Crazy town, meet the wedge — the SpaceX print splits the tape between 'no dark GPUs' and 'full-on crazy town', while Ideogram, Addi and Pixar show the only durable edge is focus, taste and candour

A week ago, in Issue 09, the cohort traded the print as it happened: SpaceX clearing the largest IPO in history at ~100× sales, with the people closest to it admitting they could not price what they had just bought. This week the tape stays on the same event but the room divides cleanly into two conversations that never quite meet. In one, the SpaceX IPO is a live market — Brad Gerstner on the floor calling it a no-brainer, a skeptic beside him calling it “crazy town.” In the other, three builders who have nothing to do with the IPO — an image-model founder in Toronto, a fintech founder in Bogotá, and Ed Catmull recounting how Pixar actually worked — quietly describe the only edge that survives a froth: a wedge. Put the two conversations on one page and the issue writes itself. The macro is unpriceable; the micro is teachable. Five threads.

1. The print lands — and the froth finally gets a name

The cleanest way to read the SpaceX week is as a bull and a bear sharing a desk. On the CNBC Halftime panel, Brad Gerstner — a pre-IPO holder watching the stock open up ~16% and run to ~29%, vaulting SpaceX above a $2 trillion cap and into the top six or seven US companies — makes the structural bull case: the stock re-rated the moment SpaceX “moved into the category of being an AI hyperscaler,” going from zero to ~$27bn of AI-hyperscaler revenue in six weeks on Anthropic and Google deals, with bank targets of $160bn revenue by 2028 (versus ~$18-20bn last year). His investing creed is the part worth keeping regardless of the name: “a lot more money has been lost sitting on the sidelines wringing your hands about all the things that can go wrong rather than betting on the upside,” and the discipline beneath it — “if all I did in my career is back the best founders and the best leaders and short the others, I would have done extraordinary well.” [forecast: 2026-06-21-002]

Sitting next to him, Malcolm Etheridge supplies the inversion this newsletter keeps asking for. The valuation is “27 times roughly the value of Amazon” — roughly 100× sales against Alphabet at ~10× and Amazon at ~3× — and most of the $28.5tn TAM “is all from AI… nothing to do with launching rockets,” even though Starlink connectivity is the biggest revenue line today. His sharper point is plumbing, not multiples: “roughly $200 billion being raised this year between 3 companies… is going to suck a lot of oxygen out of this market before the end of the year” (against ~$43bn for all of last year’s non-SPAC IPOs), with retail facing 15-to-30-day lockups and the real risk of “getting crushed by that wave.” His verdict — “I think we’ve reached full-on crazy town personally” — and he is rotating into REITs and financials. [forecast: 2026-06-21-001]

The iconnections panel supplies the market-structure spine that adjudicates between them. A late-stage growth investor (“Paul,” whose fund’s single biggest position is SpaceX) frames the bull mechanics — a “rule of 60” company (30% growth + 30% margin), three subsidiaries each above $1bn revenue, sell-side modelling a “100% revenue CAGR over the next 5 years… to $1 trillion of revenue in 5 years.” But the Deutsche Bank macro seat defuses the liquidity panic with the right ratios — equity raised ÷ S&P market cap was ~1.5% in the 2020-21 boom versus ~1% now — and then hands over the single most useful number for anyone tempted to chase the open: in ‘25 “day 1 pop on average for the US IPO market was about 22%, 23%… by day 30, it was about anywhere between 11% and 12%.” The pop is mean-reverting; the marginal buyer fades. And the structural fact under all of it — “the biggest equity capital markets transaction in history was in the private market” — OpenAI’s ~$120bn raise, “9 times” the $14bn Ant raise of 2018 — is why the supply queue (SpaceX, then Anthropic, then OpenAI) keeps coming. [forecast: 2026-06-21-005]

Gerstner’s rebuttal to the froth charge is the one genuinely structural bull argument: “There is not a dark GPU. This is not dark fiber” — supply is wafer-constrained through ‘28-‘29, so the whole edifice holds only as long as compute demand does, and “the second anybody starts questioning whether or not we need this much compute will be really hard for the market.” That is the real bet, stated honestly by the bull himself. He also flags the financing fragility the bears press on — as hyperscalers move “cap-light to cap-intensive,” the buybacks the bulls bank on should slow as data centres get funded with debt and off-book SPVs (Google’s $80bn equity raise, Oracle’s $40bn). [forecast: 2026-06-21-003] [forecast: 2026-06-21-006]

The wisest frame, though, comes from outside the IPO entirely. On his 13-minute monologue, Chamath Palihapitiya reaches back to his own SPAC post-mortem and lands exactly where this week’s tape is pointed: the run “was a moment being amplified by a lot of free money sloshing around, stimmy checks… pumping, pumping, pumping” — and he mistook a liquidity regime for personal skill. The whole SpaceX-IPO desk is, in his language, a test of whether the participants can tell the regime from the edge. [forecast: 2026-06-21-011]

2. The only alpha left is a wedge

Step off the IPO desk and the week’s most valuable material is a pair of founders who win precisely by refusing the consensus. Santiago Suarez, building the Colombian payments-and-bank platform Addi, gives the thesis its cleanest statement on a16z: “Don’t let your ambition fall prey to conventional wisdom. If you’re a consensus play, there’s just no alpha.” The contrarian choices compound — he built in Colombia rather than the Brazil-then-Mexico consensus, and credits being outside the US matrix: deep US-fintech expertise had only taught him “the 10 ways your startup’s gonna fail,” whereas local “ignorance is bliss” let him act. He read the opening through customer pain, not decks — smartphones going from years-behind to universal, against a legacy UX so broken that “I went to buy a t-shirt on installments and it took me 22 minutes because they had to fingerprint me, they had to take two photos of me, they had to call two of my friends.” [forecast: 2026-06-21-008]

The Ideogram founder Mohamad, on a16z, runs the same play in foundation models — the place you would least expect a wedge to survive. He shipped an open-weights image model at 9.3B parameters, roughly 9× smaller than the ~80B “SOTA,” runnable on a single GPU, and is explicit about why: “We focused on the details of the model and we know we can’t win on scaling.” As an ex-Googler he is blunt that even a 10× raise can’t beat Google on chips, so the strategy is the inverse — focus on a niche the big labs ignore (graphic design, editable text, typography) and treat taste as a measured moat: “One element of taste is kind of being— going outside of the norm a little bit and not conforming to the average opinion, which is a little against being on top of the leaderboard.” He even ran deliberately little reinforcement learning so the model stays stylistically diverse rather than converging on the RL-flattened sameness every frontier model produces. This is last week’s Tony Fadell thesis — as building commoditises, taste and use-case fit become the differentiator — proven from inside a model lab. [forecast: 2026-06-21-007]

The common structure is worth naming because it is portable. Both founders pick a domain the giants have written off (a “too hard” country; a “too small” design niche), use proximity to a real customer’s pain as the unfair information edge, and accept that “it’s not about how good a model is in the general sense. It’s about how good is this model for my use case.” In a week where the headline asset trades at 100× sales on consensus optimism, the builders are a standing reminder that the consensus trade is, by construction, the one with no alpha left in it.

3. How the wedge gets built now — AI-native from a clean foundation

The builders also answer the question the IPO desk never touches: what does it actually look like to build with this technology today? Addi is the most concrete answer in months. Santiago’s real moat is a decision made seven years ago — a monorepo plus an event-sourcing architecture logging 10M+ events a day into Databricks — which made the company genuinely AI-ready when LLMs arrived. The payoff: 200+ agents in production, in-house agents that “handle 100% of all customer service queries” and fully resolve ~80% with no human in the loop, a merchant-onboarding agent bringing on 2,000-3,000 merchants a month, and a web rebuild that took “2 engineers, 2 months” versus the old “6-9 months, 5 engineers.” [forecast: 2026-06-21-010]

The counterintuitive lesson is the sequencing. They started AI not with customer support but with legal — Colombian “tutela” lawsuits demand a 48-hour response or the CEO is personally liable to jail — because “if you could resolve a lawsuit, you could resolve most customer service interactions.” Build the hardest pipeline first and the easy ones fall out for free. And the leadership move beneath the transformation: before mandating AI across the company, Santiago built his own stack from an empty EC2 instance — DevOps, API keys, the lot — so he could say from experience, “if this is helpful for me at this scale… the entire company should be able to do this.”

Ideogram shows the same loop from the model side: an internal MCP and agents that let a team “ask it to connect to the API and generate a bunch of images… in a couple hours you have your landing page up and running,” with the open problem being automated evaluation so a human needn’t inspect every output. And the through-line reaches all the way back to Ed Catmull, who notes that technology was in Walt Disney’s DNA but never in the company’s — the institutional version of the same lesson Santiago states plainly: “remember you’re a technology company,” because that is where the equity value compounds, and you “cannot just one day show up and be like, oh, I’m a technology company now.”

4. The machinery of candour — and quality over the growth rate

If thread 3 is the technical foundation, this is the human one, and Catmull is the master text. His central institution, the Brain Trust, is a mechanism for honesty, built because most companies only pretend to have one: “Every company says they do that. Most of them are full of shit… What they’ve got are people around them who are telling them what the leader wants to hear.” It works because the discussion stays on the problem rather than on who is right, and because the person with status has to go quiet first — “the people with power, either real or perceived power, need to shut the hell up for the first 10 to 15 minutes,” lest they set the tone. Steve Jobs is the recurring case study in productive disagreement: he fired two Pixar board members for never disagreeing — “if they don’t disagree with me, then they aren’t bringing any value to the company” — and was deliberately kept out of the Brain Trust because his voice was simply too strong for the dynamics to survive.

The same candour discipline shows up, independently, at Addi: write everything down, and always be able to “articulate the why even if it’s the most obvious why” — which, as a side effect, made the company’s SOPs agent-ready years before agents existed. And both founders replace sprawling goal-sets with a single nameable focus — Santiago’s North Star metric over a wall of OKRs (citing a Sequoia note that Elon “spends all his time on the most important thing”), Catmull’s refusal of a mission statement on the grounds that “a mission statement is an answer when typically we should always be asking questions.”

Underneath it all is a stance on quality that directly contradicts the IPO desk’s growth-rate framing. Catmull’s mantra — quality is the best business plan — comes with a villain: Jack Welch and GE in The Man Who Broke Capitalism, whose short-term-growth disciples went on to damage Boeing. Pixar was the highest-cost producer, not the lowest, and pushed films through on the strength of the team rather than the schedule, completing 21 of 22 films it started. The contrast with a market paying 100× sales on a five-year revenue extrapolation is the whole issue in miniature: one room optimises the multiple, the other optimises the work. [forecast: 2026-06-21-012]

5. The interior ledger — what the success is even for

The two episodes furthest from the tape are the ones that hit closest to the bone, and they rhyme. Chamath’s argument is that the trappings of success are a tax on excellence, not a sign of it — “Possessions are bullshit… a sign of insecurity, a rampant ego, and an unsettled mind” — and that the only durable test of the right work is energy, not money: “success is finding the thing where the 9 to 5 empowers you for the 5 to 9.” To find that work you have to excavate the true self from “all kinds of indirection and bullshit from everybody else”; his own durable thread is games, risk and problem-solving, the same game whether it is poker or company-building. And the warning that pairs with thread 1: money is a lagging indicator, “you can’t be optimized for the money,” because the field that pays today (the AI scientist) may not pay tomorrow.

Catmull supplies the same lesson from the far side of achievement. Having delivered his 20-year goal — the first computer-animated feature — he fell into a year of “now what?”, having watched founders make the Fortune cover and then “do something obviously stupid and pop like a bubble,” and chose a new mission: build a culture that outlives him. Then the most quietly radical line of the week. Wrestling with “how much of this was me?”, he concluded that the question itself is a trap — “trying to answer it is an act of separation” — and reframed it: “how much can I do on my own? It’s not the right question. It’s how much can I do with others.” The calibration habit that makes it workable, and that this newsletter should adopt wholesale: “if I know that I’m wrong half the time, that I catch it earlier. I spend less time on the wrong decision.”

Founder Action Ledger

The builder read of the week — three buildable wedges drawn from these episodes, each with a concrete next external step rather than something to merely think about. The intersection to exploit is the rare one: having actually papered AI-infra / data-centre deals, being able to build the AI systems, and carrying an investor’s instinct for what compounds. (Log the one you’ll actually move on — wedge: <what it is> | <next external step>.)

1. Start with the hardest regulated workflow. Addi began its AI rollout with legal — 48-hour, CEO-liable lawsuits — because the pipeline that resolves a lawsuit resolves everything downstream. The transferable wedge: an AI-ops build that enters a regulated mid-market vertical (lending, insurance, healthcare admin) through its highest-stakes, deadline-driven document workflow, where the buyer feels personal risk. Next step: find one GC or COO at a mid-market lender or insurer and ask, specifically, what their most painful deadline-with-liability workflow is — don’t pitch, just map the pain.

2. The clean foundation is the moat. Addi and Ideogram both won on a decision made years before LLMs (monorepo + event-sourcing; image→text→image data) that incumbents can’t retrofit. The wedge isn’t a chatbot — it’s getting a specific industry’s data into an agent-ready state, the un-sexy plumbing that is the actual durable advantage. Next step: pick one company from your deal network whose data you know is a mess, and offer to map the agent-readiness gap for free as a wedge-finder.

3. A narrow open-weights model on a corpus only you can assemble. Ideogram’s 9.3B model proves a small, focused model beats the frontier for a use case — ‘it’s not about how good a model is in the general sense. It’s about how good is this model for my use case.’ The corpus you have unusual access to is the language of CRE / data-centre leases and power-purchase agreements. Next step: assemble 50-100 real documents and run the cheapest possible fine-tune to test whether the niche even holds before committing to anything bigger.

The honest caveat, in Catmull’s frame: writing these down is the easy, comfortable part — the avoidance the whole system is at risk of helping with. A wedge is only alive once it has a next external action with a name attached to it.

It is not an accident that the week’s most expensive asset and its wisest counsel sit so far apart. The IPO desk is optimising a price it admits it cannot compute. The builders are optimising a wedge, a foundation, a feedback mechanism, and — in Chamath and Catmull — an honest answer to what the building is for. Next week: whether SpaceX’s day-30 print confirms the panel’s mean-reversion math, whether the Anthropic filing moves the IPO queue, and whether anyone on the tape starts pricing the compute-demand question Gerstner says the whole market is leaning on.