Issue No. 06 24 May 2026

Anthropic added $11B of ARR in ONE month, EBIT-positive on the print, and compute-rationing per token by 70% — unconstrained = $100-200B ARR today

Gavin Baker triangulated the Anthropic story across two podcasts in a single week (ILTB + All-In as guest besty). The print: $11B incremental ARR added in one month = combined-business-history Palantir + Snowflake + Databricks. EBIT-positive per the WSJ. The operating constraint disclosure: Opus is producing 70% fewer tokens per question — they're rationing intelligence to ration compute. Unconstrained run-rate = $100-150-200B today, meaning the $900B / $30B-raise at 18x revenue is actually ~5x unconstrained. SpaceX S1 filed: $1.75T valuation, $75B raise, June 12 IPO, Anthropic paying $15B/yr to rent Colossus. Andrej Karpathy joins Anthropic for recursive self-improvement. Cerebras opens day one at $63B as the only pure-play public AI-infra stock. Cursor Composer 2.5 (3 weeks of RL on Colossus 2) goes Pareto-dominant overnight = xAI/Cursor become the credible fourth frontier-lab. OpenAI may file S1 Friday. Orbital compute (Starlink V3 racks-in-space) + TerraFab (SpaceX/Tesla US fab) + DC-to-DC redesign = the structural answer to the power bottleneck. And the political-physics warning: 'for every Anthropic employee making $20-100M, 7,000 Block employees are being laid off' (Josh Browder).

7 episodes · 8.9 hours

The Threads

Issue 06 audit note (2026-05-28): This issue was patched after a Codex review showed it had buried the highest-conviction trade signal of the week (Gavin Baker’s cross-sectional pair trade) and that the deep-dive supporting it contained two fabricated specifics. The thread immediately below has been added, the Trade Docket section has been added, and the orbital-compute date claim has been corrected to match the source transcript. See data/editorial/2026-05-24/weekly_allocation_ledger.json for the full signal allocation.

The 1,000% multiple gap — Gavin’s cross-sectional pair trade

This is the highest-conviction single trade signal in Issue 06. It was buried in the original draft. The audit restored it.

Gavin Baker on ILTB (utt #1637-1666, ~63 min) — the cleanest investable observation of the week, in his own words:

“Cross-sectionally the valuations do not make sense. They just flat out do not make sense. They cannot all be true. You have semicap equipment companies trading at 40 times next quarter’s annualised earnings and DRAM companies trading at mid-single digit. At the peak of the last cycle that was like 5 verse 12. At one point it was like 3 verse 45. Those can’t both be true.”

And the named short leg (utt #1664-1665):

“It’s very hard to square that valuation with something like GE Vernova’s valuation because it builds in like an unfathomable amount of share loss for Nvidia.”

The pair structure:

CohortForward multiple (per Gavin)What it implies
DRAM (SK Hynix, Samsung, MU)mid-single-digitBoom-bust cycle about to break
NVDA”essentially as cheap as it gets relative to the market in the last 10-12 years” (early-April reference)Permanent share loss to Broadcom + AMD + custom silicon (per the GEV-implied math)
GEV + power · cooling · opticalimplied “unfathomable Nvidia share loss”AI shortage is permanent AND Nvidia loses share
Semicap basket (ASML / KLA / LRCX / AMAT)40x next-Q annualisedSame as above, more aggressive

Historical baseline (verbatim from utt #1645-1646): “5 verse 12” at the last peak; “3 verse 45” at one point in history. Gavin’s own framing: “it’s not worth a 1,000% multiple gap” (utt #1654 · 3805s).

Why this is symmetric:

Long Nvidia + long memory wins both scenarios. GE Vernova is the named single-ticker short candidate; the broader power-cooling-optical-semicap cohort is the basket short.

Thesis machinery (the discipline this issue audit installs):

Reinforcing signal from the same episode (utt #402 · 963s): “watts shortage will probably begin to alleviate 27/28 and then I think orbital compute will really solve that.” The 2027-2028 watts-relief window adds a structural headwind to the short leg of this pair trade — power/cooling/optical names are not just over-multiple cross-sectionally, they’re at risk from the orbital alternative becoming credible during the buildout window.

Anthropic added $11B ARR in ONE month — and just turned EBIT-positive

Last week the Anthropic-trajectory story was triangulated three ways (Krishna Rao + Brad + Sacks). This week the triangulation deepened — same data point, three more vantage points, plus a new disclosure that collapses the bear case overnight.

Gavin Baker on ILTB — the strongest framing on the tape this year:

‘Anthropic added $11 billion of ARR. The three highest-profile SaaS companies of the last 10-12 years are Palantir, Snowflake, Databricks. Anthropic added their combined businesses in one month. Nothing like this has ever happened in the history of capitalism.’

Gavin Baker on All-In — 48 hours later — drops the bear-case-killer:

‘Anthropic was EBIT-positive per the WSJ in the most recent quarter. That’s a really important fact for the whole AI narrative — now you could see $200-300-400 billion of ARR at high margins across all the language models at the end of this year.’

EBIT-positive is the disclosure that retires the ‘circular funding / no real ROI’ bear case for the entire AI trade. The compound effect: OpenAI + Anthropic + Gemini + Cursor + xAI + open source aggregating to $200-400B of ARR at high margins this year. Andrew Feldman on No Priors confirms the demand side: OpenAI signed a $20B+ deal with Cerebras in 24 days (Thanksgiving eve → Dec 24 master agreement). ‘The art of the possible has been expanded by this push in a way I’d never have expected.’

Rory Stirling on 20VC translates this into the venture math: ‘Anthropic at $900B for $30B = 18x June revenue. If ARR multiples are still fair, a 10x/year grower that will clearly IPO is the best deal in the venture universe.’ Plus the deeper structural framing: for $1T of Anthropic+OpenAI ARR, you need 20% of every engineering salary OR 5-7% of every knowledge worker salary going to tokens. Salesforce at $300M is only 1/4 of the way there. Either token consumption 4x’s from here OR the model-company valuations are wrong.

Where I’d put numbers on this:

Opus is producing 70% fewer tokens per question — Anthropic is compute-rationed at $100-200B unconstrained ARR

The single most underdiscussed disclosure on the tape this week. Gavin Baker on ILTB:

‘Anthropic Opus is generating 70% less tokens for the exact same question. They have clearly deprecated the intelligence of Claude. Unconstrained, they’d be doing $100-150-200B today. So you might be buying it at more like 5x unconstrained run-rate revenue.’

This is the rosetta-stone disclosure that operationalises the Issue 05 ILTB Krishna-Rao ‘pricing-stability + Jevons-paradox Opus 4.5 price cut’. Reading them together: Anthropic dropped the Opus price at 4.5 launch because they were rationing the actual tokens delivered per question to ration compute. The price cut was paired with a quality cut. The compute floor is binding.

Implication for the $1T-ARR investment math: if Anthropic is already at $100-150-200B unconstrained, then the Issue 05 ‘Anthropic monopoly’ framing wasn’t speculation — it was understatement. The 10x/year trajectory through 2027 ($1T Sacks call) requires either compute uncorking (Colossus/Broadcom/Trainium ramps deliver) OR the rationing tightens further (Sonnet replaces Opus permanently for most workloads).

Rory Stirling’s 20VC framing is the demand-side mirror: Sam Altman’s $2M OpenAI-tokens-for-equity offer to every YC startup tells you OpenAI has surplus tokens; Anthropic doesn’t. The $300M/batch × 4 batches/year = $1.2B/year of foregone revenue = $20-30B of implied valuation hit. Anthropic can’t afford to make this offer. OpenAI can.

Brad’s complementary disclosure on the Issue 05 Altimeter CNBC hit — 30-40% of compute is delayed this year — closes the loop. The compute bottleneck is preventing $40-100B of run-rate revenue from materialising at Anthropic alone.

Where I’d put numbers on this:

SpaceX S1 + Cursor Composer 2.5 + Karpathy = xAI/Cursor as the fourth frontier lab

The Issue 04 Anthropic-OpenAI-Google-everyone-else AI-lab market structure just got a fourth credible entrant — and the timeline was 3-4 weeks.

The All-In disclosure: ‘Cursor Composer 2.5 — 3 to 4 weeks of reinforcement learning on Colossus 2 with Cursor’s coding-data corpus → Pareto-dominant. Cursor allegedly has more tokens of coding data than exist on the public internet. Same base model as 2 (Kimmy K.25). This is amazing.’ (Gavin Baker)

The structural read: xAI was ‘dead in the water’ on compute six months ago. Elon let Cursor onto Colossus. 3 weeks later, Composer 2.5 is on the Pareto frontier (the intelligence-vs-cost curve from Gavin’s ILTB framing). Add Grok Build (Grok’s harness, finally caught up to Claude Code + Codex). The frontier-lab list expanded from 3 to 4 inside one quarter, attributable entirely to compute-access shifts.

The SpaceX S1 disclosure locks it in. All-In coverage of the S1: $1.75T valuation, $75B raise, June 12 IPO (ticker SPCX, largest IPO ever). The S1’s Elon Web Services line — Anthropic paying SpaceX $1.25B/month = $15B/yr = $45B over 3 years for Colossus — confirms that the compute-supply side is now a 4-way market (Microsoft, Amazon, Google, SpaceX), not 3-way. Plus the Cursor acquisition adds $2-3B more.

Andrej Karpathy joining Anthropic is the matched move on the other side — the highest-profile mid-career hire to validate the Issue 05 Eric-Ries ‘LTBT is the talent moat’ thesis. Chamath: ‘Karpathy is one of the rare people that can be sent off and they just invent new things.’ Recursive self-improvement is the explicit mandate.

OpenAI may file S1 as soon as Friday (breaking news during the 20VC recording). If true, OpenAI + SpaceX + Anthropic + Cerebras + Cursor + xAI are all queueing public-market access in H2 2026 — the most concentrated cohort of AI IPOs in history.

Where I’d put numbers on this:

Orbital compute + TerraFab + DC-to-DC = the structural answers to the power bottleneck

Last week’s Issue 05 Brad-on-Altimeter ‘30-40% of compute is delayed; Tom Brady launches a data centre’ framing had the power bottleneck as the cycle-killer candidate. This week’s tape gives us three concrete structural answers — all from the same Gavin Baker + Andrew Feldman + Chamath orbit.

1. Orbital compute — Gavin’s most concrete reframe of the Watts problem:

‘A Blackwell rack weighs 3,000 lbs. 8 ft × 4 ft × 3 ft. SpaceX has shown an illustration — that’s the satellite. It’s a rack with 500-ft solar wings on each side, in a sun-synchronous orbit so the panels are always in the sun and the radiator is always in shadow. Racks in space connected by lasers through vacuum.’

Why it works: SpaceX operates 98-99% of all satellites in orbit; Starlink V3 = 20 kW (cooled today); scales to 60-100 kW = Blackwell rack equivalent; lasers between racks already on every Starlink. ‘Until you have floating Optimus, you don’t repair — but Starship will change the space economy in ways we can’t imagine.’ Gavin’s actual timing (utt #402 · 963s): “watts shortage will probably begin to alleviate 27/28 and then orbital compute will really solve that.” (A prior draft of this issue and the original deep-dive attributed an “H2 2028 to H1 2030” point prediction to Gavin — that quote does not appear in the source transcript and has been retracted.)

2. TerraFab — Gavin’s other disclosure: SpaceX + Tesla joint venture for world’s largest US fab, partnership with Intel for 50 years of institutional knowledge. ‘Elon will pull the A-teams at ASML / KLA Tencor / Lam Research / Applied Materials, recruit talent via Taiwan Town + Japan Town + Korea Town next to TerraFab, and bring manufacturing back. That’s not how Intel or Samsung think.’ Direct domestic complement to Helberg’s Pax Silica forward-deployed industrial base in the Philippines from Issue 05.

3. DC-to-DC power redesign — Chamath on All-In: ‘Forget all this DC-to-AC-to-DC nonsense that goes inside a data centre. If you could just do DC-to-DC, it requires a fundamental rearchitecture. Jensen needs a design partner and Elon becomes a natural partner.’ The TerraFab + Colossus + xAI + Nvidia integrated stack is exactly the kind of partner the rearchitecture requires.

The technical context comes from Reiner Pope on Dwarkesh — a teaching episode that walks the chip stack from gates to systolic arrays. Key disclosures: multiply-accumulate area scales quadratically with bit precision (Nvidia B300 acknowledged FP4 = 3x FP8); ‘a GPU is essentially a lot of tiny TPUs tiled across the whole chip’; MatX’s ‘splittable systolic array’ sits between Cerebras wafer-scale and Nvidia small-SM; scratchpad (TPU/Groq) vs cache (CPU) is the determinism-vs-flexibility trade. This is the engineering foundation for why TPU + Trainium + Nvidia all have a role in the Krishna Rao three-platform compute strategy from Issue 05.

Where I’d put numbers on this:

The political physics of AI wealth concentration

Two voices this week — Josh Browder on 20VC and Rory Stirling on the same 20VC news anchor — independently articulated the same political-physics risk that the AI trade has so far ignored. The compound thesis is the most-underpriced exogenous risk to the ‘Anthropic at $1T ARR’ scenario.

Josh Browder’s clearest line:

‘For every Anthropic employee who’s making 20 to 100 million, there’s 7,000 Block employees being laid off. It’s not sustainable. You can’t have 50,000 people with all the money. There could be a revolution in our lifetime.’

Rory’s parallel framing on the news anchor:

‘How do you think those 8,000 ex-Facebook employees are going to vote on the wealth tax next week? When you’ve been laid off by email at 4 AM because the CEO would prefer to buy $100M of machines than $100M of people — that politics is very different. You’re not thinking as much. You’re pretty pissed off.’

The receipts this week:

The ‘reflation hiring’ charter framing from Jason on the 20VC news anchor: ‘We’ll get fit. We’ll replace our workflows. Then we’ll have a social obligation. We’ll have to reflate and hire thousands per tech leader to avoid social unrest.’ Andrew Feldman on No Priors quotes the dystopian alternative: ‘I have empathy for them. Being a CEO is extraordinarily lonely.’ But the political math is asymmetric — voters don’t care if CEOs are lonely.

Josh Browder’s positional-goods reframe is the cleanest investment implication: ‘Absolute goods stay the same — apartments, food. Positional goods explode — 8 houses on the best road in SF, 8 Delta first-class seats.’ The AI-era wealth differential will show up as a positional-goods price explosion, not broad real-estate inflation. Plus: ‘Take all your money and buy land.’

Where I’d put numbers on this:

Cerebras as the only public-market AI-infra pure-play — execution becomes the risk

The Issue 06 audit identified that Andrew Feldman on No Priors was under-extracted in the original draft — Cerebras went straight from episode summary to weekly-issue support colour. The deep-dive that should have existed would have surfaced these public-market monitor items:

The trade thesis (monitor stance): Cerebras is the cleanest single-name exposure to the AI inference build-out, but the execution risk (10x manufacturing ramp against $20B backlog with named customer concentration) is the actual variable to watch. Wrong-if: Cerebras gross margins compress below 50% in next 2 quarters. Watch-trigger: first post-IPO earnings call; AWS deployment scale-up specifics; G42 follow-on orders. Review date: 2026-08-24.

Trade Docket

The Weekly Allocation Ledger surfaced two single-episode trade signals this week that didn’t fit a convergence thread but are too high-conviction to omit.

ANTHROPIC-SEC (secondaries) · long · source: Gavin Baker ILTB

ASTL (Astera Labs) · long · source: Gavin Baker ILTB

Three short notes worth keeping

Toby Lütke’s ‘founder slot as infrastructure’ framing on Uncapped. Direct architectural complement to Eric Ries’s LTBT thesis from Issue 05: ‘It’s not about the individual founder. It’s about the slot of having the founder slot filled. You get so much social credit as having started the company that it’s a bank. Every onboarding deposits a credibility token. You spend it on big important changes.’ Plus the Net Impact reviews disclosure: Shopify formally writes the AI-with vs AI-without engineer differential in the performance doc. Three operator confirmations in three weeks (Krishna Rao 90%+ Claude Code; Andrew Feldman $25-30K/dev/mo; Tobi Net Impact reviews) of the same 10x→100x engineer-productivity discontinuity.

Andrew Feldman’s ‘Netflix-becomes-a-studio’ frame on No Priors. ‘Speed doesn’t make existing business models a little better. Netflix delivered DVDs and thought competition was Blockbuster. When the internet got fast, they became a movie studio. Right now we’re replacing things everyone can see — coding, design, SaaS tools. Once we fundamentally reorganise around this, you’ll see new business models and fundamental jumps in productivity.’ Companion to Toby’s ‘build me a million-dollar business’ framing from Uncapped this week — the consumer-side equivalent. The strongest 12-18-month forward framing on the tape.

Cerebras’s ‘professional David’ positioning. Andrew Feldman: ‘I’m a professional David — this is my 5th startup. I compete against Goliath. Every dollar we sell, if not for our brains, their muscle would have taken it in a heartbeat.’ Plus the discipline-of-when-to-give-up framework: ‘pre-commit exit hypotheses, then have other former CEOs hold you accountable to them when the slippery slope kicks in.’ The most operationalisable founder-mode framing of the issue — pairs with Dana White’s ‘I don’t have a Plan B’ from Issue 05.

Seven episodes, 8.9 hours, 2 of 7 speaker-labelled (Andrew Feldman + Tobi Lütke). The highest-conviction trade of the issue is Gavin’s cross-sectional pair trade (long Nvidia + memory vs short GE Vernova / semicap / power-cooling-optical), entered on the 1,000% multiple gap with a 90-day review window. Secondary: Anthropic-secondaries at ~5x unconstrained run-rate based on the Opus 70%-fewer-tokens rationing disclosure. Cerebras is now a public-market monitor (execution risk against $20B backlog + 10x manufacturing ramp). Orbital compute alleviates the Watts bottleneck in 2027-2028 (per Gavin’s actual transcript, not the H2-2028-H1-2030 number a prior draft fabricated). SpaceX S1 + Cursor Composer 2.5 + Karpathy adds a fourth credible frontier lab. Political-physics warning remains the under-priced exogenous risk to the entire trade. Next week: TSMC capex guidance (Gavin’s named bubble watch), SpaceX June 12 IPO price action against Polymarket’s 71%-above-$2T, whether OpenAI files an S1, and whether the Cerebras 180-day window holds. This issue was patched 2026-05-28 after a Codex review identified the buried Gavin pair trade and two fabricated specifics in the supporting deep-dive — the Weekly Allocation Ledger at data/editorial/2026-05-24/weekly_allocation_ledger.json is the discipline that prevents recurrence.

This Week's Episodes