Brian Armstrong on Bitcoin, Anthropic Drops Fable 5 & Mythos 5, NewLimit's $435M Age-Reversal | 264
Summary
Armstrong’s Bitcoin call is a probable $60,000 bottom—not a certainty—and a slower transition from risk asset to “the new digital gold.” He estimates roughly 30% of capital already treats BTC as an uncertainty or inflation hedge, while 70% still trades it like volatile tech, explaining why geopolitical stress has not reliably made it countercyclical. Against Citi’s $189,000 end-2026 projection and Polymarket’s $84,000, Armstrong called $100,000-$200,000 plausible and expects a much higher price by 2030.
Coinbase wants to become “the financial account for AI,” and Armstrong says usage has already reached about 100 million agent transactions and $50 million of value. The product stack runs from connecting an LLM to a Coinbase account through MCP and a CLI, to Coinbase Advisor, to self-custodial Base wallets that agents can open without KYC. Stablecoins will probably be the default payment layer for the agent economy.
Quantum computing is not an imminent Bitcoin threat, but Armstrong regards eventual cryptographic breakage as “almost certain.” BIP 360 would introduce quantum-resistant cryptography at the cost of larger blocks, while other major chains are also working toward upgrades. The hardest issue is the estimated 5%-10% of early Bitcoin potentially at risk: freeze unmigrated coins, preserve them as a quantum “bounty,” or freeze them with an appeal mechanism.
The episode’s central capital-markets warning is a collision between quasi-nationalized AI and three unprecedented trillion-dollar IPOs. Proposals range from 5%-10% government “golden shares” to Bernie Sanders’s suggested 50% transfer, while one UBS investor reportedly sees only $75 billion of liquid capacity against offerings seeking hundreds of billions. Armstrong worries retail could “take the hit,” while Diamandis argued that within 12-18 months, 80% of workloads could move to models that are 99% cheaper.
SpaceX is being valued less like a rocket maker than a vertically integrated AI utility. Google is reportedly paying $11 billion annually through 2029 for 110,000 NVIDIA GPUs, while the AI1 satellite is described as delivering 150 kW of peak compute in a two-ton platform with a 70-meter wingspan. SpaceX plans a 1,000-acre GigaSat factory and ultimately up to one million satellites—turning AI infrastructure from “a real estate problem” into “a launch problem.”
NewLimit’s $435 million raise funds multiple clinical shots after cellular reprogramming advanced faster than Armstrong’s expected five-to-ten-year research timeline. Its AI-guided platform searches 10 quadrillion protein combinations, tests hits in wet labs, and checks whether treated cells regain functions such as processing caffeine, acetaminophen, and alcohol. The first drug candidates are expected to enter the clinic next year: “We’re trying to do half of what Shinya Yamanaka did”—change cellular age without changing cell type.
Armstrong expects embryo editing to move from disease prevention toward enhancement because the boundary is intrinsically blurry. He cited roughly 80% support for preventing disease versus 20% for enhancement, then asked: “Not having a disease sounds like a pretty good enhancement to me.” Jurisdictional arbitrage and economic incentives—from stronger bones to Diamandis’s deliberately provocative seven-foot NBA example—make local prohibitions unlikely to settle the issue.
Fable 5 and Mythos 5 put Anthropic back in the model lead for the moment, while Apple’s Gemini-powered Siri concedes that personal context—not model ownership—is the immediate interface moat. Mythos 5 is the less inhibited model; Fable 5 adds broad safeguards, sometimes falling back to Opus, while the frontier product’s price reportedly doubled. Apple’s upside is eventual localization and on-device inference; its bear case is losing both the user interface and chip capacity: “I’ll believe that it’s good when I see it.”
Deep dive
1. Bitcoin’s $60,000 floor call rests on a split identity
Armstrong’s explanation for the drawdown had three legs: AI absorbed risk capital and attention; regulatory clarity after the GENIUS Act made stablecoins “the new meta”; and optimism that growth might contain inflation weakened Bitcoin’s familiar inflation trade. His discipline was deliberately anti-cycle: “It’s never as good as it seems, never as bad as it seems.”
Diamandis pressed the original countercyclical promise: shouldn’t wars or stock-market crashes lift Bitcoin? Armstrong still believes that thesis, but said it has taken longer than even he expected. Perhaps 30% of capital treats BTC like gold, while 70% still treats it like “some higher-volatility tech stock”; those ratios must shift before the hedge behaves consistently.
Armstrong’s hedged price call was that Bitcoin has “probably” bottomed near $60,000, though “nobody can say for sure.” Citi projected as much as $189,000 by end-2026, Polymarket implied $84,000, and Armstrong judged roughly $100,000-$200,000 plausible by year-end. He also said Bitcoin should be much higher by 2030.
AI and crypto also compete physically. Bitcoin ASICs cannot simply run AI workloads, but both systems draw on scarce energy and future fabrication capacity at TSMC; Ethereum, by contrast, became “99.9% more energy efficient” after moving to proof of stake. Jens Nielsen argued that enormous AI offerings and US data-center spending are pulling global risk capital toward the US, even as Bitcoin remains practical in hyperinflating economies.
2. Quantum migration forces Bitcoin to choose between property and stability
Armstrong called the danger non-imminent but eventual quantum capability “almost certain,” with consequences for cryptography across the internet, not Bitcoin alone. Bitcoin Core developers have BIP 360; Ethereum has a roadmap, and Armstrong estimated the upgrade work was roughly 20% along. Solana is pursuing a similar upgrade.
Coinbase’s quantum advisory council includes Stanford cryptographer Dan Boneh, Scott Aaronson, Justin Drake, Yehuda Lindell, and others. Armstrong likes BIP 360’s quantum-resistant approach, but it enlarges blocks—reopening one of Bitcoin’s most historically contentious design questions.
The deeper dispute concerns early addresses. One camp would impose a migration deadline and freeze anything left behind, treating those coins like “a ship full of gold sinking to the bottom of the ocean” rather than letting a quantum operator seize and dump them. The opposing camp says seizure resistance is so fundamental that vulnerable coins must remain a bounty, even if their owners failed to migrate.
A third proposal would freeze old coins but leave an appeal mechanism for owners to prove their claim later; Armstrong said those details remain ambiguous and no hard transition date exists. He estimated the exposed early holdings, including coins associated with Satoshi, at roughly 5%-10% of supply—not 80%—and noted that many people already assume their keys are lost.
3. AI agents are becoming financial actors before they become legal persons
Diamandis began with stale figures—3.1 million agent transactions and just over $1 million transferred—but Armstrong updated them to approximately 100 million transactions and $50 million. His instruction to businesses was direct: prepare to accept AI agents as customers because this activity is “growing quickly.”
Coinbase’s first layer lets users connect an LLM to their account through an MCP API and command-line interface, giving it account context and the ability to make changes, trades, or payments. Coinbase Advisor is the second layer, handling requests such as rebalancing and tax-loss harvesting or suggesting a higher-yielding DeFi protocol. The third gives each agent its own instantly created, self-custodial Base wallet without conventional KYC.
Armstrong favors the “polytheist” AI future: specialist models with limited context will handle coding, science, design, manufacturing, or robotics, then report upward through orchestrators. Swarms therefore need payroll, contracting, and value transfer; stablecoins will probably become their default settlement layer, and the AI economy may eventually exceed the human economy. Even interplanetary settlement could remain on-chain, though Earth-Mars transfers would be delayed rather than real time.
Alex asked whether changed KYC rules would push Coinbase toward conventional banking for agents. Armstrong said liability and legal precedent remain open: initially, responsibility might roll back to the controlling human or company, while truly autonomous agents could force a new legal category. His fraud-control idea is an on-chain reputation graph; Diamandis compared it with PageRank and suggested signals such as refund history and merchant reputation.
4. Government equity in frontier labs could become an emergency tool with permanent costs
Diamandis saw a 10% precedent in existing federal holdings: 10% of Intel, Lithium Americas, Trilogy Metals, USA Rare Earth, and Korea Zinc, plus 15% of MP Materials. Trump had called stakes in AI giants “a beautiful thing” and suggested distributing some economic gains to citizens.
Dave’s position was intentionally conflicted: strategic investment may be justified by an AI race resembling wartime mobilization, but the long-term precedent is “horrible, horrible.” A later administration would inherit both the stock-picking power and the decision of when to sell; dumping giant positions could damage the same companies taxpayers supposedly own.
Wissner-Gross argued that if OpenAI and Anthropic become larger than the rest of the US economy, some quasi-nationalization, golden share, or formal public-private structure may be inevitable. He distinguished AI from nuclear regulation: superintelligence originated in private companies, not Hiroshima, and the government reportedly declined an earlier opportunity to fund OpenAI.
Armstrong prefers government as rule-setter and customer: Coinbase already serves about 140 federal, state, and local agencies. Equity might make sense if national-security funding were necessary for a company to exist, but routine ownership would turn government into a capital allocator. Diamandis sharpened the political problem: government investment followed by corporate campaign donations creates “the most toxic circle” imaginable.
5. A sovereign wealth fund promises cohesion but invites politicized allocation
Sanders’s proposal would transfer 50% of leading AI-company equity into a public fund; Sam Altman’s quoted response was, “That’s not going to happen. That’s way too much.” Diamandis nevertheless thought a negotiation had begun around 10%. Dave called 50% “utterly insane” and stressed that government can already extract AI wealth through taxation without owning shares.
David Friedberg supplied the labs’ pragmatic case: a one-time 10% contribution around an IPO might politically inoculate them against recurring UBI claims—“I already gave up my pound of flesh.” He also noted that a regulator profiting from incumbents could disadvantage startups outside the portfolio.
Alex Karp imagined 5%-10% golden shares later diversified into an S&P 500 or total-market fund, but described the underlying problem as a structural incompatibility: the government may need to treat AI labs like civilization-scale utilities without having a governance structure suited to them.
Armstrong liked citizens receiving “skin in the game,” perhaps through a share at birth, because common ownership might reinforce social cohesion and free-market alignment. His concern was durable management quality: later governments could redirect capital toward political projects. A strict indexing rule such as “keep it in the S&P 500” would reduce, though not eliminate, that risk.
6. OpenAI may need its IPO before the market has room for it
OpenAI had filed its S-1, with Polymarket assigning 46% probability to a valuation of at least $1.5 trillion and 26% to no offering this year. It would follow Anthropic and SpaceX, the latter expected around $1.77 trillion. The earlier anxiety centered on roughly $600 billion of compute contracts and whether OpenAI was ready for public-company scrutiny.
David Friedberg’s answer was that predictability itself has compressed: CFOs once wanted three or four quarters of visibility, but in “singularity time” three to five months may be the best available. Going third during a liquidity crunch is dangerous, yet waiting for two enormous offerings to digest could take years while funded rivals accelerate. “You’ve got to go.”
The logistical comparison was Facebook, whose shares fell roughly by half after its IPO despite later success because liquidity was inadequate. A UBS investment chief reportedly described only $75 billion as liquid while forthcoming offerings seek hundreds of billions at multi-trillion-dollar valuations. The panel’s question was not merely whether these are good companies, but whether enough money exists at the required moment.
Alex Karp inferred that OpenAI had improved its prospects by emerging from “code red,” cutting Sora and its AI-for-Science work, emphasizing Codex, and converting Stargate from owned infrastructure toward leasing. Armstrong defended healthy founder-CFO tension and 70%-attainment stretch goals, but warned retail may absorb valuation risk. Diamandis argued that within 12-18 months, 80% of workloads could use models that are 99% cheaper.
7. SpaceX is monetizing compute now to buy a route back to the frontier
Google was said to be paying SpaceX $11 billion annually through 2029 for 110,000 NVIDIA GPUs in an xAI data center—despite Google’s own TPUs and infrastructure. One observer’s summary captured the valuation effect: a single contract moved the SpaceX IPO story from roughly 100 times revenue to 50 times revenue.
At recording, Grok 5 had slipped from a promised March arrival into June while compute was rented to Anthropic and Google. Wissner-Gross had called Grok “on life support,” though he expects Elon and xAI to return to the frontier. His speculative strategy: become a hyperscaler, let today’s algorithmic wars burn out, accumulate hardware and revenue, then regain the frontier with 10 or 100 times competitors’ compute.
The panel cautioned that the pools differ: Colossus 1 mixed H100s and other generations, while Colossus 2 was intended as a purer cluster; Google also needs CUDA capacity for customers unwilling to use TPUs. The Anthropic arrangement was described as cancellable month to month and Google’s as three years, leaving Elon strategic flexibility. Armstrong’s conclusion was that the bottleneck has shifted from model design to infrastructure.
8. AI1 turns data-center expansion into a launch cadence problem
SpaceX’s AI1 design was presented as a two-ton satellite delivering 150 kW of peak compute and roughly 70 kW per ton. Its physical scale is equally striking: a 70-meter wingspan, 110 square meters of deployable radiative cooling, laser links, and an integrated micrometeorite shield.
Elon Musk’s description was that an AI satellite is “essentially a lot of solar cells,” with fewer complex antennas than Starlink and much of the enabling technology already present in Starlink V3. Armstrong’s hardware questions were practical: compartmentalized redundancy against micrometeorite punctures and the “beautiful origami” required to fold the structure into Starship.
Wissner-Gross saw a first-generation, “mainframe-era node”: most surface area serves power and heat rejection rather than compute. He contrasted its ton-scale form with a roughly 20 kW Cerebras wafer-scale engine and imagined compact fusion, power beaming, and better radiators radically shrinking later generations. The present design demonstrates how far the system remains from efficiently turning solar-system matter into computation.
SpaceX’s original filing contemplated 500,000 satellites before rising to one million; Diamandis translated that into approximately one Starship launch per hour, 24/7, with a target launch cost of $100 per kilogram. Armstrong’s category shift was the cleanest: AI infrastructure stops being a real-estate problem and becomes a launch problem, making SpaceX “civilizational infrastructure” rather than merely rockets or internet satellites.
9. Vertical integration expands the SpaceX thesis from Texas to the Moon
The announced Texas GigaSat campus covers 1,000 acres with capacity for 11 million square feet of facilities and AI1 production targeted for late 2027. It would integrate solar ingots, wafers, cells, and finished satellites. Diamandis’s favorite specimen of Musk’s approach was the PICA supplier dispute: “I’m going to make it myself and put you out of business.”
Armstrong argued that integration is preparation for off-Earth manufacturing, not just supply-chain control. SpaceX concept videos showed an electromagnetic lunar launcher sending orbital data centers out “bam, bam, bam.” Manufacturing solar and compute systems on the Moon would exploit better delta-v economics and build a true Dyson swarm without lifting every kilogram from Earth.
Armstrong saw a powerful loop if Dario wins the race to self-improvement while Musk supplies factories, launch, and orbital compute. He was skeptical that Musk has ever partnered for the long term, while Wissner-Gross thought future events may make the five-year-handshake question itself look quaint.
Morgan Stanley’s projection took SpaceX revenue from $18.7 billion in 2025 to $3.4 trillion in 2040, implying perhaps $50-$100 trillion of value; Polymarket put first-day value near $2.13 trillion. Diamandis believed the long-term thesis but flagged the forecast as IPO fundraising material from a bank positioned to profit—and again worried that retail could bear the path volatility.
10. NewLimit uses AI to separate cellular age from cellular identity
Armstrong traced NewLimit to Coinbase’s 2021 IPO. With liquidity and a ten-year horizon, he saw strong teams already attacking AI, fusion, space, and brain-machine interfaces, but longevity looked underfunded. Dinners with scientists and biotech executives surfaced epigenetic reprogramming; Armstrong, Blake, and Jacob Kimmel co-founded the company, with Kimmel now serving as CEO.
Shinya Yamanaka’s 2012 Nobel-winning work showed that four proteins could turn an old skin cell into a young embryonic stem cell, changing both age and type. NewLimit’s framing is “half of what Shinya Yamanaka did”: preserve cellular identity while restoring the function the same cell possessed when younger.
The discovery engine begins with AI exploring approximately 10 quadrillion protein combinations. It first learned from published literature and now ingests NewLimit’s wet-lab data, which Armstrong called the largest dataset in the field. Pooled screens identify cells that look younger; slower functional assays test whether a liver cell can again process caffeine, acetaminophen, and alcohol like a young cell.
Successful hits progress toward nonhuman-primate and human testing. Armstrong expected five to ten years of basic research, but said human-cell reprogramming is already demonstrated and first drug candidates should enter the clinic next year. The $435 million raise buys multiple shots: a typical phase-one trial may cost $10-$20 million, with phases two and three rising from there.
11. Longevity escape velocity may arrive before consensus recognizes it
Kurzweil’s cited LEV date is 2033. Armstrong had not independently modeled it and does not expect escape velocity this year or next, but considered 2033 possible because Kurzweil has “been right on so many things.” His constraint was regulatory rather than computational: “The AI part is going really fast”; FDA approval is slow.
Wissner-Gross expects LEV to be “spiky,” like AGI. He cited, with explicit caveats, a double-blind, placebo-controlled study of HIV patients receiving GLP-1 drugs that appeared to show double-digit movement on epigenetic clocks. His question was whether retatrutide and later GLP-1 generations might carry a subpopulation past LEV before society notices, much as models passed the Turing test without ceremony.
Diamandis pushed back on clocks: morning and evening tests can disagree, so a page of biomarkers matters less than restored liver, cognition, immune, and muscle function. He cited the $101 million Healthspan Prize as an effort to measure functional reversal rather than waiting decades for mortality data. Armstrong agreed that society may cross the threshold and continue arguing over whether it happened.
The group’s best reminder of evidentiary lag was that “the baby that’s going to live to a thousand years old is already alive,” yet confirmation could take generations.
12. Embryo editing makes “disease” and “enhancement” unstable categories
The immediate research cited was in-vitro editing of PCSK9, associated with LDL cholesterol, and HBG, associated with hemoglobin, in embryos. Diamandis also invoked He Jiankui’s 2018 CCR5 editing of two children intended to confer HIV resistance, placing today’s base-editing discussion against an already breached ethical boundary.
Armstrong called base editing—single-nucleotide changes with lower error rates than first-generation CRISPR—more promising than first-generation gene editing. He cited a survey in which about 80% of Americans supported embryo editing for disease prevention but only 20% supported enhancement.
Armstrong found that line incoherent: preventing osteoporosis creates stronger bones, which is also an enhancement. His expectation is that treatment begins with disease and “upgrades over time,” potentially reaching traits such as IQ.
Diamandis supplied the uncomfortable incentive case: if designing a seven-foot child implied a 50% chance of reaching the NBA and a $20 million annual salary, parents would travel for it. Jurisdiction compounds the issue because edited children can return from permissive countries. Diamandis distinguished editing from current embryo selection, where Nucleus sequences perhaps 20 embryos and lets parents choose among existing combinations.
13. Fable 5 and Mythos 5 separate frontier capability from permission
Anthropic’s release displaced GPT-5.5 across many benchmarks—“for probably about five minutes,” Wissner-Gross joked, pending something like GPT-5.6. Fable 5 and Mythos 5 share an underlying model; Mythos is less inhibited, while Fable adds scaffolding and restrictions around biology, chemistry, cybersecurity, and other potentially dangerous domains.
Wissner-Gross’s coding test produced a visually polished cyberpunk first-person shooter in one shot, without errors and with a soundtrack. More important to him was pure-vision play of Pokémon games: winning requires spatial and long-horizon reasoning. He inferred aggressive reinforcement learning with verifiable rewards across long codebases, capture-the-flag hunts, games, and other challenges sprawling through both space and time.
David Friedberg’s four-hour verdict was that Anthropic had the stronger model ready and released it after OpenAI showed its hand. The frontier price reportedly doubled, undermining the idea that top intelligence is already a commodity. The output can also be so sophisticated that a human remains reading while the model effectively says, “I’ll just move on while you’re trying to catch up.”
Safety buffers remained broad: innocent biology questions could cause Fable 5 to fall back to Opus, while Friedberg’s attempt to reach remote agents over airplane Wi-Fi triggered a cybersecurity refusal. Diamandis predicted faster leapfrogging as OpenAI and Anthropic approach their IPOs. Wissner-Gross’s consolation was competition itself: two frontier labs are preferable to an uncontested singleton.
14. Apple rents Gemini to protect the interface it failed to modernize
Apple’s multi-year Google partnership rebuilds Siri around Gemini as an agent with access to messages, email, notes, photos, and persistent personal context. An English-focused beta was expected later in the year. Diamandis read the choice as an admission that Apple lost the foundation-model race but may still control the more durable moat: the user’s private context and ability to act on it.
Alex Hormozi’s “glass half empty” view was that Apple has outsourced a core layer of its stack, contrary to its culture of integration. The alternative is to treat models like localized search engines rather than operating systems: Google in the US, potentially a Chinese provider in China, and other regional arrangements elsewhere.
The new Siri still needs Google’s cloud because inference is too intensive for Apple’s own cloud-plus-device stack. Diamandis argued that if model costs fall by roughly 40 times year over year, equivalent capabilities could eventually be distilled onto future iPhones and run at the edge.
Dave Blendon’s bear case linked software and hardware: if Siri’s interface becomes Google while higher-paying AI customers pull TSMC capacity away from Apple chips, Apple could lose both ends of the product and push users toward Android and Samsung manufacturing. After years of failed Siri promises, Salim Ismail closed with the appropriate evidentiary standard: “I’ll believe that it’s good when I see it.”