Bottlenecks, burning houses, and the missing number — AI bends the data-centre cost curve up and makes valuations invisible, while Cohen, Liberty Mutual and Databricks show the operator's edge is contrarian, founder-run and honest about what's real

Last week, in Issue 10, the tape was split between an unpriceable SpaceX IPO and a quieter signal from the builders. This week the split sharpens into a single useful tension. On one side, the economics of AI are moving in ways that should reset everyone’s mental model — the binding constraint is now memory, the cost of a data centre is bending up not down, and the people closest to capital admit the future has gone genuinely invisible. On the other, a run of operators — Ryan Cohen, Liberty Mutual’s Vlad Barbalat, Databricks’ Reynold Xin — quietly demonstrate that when the macro is unpriceable, the edge retreats to things you can control: a contrarian entry, skin in the game, a clean foundation, and an honest reading of the real number. Four threads, then a builder’s ledger.

1. The AI-infra bet moves to memory and orbit

The most important re-rating this week isn’t a stock — it’s the cost curve. On All-In, Gavin Baker plants a flag: forget the exotic shortages, ‘the bottleneck that matters is DRAM… memory capacity and bandwidth are foundational to the performance of every AI model’ [forecast: 2026-06-28-001]. Micron’s blowout (HBM sold out for 2026, the stock up ~14×) is the symptom; the cause is that DRAM is heading toward 30-40% of hyperscaler capex, and only three firms can make the hard kind. The second-order effect has reached your desk — ‘AI-flation’ is now pushing up Apple and console prices as data centres hoover up the supply.

That feeds the counter-intuitive headline: data centres are getting more expensive, not cheaper. Baker does the first-principles math — a gigawatt is ~$35bn of silicon plus ~$25bn of (labour-heavy, inflationary) power and cooling — which is exactly what makes orbital compute pencil: ‘when Starship is reusable, it’s going to cost $5 billion to put a gigawatt of compute into space’ — ‘racks in space linked with lasers,’ not Pentagon-sized buildings [forecast: 2026-06-28-002]. The terrestrial sites that remain become scarce ‘diamonds,’ and the whole equation quietly favours SpaceX’s stack on token cost — a direct continuation of last week’s Gerstner thesis that ‘there is not a dark GPU’ [forecast: 2026-06-28-003].

The model layer is moving in the opposite, deflationary direction. China’s open-weight GLM-5.2 lands a tick below Opus 4.8 on coding at ~85% lower cost (reportedly trained on Huawei chips); Baker concedes heavy distillation but argues the destination is composable models — Andrej Karpathy’s ‘council of LLMs’ — routing most queries to a cheap open model and only the hardest to frontier checkers, which ‘shifts economic value from the frontier labs to the infrastructure’ [forecast: 2026-06-28-004] [forecast: 2026-06-28-005]. David Sacks, now running PCAST, supplies the urgency: GLM is as good as the currently available US models, so ‘we are going to lose if we keep doing this stuff to ourselves.’ And on the demand side, Baker pegs Anthropic at ‘$3 trillion today’ — ending the year above $100bn revenue at ~85% inference margins — with the IPO wave read as a manageable private-to-public reshuffle [forecast: 2026-06-28-006].

2. AI is rewiring the company itself

If thread 1 is where the money goes, this is what AI does to how you build. The sharpest articulation comes from Databricks’ chief architect Reynold Xin on Y Combinator: working coding agents are collapsing the engineering pyramid (manager → seniors → an army of juniors) into an ‘I-shape’ — small, top-heavy teams of people who know what to build and how, with agents doing the grunt work [forecast: 2026-06-28-009]. His analogy is the one to keep: early factories swapped their one giant steam engine for one big electric motor and got incremental gains; the real unlock came decades later when they redesigned the factory around many small motors. Software is at the same juncture — ‘slap a bunch of AI in’ a legacy system and you get crumbs; build a new AI-native ‘software factory’ from scratch and you get the step-change. The practical counsel for any incumbent: don’t just retrofit, ‘create new teams, new efforts, new product lines… new organizations to be more AI native.’

That rewiring needs a different substrate, which is where Databricks’ Neon (serverless Postgres, revenue up 10× in under a year) comes in: agents run many cheap parallel experiments, so infra ‘needs to be able to start super lightweight’ at near-zero cost and scale only on success — ‘a great time right now for disruption in infrastructure,’ targeting the long tail incumbents can’t serve [forecast: 2026-06-28-010]. And the individual version of the same shift comes from Liberty Mutual’s Vlad Barbalat on Invest Like the Best: AI is a craft tool, not a software install — engage it as a sparring partner and ‘become an editor… because that’s where slop tends to live.’ Chamath’s frame on All-In ties it together — AI as ‘the greatest economic leveler’ that gives every person ‘an equivalent co-founder’ — the optimistic mirror of the political backlash sweeping it up.

3. The operator’s edge: contrarian, skin-in-the-game, founder-run

When the macro can’t be priced, the operators in this issue all retreat to controllable edges — and they rhyme with last week’s Catmull and Gerstner. Ryan Cohen, on All-In, is the purest specimen: he buys what everyone hates (‘I like… running into a burning house’), hires for ‘will over skill,’ and negotiates so hard that ‘if our suppliers are sending us gifts in the mail… we’re overpaying.’ His eBay thesis is a founder-vs-management diagnosis — a marketplace with a real moat that stagnated once it went from founder-operated to professional management, alienating the sellers who are ‘the customer’ — and his pitch rests entirely on skin in the game: he’s putting $500m of his own money in, against a board that ‘don’t buy stock with their own money.’ As he puts it, ‘there’s nothing more American than basically risking your own capital’ [forecast: 2026-06-28-007].

Barbalat supplies the institutional version of the same principle: permanent, single-LP capital (a mutual insurer, no shareholders demanding buybacks) lets Liberty ‘make decisions that are the right decisions, not expedient decisions’ and keep ‘investment hygiene’ — because in a third-party fund ‘your business strategy is going to always dwarf your investment process.’ It’s the same ‘free from the wrong pressure’ that made Catmull’s Pixar and Gerstner’s founder-backing work. And the football desk on 20VC lands the management nuance: ‘football managers betray themselves by their weak choices of assistant’ — the best pick value-additive deputies, not loyal ones, are ‘much more focused on performance than result,’ and know that ‘amateurishness infects an entire elite sporting organization internally.’ Different arena, same Brain-Trust truth from last week: candour and the right lieutenants are the system.

4. Follow the real number

The throughline that should outlast this issue is epistemic: in a week where the future went invisible, the operators won by separating the reported number from the real one. The cleanest case is almost comic — FIFA’s headline ‘99% stadium occupancy’ is a tickets-sold figure, and as a former club CEO explains, ‘there’s absolutely no chance the stadiums have been 99% full’; the difference between reported and in-stadium attendance (really ~85-90%) is exactly the gap a careless reader misses. Tickets shifted to agencies still count as sold even when the seats are empty.

The same discipline shows up where it actually moves money. Barbalat’s most valuable section is his ‘invisible future’ thesis: AI makes it genuinely hard to know which businesses survive 10-15 years, so for the first time he’s questioning multiples on technological rather than macro grounds — ‘you’ll likely have trillion-dollar companies in 2030 that currently don’t exist, and… companies that will not exist.’ His Salesforce frame is the template: the risk isn’t enterprises vibe-coding their own CRM, it’s whether the trillion-dollar company that doesn’t exist yet ever adopts it — if not, that’s ‘a massive headwind’ even if every Fortune 500 uses it forever. And on All-In, the Cerebras post-IPO break is the market’s version of the same honesty problem: ‘sell-no-matter-what’ PMs dump anything that breaks deal price, so the lesson is structural — price via auction, don’t price to perfection. Distillation, too, is a reported-vs-real gap (a ‘cheat sheet’ that flatters a model’s apparent capability). Across all of it: trust the in-stadium count, not the press release.

Founder Action Ledger

The builder read — three buildable wedges from this issue, each with a concrete next external step. The intersection to exploit stays the rare one: having 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. A liquidity marketplace for in-game digital items. Ryan Cohen named it out loud as eBay’s missing third leg — ‘no marketplace is providing liquidity’ for skins/weapons/AAA items, ‘what NFTs could have been but had no utility.’ It sits at the intersection of marketplace mechanics and gaming, and no incumbent is doing it. Next step: talk to 3 power-sellers of in-game items (Discord/marketplaces) and find the single biggest friction in cashing out today.

2. Agent-native infrastructure for a regulated vertical. Databricks’ Neon thesis — infra that starts at ~zero cost and scales on success — collides with Addi’s lesson from last week (start with the hardest regulated workflow). The wedge: lightweight, agent-ready data/infra for a regulated mid-market vertical you understand from your deal work (lending, insurance, property). Next step: pick one such company and map what it would take to make their data agent-ready — offer the gap analysis free.

3. The picks-and-shovels of the memory/power squeeze. The non-consensus read from All-In is that data centres get more expensive (DRAM + power + labour), and modular ‘Megapods’ / disaggregated inference (cheap decode chips in front of old GPUs) reshape where compute lives. Your edge is the deal layer, not the silicon. Next step: write one page on who actually finances and sites the next wave of modular/edge compute, and send it to one operator in your network for holes.

The Catmull caveat from last week still applies: writing these down is the comfortable part — the avoidance the whole system risks enabling. A wedge is only alive once it has a next external action with a name attached.

The two halves of this issue are really one instruction. The economics of AI are now changing fast enough that the honest answer to ‘what’s it worth’ is often ‘we don’t know’ — Barbalat says so explicitly, Baker prices orbit from first principles because the ground is moving, and even FIFA can’t tell you how full its own stadiums are. In that fog, the operators who win aren’t the ones with the best forecast; they’re the ones with skin in the game, a clean foundation, the right lieutenants, and the discipline to follow the real number. Next week: whether Cohen escalates the eBay bid, whether the DRAM squeeze starts showing up in hyperscaler guidance, and whether anyone ships the orbital-compute or in-game-liquidity wedge before the obvious incumbents wake up.