Everyone Is Watching AI While Crypto’s Bull Market Builds | Weekly Roundup
Summary
The central paradox is that AI owns the narrative while crypto may already be in a bull market. LPs remain preoccupied with AI, even as Yano says crypto interest has revived after 18 barren months and some tokens have moved dramatically: “No one believes this is a bull market, yet tokens have increased 50 times.” His renewed watchlists show returning appetite, but also the risk that sidelined investors chase after the move.
Anthropic became the proxy for a broader question: can frontier-model revenue ever outrun compute expense and relentless customer switching? The hosts cite reported economics of $1 in revenue per $60 of compute, a recent absence from OpenRouter’s top 10, and possibly flattening ARR—but repeatedly admit they have not seen the full S-1. The disagreement is unresolved: Santiago presents the demand and usage evidence, while Jason attributes the plateau to compute constraints, citing Anthropic’s usage limits, but says capacity may constrain other businesses more than Anthropic itself.
AI infrastructure is attractive only if compute stays scarce long enough to repay an extraordinary fixed-cost buildout. The episode estimates that 1 gigawatt of current-generation AI capacity costs $50–60 billion, roughly half in Nvidia chips, and could generate $10–20 billion annually at $2.50–$3.50 hourly rental rates and 80–90% utilization. A private-credit executive’s harsher verdict was that the projects require “a venture-capital view of private credit” and that lenders underestimate downside protection.
Falling inference costs could strand capacity—or unlock enough usage to validate Jevons’ paradox. Santiago was unsure whether the model discussed was Jev or Type Safe; for Jev, he said it was 238 times cheaper than Fable, 70 times cheaper than Opus, and 38 times cheaper than Sonnet. That threatens forecasts built on today’s expensive workloads. Yet AI penetration remains below 5%, while one Blackstone portfolio reportedly expects token spending to rise from $25 million to $500 million in a year: “You’re starting from a very low level of penetration.”
Crypto’s investable expression has shifted from “long stablecoins” to “how do I go long a token?” The prior answer included Coinbase and Circle; today the hosts point to Backpack and Securitize, with Robinhood potentially offering a cleaner public-market expression. Legal structure and tax treatment matter alongside technology. Yano admits Backpack grew 50x while he watched—“What an idiot I am”—and says he still plans to buy Backpack and Pump, while Santi discloses buying more Zcash.
The hypothetical $200 billion FTX portfolio is an underwriting warning, not evidence of venture skill. Anthropic, SpaceX, Cursor, Solana, and Robinhood might have created one of the top five venture-capital funds of all time, but Santiago rejects the comparison: “If I had an unlimited money faucet” and invested in everyone, a few winners would prove little. The sharper bubble signal was that “the person who did the most due diligence was Taylor Swift and her team.”
LPs are beginning to treat crypto as the countercyclical alternative to crowded AI, particularly in growth-stage fintech. The hosts estimate roughly 60 crypto-related growth companies, many using stablecoins or tokenized assets, versus perhaps 15 suitable targets for Coatue’s new $8 billion fund. Capital looks especially scarce between $500 million and $5 billion valuations, while accessible early-stage AI funds increasingly receive the answer: “It’s too late.”
Financial agents may compress some friction rents without overturning wealth management as quickly as expected. Jason’s framing is that personal agents could move consumers toward cheaper or better financial products, while “by definition LLMs give you mediocre advice”; median advice could still improve outcomes for millions without basic financial literacy. Agents may accelerate active trading and movement toward cheaper products, but younger users appear to trade faster while older, wealthier Americans remain more buy-and-hold, and incumbents still control the available product menu.
Deep dive
1. Frontier-model growth has not resolved the unit-economics question
Jason separates AI’s undeniable product progress from its private-market pricing: some companies could become “literally trillion-dollar companies,” while others are “complete zeros, and we have no idea which one it is.” Unlike crypto in 2020–21, AI has not yet produced a Celsius-like failure that reveals where the weakest underwriting sits.
Anthropic is the test case. The hosts cite reporting that it generated $1 of revenue for every $60 of compute, disappeared from OpenRouter’s top 10 after early August, and may have seen ARR stabilize. They nonetheless emphasize that they have not reviewed the complete S-1 and are “speculating without understanding what’s going on.”
The disagreement is unresolved. Santiago presents the compute-cost, usage-ranking, and ARR evidence, then asks why Jason attributes the plateau to computing power. Jason points to Anthropic actively limiting his usage, while also saying compute may constrain many other businesses more than Anthropic itself. The available data are insufficient to settle the question.
A further concentration risk is offered cautiously: roughly one-quarter of revenue reportedly comes from two customers, “Meta and Alibaba, I think.”
Competition makes the economics harder to defend. Users switch when Codex becomes better than Claude Code, OpenAI appears to catch Anthropic in ARR, and open-source alternatives keep improving. Santi’s valuation discipline is blunt: “I don’t know if Anthropic is worth $2 trillion. I’m not a buyer at that level.”
2. The data-center boom can satisfy real demand and still destroy capital
Santiago recounts that a private-credit executive told him data-center prerequisites, payback periods, and cash-flow profiles “make absolutely no sense” on conventional risk-return terms. His criticism of venture capitalists was sharper: “These guys know absolutely nothing about loans. They know nothing about downside protection,” so he expects much of the financing to blow up.
The hosts’ preferred analogy is the railroad boom: demand can be genuine, forecasts broadly correct, and the eventual infrastructure socially valuable—yet excessive debt-funded construction can still bankrupt operators. The load-bearing questions are: “Will we be limited in computing resources over the next 5 years or not?” and will demand continue growing exponentially?
Their rough build economics put 1 gigawatt of current-generation AI compute at $50–60 billion, about half for Nvidia chips. At rental rates of $2.50–$3.50 per hour and 80–90% utilization, it could generate $10–20 billion annually, implying roughly a 3–5-year payback—but only if pricing, utilization, and chip usefulness hold.
Residual value has so far been better than skeptics expected: aging H100s and Blackwell GPUs can serve open-source models and less demanding workloads, unlike obsolete Bitcoin-mining ASICs. The longer-dated risk is space-based compute; as relayed from a Blackstone investor in CoreWeave, efficient orbital data centers could impair terrestrial assets within a 5–10-year window.
3. Model efficiency cuts both ways for compute demand
As models become cheaper and more workloads shift toward open-source or less advanced models, the hosts question whether demand will remain tied to today’s compute intensity. Santiago was unsure whether the system discussed was “Jev” or “Type Safe”; he said Jev was 238 times cheaper than Fable, 70 times cheaper than Opus, and 38 times cheaper than Sonnet. OpenAI released something similar a week later, though its precise architecture and savings were unclear.
That efficiency threatens capacity forecasts built on today’s compute intensity, but Santiago invokes the blockchain analogy: gas prices fall, and Jevons’ paradox requires believing that usage will expand enough to offset cheaper computation. With “less than 5% of the population” connected to AI, cheaper inference might grow total consumption from a very low base.
One Blackstone allocator reportedly expects token spending across portfolio companies to jump from $25 million last year to $500 million this year, a 20x increase. Much of that workload is migrating to cheaper open-source models—the tension in one statistic: explosive enterprise adoption alongside relentless substitution toward lower-cost compute.
4. Crypto’s bull market is moving faster than its participants
Yano says nobody wanted tokens for 18 months; now he has watchlists and says he has 10 potential purchases. Santi’s pushback is the cycle in miniature: making a list is not buying. Yano concedes he ignored the Druckenmiller lesson that fast markets reward stating a thesis first and revising it later.
Backpack is his painful specimen: Blockworks’ research analysts, Armani Ferrante, and the show’s own conversations repeatedly surfaced it, yet Yano watched it grow 50x. He now intends to buy “for 10 times more than I should have,” while timestamping that he owned none during recording. He also plans to capitulate on Pump; Santi discloses adding Zcash and having venture exposure to both Pump and Backpack.
The thematic trade has evolved. Last cycle’s question—“How do I open a long position in stablecoins?”—led toward Coinbase, Circle, and perhaps Tether exposure. Today outsiders ask how to go long a token, with the hosts naming Backpack and Securitize; for tokenized stocks, Santi says Robinhood may be the cleaner public-market expression. Backpack’s appeal is framed as technology plus a legal framework that Armani focused on from day one.
Market plumbing is opening too. Robinhood’s U.S.-regulated perpetuals use Bitstamp and the foreign-board-of-trade route rather than Lighter. SEC guidance also appears to say token redemptions under “functional protocols” are not management efforts. The hosts discuss liquid staking and whether staking commissions might be viable, but admit they do not understand the boundaries and want Miles Jennings and Gabe Shapiro to explain the updates.
5. FTX turns a spectacular paper portfolio into an underwriting warning
The producer calculates that FTX’s early positions in Anthropic, SpaceX, Cursor, Solana, and Robinhood could now be worth about $200 billion, potentially making it one of the five largest venture-capital funds of all time in absolute dollars. The statistic is striking, but Santiago says it answers the wrong question.
His rebuttal is that a venture fund must compound a fixed pool through sizing, selection, and accountability. FTX instead had an effectively unlimited pool of customer money and invested across nearly everything it encountered. Jason sharpens the point: “If you give me all your money” and one of many indiscriminate checks succeeds, does that make him a good venture capitalist?
Jason notes that FTX’s access itself was extraordinary. FTX became so prestigious that declining to invest made institutions look eccentric, while its backers relaxed underwriting despite available financial and biographical checks. The hosts see an AI echo when rounds rise from $1 billion to $5 billion—or $10 billion—within three months as funds scramble to compensate for missing SpaceX, OpenAI, or Anthropic.
6. Capital is rotating toward crypto’s neglected growth-stage middle
Santiago says he began hearing LP doubts about AI froth in late July or August, and that the conversation accelerated over the past month. Support in major crypto assets, continued stablecoin growth, and rapid expansion at Rain and prediction-market companies are encouraging allocators to consider a countercyclical move while AI rounds look increasingly difficult to justify.
The opening is primarily late-stage. The hosts expect larger funds aimed at Series B through D and pre-IPO rounds, but stress that only a few crypto managers “have earned the right” after the disastrous 2021 vintage. Then, growth was rapid but customer quality, spending, and fundamentals were weak; today they argue that the fundamentals are substantially better.
Their internal map contains about 60 crypto-adjacent growth companies, mostly fintechs built around stablecoins or tokenized assets. Examples include Rain, companies building on prediction-market infrastructure, Jeeves—which announced another $10 million—and neobanks Flex and Slash. By comparison, Coatue may have only about 15 targets for its new $8 billion fund.
The capital shortage appears most acute between roughly $500 million and $5 billion of valuation, below the scale where enormous generalist funds concentrate. One unnamed company reportedly grew 55% in the third quarter, accelerated above $80 million in revenue, and still struggles for attention because investors are asking, “Should we do more Anthropic?”
7. Financial agents improve the median before they dismantle incumbents
Jason expects personal agents to move consumers toward cheaper or better financial products, making them “probably really bad” for some incumbent economics. Yet he thinks agent-to-agent settlement may favor stablecoins while ordinary user transactions remain less affected, limiting how directly the change accrues to crypto.
His deliberately unglamorous claim is that “by definition LLMs give you mediocre advice.” That can still be transformative when the baseline is poor financial literacy, but it also produces clusters of people receiving similar recommendations. A large private-fintech COO questioned whether conversational advice improves on pre-filled direct-indexing products that already deliver essentially the same portfolio.
Agents could stimulate more active trading, but the hosts distinguish younger behavior from the wealth-owning population: Santiago first says people over roughly 35 are buy-and-hold, then qualifies that younger segments trade faster and that the precise age split is uncertain. Wealth managers will still curate restricted asset menus, agents may only optimize movement among Treasury and money-market products, and checking or savings yields may rise over time—making disruption real, but less sweeping than advertised.