Cursor at sixty, Anthropic at a trillion, and the year of the agent operator

The Threads

The week of $60B Cursor and $1T Anthropic

Two numbers reframed the AI cycle this week. Cursor sold to SpaceX/xAI for $60B with a $10B break clause if SpaceX’s IPO doesn’t close — the largest private VC exit in history, beating Wiz at $32B and WhatsApp at $16B. Three years from seed to $60B exit. Anthropic crossed $1T in secondary markets after declining $800B funding offers — a doubling of last week’s $850B-OpenAI flippening number — and Tim Cook announced he’s stepping down from Apple after 14 years.

Rory’s macro frame on 20VC: ‘when your stock is valued at 100 times revenues, you can buy things trading at 10 or 15 times revenue all day long.’ SpaceX trading at ~$2T (≈100x revenue) acquiring Cursor at ≈10x year-end revenue is 3% of market cap for ~15-20% of revenue. Free arbitrage if you can move quickly. Chamath on All-In read it the same way: ‘effectively Elon got a 50% discount.’ Jason Lemkin’s prediction — ‘there’ll be a $100B deal in the next 12 months’ — is no longer obviously hyperbolic.

The mechanism behind both numbers is the same: AI capability scarcity is so concentrated, and the top labs’ demand-side leverage is so total, that the aggregate market cap of the world’s seven $2T+ companies is now denominated in defensive M&A budgets. Boards meeting weekly to ask ‘who do we buy to not fall behind’ is, per Lemkin, real and routine. The Overton window for M&A pricing has expanded. Once $60B exists as a comp, $10-20B feels like a tuck-in.

Where I’d put numbers on this: SpaceX completes its IPO by end of August at ≥$1.5T [forecast: 2026-04-26-002] — Polymarket has it at 80%; my read is the regulatory + market window is good but the valuation bar is ambitious. The Cursor acquisition closes rather than triggering the break clause [forecast: 2026-04-26-001] — Polymarket’s 74% feels right; Elon walked away from Twitter once but he doesn’t pre-promote a deal he’s planning to walk from. And further up the M&A escalator: at least one more $50B+ private acquisition lands within twelve months [forecast: 2026-04-26-003], driven by the seven-trillion-dollar-market-caps-buying-defensive thesis.

Dylan Patel on Invest Like the Best put a number on the demand side that explains why: Anthropic’s gross margin moved from the 30s% at the start of the year to a floor of ~72% now — the steepest margin step-up in commercial software history. Revenue $9B → $40-45B in a few months. They are throttling rate limits, capping access, and demand still exceeds supply. That run-rate continues — Anthropic crosses $80B annualised within twelve months [forecast: 2026-04-26-005] is my call, with confidence around half. The scenario where it doesn’t is that Mythos forces a pricing crisis or governance event before the curve compounds.

The SaaS apocalypse counter-attack

Last week’s Lemkin 60%-death-spiral thesis finally got serious responders. Three this week — and they cancel each other out interestingly.

Aaron Levie (Box) on 20VC sided ~80% with Jensen Huang against the doomers. His thesis from outside the platform: ‘There will be more lawyers in five years than today. More engineers, not fewer. Tech is only 8-12% of GDP. What happens when the other 85% finally gets engineering capability?’ And his single best new framing: a new role called ‘agent operator’ — technical, but not core engineering. Knows MCPs, CLIs, agents.md files. Sits inside marketing, legal, ops. 500k-1m of these jobs by Levie’s estimate. Within twelve months that title (or a near-synonym) shows up in a Fortune-500 job posting outside tech [forecast: 2026-04-26-010] — confidence 0.6.

Amjad Masad (Replit) on 20VC said the SaaS apocalypse ‘is justified.’ But added a crucial twist: coding models are approaching plateau on the S-curve. Plateau opens room for fine-tuned vertical models — Intercom’s customer-support model already beats the frontier. ‘Cost question is secondary to performance — until you reach the asymptote, then cost becomes everything.’ Replit’s next ICP target after product teams: operations. Whether the plateau is real-real or marketing-real is testable: in six months top-of-leaderboard improvement on a representative coding benchmark is below 10 percentage points [forecast: 2026-04-26-007]. I have this at 0.4 — I think the plateau is rhetorical more than mechanical right now.

Philip Herzig (SAP CTO) on No Priors delivered the most thorough enterprise-incumbent rebuttal yet. ‘AI is a business-model transition, not a technology transition.’ Three layers being re-engineered: UI (generative replaces click-to-teach), processes (structured + unstructured blend), data (harmonised semantic model). The hard part isn’t building agents — it’s scaling them. Reasoning at 10 documents is a CEO-impressing demo. At 100,000 documents and 20,000 APIs with master-data coupling (Sarah is a US employee, Philip is German — same query, different correct answer), it’s an actual engineering challenge.

The reconciliation: the dying-SaaS thesis is right about the point-tool incumbents (workflow vendors with thin business logic). It’s wrong about platforms with deep state, regulatory weight, and end-to-end orchestration (SAP, Box, ServiceNow). My split-the-difference forecast: at least two of {ServiceNow, Salesforce, Workday, HubSpot, Adobe} see forward P/E compression of ≥20% within twelve months [forecast: 2026-04-26-004], and at least one major SaaS incumbent does an acquihire or product-line wind-down explicitly attributed to LLM-native displacement within nine months [forecast: 2026-04-26-009]. Confidence ~0.5 and ~0.45 respectively. The narrow-death-spiral version of Lemkin, not the broad apocalypse.

The PM debate: dying role or renaissance?

Last week Keith Rabois said ‘the idea of a PM makes no sense.’ This week two responses, from different angles, both right.

Nikhyl Singhal (Skip Community) on Lenny’s ran the data: ‘we have the most open PM roles globally in 3+ years.’ Forecast for the next 12-24 months: ‘companies will shed 30,000 and hire 8,000 — but the 8,000 are AI-first.’ The classical information-mover PM is dying. The judgment-and-builder PM is in a complete renaissance with comp at all-time highs. ‘Builders are going to have the time of their lives. If you don’t love building stuff, you’re in trouble.’

Cat Wu (Anthropic, Head of Product, Claude Code) on Lenny’s gave the inside view of what ‘AI-first’ looks like in practice. Feature timelines collapsed from 6 months → 1 month → 1 day. Most external PM candidates ‘approach it incorrectly’ — they optimise for multi-quarter alignment when the real job is shortening idea-to-launch. Anthropic’s hiring bias: engineers with great product taste over more PMs. ‘Product taste is rare. We’ll hire anyone who has demonstrated it strongly.’

The synthesis across Rabois + Singhal + Wu: the role is dying and reborn. Whichever company you think you work at, the PM job description from 18 months ago is largely obsolete. The new job is closer to ‘chef-mode’ from last week — what are we building, why, who’s it for, how do we differentiate. People are mostly not hiring more PMs; they are reshaping the hiring bar toward people who can ship and judge in the same head.

Token economics: from compute scarcity to value arbitrage

Last week Anj Midha called this ‘the GPU wastage bubble’ — compute is not fungible, billions sit stranded. This week the framing shifted to the demand side.

Altimeter Capital ran a 3-minute clip stating the headline number: inference cost is down 99% in 2.5 years. Yet H100 prices are still rising. Reason: model size (1T → 10T parameters) and demand are both growing faster than per-token cost falls. Cost-per-unit-intelligence collapses; aggregate compute spend goes up.

Dylan Patel (Semianalysis) showed the operational consequence at the firm scale. Semianalysis pays $7M/year for Claude code against a $25M salary base — 28% of payroll on AI. One analyst built a US-grid energy data set in three weeks ($6k/day of tokens) that would have taken a 100-person incumbent vendor a decade. Patel’s frame for the new economy: ‘A year from now the business is just arbitraging tokens. The model will know what to do with them three years from now.’ Token-direction is the new alpha.

Mythos, Patel says, is ‘potentially the biggest step up in model capabilities in two years’ — and Anthropic deliberately released a weaker version (Opus 4.7, ‘preferentially worse at cyber’) to the public. The $20k-of-credits-per-run cost of Mythos at full capability creates a clear gap for cheaper specialised cyber models. Within six months, at least one frontier lab releases a smaller dedicated cyber model competing with Mythos at meaningfully lower token cost [forecast: 2026-04-26-006] — that’s my 0.55 call, driven by Sacks’ thesis on All-In that cyber is now the white-hot subcategory of coding.

The thread that runs through this whole conversation: in Issue 01, Anj Midha’s ‘four bottlenecks’ framing put compute as one of four. By Issue 02 the bottleneck is more specific — it’s frontier-model inference at the asymptote, and the people with the most leverage are the ones who can direct those scarce tokens at the highest-value problem. The skill isn’t writing the code; it’s choosing the prompt.

GLP-1s and the trillion-dollar health revolution

Brand new theme this issue. Alex Karnal on Invest Like the Best calls 2025 the single most exciting year in 20 years of biotech. GLP-1s are the proof point: 94% reduction in pre-diabetic-to-diabetic conversion (Lilly data), heart attack and stroke risk down materially, kidney protection, addiction signal. $100B+ revenue easily — but the bigger story is that GLP-1s are the first commercial proof we’re ready for proactive medicine.

His ‘Health Stack’ has five layers, each with existing medicines if people are proactive: lipid optimisation (statins, PCSK9), cardiometabolic (GLP-1s), neurocognitive (anti-amyloid medicines, next Lilly readout this year), inflammatory (food-driven), blood pressure. ‘The gap is not needing more medicines — it’s pointing them at the impact they can have.’ Lifespan curves have been flat since the antibiotics era. The five-layer proactive stack could add a decade.

Three commercial discoveries from 2025: (1) people optimise for tolerability, not weight-loss-maximisation; (2) Lilly Direct crossed >50% of new joiners — pharma is becoming consumer-software-shaped; (3) compounded GLP-1s revealed massive price elasticity — at $200/month vs $400-500, an extra 15-20% of the market opens up. Within twelve months that direct-to-consumer share crosses 70% across the major players [forecast: 2026-04-26-008]. Pharma’s distribution model is changing in real time, in the same way SaaS’s did 15 years ago.

Founder mode and the relentless application of force (continuing)

Last week’s thread on complacency-management runs through this week too. Lemkin’s line on the SaaS-debt-bomb has the same shape: companies coasting on financial engineering rather than product reacceleration get punished. Cat Wu’s ‘remove every single barrier to shipping’ is the operational version of Mike Moritz’s ‘relentless application of force.’ Singhal on velocity-stress: ‘once you figured out your job, you used to be fine for a decade. Now if you don’t stay up in the next three months, they’ll tell you you’re doing the thing we stopped doing three months ago.’ The compounding cost of complacency just shrank from years to weeks.

Predictions filed this issue (the ledger)

Ten forecasts have been written to data/evals/predictions/2026-04-26.json with confidence and horizon. They will be reviewed via podcast-digest review review-outcomes as their horizons mature. Brief table for the curious:

If the system is honest, half of these resolve hit and half resolve miss. The whole point of the ledger is to find out.

Issue 02 was 10 episodes (~11 hours of audio) and the concept graph grew to 520 entities (128 people · 169 companies · 223 ideas). Top mentions across both issues: Anthropic (13×), OpenAI (12×), Elon Musk (9×), Nvidia (8×), Cursor (7×). Browse the concept index or search the archive for ‘agent’, ‘cursor’, ‘GLP-1’, or anything else.