whatimreading

Monday, August 3, 2026 · 77 newsletters

The AI Trade Cracks

ai · fintech · prediction-markets · politics · data-centers · marketing · media · china · macro · culture

Published on Monday, August 3, 2026.

Pulled from 77 newsletters in yesterday's inbox. A blown up AI hedge fund, Korean circuit breakers, cheaper Chinese open weight models, and a states rights fight over prediction markets all showed up on the same Sunday, all pointing at the same thing: the trade has priced in the future faster than the future is arriving.

AI: The Party Ends Loud

The blowup. Leopold Aschenbrenner's Situational Awareness fund, at one point a $45B AI focused book, was up 439% net through June, lost roughly 67% in July, and, because it was 4x leveraged, sold every public position it owned to Ken Griffin's Citadel. Simon Taylor at Fintech Brainfood led with it Sunday morning. Matt Stoller picked up the same story and pushed it a step further: it is time to force big AI firms to open their finances, since the marks Aschenbrenner was leveraging into were partly private (Anthropic among them) and effectively unauditable. Rich Turrin's Cashless newsletter framed it as "The Summer the AI Party Ended as Blank Checks Dry Up," pointing to WSJ reporting on Chinese open weight migration at a tenth of Western pricing.

Korea's circuit breaker week. Chamath's frame did the most work of any single piece Sunday. The KOSPI triggered consecutive circuit breakers, wiping roughly 40% from a July record and $2.18T in Seoul market value. Nine weeks earlier Seoul had greenlit single stock 2x leveraged ETFs. Retail crowded into the Samsung and SK Hynix versions; when SK Hynix fell nearly 20% intraday on July 29 the ETF mechanics forced further selling into the drop. CXMT, China's largest DRAM producer, meanwhile rose 466% on its Shanghai debut on July 27. Chip supply, memory demand, and Chinese domestic capacity all crossed the same tape.

Open weights versus American infrastructure. AlphaSignal walked through the engineering behind Moonshot's Kimi K3: 2.8T parameters, a 5TB weights file, and KV cache and LatentMoE tricks to keep the model from crushing the hardware. Linas Beliūnas put Kimi K3 at a third of Claude's price for frontend coding and agents, and framed it as Stripe's OpenRouter math getting better and its Anthropic problem getting worse. Alex Banks at The Signal covered DeepSeek's flash sale V4 release, plus Gemini finding its feet and a fresh music copyright hit. Luke Sophinos at Linear called open source "a build lever, not a religion," which is the right frame: use open weights and open cores where they win, quietly route the hard parts through paid closed models. Nate at Executive Briefing put the pressure on builders directly with his 5 Levels of AI Builder framework, arguing the minimum level for building is rising as labs absorb the point solution layer. The through line: the marginal open weight model is now Chinese, the marginal closed model bill is American, and both are compressing margins on the same trade.

AI Infrastructure: The Domestic Bill Comes Due

The Information's Ann Davis Vaughan broke that four states have already rolled back data center sales tax breaks and nine more are considering it, potentially adding 7% or more to every gigawatt of equipment cost. Governors from both parties who competed to land Microsoft, Google, Meta, Amazon, and Oracle now say the abatements are draining revenue without delivering the promised jobs; the next round may require guaranteed hiring or public infrastructure contributions to keep the exemption. The Daily Upside's "Prepping for RAMageddon" deep dive tied the memory squeeze back to rare earth control and Chinese DRAM breakout. MIT Technology Review led its Week in Review with a fundamental LLM flaw that leaves models strikingly vulnerable to attack, alongside OpenAI calling the Hugging Face attack unprecedented and Samsung chip workers jumping ship to SK Hynix. Techmeme surfaced OpenAI's internal version of Astra claiming results on 10 open problems in math and quantum complexity, with Gary Marcus noting the 249 page paper says nothing about how the model works or how the proofs were verified. The commonality: the AI build out's hidden costs, tax, security, supply chain, and scientific verifiability are all catching up with the announcement cycle at once.

Prediction Markets: Big Enough to Sue Now

Simon Taylor at Fintech Brainfood led with the number that anchored much of Sunday's fintech chat: Robinhood now makes more revenue from predictions than from stocks or crypto. The Information's Yueqi Yang put it in numbers: prediction markets revenue at Robinhood surged more than tenfold year over year in Q2 to $156M, or 20% of trading revenue, eclipsing both stock and crypto for the first time and becoming the second largest trading business after options. Meanwhile 20 states are suing prediction markets, with New York the latest to file. Bruce Mehlman's Six Chart Sunday flagged the WSJ's "Why Almost Everyone Loses, Except a Few Sharks, on Prediction Markets" as the second order read. The category crossed a line this year: no longer novelty, now large enough that regulators are being forced to decide whether prediction markets get built as gambling or as derivatives. The next twelve months will settle that.

Fintech: Sponsor Bank Cleanup, African Credit Signal

Jason Mikula at Fintech Business Weekly covered a fresh consent order for Lineage, a one time Synapse partner bank, extending a years long drama over BaaS sponsor oversight and forcing new leadership and a sale of a controlling stake. Tearsheet's weekly framed the Digital Asset Market CLARITY Act as narrower than the industry expected: stablecoins as settlement infrastructure, not interest bearing savings substitutes, with MiCA now operating in Europe and the UK FCA folding crypto into existing rules rather than building a parallel regime. Sam Boboev at Fintech Wrap Up shipped a 22 platform directory of EU and UK spend management companies segmented by license type, card network, and coverage. And Frontier Fintech's Samora Kariuki ran a guest post from Okestreta's Ray Besiga on customer value management for African digital financial services, opening on a boda boda rider in Kampala whose partial May payment does not trip a system that stays blind until his June miss. Every fintech story in yesterday's inbox was about the rules getting drawn tighter, not the product getting broader.

Politics and Democracy: Losing Battles, Real Fights

JVL at The Bulwark argued for fighting losing battles anyway, running the Triad Mailbag through DSA candidates, the 2030 census, and Ross Douthat joining '60 Minutes'. Marc Elias ran two pieces at Democracy Docket: one on Paul Weiss folding under political pressure while Democracy Docket refuses to, another on Todd Blanche and the fallacy of grading the lesser evil on a curve. Lauren Egan at The Bulwark reported the sharpest political shift of the week: Democrats are villainizing data centers as an issue, with Sherrod Brown running two ads in a month attacking Ohio's Jon Husted for courting them. Noah Smith called out state level self checkout restrictions as "slopulism," well meaning bills that would incentivize the exact automation they claim to slow. George Bounacos at Gov Brief Today noted California became the first state to require AI content provenance and Trump cancelled a planned weekend Iran bombing after the Saudi crown prince urged restraint. Paul Krugman opened a geoeconomics series ranking economic superpowers now that growth abroad is no longer the point. SpyTalk's Jeff Stein covered Bill Pulte resigning himself to vacating acting DNI. The through line: Democrats found their midterm villain, and it is a place, not a person.

Marketing and Media: LinkedIn's Slop Button and EU Labels

Jaskaran at The Social Juice rounded up the EU's compulsory AI labels, LinkedIn's new call out AI slop button, and X's WFA lawsuit settlement with a bar on reviving GARM or anything like it. Daniel Murray at The Marketing Millennials ran six AI marketing questions for legal teams as the EU AI Act's high risk obligations take full effect this month. Marketing Letter's weekend roundup: Google keeps replacing clicks with AI, OpenAI is quietly building an ad business, Meta is giving everyone an AI best friend, Reddit found a new way to monetize the robots, and LinkedIn declared war on slop. Nik Sharma pushed back on paid landing page proliferation: an established brand's 20 to 30 pages barely beat the PDP, because a new URL is not a new test and page count is not learning. Steve Bryant, writing from Mexico City, named the aesthetic that will not die: the "CD Vacation Wear" of tourists who look like they were dressed by an algorithm. Lucas Shaw at Bloomberg reported Netflix chasing Hot Ones as YouTube keeps expanding into TV, and Spider Man narrowly missing a records list. Influence Weekly asked why the new Larry Kings, Steven Bartlett most of all, bother old media, when the same soft interview technique that earned King a Peabody now earns Bartlett a BBC investigation. The tell across all of it: the platform layer is quietly rebuilding itself around agentic search and AI authored content, and regulators are the first to show up.

Macro: Yen Watch, Aussie Housing

Bloomberg's Asia Morning Briefing had yen traders on high alert for more US-Japan intervention after Friday's 1% jump. Bloomberg Australia reported the housing slump deepening across Australia and New Zealand, with Aussie home prices falling by the most in years. Tobin Harshaw at Bloomberg Opinion argued Mamdani's early pragmatism is doing more work than the socialism label. Bloomberg CityLab covered United's Dulles proposal winning White House approval. Dexter Roberts at Trade War returned from three weeks in Beijing and Taipei with Taiwan's H1 GDP growing 13.72%, the fastest since 1976, and Taiwanese exports to the US surpassing China's for the first five months. R.C. Whalen ran David Kotok on gold and US CDS in euros, arguing every institutional portfolio now needs a permanent gold allocation as a US default hedge.

Ideas Worth Reading

Outside Interests

Data Worth Noting

Three Takeaways for You

The AI trade is repricing in real time, and the newsletters caught it before the tape did. When Simon Taylor at Fintech Brainfood, Chamath, Rich Turrin, and Matt Stoller all lead with the same story on the same Sunday, complete with the same $2.18T in Seoul number and the same Aschenbrenner blowup, that is not four opinions, that is a consensus getting priced in. The pressure is not just on hedge funds; it is on the entire AI CapEx model that assumed compute demand was permanent, one directional, and worth 4x leverage.

The domestic AI stack is being taxed and villainized at the exact moment the marginal open weight model became Chinese and cheaper. Ann Davis Vaughan at The Information on state tax break rollbacks, Lauren Egan at The Bulwark on Democrats making data centers the midterm bogeyman, and Linas Beliūnas on Kimi K3 at a third of Claude's price all describe the same squeeze from different sides. If you are a builder, Luke Sophinos at Linear has the right frame: open source is not a movement, it is a build lever. Use it where it wins.

If you only read three pieces: Chamath on Korea's memory trade for the mechanics of how AI froth actually cracks, Lauren Egan on Democrats and data centers for the political counter current, and Abby Falik's "Ode to the Immeasurable" to remember that not everything worth building can be counted.