Issue No. 08 7 June 2026

The $4 trillion question — AI floods the public market with paper, but nobody agrees who keeps the surplus

The supply wave is now quantified: SpaceX, Anthropic and OpenAI together will return more capital than the prior decade of exits (Laffont), Anthropic prints a $47B run-rate at a multiple 'cheaper than the last round' (Yang), and OpenAI's CFO confirms $122B raised and 'not enough compute through 2026' (Friar). But three independent minds — Benedict Evans, the Dwarkesh economists, and Bill Ackman — land on the same heresy in the same week: the labs may be commodity infrastructure, value migrates up the stack or to the left-for-dead megacaps, and 'the better AI gets, the smaller its share of the economy might get.' The one unambiguous trade is the landlord's — Jon Gray signs 6GW / $300B of data-centre leases into a power/turbine/memory shortage and calls real estate 'the least bubbly part of the economy.'

12 episodes · 10.2 hours

The Threads

The $4 trillion supply wave — cash-consuming to cash-returning

For eighteen months the AI cohort has been a balance-of-payments problem: enormous capital going in, very little coming out. This week the second half of that equation arrived, and it arrived quantified. On All-In, Coatue’s Thomas Laffont put a number on the paper about to hit public tapes: SpaceX “in the next few weeks,” Anthropic’s confidential S1 “just today,” OpenAI publicly committed — and “if you add up just those three companies… it’s basically going to be more than the 10 years combined.” His “Magnificent Ain’t” basket (SpaceX, Stripe, Anthropic, Databricks, Revolut, ByteDance, Anduril) already “represents almost $4 trillion of value and has really crushed the traditional mag 7.” The reframe is the whole point: the cohort flips from cash-consuming to cash-returning the moment those listings clear.

The primary source sat one seat over. OpenAI CFO Sarah Friar confirmed the $122 billion March raise — “the largest IPO to date was the Aramco, which was about $30 billion,” so a single private round dwarfed the largest public offering in history — and was careful to demote the IPO itself to “a milestone, not a destination… just another way to fundraise.” The load-bearing disclosure was on the constraint, not the capital: “compute is a very scarce resource… there’s just not enough tokens available,” and “in ‘26 we still won’t have enough compute.” Her financing playbook is the tell for how this gets funded without sinking the equity story — ride multiple CSPs to “shift CapEx into OpEx,” run multi-chip (Nvidia Vera Rubin, AMD, Cerebras, a Broadcom co-design) to stay on the frontier, and keep “maximum optionality… in a moment where I’m not yet an investment grade type of entity.”

The valuation anxiety the 20VC roundtable kept circling — Rory O’Driscoll’s worry that “all these businesses have gone from capex-light cash-flow machines to capex-heavy cash-consumptive machines,” and “across history, things that eat money tend to be bad investments” — got its cleanest rebuttal from the investor closest to the paper. On her 11-minute CNBC hit, Altimeter’s Pauline Yang, who led Anthropic’s Series H, defused the bubble read with a multiple rather than a vibe: at the $380B round Anthropic was ~$9B run-rate; it just announced $47B run-rate, so “the revenue multiple for this round is actually cheaper than the last round, sitting around 20x.” A 5x in run-rate in a few months is doing the work — and she was explicit that 5x is not her base case for the rest of the year. Laffont’s frame rhymes: the labs trade at “the lowest multiple of earnings of the S&P 500,” and he models OpenAI + Anthropic ARR bigger than AWS, and potentially all of Microsoft, by 2028.

Where I’d put numbers on this. The supply is now a scheduling question, not a maybe: at least two of {SpaceX, Anthropic, OpenAI} complete a public listing by 31 December 2026 [forecast: 2026-06-07-001]. The pricing discipline is the more interesting bet — I’d take the under on the bubble headline: if Anthropic lists, it prices at a forward run-rate revenue multiple at or below ~25x (i.e., closer to Yang’s 20x than the “100x sales” the bears keep quoting) [forecast: 2026-06-07-002]. And the structural call worth tracking for years: combined OpenAI + Anthropic run-rate revenue exceeds AWS’s by the end of 2028 [forecast: 2026-06-07-003] — lower confidence, but it is the load-bearing assumption under every valuation in this thread, and the first one that will visibly bend if the ROI reckoning bites.

Who keeps the surplus? — electricity, platform, or the left-for-dead megacaps

Here is the week’s genuine surprise. In the same seven days that Laffont sold a $4T wave and Yang defended a near-trillion mark, three independent thinkers — none of them perma-bears — converged on the same heresy: it doesn’t matter how big AI gets if no single company gets to keep the surplus.

The sharpest version came from Dwarkesh’s economists. Alex Imas (DeepMind / Chicago) and Phil Trammell (Epoch / Stanford) ran the title argument — the better AI gets, the smaller its share of the economy might get — straight at the cohort’s euphoria. As automation drives the price of machine-made goods toward zero, demand for them satiates, and a constant-or-rising fraction of spending flows to the human-intrinsic “relational sector” AI can’t replace. The receipt is two centuries deep: through the entire history of automation, labour share has stayed “over 60%.” They reduced the whole IPO cohort to one question — is AGI “like electricity or social media”? If it’s a commoditised utility whose gains diffuse to users, “you can’t capture it, so just buy the index.” The single counter-case, and the only durable bull argument they’d grant, is compute non-satiation: “an H100 costs more to rent now than it did three years ago… because as models get smarter, the opportunity cost of compute gets higher.” That sentence is the actual trade under this week’s noise — own the one input whose demand might never satiate, not the labs whose output commoditises.

Benedict Evans, on Lenny’s, arrived at the identical place from the strategy side and named the comp: foundation models “look more like cloud than they look like Windows.” No network effects means “you should have competition indefinitely,” so “why would the model companies have pricing power and wouldn’t all the value be further up the stack?” His telecoms grounding is the part to keep: mobile data volumes are up ~1,500–2,000x since 2010 “and the stocks have gone nowhere in 25 years because it’s an ex-growth, low-margin commodity utility.” Crucially, this is bear-on-margins, not bear-on-tech — “you don’t have to believe in any of that to believe this is a giant deal” — and it dissolves the very ROI panic the bulls fear: the user who “spent one and a half million dollars on tokens last month” is “like somebody getting a 50 grand mobile data bill in 2010 — that’s temporary.” The bill normalises; the margin compresses; the buyer wins.

And then the cleanest expression of all, because it came with positions attached. Bill Ackman answered the title of his own All-In appearance — what is the market missing? — by pointing at the boring side of the ledger: while short-term money crowds into “the new new thing… chips and semiconductors and energy,” the durable compounders get “left for dead.” He owns Microsoft, Meta and Amazon and calls them “really cheap,” reaching for the 2000 analogy where “Berkshire traded at the lowest valuation in its history” as the crowd dumped “old stuff.” This is the precise inverse of Loeb’s “SOX +40%” regime call and Gerstner’s memory cohort from last week — same tape, opposite trade.

Loeb himself supplied the contrarian capstone in his own All-In sit-down: Nvidia is an undervalued long “on earnings over the next two or three years,” and it only screens as a crowded short for plumbing reasons — “the long-short pods are structured such that they have to be short something. So Nvidia feels like a safe short. By the way, Google was a safe short. Amazon was a safe short.”

Where I’d put numbers on this. The commoditisation thesis is falsifiable on price: at least one major lab cuts flagship per-token API pricing by ≥30% within 12 months as competition bites (Friar already concedes a “20–30% cost reduction per token” even while raising headline prices) [forecast: 2026-06-07-004]. On the megacap-vs-”new new thing” rotation, I’d side with Ackman on a 12-month view: an equal-weight MSFT/META/AMZN basket outperforms the PHLX Semiconductor Index (SOX) over the next 12 months [forecast: 2026-06-07-005]. And the narrower contrarian tell: Nvidia delivers a positive total return over the next 12 months — i.e., the “safe short” gets run over one more time [forecast: 2026-06-07-006].

Commoditisation is real — the moat just moved to the eval and the harness

If the labs commoditise, the obvious next question is where defensibility goes — and this week answered it with unusual precision, from both the seller’s and the buyer’s chair. The answer is not “nowhere.” It’s the eval and the harness.

Mercor’s Brendan Foody, on 20VC, made the brutal case that the application layer has no moat — “2025 was the year of how do you get a model to make a PR… 2026 is the year of how do you get the model to clone Slack end to end,” so a customer paying “a million dollars a year” on a SaaS product “could just tell Claude to copy it.” But the load-bearing datapoint he dropped is the one to carry into every enterprise-software model from here: “right now we’re spending more on tokens for our internal agents than we are on employee headcount,” and “in five years the average enterprise spends more on compute than headcount.” Token spend has crossed the wage bill — at the frontier, today.

Satya Nadella, on No Priors, supplied the enterprise mirror and the reframe that matters most for anyone valuing software. The durable IP in an AI-native company is no longer years of tenure — it’s the private eval: “every company having private evals may be the biggest IP… can you switch from model A to model B and climb up? If you can, then you’re in control. If you can’t, you’re not in control.” That is model-commoditisation restated as an operating asset — own the eval, the harness and the context; rent the model. His internal proof is the supply-side mechanic under the whole cohort’s demand-side numbers: Microsoft “built in the last 15 months more Azure capacity than in the first 15 years,” and the networking team, instead of adding heads, “started screaming for more tokens — we don’t need headcount, we need tokens.” Pricing follows function: per-seat gives way to a consumption meter because “I launched 10,000 agents going on all day.” Pauline Yang’s harness argument is the same claim from the investor’s side — models “can be swapped in or out,” but integrations and context drive net retention over 500%.

The live proof sits in Uber’s P&L. Dara Khosrowshahi, on ILTB, described the reckoning landing for real: “We blew through our AI budget in a quarter — for the whole year, essentially. And it is forcing us to adjust.” Uber now meters headcount as engineers go “superhuman,” and tiers models explicitly — “the more expensive models to explore… once we scale, we bring in more efficient models, more efficient on a token basis or open source.” That is the model-commoditisation hedge spoken by the buyer, and it is exactly the behaviour Foody is selling into and Evans is pricing on.

Where I’d put numbers on this. The token-over-humans inflection is now observable: by the end of 2026, a named large enterprise (Uber, Microsoft, Anthropic or peer) publicly states that AI/token spend has displaced or frozen a measurable share of its engineering or headcount budget [forecast: 2026-06-07-007]. The commoditisation-at-scale call is testable too: within 18 months a named enterprise shifts the majority of its production inference to cheaper/open-source/distilled models, reserving frontier models for exploration [forecast: 2026-06-07-008]. And the ramp the whole thread rests on: Anthropic’s run-rate revenue exceeds $70B by year-end 2026 [forecast: 2026-06-07-009] — the continuation that has to hold for the harness-stickiness defence to survive contact with the ROI gate.

The landlord’s read — a shortage, not a bubble — and the CRE tailwind

After three threads of “who keeps the surplus,” the most directly actionable read of the week came from the one person who doesn’t have to care: the landlord. Jon Gray, addressing 700+ Blackstone LPs, inverted the entire bubble debate from the supply side. “Everybody talks about is there overbuilding… I would argue today the opposite is the risk.” The binding constraints are physical — “the shortage of power, the shortage of turbines, the shortage of memory chips and the almost exponential growth in demand” — the same multi-year memory/power shortage Cerebras and the memory cohort flagged last week, now confirmed by the largest owner of the picks-and-shovels layer.

He sized it in terms a balance sheet can hold: “$800 billion is going to be spent by five companies,” and Blackstone alone expects to sign 6 gigawatts of data-centre leasing this year — “6 gigawatts is $100 billion of data centers. Another 200 billion of chips… $300 billion is the size of the Finland economy.” Laffont’s “no TSMC for memory” line and Friar’s “not enough compute through ‘26” are the same fact viewed from the demand side; Gray is the operator confirming the supply side won’t catch up soon.

For a CRE desk this is the cleanest signal in the issue. Gray called real estate “one of the least bubbly parts of the economy” after a four-year drought and said it “is going to really get a tailwind here” as “people look for terra firma,” debt costs fall, and capital “rediscovers” a supply-starved development pipeline — logistics his favourite segment. The honest counterweight he volunteered is the one Jack should sit with: software multiples are “resetting lower” as “a fact of life” (Blackstone took significant Q1 markdowns on its growth-software book, 6.5% of the firm), and — pointedly — “what is a billable hour at a law firm going to be? I’m not sure.” The professional-services repricing the whole cohort keeps gesturing at lands closest to home here. On financing, Gray defended private credit as structurally less levered than the bank-to-CLO chain it replaces and tipped investment-grade AI-infrastructure financing as the “most explosive growth” ahead — the data centre as a new institutional asset class.

Where I’d put numbers on this. The shortage is the highest-conviction call in the issue: DRAM/HBM memory stays in shortage — contract pricing elevated, not collapsing — through year-end 2026 [forecast: 2026-06-07-010], and US data-centre power/leasing remains supply-constrained through 2026, with vacancy near record lows and interconnect queues uncleared [forecast: 2026-06-07-011]. The rotation Gray is teeing up is the harder bet but the most relevant: listed logistics/industrial real estate (proxy: Prologis or a logistics-REIT basket) outperforms the S&P 500 software index over the next 12 months [forecast: 2026-06-07-012]. And the financing wave: at least one investment-grade data-centre/AI-infrastructure financing of $10B+ prints within 12 months [forecast: 2026-06-07-013].

Short notes worth keeping

What got compressed

The whole issue is one argument with the sign flipped four times: AI’s revenue is real, its supply wave is real, and both the bull and the bear now agree the open question is margin capture, not adoption. The bulls (Laffont, Yang, Friar) say the harness and compute-non-satiation let the labs keep the rents; the bears (Evans, the Dwarkesh economists, implicitly Ackman) say models are commodity infrastructure and the surplus diffuses up-stack or to users. The single fact that adjudicates it is per-token gross margin at the frontier labs over the next 12 months — everything else is narrative. The one trade nobody disputes is physical scarcity: power, turbines, memory, and the real estate that houses them.

What to watch next week: the Anthropic S1 numbers themselves (gross margin and the cost-of-revenue line are the adjudication of this entire issue), any SpaceX pricing/allocation detail as it lists, the first per-token price cut from a major lab, and whether the DRAM/memory contract prints Gray and Laffont are leaning on actually roll over. The bull case and the bear case are now the same sentence — they just disagree about the margin line.

This Week's Episodes