and Crypto-AI efforts from Near, Venice and maybe Pearl. But serious AI researchers mostly do not care about Crypto AI; how the two will come together—“But nobody fucking knows.” Crypto natives want to trade the theme, and these assets may get another run when the next tech wave arrives.

8. AI employee cash-outs could become a major source of capital for Crypto

• Jason’s estimate is conditional: assume AI company employees make up 15%–20% of Silicon Valley’s workforce and cash out through listings or other liquidity events over the next 2–3 years, then allocate 2%–15% to Crypto. That would mean potential inflows of $10B–$60B. He compares that with roughly $55B in cumulative BTC ETF inflows to show how significant the employee wealth effect alone could be.
• The reverse is also true: if AI crashes, he does not think Crypto will be spared; “it will definitely fall.” If monetary easing later pushes Bitcoin higher, that would be an “everything goes up” phase. That is why he likes the 12–18-month window but is unwilling to go all in right away.

9. The AI inflection point is revenue per GPU, not token prices alone

• Looking back at his earlier edge, he says the market underestimated Claude Ko at the start of the year. In March and April, Claude started doing agent work and GPU demand became more visible. In July, Claude was hard to use and Codex worked well, so the market may have mistaken a shift in market share for an industry slowdown. He adds that Claude may in fact have slowed later; it was not all a false signal.
• The hardest question now is whether lower token prices mean lower revenue per GPU. Model efficiency and chip optimization are also reducing the cost of generating tokens, so token prices and revenue per card are not directly linked. He says nobody really knows the answer yet.
• The second unknown is marginal demand: can the rise in semiconductor demand from open-source models and post-training offset the demand lost as OpenAI and Anthropic growth slows? “Nobody knows.” The answer may also be where the timing of AI’s peak lies.
• Waiting to turn bearish until GPU rental yields actually fall could be too late; smart money may already be out. For now, he can only do channel checks every quarter and watch hyperscaler earnings. “The bear case is very real,” but it is not as directly observable as model company revenue was before.

10. Cloud offers more visible certainty for now, and capital could return to Crypto

• The host’s view is that the semiconductor bull market may not be over, but investors can no longer buy anything and expect it to rise as they could in February, March and April. As open-source models catch up with closed-source models, CSPs can take on enterprise deployment and improve efficiency, shifting capital gradually from GPUs, CPUs, optics and memory toward cloud.
• He also warns against extrapolating total industry demand from OpenAI and Anthropic revenue alone: in his view, the two account for about 30%–40% of compute. Closed-source models capture more than 90% of the economic value, while open source may account for 50%–60% of usage; everyday tasks do not all need the strongest model.
• Jason agrees that Microsoft, Amazon and Google, among other hyperscalers/CSPs, are easier for him to consider right now. He might even leave Oracle alone, and mentions “SKY ETF” and other cloud ETFs. But the pace of open-source progress and revenue per card are still unknown, while semiconductor expectations are already high. He does not think these positions can be held indefinitely.
• He sees some of his smartest Crypto friends reaching the same conclusion and going back to trade Crypto. So if the broader market does not break down and still performs well, he thinks there could be a major Crypto rally in the 1–2 weeks after the recording. The investable stories are still concentrated in those roughly 10 coins.

11. BTC’s floor rests on holders’ reluctance to sell, but three shocks could still interrupt the rally

• The host would rather consider putting 70%–80% of his allocation into Crypto, which he knows better, than compete with people who have spent 20 or 30 years studying semiconductors. He sees BTC holding firm around 80K; if it breaks back above 88K, the next leg could be 90K–97K.
• Jason puts more weight on holder behavior: “Unless people are forced to sell,” medium- and long-term holders will not want to sell at these prices. He sees the market holding up after MicroStrategy risks were removed and BTC’s correlation with gold increased as signs of a floor. With risk-taking capital out, holders’ reluctance to sell means even modest buying can move the price.
• The first shock would be AI productivity getting so strong that it weakens the case for holding nonproductive BTC. The second is quantum Q day. He thinks a wave of quantum FUD will “definitely come,” but admits he does not understand the technology and only guesses that better-resourced people might find a solution, such as a fork.
• The third is an AI cycle ending alongside leverage, triggering a financial crisis and forced liquidations. In that case, BTC will “unquestionably fall.” Even so, he puts these risks inside his 12–18-month buy-the-dip framework, not a case for going all in at once—and he does not guarantee that BTC will outperform AI.

12. China AI offers three bets with distinct sources of advantage

• Jason leans toward DeepSeek, Tencent and ByteDance. Chip and compute constraints limit near-term progress, but he thinks they will be resolved over a longer horizon. Existing open-source models plus post-training are already enough to change a lot in personal agents and specialized use cases.
• DeepSeek reminds him of Costco and PDD: be the most open, lowest cost and most universal; win first on cost and performance, then figure out what comes next. He sees a very clear strategy and would be interested in buying if it listed.
• Tencent trades at about 13x PE. AI can improve ad targeting efficiency and lower the cost of making games; Tencent still owns its WeChat distribution, game publishing and monetization know-how, which AI cannot replicate for now. He explicitly says he does not know whether users will keep spending their time on games in the future.
• ByteDance’s private valuation is around $600B, with net profit of about $50B, or 12–13x PE. He estimates overall growth at 15%–20%, and says ByteDance is ahead of Chinese peers in spending on chips. Its ad business provides a base; he sees it as China’s most liquid private-market asset, one anyone can buy.

13. ByteDance’s video flywheel and app factory offer additional paths to growth

• Jason thinks ByteDance has built a commercial flywheel beyond coding agents, centered on Seedance. He says it holds 70% of the generation market and expects that after another 6 months to 1–2 years of iteration, AI video will be indistinguishable from Hollywood production.
• His scenario: one or two people, or a few directors, plus a pile of compute cards could make excellent short-form and long-form shows. That would break the old moats around film and TV content and actors’ pricing power. The monetization ceiling is set by Netflix and movies, while production costs could be zero or very low.
• Another ByteDance advantage is its mature product portfolio and growth playbook; it already resembles an “app factory.” AI makes it easier to build and distribute more products. He sees particular room in overseas markets outside the US, while acknowledging the risk of war.
• The host brings up AI-generated adult content. Jason sees it as a clear use case for real-time, customized video and is looking at related investments. But he points to product friction: users may want to consume directly rather than write their own plots. Real-time generation is only a matter of time; that does not mean users want to do more thinking.

14. Cheap Chinese stocks lack a catalyst; PDD is a recession hedge

• The host thinks A-shares and Hong Kong stocks are already cheap, but Jason does not equate cheap with an imminent rally. A-shares have financing needs, while Hong Kong has governance issues that concern foreign investors. When capital is scarce, money will move from companies like PDD to the fastest-growing assets unless liquidity expands or the AI trend reverses.
• Jason owns PDD, values it at about 3.5x PE and thinks it could still grow 10%–15% a year. Its large cash balance is not simply idle: the company needs to defend against competitors such as Alibaba and may have social and economic responsibilities to meet, so it has to keep reserves.
• If AI blows up and the economy heads into a K-shaped recession, he thinks PDD could be very strong as it connects with agents: it is the cheapest and most efficient, with the ability to deliver to corners of the world. That makes it one of his best picks for that kind of environment.
• He draws on his research into the 2007–2009 US crisis: Dollar Store and Dollar General did not fall; EPS and valuations rose, and their shares gained 10%–20%. The host calls this “downside protection,” and Jason agrees.

15. BidClub shifts from an investment-opinion community to a podcast tool for its founder

• Jason says the original idea was a platform for mid-to-long-form investment views on individual tokens, which would also help him meet more investors. Later, he realized everyone already knew the good assets, so the platform was not essential. He turned it into a podcast summary tool to capture new ideas and compare views across episodes, with the hope that more capable AI could make use of the database.
• He considered issuing a token and using platform revenue for buybacks, but still has not worked out what the product should ultimately become. He also admits that content aggregation alone has limited value. The host counters that media businesses have low margins and struggle to earn high valuations; Jason agrees, so for now the product addresses his own problem: too many podcasts to listen to and too much effort to summarize them.

16. The next research questions are who can monetize cheap intelligence and where robotics fits

• Jason most wants to understand who will “monetize the spread” as tokens and intelligence get cheaper. Could a PE firm such as Long Lake acquire businesses, transform them with cheap intelligence and turn lower costs into substantial enterprise value? He does not have an answer yet.
• His second question is what to buy if the AI cycle eventually weakens because revenue per GPU falls. “What should I buy?” He repeats the question and says plainly, “I don’t know.”
• The third area is the one he has the most confidence in: AI is the key to bringing robotics into the physical world, and “robotics is definitely the next thing to get hot.” But he is not nearly as sure about what to buy.
• He says he invested in Robot Strategy and confirms there were exits in May and June. Based on current market prices, he estimates Andrew Kang’s NAV could be 20 or 22–23. But related assets could still get crushed and trade below NAV if AI breaks down. Confidence in the direction does not settle the price or the timing of an entry.