Anthropic Buys Compute From Elon & Commits $200BN to Google | Cerebras IPO | Ramp Raises at $40BN

Key points

• **Anthropic SPV crackdown — the secondary market cleanup.** Anthropic publicly named bad-actor entities operating unauthorised SPVs and 'beneficial ownership' contracts. Secondaries trading $200-400B reportedly. The mechanic Rory walks through: legitimate SPVs (VC over-allocation vehicles) vs. illegitimate ones (contractual transfers of economic interest without share custody — 'Jason kind of has the economic right but no shares on the cap table'). Lemkin: 'nothing-burger — Anthropic warned everyone last year, even called out a top-investor publicly. People wouldn't listen because the greed was too high.' **Pre-IPO sweep: Anthropic going public soon and cleaning up the equitable-remedy risk in any cap-table mess.**
• **Anthropic ↔ Elon — the deal three months in the making.** Elon's xAI/Colossus had 11% data-centre utilisation; Anthropic leases all available capacity (the same Colossus 1 deal we covered in Issue 04, but with new Elon-only side and price details). **$3-5B/year in revenue to SpaceX = ~15% lift to SpaceX's $20B run rate** for an asset that was a money pit 60 days ago. Same week, Elon was calling Anthropic 'evil, anti-white, anti-Chinese.' **Elon's xAI/Grok essentially concedes frontier-model race — pivots from net-buyer of capex to net-seller.** 'Their own little CoreWeave.' Now performa SpaceX IPO has 4 distinct businesses to combine.
• **Anthropic commits $200B to Google over 5 years = ~40% of Google's total future revenue backlog.** Google: model competitor with Gemini AND lead compute supplier. **'I'm sure Sundar would prefer Gemini won, but happy to be paid $200B by your competitor.'** Stat from the WSJ this week: in 9 months Gemini enterprise market share went 27% → 40%, Claude went 21% → 48% (multimodal, doesn't sum to 100). Google now has the fastest-growing #1 (Gemini) AND a large share of the leader (Claude) feeding compute revenue.
• **'Anthropic figuring out a way to have more compute than OpenAI.'** Lemkin frame: '2 weeks ago Anthropic couldn't launch Mythos because of compute. Now they're hoovering up everything — CoreWeave, xAI, anything orbiting.' Sub-text for the SpaceX roadshow: bankers have to explain a quarter with 4 different businesses combined (legacy SpaceX + Starlink + xAI/Colossus + Cursor).
• **Cerebras IPO 20x oversubscribed** — range $115-125 → $150-160, raising **$4.8B at $48B fully-diluted**. Foundation Capital, Eclipse, and Benchmark each have ~8-9% (not 20% as Twitter rumour). Lemkin admires the Andrew Feldman story: 2016 incubation by Foundation, surviving 2022-23, pulled previous IPO, now riding the OpenAI commit + Amazon commit. **'This is actual venture capital — what the industry is supposed to do, but no one does it anymore.'** Rory: expect a pop, but 'when bankers raise the range this much, they're highly confident.' Companion to the [Issue 04 'inference is the last market' Baseten thesis from Issue 03](/issues/2026-05-03).
• **Ramp at $40B valuation on $1B revenue = 40x run-rate** (Brex comp ~6x; Brex even on slower growth). Ramp's agentic procurement announcement — agents that negotiate with suppliers automatically. **Lemkin commits publicly: 'I'm going to leave Brex or stay with Brex based on which has the best agents.'** Sub-text for the multiple: 'pricing in 2.5 years of doubling before you're at the Brex multiple — outer edge of terrifying.' But the preference stack + same-investors signal = they think across the round it's fine.
• **Earnings divergence — what wins now is 'accelerated guidance,' not just numbers.** Monday +20% (decelerated growth but RAISED guidance for next quarter); HubSpot -18% (decel + didn't raise); AppLovin **crashed at 30% growth** because trading at high multiple — 'people are no longer willing to give 20x; they'll give 5x for 30%+ growth + profits + a story.' Cloudflare beat numbers + cut 20% of base = stock down (concern: 'why cuts now, what's wrong'). Anti-Issue-03 ZoomInfo case study: 1% growth, **35% adjusted operating income, 1x revenue** = classic PE take-private setup. 'Clay and friends stole all of ZoomInfo's growth from it.'
• **Token consumption debate sharpens** — Goldman: agents push tokens 24x by 2030 (Lemkin: 250x easily, parallel agents). But the *counter-counter:* 'we don't need as many tokens as we think. Web devs are token-trashing. 10x and 100x engineers will use lots; the average is wasting.' Rory's Goodhart's Law observation: monitoring tokens distorts production (the Amazon/Meta token-leaderboard cases). The two truths: best engineers leverage massive token spend productively; mediocre engineers burn tokens performatively. Cross-references the Lemkin token-intensity surprise from [Issue 04](/issues/2026-05-10) ($254/mo for 2 autonomous AI VPs).
• **Anthropic releasing 10 financial agent templates** (killing some YC companies) and **about to release a legal product** (Harvey + Lagora threat). Rory's frame: this is the classic Microsoft-Office-vs-vertical-app dynamic. Microsoft tried (Great Plains acquisition) and failed. Snowflake survived AWS Redshift despite VCs writing it off. **Sub-text: most of the app layer will survive but with massively compressed growth and stronger evolutionary pressure than any prior platform shift (operating systems took a decade to disrupt; agentic models could do it in 18 months).**
• **'Terminal state of decay' framing — Lemkin's strongest call on legacy SaaS.** 'In a year we won't need any traditional marketing automation software. Agents do not need HubSpot or Marketo or Salesforce. They have no need to hand-compose an email in a third-party template.' Direct continuation of the [bucket-1 'melting iceberg' SaaS framing from Issue 03](/issues/2026-05-03). The decay rate is 18 months in the agentic era, not 10 years.
• **Memory chips — Lemkin's 'is it too late' question.** Lemkin explicitly asks Rory if he should buy SK Hynix / Micron now. Rory's frame: 'how long does the CapEx boom last vs how fast can they build new fabs.' Korea is digging holes in the ground now that come online 'just as Anthropic starts to cut orders in 2028.' **Honest tension on this week's highest-conviction trade from [Brad Gerstner](/concepts/people/brad-gerstner) at Milken — the trade works only if you believe the boom outlasts the fab-supply response.**
• **MrBeast rage-bait conversation — Lemkin's strongest founder-disclosure on the show.** MrBeast tweeted: 'sacrifice of mental health is required to have the success.' Both Lemkin and Rory agreed. **Key insight from Lemkin: 'after 4-5 years your brain is permanently rewired. You can never go back. You're not the same person. Vacation doesn't fix it. Buying a yacht doesn't fix it.'** 'Any normal human would have sold for $50M / $200M / $1B. You have to be Daniel Dines or Adam Foroughi or crazier to do this journey.' Rory adds important nuance: 'you still have to find a way in that maelstrom to keep perspective — when you're not making good decisions, you actually owe it to yourself to pull back a little.' **Cross-references Adam Foroughi's biographical interview from [Issue 04](/issues/2026-05-10) directly: 'his brain's been rewired. The things he's saying make total sense to me after my journey.'**
• **Lime announces IPO prep** — the bike-rental company. Solid healthy business turnaround. Worth tracking as another mid-priority IPO calendar entry alongside SpaceX, Cerebras, and the broader 2026 cohort.

Notable quotes

Themes

• Anthropic capital-stack consolidation week (SPV crackdown + Elon deal + Google $200B)
• Cerebras IPO 20x oversubscribed at $48B as the first AI-infra IPO of the cycle
• Earnings rule shift: accelerated guidance > absolute numbers; AppLovin crashes at 30% growth
• Terminal-state-of-decay framing for marketing-automation SaaS (Lemkin)
• 'Founder brain rewired' — Lemkin's strongest disclosure on the show