Issue No. 05 17 May 2026

Anthropic added $14B of ARR in April, Cerebras opened the AI-infra IPO window at $100B day-one, and the governance question quietly became the moat — featuring Krishna Rao direct

The Anthropic CFO went on ILTB and confirmed the Sacks-Altimeter triangulation from Issue 04: $9B → $30B Q1 ARR, NDR >500%, $100B+ closed in May (Google + Broadcom + Trainium + Colossus). Brad on CNBC added $14B incremental ARR in April alone = '2 Databricks + 1 Palantir in one month.' Cerebras opened 90% above the IPO price, $100B market cap day one, and introduced the 'dribble lockup' that quants can't arbitrage — the IPO-structural innovation worth watching as SpaceX/Anthropic/OpenAI line up. Eric Ries on Lennys argues the Anthropic Long-Term Benefit Trust is the *reason* the talent flocked to Dario — and that the standard VC governance docs every founder gets are statistical extinction (Harvard Law: only 20% of founders are still CEO 3 years post-IPO). Software 're-rated to a market multiple' (Brad) — Salesforce -37%, ServiceNow -42%, Workday -45%, ~$180B of mkt cap evaporated; Benioff: 'not my first SaaSpocalypse, and I'll probably spend $300M on Anthropic this year.' Pax Silica gets 4,000 acres in the Philippines — the first 'forward-deployed industrial base.' Trump-Xi summit and Friedberg's call: 'Taiwan is no longer an important conversation 18 months from now.' Memory at 5-6x earnings holds. Brad's sell signal: when Tom Brady launches a NeoCloud.

10 episodes · 12.1 hours

The Threads

Anthropic added $14B of ARR in April — and the CFO confirmed it on tape

The single sharpest data-point of the issue, and we got it triangulated three ways within the same week:

If you back-of-envelope it: $30B Q1 end-of-quarter + $14B incremental in April = ~$44B run-rate exiting April — directly consistent with Sacks’s $44B April number from Issue 04. The Sacks framing of ‘biggest monopoly in human history’ has now had a 4-week stress test against fresh primary data and the trajectory hardened, not softened.

The most quietly important Krishna Rao disclosure is the three-platform compute strategy as a multi-year flexibility moat. Anthropic is the only frontier lab on all three clouds (AWS / GCP / Azure) and the only one using all three chip platforms (Nvidia / TPU / Trainium) at scale. Rao stays at 30-40% of his time on compute alone. The Broadcom 5 GW TPU deal starts 2027; Amazon Trainium up to 5 GW; SpaceX/Colossus carries the consumer and prosumer load. Reads as a direct hedge against the Issue 04 ‘three floors of compute’ constraint thesis — capital floor solved with $75B raised since Rao joined plus $50B more inbound; power floor still tight (see Pax Silica thread); chip floor solved by being the customer everyone wants.

The most disclosed operating data we’ve ever heard from a frontier lab also dropped here: 90%+ of Anthropic’s code is written by Claude Code, the finance team has a 70+ skill library, the monthly financial review is 90-95% Claude-produced. The compute floor for model development is sacrosanct — Rao explicitly will not flex internal training capacity for customer revenue, period. Lost only 2 people to Meta during the compensation poaching war (other labs lost dozens) — the talent retention story is mission-driven, which sets up the Eric Ries thread below.

Pricing-stability disclosure worth the whole episode: Rao dropped the Opus price at the 4.5 launch because customers were ‘fitting Opus problems into Sonnet workloads’ — pure Jevons-paradox bet. Worked. This is the operator-side proof of the Issue 04 token-intensity thesis and the Goldman 24x / Lemkin 250x agentic-token forecasts from this week’s 20VC anchor.

Where I’d put numbers on this:

Cerebras opened the AI-infra IPO window — and the dribble lockup is the structural innovation

Brad Gerstner did the CNBC hit live on the Cerebras IPO day (May 14), which means we have a date-stamped real-time first trade. The numbers:

But the technically interesting part of the IPO is the dribble lockup — a structural innovation worth tracking because SpaceX, Anthropic, OpenAI and the queue of AI-era IPOs are all going to look at it. Traditional structure: 6-month cliff = all insider shares unlock day 180 = massive overhang quant firms arbitrage against. Cerebras’s structure: rolling release. Employees can sell within days; investors after the first earnings report (~5-7 weeks); ~15% released; 6% every few weeks over 6 months. Two consequences: (a) avoids the post-cliff overhang quants front-run; (b) gets supply into a clearly-supply-constrained market faster. Brad posted the chart on Twitter live.

The 20VC anchor (recorded earlier in the week before the actual open) called this almost exactly: range bumped $115-125 → $150-160, $4.8B raise at $48B fully-diluted, 20x oversubscribed. The first trade exceeded their upper bound by ~2x. This is the cleanest signal yet that the IPO-window vibe has flipped from ‘wait for clarity’ to ‘investors will pay for scarcity’ — directly contradicting the Issue 03 / Issue 04 ‘IPOs may slip to 2027’ framing for the mid-cap cohort.

Cerebras also gives the memory-as-the-real-bottleneck thesis from Issue 04 a structural confirmation. Brad explicitly rejects the Cisco-2002 comparison: ‘Nvidia is growing 70% and was trading below market multiple. People propagating the Cisco myth for 2.5 years have missed all the upside. Memory at 5-6x earnings is structural mispricing, not cyclical.’ Sustaining the Issue 04 memory thesis — Nvidia and the memory cohort remain his highest-conviction trades, with Micron CEO booked for the next BG2 pod.

Where I’d put numbers on this:

Incorruptible: the governance moat underwriting AI founder mode

Last week’s framing centred on AI founder mode as the converged operating model (Chesky / Toby / Schoening / Foroughi / Armstrong all saying the same five things). This week the question gets pushed one level deeper: what’s the structural prerequisite that lets founder mode survive contact with capital? Eric Ries on Lennys argues the answer is governance — and the four most operator-rich episodes this week independently rhyme with the thesis.

Eric Ries (Lennys): The Harvard Law statistic that should make every founder file a Public Benefit Corp this week — only 20% of venture-backed founders are still CEO 3 years post-IPO. Standard governance docs that every lawyer hands you are statistical extinction. Centrepiece case study: a hot AI-era pre-IPO company that ignored his warning, went public, competitor got acquired 5 months later, founder ousted. Same bankers/VCs who said ‘wait, this is too early’ kept all the fees.

The book’s structural prescription = ethos + integrity. The architecture Eric points to as the gold standard is Anthropic’s Long-Term Benefit Trust — outside trustees who are AI safety experts, no equity, who appoint directors to the for-profit board. Critically: ‘Whenever Anthropic does the right thing — refuses to release a model, turns down the Pentagon’s $200M contract — think about how much that’s costing them. People say they do it for publicity. Publicity is nice. You know what’s nicer? Having the number-one model that everyone has to pay you to use.’ The Anthropic talent retention story (Rao: lost only 2 to Meta) and the Issue 04 ‘Anthropic monopoly’ trajectory are downstream of this structural choice — it’s why the LTBT exists at all. Cross-reference: at the Vatican AI governance panel, Eric realised that not a single major lab uses standard Delaware governance. That’s how universally bad the standard model is at this scale.

Krishna Rao (ILTB): The mission-aligned execution underneath the LTBT — Anthropic stays at the model-dev compute floor regardless of customer-serving pressure; pricing decisions optimised for Jevons paradox not short-run revenue; 9 of the Fortune 10 on the customer roster; only 2 lost to Meta during the recent comp war while other labs lost dozens. ‘Talent density beats talent mass. People want to work for the good guys.’ This is the operating-side proof of Eric’s thesis — the mission-protection structure is the moat that compounds the talent moat.

Shiv Rao (Abridge, 20VC): The vertical-AI inflection underneath the same operating philosophy. 5-year wilderness (2018-2023), zero pivots on the underlying thesis (‘healthcare is about people having conversations’), ~40% of model outputs in-house, won’t sell data (‘trust is everything’). The counter-positioning frame is identical to Anthropic’s: ‘Build in a way the incumbent can’t because it would impact their current business.’ Microsoft/Nuance is the incumbent. When asked talent vs frontier-model access: ‘B for 6 months, without question. It’s always about the people.’

Charles & Chase Koch (All-In): The industrial-scale 60-year compounding case study — 300 employees and crude-oil gathering 1961 → 130,000 employees across 60 countries, ~9,000x value increase. Operating frame: ‘capability-bounded, not industry-bounded.’ Reinvest 90% of profits. Values-first / skills-second / credentials-last hiring (CIO Jared Benson started painting parking-lot lines, no college degree). Why they stayed private: ‘We never could have built principle-based framework or capability-bounded approach if we’d had analysts forcing us into an industry box.’ Direct rhyme with Eric’s ‘companies with foundation structures are 6x more likely to live to year 50, with superior ROIC.’ The Koch operating frame is what an industrial-foundation governance structure looks like after 60 years of compounding without analyst pressure.

Dana White (Founders / Senra): The single-founder version of the same thesis. ‘I don’t have a Plan B’ as the operating mode. The Spike TV ‘we pay for production, you keep distribution’ deal that retained UFC 100% of content rights — single highest-leverage capital-allocation decision in the company’s history. COVID-as-opportunity: the only sport that kept operating, Iger gave them the paid-no-matter-what guarantee. Joe Rogan did the first 12 fights for free. The Senra/Founders framing — ‘I don’t have a Plan B’ — is the cultural-DNA expression of what Eric Ries codifies structurally in the corporate charter.

The synthesis: AI founder mode (Issue 04) is the operating model. Mission-controlled governance (Issue 05) is the structural prerequisite that makes founder mode durable past contact with capital markets. Anthropic is the working case study of both — and the talent-retention + customer-trust outcomes from the ILTB confirm it.

Where I’d put numbers on this:

Software ‘in the too hard basket’ — ~$180B re-rated to a market multiple

Two converging takes, very different vantage points, same conclusion. Brad Gerstner (Altimeter, long-only public-markets shop) and Marc Benioff (operator running the largest enterprise SaaS company on the planet) both explicitly said this week that the application-software re-rating is structural, not cyclical. The receipts:

Benioff on All-In (the ‘not my first SaaSpocalypse’ frame): Salesforce -37%, ServiceNow -42%, Workday -45%, ~$180B of market cap evaporated. ‘HubSpot at 2x sales — I’ve never seen that before. Salesforce at $46B revenue, $16B cash flow, 83,000 employees.’ And the disclosure that should reframe the entire SaaS investment thesis: ‘I’ll probably use $300M of Anthropic this year at Salesforce.’ Salesforce announces a $50B buyback. Benioff is openly turning Salesforce into an Anthropic-distribution layer — and re-rating his own cost structure accordingly.

Brad on CNBC (the ‘too hard basket’ frame): ‘Software stocks have reverted from a decade-long superior multiple to a market multiple. When you adjust for real SBC and treat them like all the other Mag-7, they’re basically at market. I can’t see out more than 2-3 years on AI’s impact on software, so paying 20x FCF feels fundamentally at odds with AI.’ Exceptions: data-infrastructure layer (Snowflake, Databricks, ClickHouse) — token consumption drives basic-services consumption. Point-solution SaaS = ‘on the front of the conveyor belt heading toward the guillotine.’

The 20VC anchor adds the earnings-divergence data this week confirms the new rule: Monday +20% (raised guidance despite decel), HubSpot -18% (didn’t raise), AppLovin crashed at 30% growth (multiple compression), ZoomInfo at 1% growth = the classic ‘Clay-friends-stole-all-the-growth’ take-private setup at 1x revenue / 35% adjusted operating income. The rule shift: accelerated guidance now matters more than absolute numbers. Companies that don’t visibly accelerate get crushed regardless of size.

This is the direct continuation of Lemkin’s ‘terminal state of decay’ from Issue 04 and the Issue 03 ‘melting iceberg’ framing — three consecutive weeks of converging analyst, operator, and earnings-tape evidence. The most under-discussed implication: the data-infrastructure layer (Snowflake / Databricks / ClickHouse) gets to stay above market multiple because token consumption forces basic-services consumption. That’s where the surviving SaaS premium lives.

Where I’d put numbers on this:

Pax Silica + Trump-Xi: the supply-chain industrial policy is happening fast

Jacob Helberg on No Priors is the most concrete supply-chain industrial-policy disclosure on the tape this year. Pax Silica = 14-country economic-security coalition for the AI supply chain. The first ‘forward-deployed industrial base’ is live: the Philippines has granted the US 4,000 acres (~1/3 the size of Manhattan), held by State Department under embassy-grade authorities. Two-year window to negotiate the multi-decade investor-protection + tax framework. June rollout: 4-5 additional lines of effort.

The strategic frame: ‘We’re not going to do government-operated supply chains because that’s not how we shine. Our superpower is the private sector. American products enchant and delight users around the world by the billions.’ Direct contrast with China’s Belt and Road — state-owned enterprises, debt-traps, equity-conversion on default. Pax Silica deliberately puts private US companies in the driver’s seat. Robotics supply chain explicitly flagged as the highest-priority target. Critical Minerals Summit Feb 4 was the largest in State Department history (55+ countries, MOUs signed). Demand-side: ‘I’m incredibly confident we resolve the pricing issue for minerals before the end of this administration.’

This dovetails with the All-In coverage of the Trump-Xi summit (first face-to-face since 2017): big sales delegation (Cargill, Visa, MasterCard, Boeing, Nvidia, Qualcomm), Polymarket ‘China invades Taiwan in 2026’ at 6% (17% by end-2027). Chamath frame: ‘they’re carving up the world — China steps back from Latin America + Middle East, US relaxes on Asia-Pac.’ Friedberg’s call that should worry every defence analyst: ‘I think 18 months from now Taiwan is no longer an important conversation’ because TSMC Arizona + Huawei mainland fabs make Taiwan less strategically critical. If Friedberg’s read is right, the dual-track de-coupling (US reshoring + China indigenous fab buildout) effectively neutralises Taiwan as the AI-supply-chain choke point sooner than markets are pricing.

The Anthropic Pentagon $200M turn-down (Eric Ries on Lennys) is the inverse data-point for this thread. Helberg explicitly wants tech companies leaning into the security-industrial complex — Anthropic just demonstrated that an LTBT-protected lab can turn down a $200M contract. The dual-mandate problem (lab wants commercial liberty; State Dept wants lab participation in Pax Silica) is the under-discussed AI-policy tension that emerges from putting the two episodes side-by-side. The model distillation IP-protection front Helberg flagged is the lever — labs that engage on IP protection get policy support; labs that don’t, don’t.

Brad’s complementary capex disclosure on the same day: ‘30 or 40% of compute is delayed this year. Keep your eye on power and compute. If delays persist, it will roll back onto revenue and ROI of capex.’ Same number Chamath quoted on Issue 04 All-In. Pax Silica is the supply-side response to exactly this — except it’s running on a 2-year diplomatic-property runway, which is geological by AI-capex standards. This is the tension to watch: industrial policy is moving fast by State Department standards but slow by AI-CFO standards.

Where I’d put numbers on this:

Three short notes worth keeping

The ‘founder brain rewired’ Lemkin disclosure. The MrBeast / rage-bait conversation on 20VC produced the strongest personal-disclosure moment Lemkin’s had on the show: ‘You can never go back. You’re not the same person.’ The thread continues from Toby Lütke and Adam Foroughi in Issue 04 — the founder operating mode has measurable cognitive side effects. Implication for the Eric Ries thesis: if ‘AI founder mode’ is statistically associated with founder-CEO burnout and the 20%-3yrs-post-IPO ouster rate, the mission-guardian structural protection becomes even more important because the founder is the most likely point of failure even before activist investors get involved.

Multi-sensory AI / Mira Murati’s Thinking Machines. The 20VC anchor and the All-In panel both flagged Murati’s product: real-time multi-sensory world model that uploads webcam + desktop + voice to two models in parallel every 200ms. ‘1000x token consumption per knowledge worker if this becomes the default work mode.’ Lemkin’s 250x agentic forecast just got upper-bounded upward. Companion to the Krishna Rao Jevons-paradox Opus pricing disclosure on ILTB — frontier labs are already operating with the assumption that token consumption is the upward sloping curve to bet against.

Vertical-AI proof points: Abridge $5.3B, Legora $100M ARR. Two of the cleanest single-vertical AI build-outs we’ve seen on the tape, both with operator-CEOs / CROs disclosing the operating data. Abridge: 5-year wilderness then sky-opens-2023, doctors refuse to join hospitals without it, ~40% in-house models. Legora: $3.5M → $100M ARR in 12 months, 8-12x quota multiples, 280% average attainment, 78% pilot conversion. The agentic-GTM playbook is now codified: demo-first not discovery-first, forward-deploy domain experts at every $100k+ deal, on-site by second meeting. Direct continuation of Project Hawaii from Issue 04 — the lean-ops Navy SEAL team is also the agentic-GTM team.

Ten episodes, 12.1 hours, 6 of them speaker-labelled. The Anthropic ARR triangulation finally closes — CFO-direct numbers match the Sacks / Brad numbers from Issue 04. The Cerebras open at $100B market-cap and dribble-lockup innovation re-opens the AI-infra IPO window earlier than Brad’s Milken timeline suggested. The AI founder mode framing from Issue 04 gets its structural sequel: Eric Ries argues the mission-guardian governance choice is the moat that compounds the talent moat and customer-trust moat. The software re-rating is now a 3-week pattern from three vantage points (Lemkin, Brad, Benioff). And Pax Silica + Trump-Xi shows the supply-chain industrial policy is moving fast — though still slow by AI-capex clocks. Next week: watch the Cerebras price action against the dribble-lockup test, whether Polymarket adds a ‘next Pax Silica geography’ contract, and whether more late-stage AI labs publicly adopt LTBT-equivalent structures in the wake of Eric’s book launch (May 26).

This Week's Episodes