Monday, May 11, 2026 · 89 newsletters

Anthropic Rents Elon's Colossus

anthropic-colossus · agent-bifurcation · anthropic-fintech · trump-xi-prep · iran-hormuz · voting-rights · ai-regulation · ai-trust · supply-chain · europe-decline

Published on Monday, May 11, 2026.

Pulled from 73 newsletters sent to read@madho.net yesterday. A Sunday inbox dominated by one deal that crossed every category at once: Anthropic leasing SpaceXAI's Colossus 1 supercluster and, in the same week, shipping ten finance agents through FIS. Around it: the Trump-Xi summit prep, Iran's tightening grip on Hormuz, and the slow procedural collapse of the Voting Rights Act after Louisiana v. Callais. Here is the signal organized by trend.

The Big AI Story: Anthropic Rents Elon's GPUs, and the Agent Story Splits in Two

By volume and by significance, the dominant thread of the day. Alex Banks at The Signal led with the deal everyone was processing: SpaceXAI, the February merger of SpaceX and xAI, handed Anthropic the full compute capacity of Colossus 1 to serve Claude inference: 220,000-plus NVIDIA GPUs and 300-plus megawatts online within the month. Claude Code's five-hour rate limits doubled across paid plans, peak-hour throttling on Pro and Max is gone, and Opus API ceilings jumped. Jamin Ball at Altimeter modelled Colossus 1 at standard rental rates as roughly $5B of annual revenue for SpaceXAI; applying Dario Amodei's training-inference framework from Dwarkesh, Anthropic could turn that $5B compute into roughly $15B of inference revenue at 60 to 70 percent margins. Three months ago Elon called Anthropic "misanthropic"; last week he called them "highly competent". What changed was the unit economics.

The lease tells you what the two companies actually are right now. SpaceXAI moved frontier training to Colossus 2, a 1.5GW cluster roughly 3x the power and 2.5x the GPUs of Colossus 1, so the old cluster was depreciating idle; Anthropic, meanwhile, cannot buy capacity fast enough to serve enterprise demand. The headline framing is rivals turning friendly. The actual framing is that compute is now a tradable commodity between labs, and the lab that monetizes inference best wins. (Wall Street has not priced this in.)

The agent story bifurcated this week, which is the more interesting structural shift. Every's weekend dispatch was on the ground at Anthropic's Code with Claude developer conference and identified the split: desktop apps for synchronous collaboration on one front, long-running agents like OpenClaw or Claude Managed Agents that teams hand off work to on the other. The same day, OpenAI ran a "low-key" GPT-5.5 launch at 5:55pm on 5/5, a time picked by the model. The Anthropic event announced three new features for Managed Agents alongside the SpaceX compute deal.

Builder skepticism on parallel agents is finding its voice. Claude Cowork argued that the hard part of parallel-agent setups is not launch, it is inspection: "Confidence in the final message doesn't prove the task survived contact with the files, sources, tests, or review process." Peter Yang's interview with Moritz Kremb walked through a personal-OS Claude Code build that handles email, content, and groceries, with Moritz explicitly framing the move off OpenClaw as a reliability problem. The picture across the day is consistent: agents that scale work are the durable bet; agents that run autonomously without inspection layers are not. The fact that the SpaceX-Anthropic deal lit the agent-economy debate on fire is a sign the conversation is no longer about whether agents work, but which agent topology survives contact with production.

Anthropic Eats Financial Services: The Other Half of the Story

A second cluster, equally important, that the AI press buried under the Colossus headlines. Simon Taylor at Fintech Brainfood called this "the week Anthropic tried to eat financial services": a partnership with FIS as Anthropic's way into banking, a private-equity sales partnership, and ten new agents for KYC and Credit. Taylor's reporting from lab sources: finance is now Anthropic's single largest enterprise category.

Linas Beliūnas at Linas's Newsletter argued the same shift in stronger language: "Anthropic stopped selling AI to Wall Street and started becoming Wall Street's Operating System," and noted Anthropic also outsourced enterprise sales to private equity. Rich Turrin at Cashless framed the inverse from the bank side: banks running out of time because of governance failures, AI agents already routing customers away before incumbents have a response. He also covered the IMF drawing a hard line on agentic payments: let agents think, do not let them pay, because AI is probabilistic and payment systems demand certainty. Sam Boboev at Fintech Wrap Up shipped a 2026 directory of stablecoin card-program enablers, anchored by MiCA's July 1 hard deadline in Europe and the GENIUS Act maturation in the US, with 38 accredited EMT issuers now operating in the Eurozone.

The supporting cast on the fintech fragility side: Jason Mikula at Fintech Business Weekly on the Parker bankruptcy, a Y Combinator and Valar Ventures-backed SMB banking startup that partnered with Patriot Bank and abruptly shut May 4 after Patriot pulled the plug; and Luke Sophinos at Linear on the brutal SaaS multiple reset, with median public EV/revenue at 3.4x per Aventis. Two narratives running in parallel: Anthropic owning the enterprise stack at one end, and the BaaS/vSaaS layer below it cracking under regulatory and valuation pressure at the other. The honest read is that the AI-finance buildout and the legacy fintech cleanup are the same story, and Anthropic is now the largest single beneficiary of both.

China: The Meeting Is Being Pre-Negotiated in Public

Three sources converged on the same read of next week's Trump-Xi Beijing summit. Bert Hofman read the Chinese media setup as deliberately modest: not "reset" but "managed coexistence under pressure," with He Lifang and Treasury Secretary Bessent doing last-minute Seoul prep this week. China's first-quarter growth came in at 5 percent year on year and April exports grew 14.1 percent per Dexter Roberts at Trade War, which is the leverage Beijing is bringing into the room. Roberts expects no breakthroughs, but a tariff truce extension, Chinese commitments to soybean and Boeing purchases, and Washington pressing Beijing on Hormuz, paired with a Beijing quid pro quo on advanced semiconductors and rare earths. American CEOs from Boeing, Mastercard, and Citigroup will accompany Trump. Bruce Mehlman's Six-Chart Sunday framed the AI overlay: Trump and Xi are reportedly setting up official US-China AI talks at the summit, which would be the first formal channel.

The interesting thing about pre-negotiating in public is that it tells you the deliverables are small. The Chinese side is comfortable: courts have contained the tariff regime, rare-earth export controls neutralized the tech-restriction lever, and exports rerouted through third countries are still growing. Trump needs a win that does not look like one to his base; Xi needs the meeting itself, not any specific outcome. The China policy of the next twelve months is being written as theater.

Iran and the Strait: Month Two of a War Nobody Is Reframing

Two notable threads. Gov Brief Today led with Iran's Revolutionary Guard threatening to attack US bases if Iranian tankers are hit, the day after a US fighter jet disabled two tankers in the Gulf of Oman. Trump is "awaiting Iran response," which has been the posture for weeks. John Ellis and Richard Haass on News Items' Alternate Shots laid out Trump's three options at this stage of the war: cut a deal, escalate, or carry on as before with sanctions and blockade. Each carries serious risk; each makes some part of the coalition unhappy. Maritime Analytica's executive brief collected the week's analyses on shipping, including whether Trump's "Project Freedom" can reopen Hormuz and how MSC is moving cargo without relying on Hormuz.

The carriers are quietly re-routing global trade around the conflict and nobody is treating that as the story. The political class is still running on the assumption that this is a discrete war with a clear endpoint; the shipping class is treating it as the new permanent topology of the Gulf. One of those two reads is right and the other is going to look badly wrong by the fall.

Politics: The Voting Rights Act Becomes Procedurally Dead

The dominant political thread on the day. Lincoln Square's Brian Daitzman wrote the cleanest piece on what Louisiana v. Callais actually did: by tightening the limits on how race can be used in redistricting, even in service of Section 2 compliance, the Court did not pick a winner, it changed the system. Plaintiffs now have to distinguish race from partisan behavior to prove vote dilution, which is essentially impossible. Lauren Egan at The Bulwark ran a long piece on Southern Democrats facing the aftermath: Tennessee Republicans moved aggressively to carve up Memphis's majority-Black district, and State Rep. Justin Jones's "history books with Bull Connor" speech was righteous and, in immediate operational terms, ineffective.

Marc Elias at Democracy Docket made the broader frame: the Roberts Court has, over a decade, "declared fire hydrants, sprinklers, and smoke detectors illegal" while Republican legislatures lit the match. Rick Wilson ran a literary parody on MAGA and idolatry that landed as a sharp counterpoint: the party is now a religious movement, and the Cassidy-Massie-Boebert purges are heresy trials.

The structural read here is that the Voting Rights Act is not yet dead by statute, but is now procedurally dead by evidentiary standard. The Tennessee map, the Louisiana map, and the Missouri map are no longer outliers. They are the new equilibrium, and the 2026 midterm map is being drawn around them now.

AI Regulation: Coming Soon, From a Direction You Did Not Expect

Bruce Mehlman's six-chart Sunday made the most coherent case of the day that AI regulation is now a when, not an if. The trigger sources are not what you expect: NEC Director Kevin Hassett floated FDA-style pre-release safety review for AI models earlier in the week, the White House walked it back, but the trial balloon went up, and The Economist ran a piece on AI granting extremely dangerous powers including bioweapon synthesis. A March NBC poll found AI less popular than ICE among the US public, which is a meaningful political signal.

Morning Consult's AI Trust Report lined up beneath that finding: AI is now among the least-trusted industries Morning Consult tracks in the US, even as Americans report more practical use of the tools. The global picture is starkly different, with trust stable or rising outside the Anglosphere. The honest read is that the regulatory window opens not because of any single capability launch but because public trust is collapsing while capability is rising, and that combination produces regulation in every prior tech cycle without exception.

Supply Chain: Amazon Goes For-Hire, the Carriers Reprice

Matthew Hertz at Sent Items covered Amazon Supply Chain Services going live to external sellers, with shipping stocks taking an immediate hit per the WSJ logistics report and Supply Chain Dive. Hertz's smart take: this looks more like a relaunch of Supply Chain by Amazon, now open to non-Amazon sellers, than a wholly new offering, more loud than material. Maritime Analytica covered COSCO's Q1 revealing a new container shipping reality and the ZIM-Hapag-Lloyd deal closing under new geopolitical conditions.

The deeper signal is that Amazon's logistics arm is now public infrastructure for sale, and the carriers are repricing for a Hormuz-permanent world. Two pieces of supply chain news that look unrelated, both pointing at the same fact: the era when global shipping was a commodity invisible layer is over.

Ideas Worth Reading

Paul Krugman pushes back on the European decline narrative. A useful contrarian read against the Lutnick-at-Davos consensus, including the Wall Street Journal "as poor as Mississippi" framing. Krugman concedes Draghi's productivity concerns but argues life expectancy, healthcare outcomes, and median living standards tell a different story.

Eric Ries on Lenny's podcast about Incorruptible. 80 percent of venture-backed founders are ousted within three years of going public. Ries's new book argues the governance structures that protect Anthropic, Costco, and Novo Nordisk are a two-page legal filing most founders never make. The Anthropic example is the timely one given the Colossus deal.

Hannah Zhang at Nonlinear News on lifestyle businesses post-corporate. An interview with NYC stylist Chelsea Hollenbeck on the move from oncology nursing to a fully-booked styling practice. The argument: taste and curation are core business skills in the AI era, not luxuries.

Shane Parrish on Brain Food, "Do Your Habits Match Your Ambitions?" Three tiny thoughts: trust habits over stated ambition, success is 50 small things in the same direction, and most beliefs are soft until challenged. The frame on disagreement, "seems like you have a reason for saying that," is the keeper.

Ben at Next Play on ten startup ideas worth building. Notable for the AI Talent Tracker concept, which monitors LinkedIn profiles of OpenAI, Anthropic, Meta AI, DeepMind, and Cohere employees for departure signals, and the Forward Deployed Engineering community pitch. The list reads like a working set of 2026 founder briefs.

George Bounacos on Gov Brief Today #464. 7,331 stories over 464 consecutive nights, free and ad-free. The Mississippi firing-squad billboard piece (Trump's 1989 Central Park Five ad in four New York papers, the through-line from there to now) is the kind of context the cable feed strips out.

Outside Interests

Brick on building your own farmers market meal plan. Reverse engineering meals from foraged ramps, fiddleheads, stinging nettles, and king trumpet mushrooms. The "plug-and-play" method is the useful part if you have ever come home from a market wondering what to do with komatsuna.

Polina Pompliano on Jake and Logan Paul at The Profile. Months of reporting on whether internet attention converts to lasting wealth and power. The crux: both brothers want to be taken seriously by the world they got rich by not taking seriously.

The Culturist on leisure as the basis of culture. A primer on Josef Pieper from a retreat in northern Italy. Leisure is not the absence of work, it is the gateway to reality, which is a more demanding claim than the Netflix-and-recharge version.

DrawTogether's Wendy MacNaughton on Thomas Curran's Perfection Trap. Perfectionism is not a character trait, it is a worldview built on a permanent sense of deficit. The drawing frame makes the abstract concrete in a way Curran's book sometimes does not.

Abby Falik on being ghosted by the LinkedIn algorithm. Includes a smart sub-essay on MrBeast as the inverse of Mister Rogers, with the operating-instructions memo line "Show no hesitation. Silence is failure." The MrBeast piece linked from Falik is worth the twenty minutes.

1440 Sunday's Mother's Day edition on pregnancy. The fact that brain areas heavily involved in processing emotions shrink by 5 percent on average during pregnancy, that feet may permanently grow, and that miscarriage rate runs as high as 20 percent in the US are facts the cultural narrative rarely centers.

Data Worth Noting

Anthropic's projected $15B inference revenue against $5B compute spend on Colossus 1, per Jamin Ball at Altimeter via The Signal. 60 to 70 percent margins, online within the month, 220,000-plus GPUs. The unit economics are why the deal happened, not the rhetoric.

Chinese exports grew 14.1% year on year in April, per Dexter Roberts at Trade War. Q1 GDP came in at 5 percent year on year. The leverage Beijing is bringing to the summit table is structural, not theatrical.

Median public SaaS EV/revenue multiple at 3.4x as of March 2026, per Aventis via Luke Sophinos at Linear. Down from peak-era 10x. ServiceNow 4.8x, HubSpot 2.7x, Shopify 8.3x. The valuation reset is the precondition for the AI-native re-rating now underway.

Three Takeaways for You

The AI story for the day was not the SpaceXAI-Anthropic compute lease in isolation; it was the lease plus the FIS partnership plus the ten finance agents plus the agent-economy bifurcation, all in one weekend. Anthropic is no longer a frontier-lab challenger; it is structurally larger than its public framing, and pricing in the equity markets has not caught up. The story to watch over the next two months is whether OpenAI responds with a comparable enterprise pivot or stays in the consumer-and-research lane, because the customer-acquisition window in regulated finance is closing fast.

American politics is operating on two clocks. The slow clock, Iran and the China summit and inflation, is producing the cable-news daily cycle. The fast clock, the Voting Rights Act procedurally dead and redistricting maps redrawn in five states and Mississippi firing-squad billboards, is reshaping the 2026 electoral map in real time. Democrats who treat the November environment as the variable are misreading; the variable is the map itself, and the map is locked in by July.

If you only read three pieces, I'd suggest Every's "AI Work Is Splitting in Two" as the clearest read on where agents are heading, Simon Taylor's Token Economy at Fintech Brainfood for the Anthropic-as-Wall-Street-OS thesis with the receipts, and Lincoln Square's Brian Daitzman on Louisiana v. Callais for the most under-covered structural shift of the spring.