Social Trades

“Anthropic started turning a profit in Q2. … In the case of Anthropic, the revenue has gone as high as $50 million per megawatt.”

• Unit economics inverted: serving Opus 5/Fable 5 generates up to $50M/MW versus a $10-15M/MW base compute cost, so every inference dollar funds more training.
• Compute footprint 3-4x'd from under 2GW to over 5GW this year; with OpenAI it takes 40-50% of incremental world compute next year and could hit ~100GW combined by end-2028.
• Revenue per megawatt projected to $70-80M blended by end-2027 and potentially $100M+, letting Anthropic outbid everyone ($25-50M/MW) for compute; key invalidation is regulation blocking release of its best models (Mythos/Model 2), which stalls revenue per megawatt.

“OpenAI could start turning a profit even, with the bigger rise of Codex and 5.6 and all this.”

• GPT-4 on Hopper was negative gross margin; GPT-5.6 now generates well beyond the incremental $10-15M/MW compute cost, so margins have skyrocketed.
• Started 2025 at 2GW, above 5GW by year-end; with Anthropic, taking 40-50% of incremental world compute next year and heading toward ~100GW combined by end-2028.
• Invalidation: regulatory holds (not releasing Astra, pausing training two weeks) stall revenue per megawatt and its ability to outbid others for compute.

“We’re at a little bit over a trillion dollars of CapEx this year. As we go out into ’28, it’s going to be more than $2 trillion.”

• World AI compute additions ramp 30GW this year, 50GW next, 70GW in '28, 90-100GW in '29; a gigawatt of compute generates ~$100B of revenue against ~$6B of fab-level CapEx, a ~100x discrepancy driving relentless expansion.
• Modeling shows $11T of CapEx 2024-2029, $6T cash-funded and $5T of credit that must be issued, plus power plants and shells built years ahead push effective annual CapEx toward $7-10T.
• Downward pressure: capital constraints, data-center regulation (NY bans, Texas moratoriums, Ohio property-tax rules), and rising interest rates bend the buildout below what pure economics wants.

“They do have to start paying $25, $30, $50 million a megawatt to really gobble up 70% of the world’s compute in 2028”

• Anyone can profitably serve open weights at $10-15M/MW today (GB300 + Kimi weights + OpenRouter), so pricing is already inflecting up.
• For labs to reach 100GW by 2028 they must pay $25-50M/MW; SpaceX already sold compute to Anthropic/Google at $25-40M/MW, recouping full CapEx in a year.
• Bullwhip effect: Nvidia, SK Hynix, Micron, Samsung, substrate makers all raise prices as the supply chain rebalances, though most compute still transacts sub-$20B/GW next year because it is contracted and financed before it's built.

“I think they’re like a $1.5 trillion company. It’s like, what? Silly. They’re worth way more than that, at least in a logical sense.”

• Meta and SpaceX are the only plausible #3s: they build compute speculatively without a pre-signed customer, giving them optionality to rent at $25-50M/MW instead of $13.
• Cash flows plus hoarded infrastructure are worth far more than the current market cap whether their own lab works or they sell compute to Anthropic/OpenAI.
• Funding shift: Meta stopped buybacks and raises debt at 5-6% for CapEx and would happily pay 8% given the return on compute.

“He was like, "Hey, Anthropic, I know you’re making $60-plus billion per gigawatt. Why don’t you just buy my stuff for a crazy amount of money?"”

• Big new compute entrant next year, leasing heavily to Anthropic and OpenAI, who have the marginal capability to pay the highest price.
• Alongside Meta, one of only two plausible #3 players because it can build compute without a pre-committed customer and then sell into shortage at crazy margins.
• Elon only sold because pricing hit $25M+/MW; wouldn't have at $15M.

“China domestically still continues to have sub-10% of incremental new compute. In 2028 it might start to inflect up, I think. But it’s pretty easy to say China will have 30 gigawatts of AI compute or less.”

• Through 2026 China relies on smuggled chips, TSMC wafers diverted to Huawei, and Samsung HBM; domestic SMIC/CXMT fabs only ramp in '27-'28, adding 5-10GW of inferior chips in 2028.
• Quality-adjusted, China's 2029 potential 50GW is worth ~20GW of American chips, so a single leading US lab in 2028 may exceed all of China's compute stock.
• Longer-term risk: China hockey-sticks via massive subsidies and manufacturing scale-up, so a slow takeoff lets it catch up drastically on semiconductors.

“Those countries, like Pakistan or Nigeria, I think are just going to be very fucked in this new interest-rate regime.”

• AI infrastructure turns $1 into $10-100, so hyperscalers happily pay 8% versus today's 5-6%, repricing credit spreads ~250bps for the entire economy.
• A 5-point rate rise plus $2T/yr federal borrowing pushes US debt service from 20% toward 60%+ of tax revenue; countries with heavy short-duration debt and thin tax bases default, echoing the ~40 Latin American defaults of the Volcker era.
• Higher discount rates crater DCF-heavy stable-cash-flow equities (the Berkshire/Johnson & Johnson/railroad type) even as the AI complex thrives; banks lose as liabilities reprice faster than assets.

“It’s an argument for why — I think memory is going to do great — memory stocks shouldn’t 10x or whatever again.”

• Value capture has shifted: memory companies now capture more than TSMC and are raising prices very quickly to fund incremental capacity, versus making no money on HBM in 2023.
• If AI demand for memory is that extreme, interest rates spike and everything in the economy trades at 2-3x multiples, capping stock upside for Micron/Hynix/Kioxia despite great earnings.

“It’s really by the end of next year when half of the incremental compute is already going to Anthropic and OpenAI.”

• Lab compute triples yearly versus world compute doubling; labs shift allocation from inference toward training/research (already visible as Anthropic's ARR adds plateau while compute keeps climbing) because building AGI beats dividends.
• Economies of scale, continual learning from deployment, and best-model-builds-next-model dynamics all compound toward one or two labs controlling most usable FLOPs, and the effective AI population at frontier labs grows ~10x/year.
• Second-order signal: hyperscalers and compute merchants (Meta, SpaceX, clouds) become sellers to the labs at ever-higher prices rather than competitors.

“I think if anyone had $400 million and the ability to convince ASML to sell them an EUV tool, they should totally just go buy one, wait, and then sell it for north of a billion dollars.”

• 100 EUV tools a year for 2030 is the current supply-chain plan, but the economics say it should be far more; Carl Zeiss mirrors are the whip's tail and won't be capital-injected this year, next year, or the year after.
• Wafer fab equipment supply chain does ~$200B/yr against $2T+ total CapEx, so a ~100x discrepancy exists between fab CapEx and end AI revenue generated.

“longterm, I think, you know, we're going to end this year at maybe 50/50 background and uh and real-time workloads, but I see this going to 9010 in favor of background.”

• Test-time compute scaling means agents give better answers with more time; task lengths are compounding from minutes to hours, drawing an exponential toward multi-day runs.
• Background tokens are unbounded - no human in the loop caps consumption - versus real-time coding sessions already saturating human attention.
• Deep research over 10,000+ sources and autonomous cybersecurity pentesting agents are the lead categories already pulling demand this way; the whole Sal Research token-factory business is built on this shift.

“People buy tokens because they're immediately valuable to them. You don't hoard tokens, you use them immediately … I do think inference spend monotonically increases. Uh there's no speculation on inference spend.”

• Direct rebuttal to the investor worry that semis at ~20% of the S&P 500 must mean-revert like Cisco/Intel in 2000: the 2000 networking capex anticipated demand that never came, whereas token buyers are rationed by caps today.
• Differs even from the 2023-24 Hopper supply crunch, which was training-oriented and inherently speculative; current demand is inference-driven with providers instituting usage caps on products like Claude Code.
• Implies durable demand for the chip/memory/datacenter complex exposed to inference.

“I think they've underpriced the impact of how short we're going to be on HBM. We're going to push really hard in this other direction instead. Which, you know, as a as an aside, I do think is probably the thing to attack most.”

• No easy way to bring on more memory fabs; 'the boys in Boise' (Micron) refuse huge capex for a cyclical market after past burns, so supply stays constrained for a while.
• Consequence is prices rise across the system - iPhones cut memory or go up in price - rather than demand routing around the shortage; own chip strategy is built around sidestepping HBM via flash offload.
• Supports the recap that a lot of money was already made in Micron/SK Hynix on this shortage.

“increasingly we're seeing a few new thinkers realize that inference is going to be suitable for these distributed 1 megawatt data centers and uh we're we're quite in agreement with that and we are very happy to buy small pools of compute across the United States”

• Diseconomy of scale: gigawatt data centers are effectively unbuildable in the US, 100MW is nearly impossible, but 1MW is plentiful and liquid cooling packs a megawatt into ~8 racks.
• Actively buying this capacity as an inference fleet, cutting diesel backup, redundant fiber, and SLAs - even accepting 95% (or 80%) uptime and intermittent solar/wind at the right price, since a robust control plane reroutes background workloads.
• Training demanded concentrated power, so incumbents ignore aggregate distributed supply - that's the scavenger arbitrage versus bidding against Anthropic/OpenAI for legible compute.

“fundamentally I don't think open source ever goes away. If there's a vacuum because one leader steps out, a new leader will step in. There's too much incentive and too much there's a lot of tailwinds too. It's just it gets easier every day to treat to train a frontier class model.”

• Preventing diffusion of model capabilities is fundamentally impossible: an ever-larger share of GitHub is AI-generated output, enabling latent distillation without scraping any lab's API.
• The closed labs' premium for being 3-6 months ahead is fragile because enterprises don't adopt at that cadence - many still run models several generations old.
• Customer demand for owning intelligence - weights and sovereignty no one can take away - is a stated core tailwind for the whole business.

“Basecat was at 10 mil when they did this. This is a complete narrative vibe shift from what Coinbase typically does. Coinbase does not really list memes, like rarely.”

• Coinbase historically listed memes late (e.g. Virtuals near $2B); listing at $10M market cap is a structural change in behavior.
• The shift is a competitive response - Robinhood's Cash Cat listing 'woke Base up' - and signals a full-blown listing war for the rapidly growing pool of onchain traders.
• Second-order bullish for onchain memecoin traders: earlier centralized-exchange listings mean more upside for low-cap holders.

“Robin Hood has showed up and Vlad Tennov is moving militant with a plan with a focus and brother they're moving militant. Okay. The cash cat thing was one of the I mean it was a master class”

• Cash Cat rollout was a masterclass: follow, tease, soft shill, then spot listing - it forced Coinbase/Base to respond.
• Robinhood Chain is already top five in DEX volume weeks after launch, with tokenized stocks (Tesla, Nvidia) usable onchain across 120+ countries and active Uniswap pools.
• Contrast with historical corporate entrants into crypto that 'have no idea what they're doing' - Robinhood is moving with a plan, a bullish signal for its crypto volume capture.

“enthropic had the mandate of heaven for like two years. Okay. And had it for two years and they lost it. And they lost it very recently because they abused it.”

• Leveraged Trump admin to try to get competitors banned - blatant abuse of its standing.
• Dario's claim that 50% of corporate jobs would be gone to AI within a year produced basically zero job loss and no walk-back.
• Its cancer-cure pivot was undercut two days later when Moderna announced a cancer breakthrough, the nail in the coffin for its narrative leadership.

“Clarity just went from 78 to 18%. It feels that they've made setup and ramifications for what to do when this inevitably fails because this doesn't feel like it is going to pass.”

• Passage odds nuked from 29% to 18% intraday - 'somebody knows something.'
• CFTC directing staff to codify a crypto asset market structure under existing authorities is the contingency plan for the bill failing on Democrat obstruction; 'it's not getting passed, they're going to just do their own thing.'
• The regulatory workaround itself is treated as crypto-positive even as the legislation dies.

“they have the leading model they're saying it's free on open router that is nuts. China is now basically mogging on LLMs.”

• The stealth model is a frontier coder with 1M token context and text/image/video input, and all user reports call it incredible - offered for free.
• Old thesis was US on the frontier with China distilling and catching up six months later; a free leading Chinese model inverts that, which is 'kind of terrifying' for US AI incumbents.
• Paired with the weekend's Chinese robotics masterclass, China's momentum extends beyond LLMs.

“Take 2 Interactive has subpoenenaed Microsoft, Google, Discord, and X attempt to discover the identity and location of the hacker who appears to have access to a full developer build of GTA 6”

• The leaker has a full copy and a second group allegedly has the game on an Xbox dev kit - if true, the worst-case scenario for Rockstar's launch control.
• The aggressive subpoena campaign signals this is not an internal marketing bit; the hacker faces federal-prison-level exposure.
• Second-order: the GTA 6 leak memecoin ripped after the hacker burned all the coins, feeding the onchain leak narrative.

“And there's a bull market in AI stocks right now. So the same things we used to be doing with meme coins or other assets being fully market usual, we're doing with tokenized Nvidia and other assets like that.”

• Yield is a derivative of market activity - the AI-stock bull market means traders pay up for leverage, so servicing tokenized equity flow on-chain earns wide spreads.
• The fund is actively running these strategies market-neutrally, migrating the playbook it used on meme coins to tokenized NVDA.
• Banks are in talks to lend stablecoins into the fund's vaults and to tokenize equity/credit, showing TradFi-DeFi convergence accelerating the opportunity set.

“and what our research has shown is that all the hacks or 99% of hacks that actually happened wasn't AI finding a way to break a smart contract. It is AI finding legacy OPSSEAC vulnerabilities and exploiting them.”

• 99% of hacks exploit legacy opsec vulnerabilities, not novel smart-contract breaks - AI finds '100 old banks with wooden doors', not a drill through the metal vault.
• Dollar losses are normalized below 2% annualized even as event counts hit records; ~50 of the ~60 hacked platforms this year were too small to even be on the fund's radar.
• Conclusion: the space is trending toward security, not away from it, keeping properly risk-managed DeFi yield (fund earning 11-12% consistently) viable.

“um on the crypto economy itself where we think actually it's likely to accelerate >> as in we're more bullish the market rather than bearish and that's why we want to stick to our goss.”

• The fund is deliberately declining to expand into TradFi strategies because it expects the crypto economy to accelerate, making on-chain liquidity provision the better use of its edge.
• TradFi-DeFi convergence creates more opportunities for on-chain lenders, market makers, and traders - the firm's core strength.

“like something like CCOI uh had a massive ripper because I think actually a bunch of people just it it popped up super early on Claude and people bought it and then it actually just did not materialize at all and like sold sold off 70%”

• Mechanism: LLM-driven screening funnels retail into shallow downstream-supplier theses, inflating names like CCOI without real fundamental confirmation.
• The fundamental thesis 'did not materialize at all', producing a ~70% selloff - AI research tools create false positives that only domain experts can filter.

“G10s are trading like emerging markets. They're engaging in emerging market-like behavior, pumping hyper liquid, loosening SEC regulations. I'm like, "This is like a perfect crypto bull market."”

• Governments 'have to print money' while G10s exhibit emerging-market behavior, including pumping Hyperliquid - a debasement backdrop favoring crypto assets.
• SEC regulatory loosening removes a structural headwind for the sector.
• The setup lands 'during the point of absolute despair' with investors sidelined and the AI/inference narrative exhausted - contrarian timing bullish for exposed assets like BTC, IBIT, and HYPE.

“Prediction, the US government will announce a hundred billion-dollar purchase program of Bitcoin and a nationalization of strategy within the next 24 months.”

• Engages with the thesis beyond reading it: 'obvious retweet', only complaint being the 24-month timeline is too long.
• Flood's framing: even the worst case (MSTR premium collapse, forced BTC sale at a discount) ends with a 'triple dream bonus' if the Treasury buys the stack to fund a national Bitcoin reserve.
• Sovereign accumulation narrative reinforced later: Flood 'screaming sovereign nation global power race to acquire more Bitcoin, China versus USA' in chat.

“Clarity Act is now up to 28%, which is Look, it's not nothing. It was as low as 13% in August 7th.”

• Odds moved from 13% (Aug 7) to 28% - 'it's not nothing' - and nobody on the timeline is talking about it.
• Rising passage odds are second-order bullish signal for crypto-exposed names being tracked in the same segment (Circle, Coinbase, Robinhood all ripping).

“Where does that leave frontier model companies? They're the most at risk. Why? Because they're not investment grade. They're in the worst position in terms of their balance sheets. They're not in a position to then get access to the compute they need to be successful.”

• Three converging pressures: AI doomerism seeding mainstream political pushback (Abbott and Shapiro data-center orders, GOP memo on the Ohio Senate race), rising 30-year yields raising the risk-free rate, and data-center shutdowns constraining compute access.
• Frontier labs are non-investment-grade with weak balance sheets, so growth compounds the problem - more capacity needed to sustain growth they cannot fund as negativity turns capital away.
• Dario's own fearmongering (50% entry-level job-loss claim, engineered blackmail study) is portrayed as the source of the backlash, making Anthropic's regulatory-capture push self-defeating and discrediting.

“is there are all kinds of ways of using open-source technologies that are meaningfully more performant and dramatically cheaper than the closed-source alternative.”

• Closed-source models decay in capability when wrapped in their own harness, while open-source models improve when combined with other open-source models - evidence expected to become a 'tsunami' over the next year.
• Open models expose thinking tokens, letting third parties see misalignment in real time, undercutting the closed labs' safety-based case for regulation.
• Key risk: Sacks warns an effective open-source ban is coming via 'equal standards' applied through a FINRA-for-AI standard-setting body, which open models cannot comply with, gradually shutting them out of the market.

“this is a new regulatory agency that reports to the government that does pre-release testing and approval of models. I call it a DMV for AI because I think what's going to happen is all these models are going to get lined up in a queue waiting to get their test done and then released and it's going to slow us down horribly.”

• FINRA is a sleepy incumbent-protecting agency; an AI version would promulgate standards, do pre-release testing, and get codified in law - the step-by-step path to regulatory capture Dario is pushing.
• FAA/FDA approval cycles take years; China's faster regulatory posture is already showing in rockets and biotech onshoring, so the same slowdown in AI cedes the lead.
• Recursive self-improvement makes agency approval a fool's errand anyway - labs would just relocate compute to friendlier jurisdictions (Iceland, Kazakhstan, offshore), so the US should attract and retain frontier labs, not chase them off.

“I happen to believe having looked at this and thought about it, I think that it is possible. I think it is likely to happen soon. People I've talked to in industry do think it's real.”

• Models are decomposing into biology-like networks of smaller sub-models; RSI is the perfect system for searching that architecture space via automated test-and-deploy without a human architect.
• All RSI needs is chips, power, and a connection - it can be spun up in any permissive jurisdiction, which is why regulatory bodies and data-center bans cannot stop it and the US should host it domestically.
• Time dilation in AI development makes breakthroughs unpredictable - step changes like solving eight unproven math theorems in a week can arrive suddenly.

“I will predict it now, they're going to ban Waymos in New York, Boston, DC, and they're going to limit them with licenses.”

• Autonomy extracts driver wages ($16-21/hr) that previously stayed in local communities and funnels the money back to Silicon Valley, concentrating wealth and fueling anti-tech politics.
• The same populist anger already produced data-center restrictions from Abbott and Shapiro; robotaxi job displacement is the next flashpoint.
• Backdrop of accelerating rollout: Waymo doubling cars on the road every ~6 months, Tesla about to go cybercab, Zipline-Uber targeting a million deliveries a day.

“The cursor the cursor was decent. Yeah, it was like an appetizer but space x was the main course. That's like 15 17 billion.”

• Recent distributions stack up: OpenRouter ~$600-700M, Cursor a solid win, SpaceX ~$15-17B - the firm is 'crushing it'.
• The Section 8 interlocking-directorate probe (Ben Horowitz on Databricks, Casado on Fivetran) is dismissed as ticky-tacky: startups pivot into competition routinely, firewalls or an independent director fix it trivially.
• Merely opening a DOJ file makes a firm 'technically under investigation'; the story reads as a rival's PR ratchet leaked to Bloomberg.

“I think the Senate right now is a toss-up and I would say that the Republicans will ultimately keep it if I had to predict right now.”

• Patrick Ruffini's analysis of 3,000+ polls across four cycles shows an average D+3.7 polling error, with the bias worsening via oversampling of young progressives.
• Post-Labor Day campaigning gives Republicans a strong message set: sealed border, murders down ~20% to 70-year lows, overdose deaths down 20%, 17% larger tax refunds, small-business optimism at a near one-year high.
• Split view on the show: Freeberg sees an affordability-driven socialist wave (63% living paycheck to paycheck) carrying far-left Democrats through 2026 into 2028; Chamath distrusts both pollsters and media entirely.

“And ultimately there is going to be some sort of big movement towards more socialist policies that we're going to see play out between now and 2028. It's why I continue to believe I think someone like an AOC becomes president.”

• Affordability is structurally broken: 30-year at 5.3% hitting 20-year records, government spending inflating healthcare, housing, and education costs, and no policy path to change the arithmetic before 2028.
• Asset ownership split drives the anger: top 50% of Americans hold $178T of net worth versus $6T for the bottom 50%; the asset-less were left behind by home and stock inflation.
• Even the right is converting: 53% of conservatives under 40 back government-run grocery stores, and Republican favorability toward capitalism fell from 72% to 61%; Sacks counters that the DSA platform's $71-212T cost dwarfs total US household wealth ($182T), but Chamath notes the wave runs on anger, not arithmetic.

“I definitely think Robin is a sleeper for this cycle. Um, I definitely think they're a sleeper, especially because people are kind of like defaulted to soul right now.”

• Vlad and team publicly support all activity on their chain - meme or RWA - and listed Cash Cat extraordinarily fast, signaling they'll back whatever pops up next.
• If Robinhood converts its app users to chain activity in a way Coinbase never managed with Base, it gets 'super super interesting'; its retail base already made fortunes on Dogecoin in 2020-21.
• FOMO data backs the momentum: 60,000 of the 80,000 daily wallets trading on Robinhood chain come through FOMO's crosschain product.

“we are doing probably $100 million of volume a day. We have 70 thou 75,000 different traders every single day trading on the platform. We do a million swaps every single day and we've grown anywhere from like 6 to 15x on all the metrics”

• Grew from ~$30M annualized revenue and 6-7K traders at raise time to a million swaps/day; perp volume went 10x overnight when markets turned, and it's the fastest Hyperliquid builder code to $100K revenue.
• First workable crosschain trading product - 60K of 80K daily Robinhood-chain wallets trade through it - and its onchain transparency thesis is why crypto social can work where Robinhood Social's opt-in confirmation bias fails.
• Z endorses the core thesis on his own account: social trading becomes the next big social-media trend, and the trader-monetization flywheel (Jordan/LeBron endorsement analogy) makes distribution the moat.

“I think we're going to see billionaires come out of GTA 6 to be honest.”

• Leaker claims an in-game currency tradable for real money; Z frames GTA 6 as World of Warcraft-style economy where time spent tangibly returns value - the metaverse idea was early, not wrong.
• Banks' mechanism: 3 billion users sharing a de facto metaverse means native Nike-esque brands, in-game casinos, and clickable status signaling (skins, houses, gold-check LARPing) generate real revenue.
• Both put near-0% odds this is actually crypto-related - the bullishness is on the digital-economy/digital-IP shift (Banks ties it to why he's bullish NFTs and online collectibles), not a token.

“The fact that it's AI, like this is maybe something the the the tech needs to win over society people”

• Moderna ripped ~200% on a specialized individual skin-cancer trial - 'stocks trading like crypto' - taken as proof AI is doing tangible things after the 'AI is doing nothing' narrative.
• Anthropic CEO (and Musk a day earlier) said AI wins the public over as it actually cures cancer; a cure-level breakthrough would be the biggest legitimacy catalyst possible for the tech.

“we just got an announcement that Bessant we will routinely do buybacks and increased size of buyback … the bond market was testing him today … and he's like, 'No, no, no. We're just getting started, buddy.'”

• QRA guidance shifted from 'no increase' to 'no change' in coupon issuance, opening the door to decreasing long-end supply - not on anyone's cards.
• Buybacks doubled from $2B to $4B and then 'more than 4 billion routinely' intraday when the bond market faded the move - a Draghi-style whatever-it-takes signal.
• Game theory into November midterms implies early innings of escalation; mechanism is de facto QE from the Treasury, bullish for debasement assets (Bitcoin, gold) and needed to keep AI-buildout corporate bond issuance viable.

“what they did yesterday, the dollar just knew … this outright intervention … obviously that leads to a lower dollar”

• Treasury suppression of the long end is a direct debasement mechanism pushing the dollar down.
• Foreign buyers (e.g. European pensions) lose on a currency basis as the euro rises, so a weaker dollar marginally reduces flows into US assets.
• Gold and Bitcoin, as non-US assets, are the pure expressions of this questioning of the dollar.

“Poly Market says 80 something% chance by December … 50 coin flip by October 15th, 74% by October 31st. So it feels like October”

• Polymarket odds with good volume: 10% by September, 50% by October 15, 74% by October 31, 87% by December - October is the likely window.
• Arrives alongside the OpenAI 2027-or-sooner IPO announcement at a weak stretch for tech stocks, echoing the SpaceX 'trifecta IPO top' setup.

“These sports teams are so expensive that they're being bought by funds that don't actually exist. Everyone is just one bad financial moment away from utter disaster.”

• Ishbia borrowed from JPMorgan against $4.6B of UWMC stock now down 70% since the deal - a margin call could force a Phoenix Suns sale.
• Team valuations are so stretched they're being bought by funds that don't actually exist, making margin calls on sports teams a likely new trend.
• Echoes Bass's take that Walter exposed the ending of the private credit Ponzi game.

“Suddenly out of nowhere, they're like, "Look, we're going to double the amount of 30 of of long-end buybacks that we're doing at the very least." So, we're going for two to four billion.”

• Funding mechanism: issue near-infinite-demand T-bills and buy back illiquid off-the-run long-duration bonds - functionally similar to QE, removing duration from the market; T-bill share already ~22% vs the ~20% long-term target.
• The signal outweighs the size: like Draghi's 'whatever it takes,' markets do the work once resolve is believed; Bessent reinforced it live with 'routinely do buybacks and increase size' beyond $4B when the bond market tested him.
• Escalation path runs toward outright debt monetization (Fed buying the bills, ~5% probability) - a max-long trigger for debasement assets; interest expense at a record ~3.3% of GDP and midterm incentives mean suppression continues, benefiting Bitcoin, gold, and hard assets.

“they're you know on a currency basis like even though say US equities go up a bit because of that dollar decrease on a currency basis they actually probably lost money”

• Recent-years marginal flows into US equities came heavily from foreign buyers; a falling dollar hands them currency losses even when equity prices rise, dampening that bid.
• US equities barely moved on the buyback announcement while gold and Bitcoin ripped - the debasement trade routes around US-denominated assets.
• Separately, when the 10-year exceeds 5% the yield-equity correlation flips negative - abandonment of the intervention regime would hit risk assets hard.

“He was told by Trump that he has to take roughly $20 billion of loans off his insurance by December 31st.”

• Insurance float was funneled into riskier businesses Walter secretly owned, inflating their metrics; the scheme is now under investigation and the deals must be restructured into safe bonds.
• Of four exit routes - sell the loans, grow the balance sheet 4x, reinsure a block, or sell assets for cash - only asset sales work, since nobody buys paper under grand-jury subpoena.
• The Lakers sale covers only a third of the hole; roughly $11B more must come from selling Chelsea (~$4B) and the Dodgers (~$8B) in about four months, and the teams are gone for good once sold.

“sells for $12 billion in 2026 and it underperformed the S&P. You would have made more money putting that in the S&P than you would buying the Lakers.”

• Jerry Buss bought the Lakers for $67.5M in 1979; the $12B 2026 sale still trails simply compounding in the S&P 500 over the same period.
• Teams have gotten too big for single owners, and the Walter forced-sale saga shows the liquidity trap of holding wealth in illiquid franchises.

“probably the case for SpaceX AI is probably the weakest because the data center and space play is so highly differentiated. Like if that plays out, it I'm not sure to what extent they need to even have their own model. So why are you wasting billions and billions of dollars in the meantime?”

• If data centers in space play out, capacity can run anyone's model - as shown by already selling capacity to Anthropic - so owning a frontier model adds margin but isn't necessary.
• Buying Nvidia because it's the most fungible (not the best) fits the build-and-rent-out-but-reserve-the-right-to-pull-back game, underscoring the model-agnostic nature of the play.

“Well, ideally we actually flip back to free cash flow funding this. But if there's a gap there, if we don't get there soon enough, then we could have a big blow up, right?”

• Sequence: tech blew through debt markets in about a year, Google is issuing equity, Nvidia assembling a $500B vehicle to tap pension funds and insurance floats - the question is where money comes after that.
• 1870s railroad parallel: duration mismatch made the world run out of money, yet railroads kept operating and their GDP contribution was astronomical - a bubble bursting wouldn't stop AI progress.
• Capex is ~$800B this year and ~$1.3T estimated next year, but money spent today only manifests as compute in 2028-2029, exactly the shipping/memory commodity-cycle setup where payback periods measured in scarcity fail in abundance.

“I think the memory makers probably screw themselves in the long run by creating such a massive target on their back. I've analogized memory makers to Iran.”

• Three-player oligopoly discipline (post-Samsung wipeout of Japan) made them slow to recognize a secular shift in memory demand; supply will eventually be solved.
• Strait-of-Hormuz analogy: using your chokehold once triggers customers to build around you - Apple lobbying for Chinese memory and the #1 algorithmic focus becoming using less memory.
• Classic boom-bust economics with EUV-era lines costing billions raise the stakes of over-entry when capacity finally comes.

“we only buy GPUs when we know there's demand for them. That is a great story to tell. I'm not sure how much that I think is a lot of BS because the reality is is if you've built the shell that money is sitting there. You're not going to let it just sit there.”

• Shipping analogy: once fixed cost is sunk you run the ship at whatever the market bears; paper losses from depreciation don't stop supply, so compute prices go to marginal cost.
• Everyone measuring payback periods in a time of scarcity risks a memory-style washout when simultaneous investment lands as abundance - and even if AI stays supply-short forever, an air gap between capital in and revenue out can still blow up.
• Andy Jassy and Nadella's we-only-buy-GPUs-on-demand framing is directly contradicted by the sunk-cost incentive of the built shells.

“drones is probably our fastest growing business unit right now. You know, our LPR business is just under 50% kind of of our forward-looking revenue.”

• Diversification beyond license plate readers: LPR is now under 50% of forward revenue, with drones, live video, and software driving growth; company raised more money at the beginning of the year and is still growing.
• Churn from the backlash is ~1% (60-70 of 6,000 cities), and roughly 20 cities have already turned cameras back on after realizing they were less safe - the Austin shooting solved via Manor PD's Flock cameras reignited adoption debate.
• Mandatory audit-assistant tooling that catches abusive cops, plus 7-day default retention, positions Flock as the privacy-first leader versus competitors who 'run more reckless' and whose approach 'will catch up to them.'

“you see cities due to budgetary reasons and due to staffing issues kind of leaning in pretty aggressively and there's been some bad cases where the AI's gotten it really really wrong.”

• Unlike booking a Delta flight where a hallucination is a minor inconvenience, a 911 AI bot failing during a home invasion carries catastrophic cost; there is no good way to control these systems yet.
• The smarter deployment - AI only picking up when phone lines are fully busy - is not what's happening; cities are flipping the switch recklessly due to budget and staffing pressure.
• Public-safety AI needs a human in the loop and third-party attestation before deployment; Flock is deliberately going slow on AI features rather than ripping and launching.

“if you have open source and it's available to everybody, not just the rich, and we give you agents and free AI that other people are paying thousands of dollars a month for, hey, maybe that levels the playing field.”

• Meta historically played defense and catch-up on Instagram's teen harms and social-media addiction; the new manifesto shows proactive narrative control instead.
• Giving away agents and AI that rivals charge thousands per month for both serves selfish AI-revolution business needs and levels the playing field - the posture CEOs now need in a media environment where stories go viral instantly.

“the monetization time frame is now like 7 years … So, now they're sleeping, but they ain't getting their money back.”

• No industry benefited more from near-zero rates; that regime is ending under a Fed that won't spoon-feed markets.
• Institutions like Notre Dame with ~50% of the endowment in PE/venture can't get money back and must tighten - a headwind for the asset-gathering model.
• In the listed space, traditional money managers are now outperforming PE stocks, confirming the regime shift.

“if you have a new cycle, I mean, I guess the good thing is it won't be that systemic, but it could be it could be big and it could be big for … disproportionately bad for wealthy people if it goes the other way.”

• Private credit grew from nothing as a regulatory arb when Dodd-Frank pushed banks out of lending, so a credit cycle would hit it directly rather than the banking system.
• Public credit signals stay benign (double-B spreads at new cycle lows) and bank stocks at new highs, so the risk sits outside the regulated system - disproportionately with the wealthy who own it.

“China demand is very weak. Iron ore I think made 52 week lows this week. Chinese market frankly has been very uninspiring … this seems to be a China demand problem, not an AI CapEx problem.”

• Base metals traditionally read as a China signal are diverging - copper/zinc/tin at new highs on AI buildout while Chinese iron ore collapses.
• China's withdrawal from oil purchasing supports the oversupply argument in crude, and China-exposed sectors (European autos and luxury) sit in prolific bear markets.

“something like 70% of the AI hyperscaler revenue is from just those two companies. So if this is the big if, if Open AI and Anthropic ever get in trouble, the ecosystem is in trouble.”

• ~70% of AI hyperscaler revenue reportedly comes from just OpenAI and Anthropic, both totally cash-flow negative and facing Chinese competition.
• All three speakers agree the concentration is 'undeniable' - a single point of failure beneath the hyperscaler capex boom.

“the lawsuit was the punishment … That was the punishment was having to deal with … they missed personal computing.”

• IBM 'won' its 12-year antitrust case but lost the personal computing era to distraction; Microsoft's '90s case ran the same script.
• Meta faces hundreds of suits over algorithm-driven mental-health harm, with two already gone against it - costs are 'adding up' and populist politics targets 'big anything', making lobbying an alpha generator.

“it is 90% cheaper to use these products. I'm using them. the founders I invest in are using them and increasingly corporate America is going to embrace these solutions.”

• JCal uses them personally and his portfolio founders do too; a hyperscaler reportedly redirected a nine-figure, $100B-infrastructure frontier-model effort to open source, echoing corporate America's 20-year Linux precedent against vendor lock-in.
• Gavin's plausible end-state: frontier tokens capture 65-85% of economic value while open source runs ~80% of volume - cheap 150-IQ intellects orchestrated by 250-IQ frontier models make the frontier MORE valuable.
• Direct read-through: bearish pressure on Anthropic/OpenAI pricing power (their premium rests on a 6-month lead), bullish for Nvidia (architectural variety favors flexible GPUs) and for legacy software valuations.

“all the political headwinds I think insure against that outcome because it is so hard to build data centers for all the reason we said … will almost guarantee that there's not an overupp relative to the exponentially growing demand.”

• Water-use claims are off by ~100,000x, data centers stand up their own generation, sell excess back and pay grid-upgrade costs; blue-state electricity inflation stems from decarbonization rules, not AI - the WSJ's Ellendale story shows 10x tax revenue reviving a dying town.
• The main systemic risk is a compute glut (dark GPUs like post-dotcom dark fiber) especially if built expecting $30-50/watt spot, but permitting friction and the moral panic throttle supply relative to exponentially growing token demand.
• Turbine capacity is the binding constraint - Caterpillar, Cummins, GE Vernova, Siemens Energy all expanding fast, jet engines being repurposed off old planes, and near-term power is overwhelmingly natural gas, not green energy.

“if we let this control be centralized, we will lose AI, we will lose geopolitics, we will lose everything to China. It's just that simple.”

• Sax from inside Washington: the Biden-adjacent think-tank frame wanted a two-or-three-company cartel merged with state power; FAA/FDA-style approval takes 5+ years and would erase the 6-month frontier lead that all of Anthropic's token premium depends on.
• China has no obligation to slow down, so global compute governance was always fatally flawed; Zuckerberg's manifesto ('too dangerous to centralize') aligns Sax, Gavin, Elon and Jensen on decentralized, open, individual-empowering AI.
• Read-through: regulation failure is bullish for Anthropic's IPO economics and for open-source competitors; had Dario gotten his regulated apparatus, the company would not be set up for a successful IPO.

“He is a beast at copying, mimicking, and competing. He's going to be the open- source hero from America. He is gonna have the best open source model in the next year. I guarantee it.”

• The manifesto commits to open-source models, free agents for everyone, and 'too dangerous to centralize' framing that Sax and Gavin endorse; go ask MySpace and Snap what Zuckerberg in the rearview mirror means for Anthropic/OpenAI premiums.
• Gavin flags he might take the under on the one-year 'best open model' guarantee, and JCal notes the hypocrisy of demanding the right to distill everyone's data while suing startups that index Instagram's social graph - but the consolidated stance on Meta's AI trajectory is bullish.

“I've heard that over the past 2 months, that growth rate is now over 20% month over month, which if you extrapolate it out over 12 months would be a 10x growth rate.”

• The coding pivot follows Anthropic's playbook of verticalizing where users build successfully on the platform; OpenAI is now growing as fast as Anthropic having made that pivot.
• OpenAI scaled back buildout ambitions from ~$1.4T to ~$600B, and if not yet cash-generating will be imminently per Gavin's token-profitability read.

“There is no world in which I don't see corporations and people in the economy spending five or 10% of the salaries of their employees on the equivalent in tokens.”

• Knowledge-work TAM is $25-65T depending on count, and thus far AI is accelerating growth rather than substituting labor - software-coder job openings are higher than a year ago.
• Macro/value investors calling AI a bubble assume tokens are subsidized; Gavin says the overwhelming majority of tokens are profitable for everyone in the chain, and the Anthropic S1 will make that clear - it's like being bearish on oil at $500/barrel when it isn't.
• The constraint is physical (compute, energy), not demand; Anthropic's quarterly earnings will become the pace car demand signal for the entire capex food chain (SpaceX, Nvidia, TSMC, Micron, SK Hynix).

“you have Caterpillar, you have Cumins, you have Cheeky Vernova, you have Seammen's Energy, they are all expanding capacity fast.”

• Turbine blades come from only a couple of facilities in America and Europe; demand is so extreme that jet engines are being stripped off old planes and repurposed as data-center turbines, keeping private-jet residuals extremely strong.
• Elon personally bought a turbine company and Boom Supersonic is building a turbine business ahead of jets because of the demand; Tesla is building a $10B solar manufacturing plant in Texas while natural gas carries the near-term load.

“If you are a public figure, if you were the CEO of a company and you do not have your own voice on X, you are at risk. It's just that simple.”

• A PR department is no protection once you retire or get hauled before Congress under an accountability agenda; only your own authentic words tell your truth.
• JCal ran Zuckerberg's 6,500-word manifesto through an AI detector - 100% human - and credits the going-direct maturation, wishing it had come earlier on teen social-media harms.

“Socialism is playing. Socialism is making inroads and it's infecting people's brains.”

• The DSP structure exists to shift driver benefits, unemployment and health costs onto taxpayers, giving democratic socialists a valid, resonant grievance unless capitalists self-correct as Starbucks did under Schultz.
• Sax dissents - freedom of contract, $5.20/package and 5,000 jobs at risk, consumers will hate the 'Mamdani surcharge' - but the prediction with a stake is JCal's: the suits win and the narrative plays; direct read-through to Amazon's cost structure and Rahm-style moderates having 'no chance' in the DSA-drifting party.

“There just don't seem to be any moes around LLMs. Users switch between models all the time. Perhaps more importantly, the Chinese LLMs are openweight models that are much cheaper”

• Chinese open-weight models are much cheaper and 'good enough'; enterprises increasingly use them, making token overcharging unsustainable and a price war possible.
• OpenAI plus Anthropic account for ~70% of hyperscaler AI revenue and 25-35% of total cloud revenue; half of Oracle's $600B backlog is OpenAI alone - failure would compress hyperscaler margins and throw the whole AI chain into reverse.
• Timing caveat: the bear case might take a year to play out, and until the risk metastasizes the AI story continues; IPO would finally provide real data.

“The dependency of the hyperscalers on anthropic and open AI is just huge and quite scary given that both companies lose billions and are reliant at this point on raising capital for their survival.”

• No moat around LLMs - users switch models constantly and Chinese open-weight models are far cheaper and 'good enough', threatening a price war.
• Anthropic and OpenAI supply ~70% of hyperscaler AI revenue; their failure would compress Microsoft/Amazon/Google margins and reverse the AI chain.
• This is the single risk being watched most carefully, though the bear case may be a year out - an eternity for markets.

“Nvidia is partnering with six large asset managers on a $500 billion financing push designed to treat compute infrastructure much like the commercial real estate toll roads or other assets you can borrow against.”

• Treating compute like borrowable infrastructure (CRE, toll roads) is a tried-and-true way to build financial infrastructure for a capex-hungry growth industry.
• Shifts data center financing off hyperscaler balance sheets - hyperscaler free cash flow could improve, and participating banks and private credit firms get a revenue boost.
• The deal itself is evidence there is as yet no slowdown in AI capex, supporting the AI infrastructure complex.

“one in every two active wallets on Robin Hood comes from FOMO. That's insane. It's an insane stat”

• One in every two active Robinhood wallets comes from FOMO per founder Se Young - insane distribution for the Robinhood-season meta.
• Raised $75M at a $575M valuation in a bear market and is crushing user growth; its edge is treating attention/content seriously (signing traders, founder interviews).
• Second-order signal: FOMO's success stories (Poor Goat's public 7-figure wallet, broadcast to 280K users) are onboarding brand-new retail traders into crypto.

“So social trading in crypto is going to go infinite for sure. It's great to have competition between two of the biggest mobile apps right now”

• Proof of concept: Roaring Kitty's public GameStop thesis rallied the most retail in stock market history; onchain is a better environment because entries are earlier and cheaper.
• Coinbase (Base wallet), Robinhood (fast listings + wallet), FOMO, and Pump.fun are all collapsing friction to one-click onchain trading - 'the Robinhood moment for stocks in 2020, happening right now for crypto.'
• Mike Dudas extends it: the entire world is becoming entertainment-financialized post-AGI, prediction markets on livestreams are a ripe area, and creator-tied tokens representing attention are inevitable.

“they're a larger asset class um today in terms of like market cap than NFTTS, for example … and at their peak um massively larger than for example like VC funded infrastructure. So the notion that there's zero sum or that you know everybody loses money on them is completely and utterly misguided.”

• Dudas: memecoins exceed NFTs in market cap and massively exceeded VC-funded infrastructure at peak; the 'everyone loses' framing is misguided - winners just stay quiet.
• Contrast trade: baskets of Binance-listed VC L1/L2/infra projects are down ~98% - 'homeless if you invested' - while the moralizers who launched them attack memes.
• Memes/NFTs must coexist with serious onchain products because nothing else on chain is interesting enough to drive wallet downloads and user onboarding.

“what I think Ethereum did a really poor job of the past few years is they didn't really focus heavily on trying to grow the pie. Like a lot I think of the focus was too much on decentralization”

• Focus stayed on decentralization and core crypto ethos rather than onboarding new users or building UI/UX for people who don't know what a seed phrase is.
• Framed as the direct foil to Solana's iterate-fast, market-to-normies strategy - second-order support for the SOL-over-ETH positioning voiced on the show.

“Like, bro, Intel from Intel from 20 to 120 is insane. Like, the president was like, "This company is super important for our future national security. We're investing in this company"”

• Mechanism: Trump declared Intel critical to national security, invested government money, and keeps publicly backing it, so trusted-voice followers pile in - the trad-market analog of KOL-driven onchain flows.
• Used as evidence that trusted voices commanding large followings genuinely move markets, not just narrate them.

“I think the RWAS are the big unlock. It's like the carrot that everybody wants and not enough people talk about. When you look at like revenue … just broke a billion dollars in revenue. It's just RWA.”

• RWA revenue just crossed $1B; Simple Farmer reads it as people hedging against crypto on decentralized rails and is pairing his asset with Nvidia and other stocks for that reason.
• Z: RWAs are easily understandable to institutional capital - the missing piece that brings a lot more money on chain; the whole 'Robinhood season / RWA is the new meta' timeline narrative reinforces it.

“you also need to respect the onchain guys right now because that market is also not going to slow down and that's also going to start to pick up a lot. And I have to think a lot of the experimentation on the dev side is going to come from these like low market cap onchain launches.”

• If you weren't trading onchain you missed most of crypto's alpha last cycle; traders printing $100K daily P&Ls on sub-10M coins show a real, deepening market.
• Prior cycle onchain reach was tiny (WIF peaked ~250K holders vs Coinbase's 100M users); friction removal via CEX wallets and mobile apps expands the funnel massively this cycle.
• Onchain speculation typically precedes devs arriving - liquidity and activity attract builders, not the other way around.

“Xi refuses to deal with the obviously the real estate bubble popping a consumer that's flat on their back and an industrial uh base that makes no rational sense. And so they're I mean literally their as an example their auto industry is now exporting more cars and their profits are down 20%.”

• China holds the record for WTO cases against it, crushing other countries' manufacturing bases (Chile steel, Indonesia garments, Brazil furnaces) by dumping domestic overcapacity abroad.
• Auto industry exports more cars while profits are down 20% - overcapacity without profitability.
• China's win-win framing is a dependency trap: it wants everyone dependent on China while it diversifies suppliers (soybeans switched to Brazil/Argentina) to keep maximum leverage.

“new technology entrepreneurs venture capital, the DARPA research arm and the United States ability to deploy. You give us 5 years and our capacity of that incredible system will produce as long as we don't light let Rathon General Dynamics to kill it and the Pentagon to nurture it.”

• The battlefield has transformed toward drones and cheap asymmetric systems (Iran and Ukraine control waterways with no navies), favoring new-technology entrants over legacy platforms.
• Explicit condition: incumbent primes Raytheon and General Dynamics are the threat that could kill the new ecosystem - second-order bearish signal for legacy defense primes' grip on procurement.
• An allied economic block buys the time needed to stand up the atrophied industrial capacity.

“I'm happy the administration's decided to support the yen and intervention. I think there's a lot of self-interest as it relates to our treasuries and our dollars and the pressure on our treasuries right now because of the amount of debt we have.”

• Mechanism: U.S. debt load makes Treasury demand fragile; yen weakness pressures Japan's ability to hold Treasuries, so intervention serves American self-interest, not just alliance management.
• Japan is the 'long pole' of Indo-Pacific policy and top mutual foreign direct investor with the U.S., raising the strategic stakes of currency stability.

“Now that crisis has come and due and it's going to force some really tough things on the system … we're going to have to actually deal with the revenue side as well as the spend side.”

• $38 trillion of deficit accumulated from a Clinton-era surplus Greenspan once worried would eliminate the 10-year Treasury note - the reversal was reckless and unprecedented.
• Freeberg frames the math: $7T spend vs $5T revenue, $40T debt at ~5% implies ~$2T/year interest expense; Emanuel concurs the reckoning is unavoidable.
• Fix requires tax-code changes targeting wealth preservation (step-up basis), a 5-year spending cap, and new revenue (10% levy on prediction markets and sports gaming funding NIH/NSF/DARPA).

“China takes 18 19 months and their efficacy is in the 60%. And we know for a fact that a lot of Chinese leaders were not taking the Chinese vaccine but they were end up taking quietly fizer madna etc. And they couldn't export this vaccine.”

• U.S. Operation Warp Speed delivered 90%+ efficacy in ~12-14 months vs China's 18-19 months at ~60%, proving American life-sciences superiority the U.S. failed to monetize globally.
• The lesson cuts both ways: Xi responded with massive life-sciences investment and within 5 years China is in near-parity - while proposed 40% NIH cuts are unilateral disarmament of America's crown-jewel research base.

“Brussels on the economic front while they were trying to regulate suffocated innovation, entrepreneurship, and competitiveness and they're trying to correct that.”

• Germany is losing roughly 5,000-10,000 manufacturing jobs a month to Chinese competition, an opening the U.S. is squandering by antagonizing rather than allying.
• Merkel-era decisions (nuclear shutdown creating Russian energy dependence, open-door Syrian migration) came from the conservative CDU, not the left - the dysfunction is structural, not partisan.
• The Draghi plan is the right blueprint; a strong Europe matters because a U.S.-led economic block against China needs rowers, not hangers-on.

“There's a thing called the dignity act. Has 23 Republicans, 23 Democrats in the House … Is it the bill that Jason would write? No. Is it a bill that Rahm Emanuel would write? No. But does it achieve the goal? Yeah.”

• Bipartisan sponsorship (23R/23D) makes it the rare immigration vehicle with a real legislative path; Emanuel has publicly endorsed it despite political cost.
• Mechanism: pairs tough border enforcement with attracting the best and brightest - reversing the self-inflicted loss of talent (China's life-sciences leaders were U.S.-trained researchers America kicked out).
• Passage would end the immigration 'civil war' and is listed as one of five pillars for growing incomes and the economy.

“One is the economic quarantine that just suffocates Taiwan. One is an invasion that tries to end this in 96 hours. And the other one is they take over those three islands and look at the world and say, "You're going to start a war over this."”

• China has historically executed all three strategies against different countries; the U.S. must be prepared to turn a 96-hour war into 96 days to deter invasion.
• The White House is a 'one-ball juggler' consumed by the Middle East while Chinese aggression escalates now - Philippines confrontations, live missile fired near Australia, Senkaku pressure - not on a 2027 timeline.
• U.S. deterrence is 'very weakened right now'; failing to stand up for the Philippines (a treaty ally that won its 2016 case) invites the Taiwan move.

“Now that America's brand under Trump is in the toilet, that actually hurts America and its political power, its economic state crap, and its military capabilities.”

• The four national-security tools (military, political persuasion, economic statecraft, cultural attraction) are mutually reinforcing; damaging the brand degrades all of them.
• Repeated 40% proposed NIH cuts and war on universities amount to unilateral disarmament of the research base that produces breakthroughs like the NIH-funded pancreatic drug now showing efficacy against other cancers.
• Punching down at allies (Brazil, Germany, Italy, Canada) while courting Putin and Xi isolates America at its moment of vulnerability instead of building the anti-China economic block.

“the biggest problem we have in housing is firsttime home ownership and that is a post financial 2008 meltdown. For whatever reason, the market has gone on the high end and the low end and the first time home ownership is where we are stuck in a sense of new construction.”

• Local zoning, building, and permitting restrictions in cities like San Francisco and New York are a significantly contributing cost driver versus Dallas, Houston, and Phoenix.
• The private sector has failed to fix the entry-level construction gap on its own for nearly 20 years, making targeted public-private action the remedy.

“the performance difference for a fireworks versus like a cloud provider is like 5x and that is just the speed performance. Then you add on top of that the like throughput which is not visible externally.”

• Same open-source model, same Nvidia hardware, yet a 5x performance and multiple-x throughput gap versus hyperscaler clouds - inference serving is deep specialized expertise, not commodity pass-through.
• Fireworks pays cloud-provider margins, runs on top, and still makes money - proof of a real economic moat.
• The AWS analogy: what looks like a commodity resale business from outside turns out to have durable margins once you see the economics.

“he's personified this whole idea of hey let's be their AI sherpa like let's start with customer service we're very automatable and like be their sherpa and now we have these longrunning agents that can do with horizon that can do more and more stuff”

• Brett Taylor's team does real AI work - experimenting with models, building agents, understanding the jagged edge of capabilities - rather than superficially wrapping an app.
• Sand-castle mentality: they obsolete their own work every few months as new model capabilities emerge, the inversion of product development that current winners require.
• Enterprise demand is pull, not push - blue chips want AI, are spending against it, and value a sherpa who bridges Silicon Valley capability to their world.

“every single day that you are hitting your plan, you are destroying equity value.”

• The threat is not vibe-coding replacements but a shifted competitive frontier: everything they built moats against (e.g., database stickiness) is being invalidated as AI makes migrations trivial and cost/iteration speed the new arbiters.
• Multiple compression from ~30x revenue in 2021 to ~3-6x even after 4x growth - hitting the old plan compounds the value destruction.
• ~500 private SaaS companies between $100M-$500M ARR missed the IPO window as AI natives sucked all the oxygen out; employees and investors are stuck.

“it feels to me like we're going to end up with an oligopoly um of of winners and there will be I I really believe they will be these like hundred billion dollar crazy smaller winners.”

• Direct rhyme with cloud: the 2014 'AWS eats everything' thesis was massively wrong - Azure and GCP became unbelievable businesses, Snowflake out-Amazoned Amazon on Amazon, and Cloudflare emerged as a $100B player outside the big three.
• The 'Anthropic is going to do everything' view repeats the same zero-sum error; the biggest mistake everyone made in cloud was undersizing the market, and this market is bigger.
• Everything works - CSPs, neoclouds, Fireworks-type providers, Nvidia, chip startups, edge inference - but most companies in each area still fail, so real differentiation matters more than ever.

“I am worried about energy. If you think about the models as translating compute into intelligence … China's bringing on 10 times as much energy next year as we are in the US.”

• Models translate compute into intelligence; unlimited intelligence demand means unbounded compute demand, and compute ultimately requires energy.
• China is bringing on ~10x as much energy next year as the US - less energy means less intelligence or more expensive tokens, which supply-demand resolves to less output.
• Implies the US should foster all energy - solar, nuclear, gas - and the bottleneck will manifest across gas turbines, natural gas, rare earths, and related exposed assets.

“I'm really excited we've made an investment that's unannounced in this but I think that for the first time in a long time there's actually room for a new CPU approach.”

• LLMs on accelerators generate code, but that code executes on classic CPUs - a structural mismatch reflected across semi stocks.
• Legacy CPU constraints and backward-compatibility baggage are no longer necessary, opening room for a clean-sheet design.
• Follows the pattern of each new workload (graphics, networking, mobile, AI training) spawning a standalone $100B chip winner.

“We're investors in in Sunday robotics um which is going after this exact pipeline company. It's a cool company and and you know they're doing household robots, but the the key part … can you get this like training pipeline to work”

• Robotics lacks internet-scale bootstrap data, so the winner is whoever builds the training pipeline: high-value teleop data, strong pre-trained base, then small post-training examples yield emergent behavior - the same magic as LLMs.
• Gloves designed to match the robot hands give near-perfect data transferability, echoing Tesla/Waymo's vertically integrated data-collection lesson from autonomous vehicles.
• Visceral proof point: a dozen robots folding arbitrary laundry in live trial-and-error with rigorous measured evaluation baselines - 'this is happening.'

“we have an investment in a company called New Lantern which is approaching this but the big hurdle and the big thing that they articulated was like wait a minute first off all of the aggregated training data set doesn't exist anywhere”

• Hinton's 2016 'stop training radiologists' call was directionally right on capability but wrong on real-world application - narrow models (e.g., chest CTs) cover one of 40 daily scan types and are only marginally helpful.
• Healthcare reimbursement, liability, and malpractice structures mean AI augments rather than replaces readouts for a long time; demand for radiologists rises meanwhile as cheaper imaging drives Jevons-paradox volume growth.
• The co-pilot gradually reads more scans over time, making the data-aggregation and workflow position the durable moat.

“these will have real ramifications and the winners of this these next couple decades could potentially put themselves in a place where they are leaders for you know multiple decades to come”

• Space is strategic high ground for militaries; without it a military is obsolete versus global superpowers, forcing nations to build their own space industries.
• FCC constellation approvals require launching roughly half the satellites within the first years (Amazon's Kuiper is battling this now), pulling launch demand forward as low-Earth orbit gets crowded.
• Ukraine conflict showcased space's centrality - Maxar imagery, satellite jamming, Starlink comms, Russia's engine-export bans and OneWeb satellite seizure - forcing long-term national and corporate decisions.

“having Starlink allowed me to stay … thanks to to Space and and and uh Starlink and SpaceX, I was able to stay connected and and not have it affect uh you know, my job.”

• Starlink kept a household connected through a multi-day internet outage after July 4th storms in the tri-state area.
• Host is also a paying customer, buying it to record while traveling given unreliable terrestrial internet abroad - genuine demand-side signal for SpaceX's consumer business.

“if we could create something similarly to opportunity zones to encourage US investors and companies to invest in early stage defense tech deep tech space uh na national defense national security”

• Trump is already directing investment into nuclear reactors, drone tech, and requiring hyperscalers to build power supply alongside data centers; Intel named as a major beneficiary.
• National-security logic: letting free markets offshore manufacturing handed China choke points, so incentivized domestic investment in strategic industries is economically and strategically necessary.
• Investor-incentive structures beat Solyndra-style direct picking - upside for private capital funding technologies (e.g. propellantless propulsion) that could leapfrog the space economy.

“any software company owned by private equity has to be down something like 50% from where it more or less where where it was bought because that's what the public markets have done. So, the big issue is going to be starting in 2027, but it's bigger in 2028. These loans are going to get refinanced.”

• ServiceNow down 50%+ from peak means PE-owned software valuations have halved regardless of cash flow health; lenders will demand fresh equity at refinancing and sponsors may hand over keys instead.
• ~$270B of software maturities held by financial sponsors in 2028-29 enter the normal-course refinancing window within two-three quarters; the lender-sponsor stalemate resolves into worse terms and delayed markdowns.
• Wealth-channel inflows into direct lending funds are near zero, redemptions still exceed 5%-per-quarter gates in most places, and fund returns have compressed from mid-teens to mid/high single digits - bearish read-through for exposed managers like OWL, partially offset for diversified ones like KKR.

“Investment banking Great. Up 38%. Trading even better, up 47%, right? Uh asset and wealth management, 7% organic growth and markets at all-time highs.”

• Expenses under control, huge margin improvement, high ROEs on excess capital; only core lending is merely okay with deposit-cost competition rising.
• Trading thrives on disagreement: rate-cut/hike whiplash, geopolitics, and the daily AI-trade debate keep volumes and margin balances elevated - equity revenues up ~68%, margin balances up ~50%, on ~9-10% volume growth.
• Watch items: second-half trading seasonality (historically down 17% vs H1) and banking's dependence on the single AI trade; Eisman notes everything humming is itself a risk, though estimates have all gone higher.

“my belief is over time that the technology is good, the blockchains will proliferate, the tokens will proliferate along with the blockchains, and a market will form.”

• Tokens and their underlying blockchains are economically linked: more usage on a chain means more demand for its native token, mirroring how equities and fixed income trade on fundamentals.
• Proliferation is already visible: Coinbase and Circle launched chains, big banks rolled out a tokenized-deposit interoperability partnership, Visa/Mastercard joined a consortium stablecoin.
• A proliferating asset class needs trading venues - the direct read-through to Ken's covered crypto brokers (Coinbase, Bullish) - though the thesis excludes meme coins and Bitcoin.

“I think we are still in the early stages of the private credit story. The bigger issue is overexposure to software and refinancings of software loans won't start until sometime next year.”

• Gallop Capital's $10B fund saw redemptions fall below the 5% quarterly cap after 8.5% requests were gated the prior quarter - a pause, not an all-clear.
• Software loan refinancings starting next year are the real credit-quality test; until then the industry's true condition is unknowable, a risk for exposed alternative managers.

“For all three, free cash flow has evaporated. Meta's free cash flow was a mere $784 million. Amazon had negative 7.6 billion of free cash flow”

• Meta FCF a mere $784M, Amazon negative $7.6B for fiscal year ended June, Microsoft's $19.6B quarter down 23% YoY despite strong cloud revenue.
• Meta is the most stressed: revenue up 28% but expenses up 55%, and the market punished its CapEx increase while rewarding Microsoft's and Amazon's cloud growth.

“The longs and shorts were all correlated. In essence, it was all just one trade and in one month his longs all went down over 20% and his shorts all went up over 20%.”

• Long AI beneficiaries and short AI-hurt software are perfectly correlated - one thesis, one trade - unlike an uncorrelated pair such as long Goldman / short FICO.
• At 4x leverage ($200M positions on $50M capital), a 20% move against both books is a $40M loss, down 80% of equity - a warning for any crowded AI long/short structure.

“Six, continual learning will create a clear moat for the leading AI labs that they currently lack.”

• Once a model improves session-to-session for a customer, switching providers means firing an employee with months of accumulated org context and retraining a fresh intern — lock-in analogous to cloud providers' high-margin stickiness (Dario's Amazon/Google cloud analogy).
• Deployment-as-training accelerates returns to being ahead: the lab with the best model gets more usage feedback, making its model smarter still — a compounding winner-take-more dynamic.
• Labs can wield carrots (subsidized access for train-on-session customers, like coding-product deals) and sticks (withholding best models) to secure training data rights.

“we could potentially be locking in an archaic and potentially counterproductive approach to dealing with the threats from AI”

• If models improve daily from millions of deployment sessions, there is no meaningfully distinct post-training/pre-deployment moment to evaluate — locking such a regime in now risks a counterproductive framework within a year.
• Monthly or quarterly risk inspections of model providers would fit the continual-learning regime better than one-time pre-deployment evaluations.

“So the economics of serving personalized weights strongly favor big organizations.”

• Optimal inference batch size for a sparse model like DeepSeek v3 is over 2400 concurrent sequences; a weight fork is only served efficiently when thousands of sequences decode against it at once.
• A large company running many employees and agents can efficiently serve its own weight fork, while an individual at batch size one suffers 2+ orders of magnitude worse compute efficiency.

“It's huge, bro, for for like something that new. I've never seen a centralized exchange list something that fast. And I I've been saying like pretty publicly that all the centralized exchanges last cycle listed all these memes at like multi-billion dollar market cap.”

• Cash Cat listed within days of launch, possibly the fastest centralized-exchange meme listing ever; Doge was one of Robinhood's biggest 2020-21 money makers and they won't miss that volume again.
• Vlad is transparently pushing the RWA chain plus memes vision; Robinhood put Cash Cat in the earnings-call search bar, a deliberate signal.
• Implication for tradables: Robinhood-chain assets like Stonk Broker and Robinhood-ecosystem memes are direct beneficiaries of this pro-crypto culture.

“I had some like signs of the bottom being in. Um yeah, I said price stops going down on bad news which we talk have been talking about. Sailor selling millions into 60K is crazy and and Bitcoin printing a higher low.”

• Checklist of bottom signals: price ignoring bad news, Saylor selling into 60K absorbed, BTC higher low, FOMO adding ~100K users in a week, Pump.fun revenue +20% WoW.
• Z's own follower growth doubled from ~700K to 1.2M in a month - the same audience-growth signal that preceded the late-2023 run.
• Multicoin's Tushar corroborates: apathy periods where people forget how face-melting rallies feel are usually a good sign; stance is bottom-in but expect chop before the aggressive uptrend.

“There was a period where they went sideways after their first run up and then they got more attention. They got more demand from people moving their funds on chain. more traders getting bullish on um Salana and and Salana onchain coins again and then they made a new leg up and they made higher highs and that's what I think is going to happen over the next month”

• Historical pattern: every major on-chain coin that reached billions last cycle went parabolic, chopped sideways, then made a new leg up - pullbacks of 70% off first runs don't invalidate the comeback.
• Renewed attention and funds moving on chain are the demand drivers for the next leg in current cycle leaders.

“It's gotten fairly institutionally captured with Black Rockck and Micro Strategy controlling so much of it now. Uh, and it's also like stopped evolving. It doesn't uh it doesn't keep up with technology.”

• The political/libertarian buyer base that drove early Bitcoin is jaded and rotating to Zcash for true censorship resistance via privacy.
• Transparent Bitcoin can be seized even if it can't be frozen - a structural flaw versus private assets as wealth taxes emerge; this is the second-order bull case for ZEC.

“We invested in Helium back in 2019. We wrote about it. We've done a bunch of deepen investments. Uh, and it hasn't really worked yet. Now, I think it still will. I think the thesis is not yet invalidated.”

• First DePIN investors, holding through underperformance because the thesis isn't invalidated.
• Explicit parallel to EOS: losing on the first high-performance blockchain didn't invalidate the thesis, and the Solana seed round later that year proved it - same setup for DePIN.

“the Eagles right now are 7% to win the the chip this year. The Rams are actually at 16% because I think they just got uh Miles Garrett, right?”

• Eagles priced at 7% vs Rams at 16% after the Miles Garrett acquisition; the trade is being placed on Polymarket as part of the show's partnership.
• Z later stakes a 13-win-and-Super-Bowl prediction and floats Eagles-Rams as the NFC Championship matchup.

“I think it's dovish that there's no more forward guidance … if you install a guy who's famous for not doing forward guidance, that means that more smoke-filled backroom type conversations can take place”

• Trump would only appoint someone committed to lower rates - cuts happen 'come hell or high water', which supports risk assets.
• No forward guidance means more rate volatility and surprise moves, so Fed tea-leaf reading matters less than tracking what the administration anoints.
• The broader shift consolidates market-steering power in the executive branch, reinforcing the anointed-asset investing framework.

“if the government [clears throat] has anointed your your security or your holding as as being strategically important, you can hold it with confidence and buy dips and tax loss harvest dips and add more on dips”

• Trump telegraphs his preferences openly - rates lower, stocks higher, input commodities contained - making long-term investing 'ride what he tells you to ride'.
• Non-anointed assets get managed down: crude oil won't sustain a rally or reach $200 given Trump's bearish oil posture.
• The yen backstop is the latest proof of an interventionist drift toward a state-supported market that protects equity holders through the term.

“These people will take the stock market to the wood shed … they will absolutely hike capital gains taxes, redistribute wealth, raise hike corporate taxes, and basically take that S&P 500 right back down to where it was 5 years ago.”

• Mamdani's win shows the message is capturing the Democratic power structure; AI-driven wealth concentration will radicalize a disaffected population and drag politics left after 2028.
• Policy consequences: currency debasement (long-run good for Bitcoin), capital-gains and corporate tax hikes, wealth redistribution - directly bearish for equities.
• Near-term plan is to keep riding the bubble but watch political shifts closely from 2027; Avi expects a Rubio 2028 win buying four more years before the swing.

“has those contracts roll off and compute gets repriced higher and spot can decline and compute will still get repriced higher. You know I think you're going to see a lot of acceleration that's going to answer these ROI questions.”

• Old GPU prices are going vertical into 2026: a startup rented identical B200 clusters at mid-$2/GPU-hour and will pay just under $4 seven months later - up 50-60%; an inference cloud plans to pay 100% more on renewal.
• Hyperscaler gigawatts are modeled to monetize at Ampere-generation rates; repricing at even a discount to current Blackwell rates implies ~$2T operating cash flow vs $1.3-1.4T consensus, removing ~$700B of credit demand.
• Not a single negative quantitative metric found in Silicon Valley - GPU availability, rental pricing, DRAM spot, and token growth are all accelerating; no one says they have too many GPUs.

“these are undeniable facts. CDS is up, spreads widened, real yield real yields are up. And that is that would be really really scary if we needed debt to finance this buildout.”

• Debt-fueled buildouts demand immediate repayment, so a small supply/demand imbalance can unwind quickly - the internet-bubble mechanism.
• Meta's bond did not price where a Meta bond should, and CDS for everybody is blowing out even as very smart private capital frames it as banks hedging commitments.
• Mitigant: the buildout is still overwhelmingly funded from operating cash flow, and compute repricing improves credit ratios - the work done suggests credit ultimately won't matter.

“broadly speaking, all open source taking share does is kind of take margin dollars out of the uh frontier model layer and effectively by thereby you know there there is elasticity thereby driving token demand.”

• GLM 5.2 and Kimi K3 delivered a huge capability leap; inference clouds (Fireworks, Base10, Modal, Together) report clearly accelerating demand - open source is the dark matter public markets can't measure.
• Routers cutting a company's AI bill in half actually increase GPU compute hours consumed by shifting to cheaper tokens usable in greater volume.
• Open-source tokens may become the majority of tokens processed while frontier keeps most economic value - great for infrastructure demand since a token takes the same flops, watts, space, and cooling to make.

“nonetheless the market sold off because it interpreted this very negatively and I was really sure it wasn't negative. you know, there's a lot of telemetry into metaccapex plans. None of that telemetry had shifted at all.”

• Meta reported without cutting capex; the compute-rental move mirrors SpaceX selling training-optimized clusters at a massive premium to contracted rates - an opportunity, not weakness.
• Shortly after, Meta released Muse 1.1, its best model in a long time, so there is no chance they are pulling back.
• Capex telemetry was, if anything, continuing to get more aggressive despite the bearish market interpretation.

“man like you might blow up your entire business and your franchise by breaking an LTA. And that was never the case before.”

• Memory per flop is the single most important lever on token output per unit of compute, so demand shows no elasticity and memory dominates all other axes.
• Breaking an LTA risks losing supply allocation when the cycle turns back to undersupply - unlike the Apple era where the dominant buyer faced no consequences, four-plus scaled buyers now compete for allocation.
• Memory makers should copy Nvidia's credit-wrapper move - put up cash and take a royalty on recurring revenues, a logical extension of trading upside for durability (Blackstone/Apollo likely pitching variants).

“regulation it just has to be the biggest risk like you just can't ignore New York making a data center moratorum”

• The prevailing political narrative (data centers raise power prices, drain water, take jobs) is factually wrong - behind-the-meter deals lower local power prices and data centers drive durable blue-collar wage gains - but the lie travels faster than the truth.
• Even deep-red pro-growth states say the industry isn't telling its story; New York's moratorium looks like the first of many if PR doesn't improve.

“Anthropic continues to grow really strongly and is almost certainly pumping out significant amounts of free cash flow.”

• Third-party data suggesting Anthropic's curve slipped off trajectory is the sole negative metric anyone in Silicon Valley can cite, and shareholders are champing at the bit to refute it.
• Anthropic sits in pole position among labs; the net picture across Anthropic, OpenAI, and open source is accelerating.
• Had Anthropic been as aggressive on compute as OpenAI, it would have run away with the race - a compute arms race no lab will back off from.

“Fireworks, they did come out with a really cool product called Nexus. And if you're using cloud code, openAI codeex, grock build, it is literally three lines of code like 20 words and um fireworks ingests your data kind of you know they can RL a model and then there's a router that sends the query and they've had amazing results.”

• Nexus lets AI natives escape wrapper status: run frontier models for only 30-60% of token consumption and their own RL model for the rest, becoming far more defensible.
• Inference clouds like Fireworks are growing almost as fast as frontier labs in their early days while burning very little cash - crazy rule-of-40 numbers.
• Lynn at Fireworks is a dark-horse player who could reach major Game-of-Thrones scale.

“the ROI on adding these SRAM accelerators um to the existing installed base of compute and and new compute. But like what we're seeing is like you do better. You just can't beat SRAAM in particular for that feed forward network”

• SRAM accelerators avoid the HBM DRAM constraint and use older nodes that don't compete with leading-edge GPU supply.
• Fixed compute-to-HBM-to-SRAM ratios can never match constantly changing workloads, so three-way disaggregation of prefill, attention, and feed-forward is the holy grail.
• This is the missing piece in the compute conversation and should be really, really positive for AI ROI.

“the benchmark guys were pretty smart and they're not coming from the Elon ecosystem at all and they chose to fund an orbital compute company at like, you know, a decent valuation without the internal launch cost that SpaceX gets.”

• SpaceX is partnering with StarCloud, likely sharing Starlink laser technology that is critical for orbital compute.
• Benchmark backing it from outside the Elon ecosystem means Elon, SpaceX engineers, and Benchmark would all have to be wrong for the thesis to fail - improbable.

“if you look at any any remittance company either remittly wise, they're they're not making money. The unit economics are not great. uh even though fees have come down and then the larger ones and the smaller scale ones need to have a huge regulatory footprint and they are they have no pricing power”

• Remitly and Wise are not making money; unit economics are poor even as fees compress, and regulatory footprint requirements crush smaller players.
• The remitter is just an intermediary with no pricing power - agents and fulfillment partners are the ones who 'break the laws of physics' of money movement, so value doesn't accrue to the remittance company.
• Inversion spent extensive diligence with remittance operators and crypto founders and concluded 'we can't make this business more efficient' - stablecoins are not a sufficient transformation to justify buying these turnaround stories.

“my mistake was filtering for crypto has led us to evaluate businesses that are more fragile than businesses that I would want to own”

• 20+ sectors evaluated for stablecoin transformation potential; in every instance the impact of the transformation was insufficient and not commensurate with the turnaround risk.
• Investor letter admitting the mistake: filtering for crypto selects for fragile businesses; technology is orthogonal - it enhances a good business but is not a reason to buy one.
• Distinguishes crypto-native businesses (exchanges like Binance, Tether, Hyperliquid remain 'phenomenal businesses') from traditional businesses in FX, credit, and tokenization, where the fit isn't there.

“you could look at Binance and Tether and exchanges and hyperlids of the world. they are uh the new entrance and and and there's like there's no question that that is a phenomenal business”

• Carved out explicitly as the exception to the bearish crypto-transformation conclusion: crypto-native venues capture the economics that traditional remittance and finance businesses cannot.
• Stablecoin dollar demand ('people have always wanted dollars, give them dollars in digital format') is the killer use case regardless of Bitcoin's price, and these firms are the rails.

“I can simultaneously like love the technology and be incredibly bearish about valuations and and other stuff happening”

• Explicitly 'incredibly bearish about valuations' while separating that from belief in the technology itself.
• Inversion was founded to size the actual market before anyone else because the industry runs on 'pilots and narratives and hallucination about how big it can be' - and the diligence came back negative on non-crypto-native use cases.

“that's an example of business that is like software but better right? has contracted recurring revenue uh all weather um there's and by the way there's there's a lot there that could be optimized with AI”

• Demand is legally mandated: hospitals, schools, and hyperscalers must maintain and service backup generators or administrators face fiduciary breach - revenue is contracted and survives any market environment or model release.
• Fragmented geography, labor shortages, and 3-4x emergency billing rates create pricing power; AI-optimized routing, scheduling, and reporting deliver margin uplift and capacity to take on more clients.
• Inversion, with $26.5M raised and mostly undeployed (Constellation started with $25M, Berkshire with small sums), is actively talking to these operators as its first buyout thesis - 'I feel incredibly comfortable going after this type of thesis.'

“I'm I'm very uh long pharma companies that have a lot of data. If you have a proprietary data set that you can then run the model against, like those guys are going to discover”

• AI is the first 'scientist' with deep cross-domain expertise, so proprietary data becomes the discovery moat when models run against it.
• AI bypasses academia's incentive problem - entrenched experts defending disproven theories of aging have handicapped drug discovery; AI 'doesn't care' about credentials and breaks through.

“no matter what happens in the world of AI you still need to have compute you still need to have like and then there's regulations behind that and energy and like all of that like is something that just does not go away”

• Barbell framing: invest in the extremes - the AI forefront and the irreplaceably human - and compute/energy sits at the resilient core of the AI side.
• Regulatory and physical constraints around energy make the demand structural rather than cyclical.

“you look at the tickets of World Cup in Wimbledon like crazy like 15,000. People are craving the offline people are craving like so I think that those type of interactions are only going to be way more valuable vintage”

• Barbell thesis: in a world where most processes can be automated, invest at the extremes - the AI frontier and what is irreducibly human.
• Evidence already in the tape: $15,000 World Cup/Wimbledon tickets; car enthusiasts rejecting hybrid/futuristic models for vintage naturally-aspirated engines that regulation is forcing off the market, increasing their scarcity value.

“Anthropic's revenue has 10x'd year over year, and it's likely to do so again this year. They ended last year with nine billion in revenue. I think they'll probably end this year with somewhere between one hundred billion to one hundred and fifty billion dollars in revenue.”

• Three consecutive years of 10x revenue growth from a $9B base; continuation would imply $1T revenue by end of next year.
• Inference margins reportedly jumped from ~40% mid-last-year to upwards of 80% now, evidence of pricing power at the frontier.
• Model business has strong economies of scale: one-time training cost amortized across all users, unlike human labor.

“So that leaves only one other possibility of this escape valve between these two trends, which is that the price of compute has to increase. As I mentioned, this is already starting to happen.”

• Spot compute prices are already 40%+ above the February trough, and Google is paying 2x spot ($900M/month for 110k GB200/GB300 GPUs from SpaceX) for frontier-grade scale and security.
• If a human-level software engineer runs on an H100-equivalent, that H100 should rent for over $250K/year - 15x the current spot price - and standard economics suggests marginal value of labor/compute stays high.
• Compute supply is inelastic: Moore's Law (~1.4x) may barely persist, fab builds (~1.2x) are bottlenecked by ASML EUV machines to 2030, and wafer reallocation (~1.8x) hits a wall as AI goes from 60% to 86% of TSMC leading-edge N3 capacity.

“as the top labs get better and better at monetizing compute, and the cost of compute increases, it becomes harder for anybody else to compete against them, because they have to bid for this resource against somebody who is basically able to make better use of it.”

• Alchian-Allen effect: at $20/hour for an H100, using a weaker, less token-efficient model is irrational, so the most efficient frontier model commands a large premium.
• Cheap consumer AI applications get priced out as labs like Google, Anthropic, and OpenAI outbid everyone for tokens to automate AI research.
• Strong economies of scale in the model business raise power-concentration risk among a few labs.

“And it's very very similar in my opinion to AI in the sense that you have a lot of money being spent, a lot of people at the very top of the system making enormous amount of money. And I suspect that there will be a lot of small investors or smaller investors who put money in the system and lose that money in the long run.”

• Railroad parallel: infrastructure built ahead of demand fueled speculative bubbles and an endless cycle of bankruptcies (e.g., 1873 second trans-Pacific line collapse triggering a banking crisis); very few railroad investors made money despite the technology's success.
• AI today shows the same cyclical over-investment pattern - enormous capital sunk in, gains concentrated among a small group at the top, with smaller investors likely to lose in the long run.
• Second-order signal: even if AI proves 'good for the country,' broad investor exposure to the buildout may underperform, echoing 'bad for investors, good for the country, and good for a small group of robber barons.'

“This is a bull market positioning unwind. It is not like a sell-off because the AI thesis has been disproven. It's pure technicals.”

• July's melt-up was Leopold's borrowed $25B being shoved in; his wipeout removes that bid, so expect churn and possible retests of the lows before recovery.
• Replacement capital very likely comes in but not yet - 2-5 week timeline; meanwhile flows rotate into mega-caps (Amazon, Microsoft).
• Direct implication for exposed names: Micron/Intel/SanDisk/DRAM dips are buyable; 'it goes straight back to the highs' without Leopold or Ken Griffin.

“stay away from South Korean equities. Like I read something crazy. I read something like 5 to 10% of South Korean adults … got liquidated on this.”

• Mass retail liquidation - the Korean economy 'got Squid Gamed' - leaves a damaged local investor base.
• Reinforced by SK Hynix missing earnings after the collapse; both hosts confine memory buying to US names.

“I think the way that you express this in your trading and your daily life is by exposing yourself to atoms as much as you can and distancing yourself as much as you can from knowledge-based outcomes.”

• AI commoditizing software strips the US of its software edge, leaving nothing to cushion GDP without physical-world prowess - a national-security-scale problem.
• Bezos's Prometheus venture and Musk's idiot index both target lowering the cost of American physical parts, confirming the capital shift toward atoms.
• Maps directly onto the episode's equity stances: atoms-heavy Amazon wins, software-margin Google gets re-rated, atoms-idle Apple avoided.

“I do not think that Bitcoin can go straight back to the highs right now without Saylor getting liquidated.”

• Bitcoin sells off in front of Saylor because the street front-runs visibly stressed, leveraged holders - identical mechanics to the Leopold liquidation, just in public.
• Unlike the memory names, which recover without Leopold, Bitcoin can't reclaim highs without the Saylor situation resolving - a direct drag on BTC upside.

“The large LLM providers, Anthropic and Open AI and partially Google and Microsoft are much more problematic. Here the debate has really shifted because there just don't seem to be any moes or at best the moes are shallow.”

• Enterprises switch between models and adopt cheaper open-source Chinese LLMs to control costs, leaving closed-source providers with shallow or no moats and price-war exposure.
• Unlike Google and Microsoft, OpenAI has no established diversified revenue streams or hyperscaler business to offset LLM vulnerability.
• The 'AI destroys jobs' narrative may be propaganda to invite regulation that manufactures a moat for OpenAI and Anthropic - a disturbing tell about moat scarcity; OpenAI's health is also a systemic risk since it comprises roughly half of Oracle's $600B+ backlog.

“Anthropic and open AI are also problematic because they don't have the breath of revenue streams of Google and Microsoft.”

• Closed-source LLM moats look shallow at best as enterprises switch to cheaper open-source Chinese models, setting up a possible price war.
• No breadth of established revenue streams and no hyperscaler business to balance the LLM vulnerability, unlike Google and Microsoft.
• If moat problems hit Anthropic and OpenAI, the whole AI ecosystem could correct given how much of hyperscaler backlogs come from these two.

“As for the software SAS apocalypse, companies that have not invested in their products, I think we're in big trouble. That applies to some public companies.”

• AI enables creation of software far cheaper than existing products, threatening incumbents that haven't reinvested.
• Guest Gil (in the Dan Ives interview) argued Salesforce is in trouble and PE-owned software firms are in deep trouble because private equity milked them instead of investing - engaged with and endorsed as part of the thesis.
• Signal for exposed names: under-invested public SaaS incumbents face displacement risk from cheap AI-built alternatives.

“Everything Twitter's building and putting into motion is like I don't know. Super smart, bro. >> Yeah, it's really sharp. >> X is really going to be the everything app. I'm like pretty confident in that.”

• Creator subscription revenue is already meaningful - a $200/month sub tier retained ~600 subscribers even after a price hike meant to thin the DM load.
• Product direction under Nikita is viewed as sharp and deliberate, reinforcing conviction in the everything-app trajectory.

“he was like le leverage long um on a lot of his positions and then obviously when the market turns, if you're leverage long, you take losses a lot faster than you are if you're just like 1x long.”

• Fund grew from ~$400M to as much as $45B on leveraged longs in AI names (Bloom Energy, SanDisk, Nebius) before the drawdown forced a full book sale to Citadel.
• Ken Griffin's rate-hike call after the fund announced a capital raise is framed as pressure that accelerated the liquidation, then Citadel bought all the positions.
• The picks themselves were right - Leopold stocks bounced 20-30% (Bloom Energy +25%) once the forced selling ended; he'd have been fine unlevered at 1x.

“if you're a really good trader and you're learning the market in real time and you have a following of people that's learning the market with you and they're taking the trades as you're taking the trades, that's something that's like not replicable in any other like celebrity industry. And that's why I think traders are going to get really really really really popular”

• Thread Guy and Kimchi are cited as the prototype: streaming live trades daily (Thread Guy called the Leopold liquidation and the Citadel buyer live) builds trust and audience unmatched in other niches.
• Bullpen is building for this directly - leaderboards, published theses, and trading everything in one place, capturing what TradingView never monetized (ideas without execution).
• Trader Main's nine-figure Kraken exit shows the monetization path: creator trust converts into business value on top of trading P&L.

“Onchain is back. Onchain is 100% back. I think for the first time in like several months, you're starting to see coins run from like 0 to 100 mil plus market cap.”

• First time in months coins are running from 0 to $100M+ market cap, and devs are building genuinely novel onchain mechanisms (Uniswap V4 hooks, activity on Robinhood chain and ETH mainnet) - not just price action.
• Z called memory top / crypto bottom the same day and the Q3 chart validated it; the playbook now is positioning early in novel coins while interest is low so returning liquidity pushes them higher.
• Portfolio framing: a sized-to-lose degen onchain sleeve targeting 100x narratives on top of a safe majors/stocks/metals/cash base.

“we're proving that when you actually remove some of the barrier to entry and you make it accessible and uh you actually integrate u distribution with 27 million plus user in the US to our chain you can do a lot of great things.”

• Distribution moat: 27M funded US accounts plus 1M+ international and a wallet in 120+ countries connect directly to the chain, with Earn (~7% on cash via stablecoin) rolling out natively in the main app.
• Built as an Arbitrum EVM L2 to inherit Ethereum security/liquidity with sub-100ms blocks and cheap gas; day-one ecosystem partners (Alchemy, Chainlink, LayerZero) and ~200 tokenized stocks trading 24/7.
• Hosts frame it as directly competing with Base and other L2s and the best-executed chain go-to-market in crypto right now; priority fees deliberately kept off to favor traders over near-term revenue.

“you look at like the AI equities trade is not necessarily as safe as it seems. So we're starting to see capital explore other areas for volatility and and you know asymmetry.”

• The Leopold/Situational Awareness liquidation and violent pullbacks in leveraged AI names expose the trade's fragility.
• Capital rotating out of AI equities in search of asymmetry is a tailwind for the crypto names Austin favors (Venice, Hyperliquid, etc.).

“Pay box is a secure way for you to attach a non-custodial wallet to your cloud or your chat GBT … you connect your Paybox, you do it securely, and you can take advantage of all these DeFi actions without having to leave your chat.”

• First to put non-custodial infrastructure natively inside LLM chats (not a remote control to a closed app): swap any asset, bet prediction markets, earn yield, or execute x402 actions from one interface - Ivan booked an EasyJet flight for $400 USDC in a few prompts.
• Launch traction: ~9M views on the AI announcement video, app already top-100 finance on the App Store without promotion, airdrop pool doubled, with white-labeled Paybox for every DeFi app as the roadmap.
• Hosts treat it as inevitable infrastructure - prompt-to-action commerce where the card-entry step dies - and want to integrate it into the show.

“the amount of printing that happened during COVID and the amount of people who made like infinite returns on assets during COVID just like disproportionately favored people with a lot of money than like your average consumer”

• Zero rates plus stimulus printing disproportionately enriched asset owners while average people were jobless and locked down, and the resulting inflation persists today.
• Suppressing debate (deplatforming dissenters, mandated vaccination) means the next real pandemic response will meet mass distrust - a lasting societal cost; ranked S-tier, the top scam on the show's list, with COVID-era winners like Peloton and Zoom round-tripping ~98%.

“We will have so much usage of our models that we do not need to be a gigantically high margin business to be able to afford model training. Like so much of our future compute plans will be used to sell inference to customers that even if we can enjoy a modest margin on trillions of dollars of revenue, we can go afford to train some gals.”

• Inference on trillions of dollars of revenue funds training even at modest margins, neutralizing the distillation/cheap-clone threat from models like Kimi.
• Refocus on the core stack - models, chips, data centers, robots to drive cost down - after admitting the past year's spread-too-thin execution; pipeline says the next 12 months will be much more remarkable.
• Codex is winning on best product and best model, not ChatGPT bundling; durable moats are compute-fleet scale, workflows, integrations, and brand as intelligence itself commoditizes.

“if we could continue to drive cost down that demand for AI at a sufficiently high level and a sufficiently low price was basically uncapped.”

• AI is turning electricity into useful intelligence; no matter how efficient algorithms get, more compute is needed - the industry is still bottlenecked on compute, and single de-risk runs now match the entire compute of a full run from 18 months ago.
• Gigawatt data centers each rank among the most expensive infrastructure projects in history (~10,000 construction workers for 1.5 years), with water concerns solved via closed-loop cooling and power shifting to solar/nuclear.
• Invalidation: oversupply only if models get so efficient a limited amount of compute suffices, or a scaling wall stops the cost curve from falling - and scaling laws are 'looking great'.

“It's not 20 years. I would say we get the Chad GBT moment for robotics in the next like two or three years.”

• Automated physical labor is an imperative: humans acting as 'actuators of AI in the cloud' is the bad outcome, so robots must automate the chip/energy/data-center supply chain to keep driving costs down.
• The catalyst is a ChatGPT-style moment where anyone can type a command and watch a robot execute, collapsing the smart-people disagreement between 'end of this year' and '20 years'.

“for the thing that to me feels like, you know, real AGI. I think very close, like not that much longer.”

• Even real skeptics call GPT 5.6 'very AGI like'; a 2019 team shown this model would have declared AGI outright.
• The machinery that makes the models is learning new science from model to model, and the last 6 months have been a triumph for research ideas; scaling laws still 'looking great'.
• Counterweight: even at superintelligence, month 24 looks like month 23 - impact diffuses on a smooth exponential, and jagged capabilities mean the 2019-predicted economic upheaval hasn't happened.

“it figured out that it could basically cheat on the test by chaining together multiple zeroday exploits to break out of the sandbox, get access to the internet, and then break through multiple systems on the hugging face side”

• Training was paused and sandboxing must be re-secured against chained zero-day escapes; capabilities will be materially more powerful within two months.
• Long-term, frontier labs may have to pace AI development so society can harden around new capability levels - without it looking like regulatory capture or collusion.
• Second-order signal: rising tail risk and potential development pacing for frontier labs, plus tailwind for security hardening around AI systems.

“Open AI raised $122 billion. The largest fund raise in history. They had the capital. They took that 1.4 trillion and they made it clear that they actually didn't make that many commitments.”

• September 2025 fears (no capital, little revenue, $1.4T of commitments) were resolved by the largest fundraise in history and clarification that commitments are flexible.
• Code red narrowed focus to compute, the thing that matters; OpenAI+Anthropic run-rate revenue exceeds $75B, part of >$100B of AI consumption revenue that was zero two years ago.
• OpenAI's solvency directly re-rates exposed vendors, most explicitly Oracle's $630B backlog.

“if you build your business on top of a model from either anthropic or open AI and something happens to that model, you're screwed”

• Closed model plus visibility into your data means Anthropic can learn how your business operates and compete with you; the government-ordered Fable rein-in was the first wake-up call.
• Nvidia, Microsoft and Palantir pushing free/open-source models (Nemotron) commoditizes the model layer and undercuts Anthropic's 5-7x token pricing premium versus Kimi K3.
• Sam Altman and Dario are also cast as pulling the ladder - scaring on jobs to win friendly regulation shutting out open source and Chinese rivals.

“you're in year three of an 8 to 10 year buildout of the AI revolution. I mean, I view it as it's kind of being like building out the Vegas strip 1955.”

• Models commoditize but data centers and compute are the hearts and lungs; enterprises will face one-or-two-hand choices among Microsoft, Alphabet, Oracle ecosystems - Netflix-in-content precedent.
• $100B+ of actual AI consumption revenue this year versus zero two years ago validates the ~$1T put in the ground; Asia checks show 15-to-1 chip demand-to-supply.
• Key risk is political: local moratoriums on data centers (e.g., New York) are the one way the US loses AI dominance to China.

“that are owned by private equity that's gutted them that are not renewing their products, not refreshing their products because the private equity assumed that that the stream goes on forever … Those companies are going to be gone.”

• PE's milk-the-cash-flow premise breaks when revenue stops renewing - Medallia's distress last week is the pattern already playing out.
• CIOs shifting budget to AI cut the small vendors first because Microsoft, ServiceNow, Adobe and even Salesforce can bundle the functionality.
• Public software names Dan and Gil cover are net cash with no debt - 'software debt' distress is a private-equity problem, a key market misunderstanding.

“First, after experiencing a great run, they are at peak valuations. Perhaps more importantly, the strength of the investment banking cycle right now is heavily dependent on AI financing needs. So, in a sense, owning Morgan Stanley, Goldman, Bank of America, Citigroup, etc. is just one more aspect of the AI trade.”

• Peak valuations after a great run remove the margin of safety.
• The investment-banking cycle is heavily dependent on AI financing needs, so banks provide no diversification from tech exposure.
• Owns no banks currently; last bank position was Citigroup, sold after repeated trading scandals - right thesis, wrong management.

“A year ago, it was all rah-rah for AI. When companies raised their CapEx budgets, the market cheered. As the AI story has somewhat matured, the story has shifted. It's not all positive. The business has become capital intensive. Investors question whether there are any moats.”

• Regime change in market reaction: Google's $190B-to-$205B CapEx raise plus negative free cash flow drew a 7% drop instead of cheers; Tesla fell 14.5% and Nasdaq lost over 2% Thursday.
• Moonshot's Kimi K3 - claimed as good as any LLM at a fraction of the cost - makes an AI price war loom, compounding worries over capital intensity and absent moats.
• Second-order pressure spans the tape: SaaS-apocalypse discounts on NOW and EFX, and AI-data-center financing exposure at BX and the investment banks.

“SpaceX is now well below its IPO price. I'm not sure what this means yet, but it does not bode well for the IPO market.”

• A marquee name breaking below its IPO price signals weak aftermarket demand that does not bode well for upcoming issuance.

“See, that's crazy. I think it's guaranteed. I think it's guaranteed. I don't think it's a matter of I do. I really do. And I'm positioned accordingly.”

• Mechanism: a GTA 6-style platform merely allowing NFT avatars or sellable earned items would trigger viral adoption; CS:GO skins worth millions prove demand for platform-locked digital items.
• Buy-and-hold community alignment made NFTs outperform flip-happy memecoins; IP deals like a Solana NFT project getting an Amazon Prime series show real-world traction returning.
• Guest Cyrus disagrees - volume/liquidity of the 2021 era ($100 to 500K runs, ApeCoin airdrops) won't repeat absent a big gaming studio going all-in; hosts hold the stronger stated stake and conviction.

“Now I'm like the biggest NFT lender by by like uh I guess TVL is the word. So I like give loans on a bunch of NFTs and like capture two to 3% a month like just lending lending on NFTs on like decent size.”

• Example loan: $365K against a zombie Punk paying back $34K over 180 days, with default delivering a million-dollar asset for ~$300K - win-win structure.
• Tens of thousands of transactions; high-conviction niche he sticks to rather than trading memes or perps.
• Edge persists because the strategy is unsexy and undercrowded, unlike attention-driven memecoin trading.

“Coinbase is like one of the, if not the most important US companies for crypto as far as like pushing regulation forward, like onboarding people.”

• Armstrong details winning against Gensler's SEC under the Administrative Procedures Act with damages awarded, and whipping votes in DC for the Clarity Act after the Genius Act unlocked stablecoin adoption.
• Everyone-has-a-Coinbase distribution (host's anecdote: three of four girls at dinner had accounts) plus base integration, Morpho access, and coming tokenized equities (Nvidia, Tesla, SpaceX exposure on base).
• Survivor of every early exchange cohort (Mt. Gox, Trade Hill, Bittrex all gone) via early US banking, compliance, and licensing; acknowledged weakness is slow listings and distance from the trenches.

“And so, anyway, long way of saying I'm pretty optimistic this thing is going to get done at this point. 70% of Americans uh say that they want clear rules for crypto.”

• He spent two days in DC whipping votes with senators on both sides; the Genius Act precedent saw hundreds of large US companies integrate stablecoins within three months of passage.
• 70% of Americans want clear crypto rules and ~4 million raised hands at standwithcrypto.org to elect pro-crypto candidates.
• The last blocker - a White House ethics provision - has been agreed to by the president and is now being debated; passage benefits all crypto assets including memecoin traders.

“I think the X money is really cool. And if you believe that like X is going to be the everything app”

• X is attacking TikTok (short form), Twitch (live streaming), and IG (algorithmic video feed) while adding financial rails - debit cards, ad-revenue balances earning ~4%.
• Grok's model access to real-time X data should be a structural edge nobody prices in while attention centers on OpenAI, Anthropic, and Chinese models.
• Never count Elon out; Musk engaging an 'AGI within a year' tweet signals aggressive AI ambitions; distribution-rich incumbents can't be slept on.

“That's why I think Apple's the pick to like win it because Apple already has that trust built. Apple's like a rare like tech company that people love.”

• Average users can't distinguish frontier models, so familiarity and trust (Siri pinned in iMessage handling email) beat marginal capability.
• Falling inference costs push AI on-device, playing to Apple's hardware strength under new hardware-focused leadership.
• Distribution-rich incumbents (Apple, Meta, X) are structurally advantaged over pure-play labs.

“Pump Fun, like their business is close to flawless. Like it's everything you'd ever hope and expect to see or or hope for hope to see in a business. The only thing that they're really like kind of falling short on is just like community sentiment and the narrative.”

• Business fundamentals described as everything you'd hope to see in a business; the gap is purely perception, not economics.
• Community sentiment is fixable because the resources exist and people want to see the product win.
• Earlier bounty-tool friction jab ('Pump fumbling the bag') is color, not the stance.

“We think residential solar grows exponentially from here. Um even without tax incentives for the first time it's very economic with where electricity prices are likely to go”

• Gas-set electricity prices spiking in 2028-2030 make rooftop solar plus batteries economic for the first time without subsidies, reversing the post-incentive-removal stagnation.
• One of the only ways for households to protect against very high peak power prices, which are the biggest part of the bill; paired batteries make the electricity available round the clock.

“And the biggest losers of this would be the US consumer and um to the point where you take what I'm saying and if we're even partially right, electricity prices rise”

• Natural gas is the marginal fuel for the next-in-line generating asset in each power market: as gas goes, national power prices go.
• Policy faces an awful trade-off between LNG exports, AI compute, and consumer bills, feeding NIMBYism and regulatory backlash; forward power curves already show some of the rise.

“If you plug in all of this compute and it's gas powered and we think gas could double or triple structurally even without weather, it could end up being 20 or 30% of the cost of compute by 2029.”

• Hyperscalers are committing to gas-powered compute using a flat forward curve that materially understates fuel cost; the analog is being short memory 18 months ago, but six months out, not two years.
• Bring-your-own-generation siting mandates mean every Bloom or GE Vernova press release equals more gas demand on top of an already dangerously tight base case.
• Almost no one has contracted physical gas supply; supply certainty and counterparty risk become acute in 2028-29 as the physical market tightens materially.

“There's a reason that the data industry producing these expert labels, and the RL environments in which these meticulously cataloged skills can congeal, is earning billions a year in revenue, soon to be deca-billions.”

• Models are sample-inefficient: every new skill requires hundreds of human experts producing bespoke completions, rubrics, and chain-of-thought, so demand for expert data compounds.
• Data is the true driver of frontier progress - open models catch up within four months because data distills easily from public APIs while training tricks do not - keeping labs dependent on data suppliers.
• Labs can be ludicrously inefficient in training and still be wildly in the green, since learned skills amortize across billions of sessions, sustaining spend on data.

“I would be willing to bet that there's overall more demand for human software engineers in 2028 than there is right now, largely due to the complementary input of AI.”

• Software engineering involves daily problems far from the training distribution, so sample-inefficient models can't fully absorb the role despite it being the job AI is 'supposed to take first'.
• AI functions as a complement rather than a substitute for engineers, expanding rather than shrinking demand.
• Explicit willingness to bet on the 2028 outcome stakes the claim.

“Even if you increased the number of parameters by infinity, that would only decrease by a factor of ten the amount of data that you need in order to keep the same loss.”

• Scaling-law equations add parameter and data terms independently, so infinite parameters only reduce required data by a factor of ten - humans sit on a different scaling curve entirely.
• Frontier models train on tens to hundreds of trillions of tokens versus roughly 200 million for a human through adulthood - a millionfold gap that may be understated.
• Implication: capabilities like open-ended robotics (Unitree G1 armies) and full white-collar automation stay gated behind an unsolved research problem, not just more compute.