The wave makes landfall — SpaceX prints the largest IPO in history at ~100× sales, and the market admits it can't price what it just bought
Last week's forecast arrived. SpaceX completed the largest IPO ever — $135 a share, a $1.77 trillion valuation, roughly 100× sales — with Elon fixing the price himself, the book only ~2× covered, and Anthropic and OpenAI filed in the queue behind it. The tell this week is not the size but the humility: ETF veteran Dave Nadig says a ~$75bn float plus accelerated index inclusion means 'there is no real price discovery' for six months; Rory O'Driscoll thinks fundamental value drags it 'nearer $1tn than $1.7tn' inside a year; and Howard Marks, pointedly refusing to call a bubble, lands the line of the week — 'if this technological innovation with its exuberance doesn't produce a money-losing bubble, it'll be the first.' Beneath the IPO theatre the capex ledger sharpens to a single number fight: Jensen Huang sizes a $50bn gigawatt factory that 'generates 300, $400 billion in intelligence,' while Bill Gurley watches the same Mag 7 drive free cash flow 'down near zero' and warns that circular deals both raise the odds of a correction and extend the runway before it hits. The one trade nobody on the tape argues with is physical — the power and the decommoditising hardware underneath the models, not the labs themselves.
A week ago, in Issue 08, the cohort was trading a forecast: Thomas Laffont’s “$4 trillion AI IPO wave,” SpaceX “in the next few weeks,” Anthropic’s confidential S-1 “just today.” This week the forecast became a print. SpaceX completed the largest initial public offering in history — “We know that the IPO is $135 a share, 1.77 trillion,” as Gavin Baker put it on bg2 two days before it cleared — at roughly 100× trailing sales, with Anthropic and OpenAI filed in the queue behind it. The supply wave is no longer a thesis. It is a tape.
What is striking is how little the people closest to it claim to know. The dominant register across eleven episodes is not euphoria and not doom; it is calibrated humility about price. The bulls and bears agree on the same uncomfortable fact — that a $1.77 trillion company can come public with no functioning price-discovery mechanism, that the ROI math depends on a revenue line nobody has audited, and that the only unarguable trades are the boring physical ones underneath the models. Five threads.
1. The unpriceable IPO
Start with the mechanics, because this week they are the story. On 20VC, Rory O’Driscoll and the panel walk through what actually happened: SpaceX ran a ~$75bn roadshow with Elon fixing the price himself at ~$135 rather than letting bankers run discovery, and the book came in only ~2× covered against the 8-10× bankers want for a clean pop. Two reads of the same fact: Jason Calacanis thinks a fixed price into thin coverage means it pops; Rory argues that short-circuiting price discovery raises the probability the stock breaks issue, and that over twelve months “fundamental value reasserts” — this “amazing company” may settle “nearer $1tn than $1.7tn,” still a generational outcome. [forecast: 2026-06-14-001]
Excess Returns supplies the market-structure spine via ETF veteran Dave Nadig, and it is the most important technical point of the week: a company this large with a tiny free float (~$75bn), accelerated index inclusion, NASDAQ-100 buying ~15 days post-IPO, offshore perps already trading and options three days after — means “there is no real price discovery” for at least six months, and “I don’t think there is an edge to be had here.” S&P refused to waive its seasoning and profitability rules, so SpaceX will not enter the S&P 500 soon; free-float adjustment collapses a notional 5%+ index weight to ~0.1%. Cameron Dawson adds the valuation-physics caveat that should be tattooed on every IPO-chaser this cycle: “when you trade at such a high multiple, even if the great news happens, what tends to happen is you kind of hit a ceiling mostly of a stock of that size.” Tesla, she notes, is the standing exception — forward earnings down ~two-thirds since 2022 while the multiple climbed past 100× — proof that for a Musk name fundamentals can simply stop mattering for a while. [forecast: 2026-06-14-002]
The supply behind SpaceX is what the panel keeps circling, and the forcing function is capital need: OpenAI’s own filing to go public is read here not as commitment but as expectations management — everyone is suddenly gunning for the public-capital door because frontier-model capital needs are orders of magnitude larger than anything else (last private rounds: OpenAI $122bn, Anthropic $30bn). The biggest winners, the panel argues, are the LPs who got the co-investment — Ontario Teachers and the savvy endowments — and the biggest risk is the rest of venture extrapolating a once-a-decade event into a new bar that quietly kills sub-$8bn outcomes. The cleanest money-path read is on Elon himself: turning xAI’s data-centre overbuild into ~$2bn/month of compute revenue ($1.25bn from Google, $950m from Anthropic) makes him, in the panel’s phrase, “the most efficient CoreWeave” with the lowest cost of capital.
The wisest frame comes from Howard Marks on Prof G Markets, recorded the day SpaceX completed at close to a $2 trillion valuation and 100× sales. Marks refuses the bubble label — “nobody, including me, should say definitively that this is a bubble” — but he reframes the entire act of participating: AI is a “concept” you cannot put numbers on, so “you have to accept the likelihood that what you’re doing is closer to speculating, and I don’t say that word pejoratively, than analytical investing.” The honest move is to place yourself deliberately on a risk spectrum — hyperscalers, then established one-product AI names, then pre-revenue lottery tickets — and size accordingly. And then the base rate, the single most-quoted line of the cohort: every prior technological revolution (railroads, radio, autos, computers, the internet) drew in too much capital, built too much infrastructure, and lost money for late providers — so “if this technological innovation with its exuberance doesn’t produce a money-losing bubble, it’ll be the first.” [forecast: 2026-06-14-003] He pairs it, characteristically, with the un-alarmist read on the index itself: the non-Shiller S&P PE is ~23 against an 80-year average of 16 — “lofty but not naughty” — a point that becomes its own trade in Thread 3.
The IPO window is open because capital is confident, not because the businesses are proven. The question every other thread is really asking is what would scare it.
2. Does the capex maths?
This is the cohort’s central argument, and this week it crystallised into a single number fight with a bull and a bear standing on the same gigawatt.
The bull is Jensen Huang on Training Data, and his receipt is specific: “that one $50 billion factory also generates 300, $400 billion in intelligence” — a 6-8× revenue-to-capex claim he uses to argue ROI is “extremely fast.” He sizes the denominator to make the number feel small: “we’re $1 trillion in of a $20 trillion a year ecosystem,” i.e. the build is ~5% deployed. He grounds it in physical units — each rack holds 72 chips, weighs two tons, costs $4m, contains 1.5m parts; NVIDIA ships ~8m chips this year. Baker, on bg2, runs the same math from the demand side and concludes it “maths”: Morgan Stanley has lifted 2027 capex to $1.1tn (Gerstner thinks ~$1.5tn including SpaceX and CoreWeave) against ~$300bn of inference revenue, but inference gross margins are 50-70%+ and revenue is being underestimated — already >$200bn this year. [forecast: 2026-06-14-004] [forecast: 2026-06-14-005]
The bear is Bill Gurley on The Knowledge Project, and he is not arguing the demand — he is arguing the system dynamics. His tell is an admission: “if you told me five years ago that … these max seven would become worth $3 trillion and then turn around and take their free cash flow from 50 to 100 billion a year down near zero because they were going to spend it all on CapEx, I’d have been like no way.” His sharpest contribution is on the circular deals everyone is uneasy about — Nvidia funding a customer who spends it back on Nvidia compute — which he argues do something genuinely two-sided: they both raise the probability of an eventual correction and extend the runway before it hits, because money handed to a company to spend back on the giver’s own service inflates growth that wouldn’t otherwise exist. That is the most useful single sentence for timing this cycle: the thing that makes it more fragile is the same thing that makes it last longer.
Sit the two side by side and the disagreement narrows to one empirical question: does realised revenue-per-deployed-gigawatt come in nearer Jensen’s $300-400bn or nearer OpenAI CFO Sarah Friar’s own ~$10bn/GW figure from Issue 08? That is a ~30-40× spread between two people building the same factories, and it is, mercifully, measurable — hyperscaler and neocloud disclosures over the next 18 months will settle it. [forecast: 2026-06-14-006] Alex Sacerdote on Invest Like the Best casts the deciding demand vote: enterprise application AI is “less than 1% penetrated,” infrastructure ~10%, and demand so far outruns supply that the world is “already sold out… there’s not enough computer in the world.” The capex bears and bulls, in other words, are not really arguing about whether the compute gets used. They are arguing about who gets paid for it — Thread 3.
3. Who keeps the surplus?
Issue 08 closed on a heresy three independent minds reached in the same week: the labs may be commodity infrastructure, and value migrates up-stack or to the left-for-dead megacaps. This week the cohort stress-tests it from four directions and lands somewhere more useful than “labs win” or “labs lose.”
Marks frames the question cleanly: profitability still ultimately governs value (he invokes Buffett’s internet-era warning that efficiency is “not the same as adding to profitability”), and the unanswered question is who captures the gains — if AI is mainly a labour-saving tool sold into a price war, the customer keeps the surplus, not the AI seller. Baker resolves the open-source half of it with a formulation worth memorising: “two things can be true” — frontier captures ~90% of the economic value while open source is ~80% of the tokens. That is bearish for frontier-lab margins and bullish for compute at the same time, and it dissolves the false binary the bears and bulls keep fighting over.
The most decisive positioning comes from Sacerdote, who has turned the commoditisation debate into a literal pair trade. He frames the foundation layer as a three-horse oligopoly (Anthropic, OpenAI, Google) — like cloud’s three-vendor structure, with real differentiation (Anthropic for finance and code, Google for documents), not pure commodity. But the application-software incumbents are the short: “We basically sold almost all of our software, almost all of our application software… entering this year, we were actually net, net short and it really helped us in the first quarter.” The thesis is that a Salesforce, whose AI is 1-2% of revenue, is exposed, not insulated. [forecast: 2026-06-14-007]
For a builder — and this is the lens worth keeping — the two most clarifying voices are Tony Fadell and Jensen. Fadell, on Lenny’s, argues that as AI collapses the cost of building, taste and craft become the only durable moat — and uses it to explain the labs themselves: OpenAI “never put product in until it was too late” (a viral demo now scrambling to hire product teams), while Anthropic is “valued more and higher revenue” because product discipline beats raw capability. His warning on vibe-coding is the technical-debt counterweight to the “90% of code is AI-written” euphoria: real architects who saw a leaked main loop “threw up” at brittle, unmaintainable code, and “software, if you’re going to build a real company, can’t be throwaway.” Jensen points at the same place from the other end: the named model labs are the small part — the real frontier is teaching AI the “language” of proteins, genes, physics and robotics, unlocking the ~$80tn physical economy, with the application layer (which absorbed $100bn of VC last year, the largest year in history) sitting on top. On jobs Huang is aggressively contrarian — “you may or may not lose a job to an AI, but you will absolutely lose a job to someone who uses AI” — dismissing singularity talk as nonsense and using radiology (declared dead twelve years ago, headcount instead rose) to separate task from purpose. The wedge both men point to is the same: not the model, but the product, the vertical, and the physical/structured domain on top of cheap abundant intelligence. Gurley completes it — workflow and data moats in verticals (legal, etc.) survive even as foundation labs climb the stack. The moat didn’t disappear. It moved.
For a founder reading this for a wedge rather than a stock, Fadell’s investor-side rules are the bluntest guidance in the cohort. VCs now “only funding companies that have atoms in their business plan with software” — hardware plus AI, not pure SaaS, which is vibe-codeable and therefore, in his word, “worthless.” The entry-price discipline that follows is savage: “if you don’t have a $5 billion round raised, you’re not anything,” which is precisely how he thinks venture returns get destroyed — you cannot buy in at nine- or ten-digit valuations and expect a real multiple, and his own portfolio (Cerebras, Groq, robotics, drug-design) was bought “not when it was hyped.” The build method underneath is the part worth stealing: start from real pain, pair it with a just-arrived enabling technology, “make the press release before you … start the project”, and accept that everything needs three generations — make the product, fix the product, then fix the business. His consumer-AI read is contrarian and tradeable: at “$20 a month or $200 a month” today’s assistants are “unsustainable… there’s just no way” consumers pay that for a Siri-1.0-grade experience, and even long term “we’re still going to need a display.” [forecast: 2026-06-14-016]
Gurley supplies the disruption that has nothing to do with AI but everything to do with where margin pools migrate next: stablecoins on USDC — dollar-for-dollar Treasuries, ~4% yield, near-instant, pennies — threaten the ~2.5% credit-card take and Visa/Mastercard’s ~60% operating margins, a duopoly he says is protected mainly by US regulatory capture while the UK, Australia, India, China and Argentina already run instant transfer. He calls the IPO process itself a “greedy power grab” by Wall Street that direct listings and auctions could fix — the structural complaint sitting underneath Thread 1’s broken price discovery. [forecast: 2026-06-14-017]
One operating fact for builders threads through 20VC and bg2: the AI-native companies run permanently lean — Lovable at ~$500m ARR on 146 staff, Cursor at ~$4bn targeting $6bn — roughly $3m of ARR per head against Salesforce’s ~$350k. The caveat that makes it real rather than hype is an accounting constraint, not a virtue: a company spending 50-70% of revenue on model tokens cannot also spend it on people. The same logic shows up in the tape — Uber cutting 23% of HR while denying AI’s role.
4. The cleanest trades are physical
If Threads 1-3 are about a thing nobody can price, this one is about the things you can. The week’s densest source of named, baselined trades is the All-In Best Ideas pitch competition on All-In — an Ira-Sohn-style format where four managers each pitch one idea and the Besties size them.
The standout, and the cleanest large-cap expression of the entire AI-capex theme, is Talon Energy (TLN), pitched by Dan Dreyfus. The frame: a data centre is a refinery, “big capital intensive asset, $50 billion per gigawatt,” and “we do not need AI demand to keep the power markets incredibly tight for the next 20 years” — AI only turbocharges it. The valuation is a hard-asset mispricing: “you could purchase this company at a $25 billion enterprise value, the replacement cost is 45 billion,” on 2GW of nuclear plus 6GW of gas baseload. That is a double just re-rating to replacement value, with a three-rung ladder — ~$50/share FCF base case, ~$70 if it signs more data-centre PPAs, $100+ if it builds new capacity — against a backdrop where PJM alone needs 106GW of new power in ten years. It topped the audience vote; MGM won the live one. [forecast: 2026-06-14-008] [forecast: 2026-06-14-009] The other three round out a risk barbell: MGM Resorts as an event-driven triple (Barry Diller at 26% and a $48 bid, hidden Osaka-2030 and Dubai options, sum-of-parts $100-150 — don’t tender to a financial buyer); Actis Oncology (AKTS), a radiopharma platform with Q1-2027 data and a China-supply moat in actinium-225; and Geodnet (GEOD), a DePIN RTK token routing ~80% of revenue into buybacks. Chamath’s verdict is the portfolio lesson: he loves all four — the differentiator is sizing and liquidity. MGM and Talon absorb real size; Actis and Geodnet are illiquid lottery tickets.
Sacerdote supplies the other half of the physical trade, and it is the sharpest cross-sectional call of the week: the “decommoditization of the hardware industry.” “The workloads are growing 10x every year… pushing every single aspect of this hardware to the physical limits,” turning former commodities — HBM, PCBs, Ethernet switching, fibre, power — into high-margin, IP-rich, high-visibility businesses. Named longs: Celestica (bought ~8× earnings, sole Google-TPU server supplier, 50-60% cloud-Ethernet share), Corning fibre, Elite Materials, Delta and Advanced Energy on power; with a ~30% short basket across DRAM/NAND/PCB commodity tiers and, as Thread 3 noted, a net-short application-software book. [forecast: 2026-06-14-010] Jensen’s “five-layer cake” ranks the same way from the bottom: energy first — “the single greatest opportunity in several generations” — naming Siemens, Mitsubishi and GE Vernova. The bg2 panel even prices the call option above the atmosphere: Andrew Fox lays out ~$5bn/GW to put compute in orbit versus ~$20-25bn/GW terrestrially, a 5× cut on half the bill of materials — if satellite reliability holds. Keep it as the option it is, not the base case. [forecast: 2026-06-14-011]
And the per-gigawatt monetisation gap on bg2 is itself a trade: Baker notes SpaceX’s leaked model implies “something like $14 billion per gigawatt per year for the AI business,” yet it just signed Anthropic at $22-23bn/GW and Google at ~$50bn/GW, and “in 30 days we went from not being an AI hyperscaler to being number four and we passed a lot of companies, including Oracle.” If the signed rate is real, the IPO model is conservative on its fastest-growing line. [forecast: 2026-06-14-012]
5. What ends it
Every prior thread bends toward the same question O’Driscoll posed — what scares the money — and two episodes answer it directly.
Luke Gromen on Forward Guidance frames Kevin Warsh’s first FOMC meeting as a forced binary the Fed has spent years dodging: the dollar or the bond market — they will have to sacrifice one. The mechanism is brutally simple: “the debt is too high and there isn’t enough balance sheet to finance it without the Fed’s help.” With debt/GDP at 122% and deficits at 6%, letting the 10-year run toward 5-7% creates near-instant Treasury dysfunction, so Warsh cannot stand aside — the Fed becomes effectively married to Treasury. [forecast: 2026-06-14-013] His catalyst is the energy shock: an Iran war and a CPI headline above 4% detonate the cut-the-front-end, sell-the-long-end plan — “like giving yourself a root canal with a shotgun” — and he expects Hormuz to stay closed through fall, oil to re-accelerate, deficits to blow toward 8-10%, and foreigners (holding ~$27tn of dollar assets) to sell Treasuries into it. [forecast: 2026-06-14-014] China is the surprise in his read: it cut oil imports four-to-five million barrels a day without collapsing and keeps buying gold — the opposite of consensus — while an under-discussed structural shift (the UAE leaving OPEC, a petro-gold/yuan settlement system, yuan swap lines with ~185 countries) quietly guts US swap-line leverage. His richest tell is a valuation one that should sober every IPO-chaser in Thread 1: his “adjusted Warren Buffett metric” (total equity market cap minus US federal debt, over GDP) is the richest in 65 years — above both 1Q2000 and 4Q21 — and gold and bitcoin falling daily are, in his read, “telling us something wicked this way comes” unless liquidity is injected fast.
The structural backdrop sits under all of it on Dwarkesh, where naval historian Sarah Paine delivers the cohort’s only genuinely off-the-tape lecture — and it is more relevant than it looks. Her thesis: maritime powers play a positive-sum game (trade, shared seas, compounding wealth) while continental powers play a negative-sum one, and “continentalists tend to be dictators… all about hemorrhaging cash to dominate their own citizens and their neighbours.” The investable through-line is the brutal economics of sea versus land logistics — containerisation cut loading from ~$6/ton to under 20 cents; the largest container ships carry over 21,000 boxes worth over $1bn — which is why she is contemptuous of Belt and Road and why the rules-based maritime order (UNCLOS, ISO container standards) is invisible infrastructure now contested for the first time since the Cold War. The point for a portfolio built on global AI supply chains: the chokepoints are physical, the order that keeps them open is political, and both are in play. It is the deepest version of “what ends it” — not a multiple, a sea lane.
What to watch next week. The cohort has handed us three falsifiable forks and a deadline. One: SpaceX’s first weeks of trading — does it break $135 (Rory/Nadig) or hold (Calacanis)? Price discovery is supposed to be absent; we will learn whether “absent” means stable or fragile. Two: the revenue-per-gigawatt disclosures that adjudicate Jensen’s $300-400bn against Friar’s $10bn — the single number the whole capex thread rests on. Three: whether either Anthropic or OpenAI actually follows its filing into a listing, or whether — as 20VC read it — the filing was “expectations management, not commitment.” [forecast: 2026-06-14-015] And underneath all of it, the macro tape Gromen is watching: the 10-year, the oil price, and whether gold’s quiet bid is, as he suspects, the smartest money already deciding what scares it. The wave is real, it is priced for perfection, and for the first time in this run the people building it are the ones telling you they can’t price it.
This Week's Episodes
- The Twenty Minute VCSpaceX Launches Largest Ever IPO, OpenAI Files to Go Public, Uber Cuts
The 20VC trio (Harry Stebbings, Jason Calacanis, Howie Lerner and Rory O'Driscoll) trade the week the IPO window blew wide open. The headline: **SpaceX begins a $75bn IPO roadshow at a ~$1.77-1.8 trillion valuation, with Elon fixing the price (~$135/share) himself rather than letting bankers run price discovery.** With the book only ~2x covered (versus the traditional 8-10x bankers want for a clean pop), the panel is split: Calacanis thinks a fixed price plus thin coverage means it pops, while Rory argues short-circuiting price discovery raises the probability it breaks issue, and that **fundamental value reasserts over 12 months — this 'amazing company' may settle nearer $1tn than $1.7tn, still a generational win.** Ontario Teachers (mistaken for Ohio on-air) and savvy endowments are flagged as the biggest co-investment winners ever; the worry is LPs extrapolating a once-a-decade event into a new bar that kills sub-$8bn outcomes. **OpenAI's filing to go public is read as expectations management, not commitment** — everyone is suddenly gunning for the public-capital door because frontier-model capital needs are 2-3 orders of magnitude larger than anything else (last private rounds: OpenAI $122bn, Anthropic $30bn). On the AI-infra money path, the panel's sharpest take is on **Elon turning xAI's data-centre overbuild into ~$2bn/month (~$24bn/yr) of compute revenue — $1.25bn from Google, $950m from Anthropic — making him 'the most efficient CoreWeave with the lowest cost of capital,' with the Cursor acquisition backfilling servers at ~10x forward revenue.** Other threads: Apple paying Google ~$1bn to power Siri (vs the ~$20bn Google pays Apple for default search) is judged pragmatic, not surrender; Uber cutting 23% of HR while denying AI's role; the leanness debate — Lovable at $500m ARR with 146 staff, Cursor at $4bn targeting $6bn, ~$3m ARR/head vs Salesforce's ~$350k — and whether startups permanently run at half the headcount, with the caveat that companies spending 50-70% of revenue on model tokens simply can't also spend it on people. Plus Ramp ($44bn), Revolut ($115bn), Bending Spoons' AOL/Evernote roll-up chasing a $20bn listing, Databricks ($165bn, staying private), and Rory's framing: **money never runs out, it gets scared.**
Read episode summary → - All-In PodcastAll-In's Best Ideas Pitch Competition: 4 Investors Present Their Top Trades
An Ira-Sohn-style pitch competition: four managers each pitch one high-conviction idea to the Besties, who then size and rank them. **The four picks are MGM Resorts (Aaron Cowen, Serretta Capital), Talon Energy (Dan Dreyfus), Actis Oncology / AKTS (Oleg Nodelman, EcoR1), and the Geodnet crypto token / GEOD (Kyle Samani, ex-Multicoin).** **MGM is pitched as an event-driven triple.** Barry Diller has accumulated 26% and just bid $48 (stock was ~$37 when the deck was built); the company bought back half its float in six years. The real value is two hidden options the market ignores: a 2030 Osaka casino license (Japan gaming ~$40bn vs Macau $30bn, Vegas $10bn; ~$2bn EBITDA, MGM owns 40% plus a management fee) and 300,000 sq ft of empty space pre-built into its Dubai property in case gambling is legalised. Sum-of-parts: Vegas ~$60, Japan ~$50, Dubai ~$40-50 = potentially $100-150. Don't tender to Diller, who is a *financial*, not strategic, buyer. MGM won the live (Besties) vote despite placing second in the audience vote. **Talon Energy is the AI-power-scarcity play.** Dreyfus frames a data center as a refinery — **'$50 billion per gigawatt'** — and argues you don't even need AI to keep power tight for 20 years. Talon (2GW nuclear + 6GW gas) trades at a ~$25bn EV vs ~$45bn replacement cost; at a ~$300 stock that's ~7x FCF (~$50/share) vs ~15x for infra peers — a double doing nothing, $70/share if it signs more data-center deals, $100+/share if it builds new capacity. PJM alone needs 106 GW of new power in 10 years. Talon topped the audience vote. **Actis (AKTS)** is a radiopharma platform (actinium-payload 'microdrones') targeting Nectin-4 and B7H3; ~$1bn cap, ~$500m EV, IPO 18x oversubscribed with a $100m Lilly backstop, key data Q1 2027. Nodelman pegs fair value at ~$10bn / $200 a share on one program working, with a China-replication moat because actinium-225 isn't available there. **Geodnet (GEOD)** is a DePIN RTK network (2cm vs GPS's ~2m precision), ~$150m FDV, ~$11m ARR growing 3x, routing 80% of revenue (~$8.8m/yr) into open-market token buybacks. **Chamath's verdict: he loves all four, the differentiator is sizing and liquidity** — MGM and Talon absorb tens of millions; Actis and Geodnet are illiquid lottery tickets.
Read episode summary → - BG2 PodThe SpaceX IPO, Fable 5, AI Capex Update & Market Check w/ Gavin Baker
Brad Gerstner hosts Gavin Baker (Atreides), Andrew Fox and Altimeter partner Clark Tang two days before SpaceX's record IPO ($135/share, **$1.77 trillion valuation**, banks modelling ~$160bn revenue by 2028). The frame: don't pump it, decompose it from first principles. The bull case rests on three terrestrial legs that all 'pencil' before you need the moonshot. (1) **Launch + Starlink** — rapid Starship reusability drives cost per kg from ~$1,500 on Falcon toward sub-$250, and Starlink direct-to-cell can plausibly 5x connectivity revenue from ~$10bn to ~$50bn (still only 0.3% of the global telecom market). (2) **AI compute ('Elon Web Services')** — the surprise of the last six weeks: SpaceX struck huge compute resale deals with Anthropic and Google and **went from non-hyperscaler to #4 in 30 days, passing Oracle**, monetising at $22-23bn/GW (Anthropic) and ~$50bn/GW (Google) vs an implied $14bn/GW in the model. (3) **The model itself — the panel's highest-conviction overlooked upside:** the Cursor acquisition (~700-800 people, ~$10bn run-rate) injects proprietary coding data into Grok 4.3 training; Composer 2.5 was Pareto-dominant on coding 12 days ago. Orbital compute is a call option, not a requirement: Fox lays out **~$5bn/GW capex to put compute in space vs ~$20-25bn/GW terrestrially** — a 5x cut on half the bill of materials, assuming satellite reliability holds. On the broader capex debate, Gerstner flags Morgan Stanley lifting 2027 capex to $1.1tn (he thinks ~$1.5tn incl. SpaceX/CoreWeave) against ~$300bn inference revenue; the panel argues the math 'maths' because inference gross margins are 50-70%+ and revenue is being underestimated (>$200bn this year). On models: Fable 5 / Mythos and ChatGPT 5.5 mark a new class defined by long-running tasks — Noam Brown's point that benchmarks should be measured in time/tokens, not snapshots, made Baker 'a lot more bullish' on compute. The open-source debate resolves to 'two things can be true': **frontier captures ~90% of economic value while open source is ~80% of tokens** — bearish for frontier-lab margins but bullish for compute. Market check: Gerstner trims from 'large' to 'medium-small' into a seasonally weak, inflation-pressured summer (CPI ~4.2%), but stays structurally long.
Read episode summary → - Dwarkesh PodcastSarah Paine — Why Putin and Xi can't escape geography
Naval War College historian Sarah Paine delivers a single-speaker lecture (no Dwarkesh Q&A) laying out a unified theory of why **continental powers (Russia, China) and maritime powers (the US, Britain) play structurally different, mutually exclusive games** — and why geography, not ideology, is the binding constraint. Her core distinction: maritime powers can defend at sea with a navy and a 360-degree moat; continental powers face contiguous land threats and 'a navy's not going to save them from Russia' (Ukraine). That single fact cascades into economics and politics. Continental powers play a **negative-sum game** — take territory, destroy wealth in the taking, surround themselves with failing buffer states, and overextend until they implode (the recurring death of empires/dynasties). Maritime powers play a **positive-sum game** — treat foreign territory as a market not a conquest, share the seas as commons, compound wealth via trade, and let allies form a collective maritime bloc. The investable through-line is the brutal economics of sea vs land logistics: containerization cut loading costs from ~$6/ton to under 20 cents; the largest container ships carry **over 21,000 boxes with cargo valued over $1 billion** versus ~600 for the longest train; ultra-large crude tankers run a quarter-million deadweight tons. This is why Paine is openly contemptuous of Xi's **Belt and Road** — it's non-continuous, mixed-gauge, and runs through the world's most unstable land, while the maritime alternative just 'goes the long way around.' She flags concrete frictions: China asserts a **200-nautical-mile freedom-of-navigation limit vs UNCLOS's 12** in the South/East China Seas; Xi is 'privileging the crony sector over the private sector'; and China's geography means its sea lanes are open **only in peacetime** — chokepoints become 'kill zones' in war. On sanctions she reframes them as 'economic chemotherapy': you can't cure the tumour, but shaving 1-2% annual growth compounds into 'the difference between north and South Korea.' Her punchline for markets and statecraft alike: continentalism correlates with dictatorship and cash-hemorrhaging, and the **only win-win is diplomats and lawyers in international forums** — the alternative is 'a third world war with nuclear follow on effects.' Most relevant builder/investor takeaway: the rules-based maritime trading order (UNCLOS, ISO container standards, post-WWII institutions) is invisible infrastructure that underwrites global trade — and it is now contested for the first time since the Cold War.
Read episode summary → - excess-returnsThe $1.75T IPO No One Can Price — 6 Things That Surprised Us This Week
Jack Forehead and Matt Zigler run a tightened, six-clip 'things that surprised us' wrap, anchored by the **SpaceX IPO arriving at roughly 100x sales for a ~$1.75 trillion market cap** (Matt: 'between one and a half and two depending on what happens at the open'). Cameron Dawson frames the valuation against Palantir — the priciest S&P 500 name at ~65x sales today, ~$340B — and warns that even when great news arrives at such multiples, stocks 'hit a ceiling,' though Musk's name (Tesla's forward earnings down ~two-thirds since 2022 while its multiple climbed past 100x) shows fundamentals can stop mattering. The hosts' sharpest contribution is the **market-structure / flows problem**, via ETF veteran Dave Nadig: a company this big with a tiny ~$75B float, accelerated index inclusion, NASDAQ-100 buying ~15 days post-IPO, perps already pricing it offshore, and options three days after — 'there is no real price discovery' for at least six months, and 'I don't think that there is an edge to be had here.' S&P refused to waive its seasoning/profitability rules, so SpaceX won't enter the S&P 500 soon; **free-float adjustment means a notional 5%+ market-cap weight collapses to ~0.1% in many indexes** — and the flows logic extends to the coming OpenAI and Anthropic IPOs. The second thread is **Kai Wu's reframing of value's 'death'**: split the market into technologically exposed vs. insulated industries, and value works fine in insulated sectors — the entire factor drawdown comes from exposed ones (retail, software), and that loss overwhelms the rest. Wu's dispersion work adds that disruption-scare stocks have fat tails — ~10% double over the next year vs ~3% market-wide, ~16% halve vs ~7% — making this potentially a career-best setup for elite software pickers (and a graveyard for the wrong ones). Jim Paulsen flags **leadership rotating to unprofitable small-cap tech beating the Mag 7**, the Goldman Sachs AI index re-rating from ~35x to over 70x earnings since March 30, and the historical pattern that **most equity pain comes after oil peaks, not during its rise**. Builder takeaway: the AI-IPO wave's mechanics, not its narratives, will set near-term prices.
Read episode summary → - forward-guidanceWarsh Must Choose The Dollar Or The Bond Market — Luke Gromen
Luke Gromen (Forest for the Trees) frames the week ahead of Kevin Warsh's first FOMC meeting as a forced binary the Fed has spent years avoiding: **the dollar or the bond market — they will have to sacrifice one.** His core thesis is unchanged and brutally simple: **'The debt is too high and there isn't enough balance sheet to finance it without the Fed's help.'** With debt/GDP at 122% and deficits at 6%, letting the 10-year run from 4.6% toward 5-7% creates near-instant Treasury-market dysfunction, so Warsh will not stand aside — meaning the Fed becomes effectively married to Treasury (he notes Bessant already doubled the pace of Treasury buybacks). Gromen reads Warsh's pre-nomination WSJ op-eds (Dec 2018 begging the Fed to stop hiking; the fall 2025 'job interview' piece) as setting up a 'disinflationary growth from AI' narrative he calls 'total bs' — pointing out no data centre is cheaper to build today than a year ago. The whole 'shrink the balance sheet' plan was, in his read, a cynical package: cut the front end, sell the long end, deregulate banks (suspend SLR-style constraints) so banks backfill the duration the Fed sells — 'It's just QE through the banks.' **The Iran war detonated that plan:** a CPI headline above 4% and an inflationary energy shock send the front end the wrong way — 'like giving yourself a root canal with a shotgun.' Gromen's call: **Hormuz stays closed through fall, the physical world starts 'kicking the financial world in the head' within one to two months**, oil re-accelerates, deficits blow from 6% toward 8-10%, and foreigners (who hold ~$27T net in dollar assets including ~$9.5T Treasuries) sell Treasuries to raise dollars — a debt spiral. China is the surprise: it cut oil imports 4-5 million bbl/day without collapsing and keeps buying gold, the opposite of consensus. He flags an under-discussed structural shift — UAE leaving OPEC plus a petro-gold/yuan settlement system that removes the need for a supply cartel — and notes China has yuan swap lines with ~185 countries, gutting US swap-line leverage. **Near-term he is very cautious:** his 'adjusted Warren Buffett metric' (equity cap minus federal debt, over GDP) is the richest in 65 years, above both 1Q2000 and 4Q21. Gold and bitcoin falling daily are, in his read, 'telling us something wicked this way comes' for risk assets unless liquidity is injected fast — which he doesn't expect until real pain forces the 'no atheists in foxholes' moment.
Read episode summary → - Invest Like the BestWhy the AI Boom Is Just Getting Started
Whale Rock's Alex Sacerdote (20 years of tech investing) lays out the firm's S-curve framework and why he thinks AI is the steepest adoption curve they've ever mapped — so vertical he calls it a backwards **'L curve, just straight up.'** The core claim for investors: **enterprise application AI is 'less than 1% penetrated'** and the infrastructure layer only ~10% penetrated, with usage today concentrated in the 'tinkerers' (Sundar's '10 bips of knowledge workers'; Anthropic ~14–15M DAUs). Sacerdote argues penetration goes from 10bps to 2–5–15% over four years, and that demand is so far ahead of supply the world is **'already sold out… there's not enough computer in the world'** — echoing Marc Andreessen that compute shortage is the one sure thing of the next four years. His highest-conviction position is **Anthropic, bought at a $180B valuation in August 2025** on a code-driven thesis: Claude Code went agentic, ~20M coders each potentially spending $20–30K/yr in tokens implies a 'half a trillion dollar market just from coding alone,' and revenue was scaling '100 to a billion on the way to nine.' He frames the foundational-model layer as a three-horse oligopoly (Anthropic, OpenAI, Google) — like cloud's three-vendor structure — with real differentiation (Anthropic for finance/PE, Google for PDFs), not pure commodity, plus recursive self-improvement from coding feeding back into the models. The sharpest investable thread is hardware: the **'decommoditization of the hardware industry.'** Data-center workloads now grow 10x/year, pushing every component to physical limits and turning former commodities (HBM, PCBs, Ethernet switching, fiber, power) into high-margin, high-visibility, IP-rich businesses. Named winners include Celestica (bought ~8x earnings, sole Google TPU server supplier, 50–60% cloud Ethernet share), Corning fiber, Elite Materials (copper-clad laminate), Delta and Advanced Energy on power. Whale Rock is structurally short the application-software layer (Salesforce-type incumbents whose AI is 1–2% of revenue) and went **net short software entering the year**, which helped Q1. Buying examples cited: Nvidia at 4x earnings (2023), Tesla 5x (2019). Risks he names: anti-AI sentiment and regulation (Maine banning data centers; only 20% public optimism), a model plateau letting open-source catch up into a 'race to the bottom,' and a major lab faltering and stranding compute.
Read episode summary → - knowledge-projectMental Models That Change How You Think — Bill Gurley
Bill Gurley (Benchmark, retired) sits with Shane Parrish for a wide-ranging session that starts on mental models and ends as the single sharpest investor read this week on the AI-capex cycle. His organising frame is **systems thinking** — multivariable, non-linear systems where a single variable flips behaviour and second/third-derivative consequences only show up much later. He applies it directly to the AI buildout: he admits he'd never have believed the **'max seven' would hit ~$3 trillion and then drive free cash flow from $50-100B/year toward zero by spending it all on CapEx**, and argues the VC community's growing faith in increasing returns / power laws is making the whole ecosystem structurally more risk-seeking (burn rates that were a million a month a decade ago are now $5B a year). His most investable claim: **the much-debated 'circular deals' both raise the probability of an eventual correction AND extend the runway before it hits** — money handed to companies to spend back on the giver's own service inflates growth that wouldn't otherwise exist. On China he is contrarian — ~10 open-source models create a faster-innovating system (the 'farmers forced to share best practices' metaphor), Western startups are quietly forking those weights 'all over Silicon Valley,' and the data wall ('painting in the corners') plus expert fine-tuning at thousands of dollars/hour is now a real constraint. On model strategy he leans against one model winning everything — **workflow and data moats in verticals (legal, etc.) survive even as foundation labs climb the stack**. He is most fired up on market structure: the **IPO process is a 'greedy power grab'** by Wall Street that direct listings/auctions could fix; **stablecoins on USDC (dollar-for-dollar Treasuries, ~4% yield, near-instant, pennies)** threaten the **~2.5% credit-card take and Visa/Mastercard's ~60% operating margins**, a duopoly bank-protected by US regulatory capture (UK, Australia, India, China, Argentina already have instant transfer). Tokenisation could legitimately fluctuate private names like Stripe — a reason great companies stay private. He defends the Tesla/Musk pay package (pay only if the stock moons), trashes ISS/Glass Lewis proxy advisers as a black-box 'heist,' and flags index-fund voting as a real governance distortion. Builder takeaways: storytelling, product instinct, and obsessive learning on the bleeding edge as founder edges; Benchmark's flat equal-partnership; and write things down because it becomes a magnet for deal flow.
Read episode summary → - Lenny's PodcastTony Fadell: How to build real taste (and why AI makes it matter more)
Tony Fadell — co-creator of the iPod, iPhone and Nest, holder of 300+ patents, now running deep-tech investor Build Collective — argues that as AI collapses the cost of building, **taste and craft become the only durable moat**. His sharpest contribution to this week's AI-IPO froth is a builder-side reality check on the leading labs. He says **OpenAI 'never put product in until it was too late'** — a viral tech demo ('Netscape Navigator' analogy) now scrambling to hire product teams — while **Anthropic is 'valued more and higher revenue'** because, in his read, product discipline beats raw model capability. He recounts the leaked Claude main-loop source code: with Dario Amodei claiming **90-100% of Anthropic's code is AI-written**, real architects who saw it 'threw up' at brittle, unmaintainable code — Fadell's frame: vibe-coding buys short-term gain for long-term technical debt; **'software, if you're going to build a real company, can't be throwaway.'** On the device race, he's contrarian: even long term **you still need a display** (voice-first, screen-secondary), the wearable-only bets (Humane) are 'different, not better,' and consumer AI at **$20-$200/month is 'unsustainable... there's just no way'** consumers pay that for what is today a Siri-1.0-grade experience. The investor takeaways are blunt: VCs now 'only fund companies that have atoms in their business plan with software' (hardware + AI, not pure SaaS, which is vibe-codeable and therefore 'worthless'); and the entry-price discipline — **'if you don't have a $5 billion round raised, you're not anything'** — destroys venture returns, since you can't buy in at nine- or ten-digit valuations and expect a real multiple. Fadell's own portfolio (Cerebras, Groq, Symbry Robotics, drug-design firm Orionis) was bought 'not when it was hyped.' The throughline is his build method: start from real pain, pair it with a just-arrived enabling technology, sell with marketing and storytelling from day one ('make the press release before you start the project'), and accept that **everything needs three generations** — make the product, fix the product, then fix the business. The closing warning to builders: use the machines, but **'don't have cognitive surrender.'**
Read episode summary → - prof-g-marketsThe IPO Frenzy Has Begun — ft. Howard Marks
Recorded on the day SpaceX is set to complete the largest IPO in history at close to a **$2 trillion valuation and more than 100x sales** — with Anthropic and OpenAI both filed to go public — Howard Marks (Oaktree co-chairman) delivers a master-class in calibrated caution rather than a call. His core move: **AI is a 'concept' you can't put numbers on, so participating is closer to speculating than analytical value investing**, and the only honest stance is to place yourself deliberately on a risk spectrum (hyperscalers → established one-product AI names → pre-revenue lottery-ticket startups) and size accordingly. Marks refuses to declare a bubble — 'nobody, including me, should say definitively that this is a bubble' — but pairs that with the historical base rate: every prior tech revolution (railroads, radio, autos, computers, the internet) drew in too much capital, built too much infrastructure, and lost money for late capital providers. His memo line: **if AI's exuberance doesn't produce a money-losing bubble, 'it'll be the first.'** The investable nuance for builders and allocators: profitability still ultimately governs value (he invokes Buffett's internet warning that efficiency is not the same as profitability), and the unanswered question is *who captures* the gains — if AI is mainly a labour-saving tool sold into a price war, the customer (the shipping/retail/warehouse user), not the AI purveyor, may keep the surplus. On valuation, he is notably un-alarmist: the non-Shiller S&P PE is ~23 vs an 80-year average of 16 (lofty, not naughty), and **the Mag 7 ex-Tesla trade in the 30s — cheap versus the Nifty Fifty's 60–90 PEs of his early career**. He flags disruption as the new permanent risk (moats like newspapers and SaaS — see the early-Feb 'saspocalypse' — evaporate fast), warns the alt-asset boom (private credit now ~$1.7T across ~700 managers, only ~3% predating the GFC) is untested by a downturn — 'it's only when the tide goes out that we find out who's been swimming naked' — yet calls private-credit fear *overblown*. Closing advice to young builders: investing rewards those who can live with a batting average far from 1.000 (Buffett's record rests on ~12 great investments in 60–70 years); if you need to be right every time, become a dentist.
Read episode summary → - training-dataNVIDIA's Jensen Huang on Building the Dynamo of the Intelligence Age
Jensen Huang's framing talk reframes AI as an industrial revolution in *generation* replacing the 60-year paradigm of *retrieval*. His central metaphor: NVIDIA builds the modern 'dynamo' — a factory that takes electrons in and produces tokens (numbers that become intelligence) out. The pitch is built to make an investor reason about scale. **Huang sizes the AI buildout at roughly $1 trillion of capex going in this year alone, against an eventual ~$20 trillion-a-year ecosystem** — i.e. the market is, by his math, ~5% deployed. He grounds the unit economics: each rack holds 72 chips, weighs 2 tons, costs $4 million and contains 1.5 million parts; NVIDIA expects to ship ~8 million chips this year. **A one-gigawatt AI factory costs ~$50 billion but 'generates 300, $400 billion in intelligence'** — a 6-8x revenue-to-capex claim he uses to argue ROI is 'extremely fast.' The investable spine is his **'five-layer cake': (1) energy, (2) chips/computers/networking, (3) infrastructure (land, power, shell, data-center ops — all in scarce supply), (4) the model layer (OpenAI, Anthropic), and (5) the application layer**, which absorbed $100B of VC last year — he calls it the single largest year of VC investment in history. Crucially for a builder audience, Huang argues the named model labs are the *small* part: the real frontier is teaching AI the 'language' of structured things — proteins, genes, cells, physics, robotics — unlocking the ~$80 trillion physical economy. **On jobs he is aggressively contrarian: 'You may or may not lose a job to an AI, but you will absolutely lose a job to someone who uses AI.'** He dismisses doom/singularity talk as 'complete nonsense,' and uses radiology (predicted dead 12 years ago, demand and headcount instead rose) and the '90% of coding will be gone' claim (NVIDIA hiring more engineers than ever) to separate *task* from *purpose*: AI elevates jobs rather than eliminating them. For a builder hunting a wedge, the load-bearing takeaway is layer-5 and the physical/vertical frontier — the application and structured-domain opportunity sitting on top of cheap, abundant generated intelligence — not the headline model labs everyone watches.
Read episode summary →