April 09, 2026
Ben Thompson · rss · 9 mins
Anthropic’s new model “Mythos” is a major benchmark leap, likely the first frontier model trained on Nvidia’s Grace Blackwell NVL72 architecture, and is priced at $25/$125 per million input/output tokens — 5x more expensive than Opus 4.6 ($5/$25) and roughly 10x GPT-5.4 ($2.5/$15). It is being previewed to ~50 critical infrastructure companies (Amazon, Microsoft, Apple, Google, Linux Foundation) under “Project Glasswing” rather than released publicly, with Anthropic citing its dangerous proficiency at finding and exploiting software vulnerabilities.
Skepticism about Anthropic’s safety rationale is warranted on multiple grounds: compute scarcity (insufficient supply to serve the public), distillation risk (DeepSeek, Moonshot AI, and MiniMax collectively created 24,000+ fraudulent accounts and sent 16M+ Claude queries to extract model knowledge), and a business preference to concentrate on enterprise customers under formal agreements who can now be conveniently rebranded as “vetted” users.
Anthropic’s pattern of withholding models on safety grounds dates to GPT-2 in 2019, when Dario Amodei co-authored a decision to not release a model trained on 40GB of internet text, citing “malicious applications.” The “Boy Who Cried Wolf” parallel is real — but so is the wolf: Mythos has already found thousands of high-severity vulnerabilities, including some in every major OS and web browser, many of which survived decades of human review and millions of automated tests.
The deeper danger is geopolitical, not just cybersecurity. A model capable of finding and exploiting vulnerabilities at scale implicitly possesses offensive cyber capabilities against any software — including U.S. government systems. That dual-use reality means Anthropic is accumulating a power base with national security implications, which Thompson previously argued forces a binary choice: Anthropic accepts U.S. government subordination, or the government acts to neutralize it.
Thompson’s bottom line: Mythos probably isn’t yet at the nuclear-weapons threshold, and a distilled public version will likely drop as soon as OpenAI ships its competing “Spud” model — Amodei’s capitalist instincts reliably override his safety rhetoric when OpenAI is the competitor. But Anthropic’s habit of positioning itself as civilization’s doom-arbiter is actively accelerating the alignment confrontation it claims to want to avoid.
Quotable:
“The punchline of ‘The Boy Who Cried Wolf’ is that eventually a wolf did come.” — on why Anthropic’s history of overcautious model releases doesn’t automatically discredit its Mythos safety claims
Matt Levine (Bloomberg Money Stuff) · rss · 14 mins
Stop-loss orders cluster at round numbers, creating predictable cascades. Retail investors set stop-loss triggers at psychologically salient prices — $50, not $51.37 — so if a stock falls to that threshold, automated sell orders pile in and push it further down. Sophisticated traders exploit this by “stop hunting”: deliberately shorting a stock to breach the cluster level, triggering the cascade, then buying back cheaper.
Institutional investor behavior is more exploitable than retail because it’s rule-based. Index funds must mechanically sell deleted names and buy added ones on rebalancing. Investment-grade bond funds must dump any bond downgraded from BBB- to BB+ (“fallen angels”), sending prices below what fundamentals justify. These mandates turn institutional psychology into near-deterministic rules that hedge funds routinely front-run. Retail investors, by contrast, act on vibes — harder to predict, impossible to front-run systematically.
Public brokerage is automating retail investing tactics with AI agents, bridging that gap. The privately held brokerage is rolling out a feature letting users type a strategy, refine parameters, and deploy an agent to trade on their behalf — buying protective puts if oil spikes, sweeping cash into higher-yielding bonds, or adding 20% stop-losses automatically. Co-CEO Jannick Malling confirmed sample prompts include “Whenever SPY drops 3% or more in a single trading day, invest $10,000 at the next market open.”
If “buy the dip” becomes a common automated retail behavior, it becomes front-runnable. Most retail investors won’t use these agents, and those who do will set idiosyncratic rules that cancel out — but the modal retail behavior (buying dips) is already stereotypical enough that if it gets automated en masse, hedge funds can anticipate billions in agentic buy orders after any 2%+ down day and front-run it at the close.
Anthony Pompliano’s ProCap Financial illustrates a second AI coordination risk: homogenized research. His company built an AI system in two weeks for “a couple thousand dollars” that generates hundreds of investment research reports per day, then sells subscriptions for $2,500/year. If many retail investors read the same AI-drafted thesis (“3 Stocks That Win From Both Tariff Refunds and the Iran Oil Shock”), they funnel toward the same trades — replicating index-like herding without being an index.
NYT journalist John Carreyrou claims Bitcoin inventor Satoshi Nakamoto is Adam Back — and the disclosure might be securities fraud. Back is CEO of Bitcoin Standard Treasury Company, planning a SPAC listing. Under U.S. securities law, material information must be disclosed to investors; a secret holding of 1.1 million Bitcoin (which could crash the market if sold) almost certainly qualifies. Carreyrou’s evidence includes listserv word choices and body language — Back “tensed up,” showed “shifty eyes” and “an awkward chuckle” — which Back denies.
The dispersion trade suffered its worst month since 2011 in March: -4.9% on JPMorgan’s index. The trade buys single-stock options and sells index options, profiting when stocks move differently from each other. February was a banner month — AI was creating huge winners and losers, so dispersion was high. Then the Iran conflict hit: macro shocks (oil, rates, commodities) moved all stocks together, correlation spiked, and the divergence the trade harvests collapsed. Citigroup’s Luca Valitutti: “The recent events had macro consequences … rather than micro impacts at the single-stock level.”
Physical inventory counting remains one of the last jobs AI hasn’t automated, and auditors hate it. Big-4 firms send junior auditors to count chickens, quarry rocks, traffic lights, and nuts-and-bolts in freezing warehouses and snake-infested pits. Blockchain was supposed to fix provenance fraud; it can’t bridge the physical-to-digital gap. AI might eventually send humanoid robots to count 162,000 bolts in seconds, but for now it’s a rite of passage documented extensively in TikTok horror stories.
Quotable:
“So they spent a couple thousand dollars to build a system that automatically generates reports with no effort, and they’re charging $2,500 for those reports?” — on Pompliano’s ProCap Insights AI research service
The Information AM · email · 7 mins
Eric Boyd — an 18-year Microsoft veteran who led Azure’s engineering of hardware and software for hosting OpenAI and Anthropic models, and who helped plan Microsoft’s first massive GPU clusters for training GPT-class models — is joining Anthropic as its head of infrastructure. His hire signals that Anthropic is now building serious internal infrastructure capacity rather than relying entirely on cloud providers: the company has been discussing leasing its own data centers and has already been recruiting former Google Cloud executives for this team.
The urgency behind Boyd’s hire is acute: Anthropic announced Monday that its annualized revenue has surged from $9 billion (end of 2025) to over $30 billion — but simultaneously disclosed a computing crunch so severe it began charging users more to connect Claude Code to high-usage services like OpenClaw. Revenue growing 3x in months while compute is a binding constraint is the pressure Boyd is walking into.
Separately, Anthropic announced “Project Glasswing,” giving 40+ organizations — including Apple, JPMorgan Chase, and the Linux Foundation — early access to its unreleased Claude Mythos model to test software for security vulnerabilities. Using Mythos internally, Anthropic claims to have identified “thousands of zero-day vulnerabilities” across every major OS and web browser, including Linux and OpenBSD. The model won’t be released publicly; after $100M in donated usage credits and $4M in cash to open-source security groups run out, partners pay $25/million input tokens and $125/million output tokens — five times more expensive than Anthropic’s current most advanced model.
AWS CEO Matt Garman, speaking at HumanX in San Francisco, pushed back on the narrative that Claude Code will cannibalize enterprise SaaS: he called the idea of using it to replace CRM software like Salesforce “overblown,” arguing incumbents have domain knowledge, existing customer bases, and proprietary data that puts them in a better position to build next-gen AI products. His caveat — that firms trying to “protect what they have and not lean in” to AI “are in trouble” — is notable given that Amazon is both an Anthropic shareholder and a cloud provider selling those same incumbent SaaS vendors’ products.
Quotable:
“They know more about their area, the edges of their software…and so they are in a better position to build the next generation of AI enabled products.” — AWS CEO Matt Garman at HumanX, defending SaaS incumbents against Claude Code disruption fears
Natalia Quintero and Mike Taylor · rss · 3 mins
The “wait for AI to mature” window has closed. When Anthropic released industry-specific Cowork plugins for legal and financial services in February 2026, the S&P 500 software index fell nearly 9% in days. Executives who held back are now making high-stakes AI decisions — procurement, headcount, strategy — about tools they’ve never touched themselves, which is a compounding liability.
Treating AI like conventional software — “evaluate, buy, integrate” — is the wrong mental model. Claude and similar tools don’t deliver value on day one like a SaaS subscription; they require explicit direction and output verification, the way a new employee does. The failure mode of the “evaluate and buy” approach is that you can’t assess an employee’s fit without actually working with them.
Using AI in practice feels like managing people, not operating software — and this is what surprises executives most when they first use it. Running Claude Cowork means coordinating 10 parallel threads (dashboard builds, inbox triage, document review), delegating clearly, and applying judgment the AI lacks: Did it pull the right data? Is the logic sound or just plausible-looking? The executives who pick this up fastest are those who already excel at delegation and output quality assessment — senior leaders, not junior employees.
Quotable:
“AI tools like Claude and Cowork aren’t products that slot into your tech stack and deliver value on day one. They’re more like a new kind of employee—one that can do enormous amounts of work, but only if you tell it exactly what to do and check whether the output is right.” — on why the standard software-procurement playbook fails for AI
Matt Levine, Bloomberg · rss · 13 mins
A federal court dismissed a securities fraud suit against Capri Holdings (Michael Kors, Versace) and Tapestry (Kate Spade, Coach) arising from their failed merger. After the FTC blocked the deal, shareholders sued claiming the companies’ confident public statements — “there is no question that this is a pro-competitive, pro-consumer deal” — were fraudulent. The judge ruled that being wrong about antitrust, even emphatically wrong, is not the same as fraud.
The Capri/Tapestry case sits in contrast to Elon Musk’s recent Twitter securities fraud loss. Musk agreed to buy Twitter, tweeted he was getting out, the stock fell, shareholders sued and won. The structural parallel is identical — public statements proved wrong, stock moved on them — but Musk lost where Capri won, arguably because Musk’s position was more clearly manufactured rather than a genuine (if incorrect) legal/regulatory opinion.
Bill Ackman’s Pershing Square proposed a €55 billion acquisition of Universal Music Group, but the headline number is constructed from three layered transactions: (1) a SPARC merger where existing UMG shareholders mostly receive stock in the reconstituted company, (2) an activist plan to borrow €5.4 billion, sell UMG’s €1.5 billion Spotify stake, and return €9.4 billion cash via buyback, and (3) Pershing Square investing €2.5 billion for ~7.4% of pro forma shares at ~€22/share — only a 29% premium to last week’s €17.10 close. The “78% premium” and “€55 billion” valuation fold in the assumed value of the activist plans themselves.
The SPARC structure (Special Purpose Acquisition Rights Company) inverts the SPAC model: instead of raising a blind pool first and then finding a target, Ackman identifies the target (UMG) and then solicits investors into the SPARC. Of the €2.5 billion equity check, €1.4 billion comes from existing Pershing Square funds; the remaining €1.1 billion must be raised from SPARC rights holders — on top of the $5–10 billion Ackman is separately raising for a new closed-end fund. Pershing Square has backstopped the full amount, so UMG’s board doesn’t bear that fundraising risk.
Goldman Sachs Private Credit Corp. reported Q1 redemption requests of exactly 4.9990% of shares — 17,281,858 shares out of a 5% tender offer for 17,285,147 — clearing the industry-wide 5% quarterly cap by 3,289 shares. Goldman framed this as evidence that its institutional investor base is more stable than the retail-heavy peers like Blue Owl (21.9% redemption requests) and Ares Strategic Income Fund (11.6%). GS Credit simultaneously reported $1.04 billion in gross new subscriptions, for net inflows of ~7.1% of NAV.
The “institutional investors are calmer” narrative cuts both ways. Other BDCs with high redemption requests noted that 90%+ of their shareholders by count did not redeem — the redemptions were concentrated among a small number of large family offices and institutions. That pattern suggests sophisticated capital is actually the one exiting, arbitraging the gap between private BDCs (redeemable at 100 cents on the dollar NAV) and publicly traded BDCs currently trading at 70–80 cents on the dollar NAV with similar underlying assets.
JPMorgan’s private credit trading desk, which ten months ago was sending monthly buy-side runs and receiving zero responses, has reversed direction: it is now circulating sell-side lists, trying to find buyers for software loans from private credit funds seeking to shed AI-disruption exposure — mostly at above 95 cents on the dollar. The desk has yet to find consistent two-way flow; the joke about JPM sending runs and getting no replies has simply flipped from the buy side to the sell side.
Quotable:
“As I’m in my room dying, I could hear them out there doing all their drills and yelling. So I’m in here thinking, This is terrible, but it sounds terrible out there, too.” — CEO of unnamed company, describing his E. coli hospitalization during a Navy SEAL–led Survivor-themed corporate retreat
Bloomberg · email · 7 mins
Retail investors are fleeing non-traded business development companies (BDCs) — private credit funds sold to everyday investors — at alarming rates. Barings Private Credit Corp. saw investors request withdrawals of 11.3% of shares in Q1 2026; Apollo Global Management faced similar pressure and, after capping redemptions at 5%, left departing investors with only 45 cents on the dollar. Goldman Sachs’ BDC narrowly avoided triggering its cap by meeting exactly 4.999% in redemptions. The trigger: a wave of high-profile corporate collapses in 2025 combined with fears that AI will destroy software companies — BDCs’ main borrowers — before their loans mature.
The redemption crisis is structurally self-reinforcing, per Moody’s Ratings, which revised its outlook on BDCs to negative from stable. Moody’s vice president Clay Montgomery warned that “proration of investors and headlines of elevated redemptions may incentivize other investors to seek redemptions” — a classic bank-run dynamic applied to illiquid private credit. The underlying fear (AI disrupting software firms) won’t dissipate in 2026, meaning the exit pressure has no near-term catalyst to reverse.
While retail flees, institutional players are aggressively positioning to buy cheap. Blackstone just hit the hard $10 billion fundraising cap on a new opportunistic credit fund designed to exploit market dislocations. Goldman Sachs told BDC shareholders that when capital becomes scarce, “spreads widen, structures tighten, and documentation improves” — positioning itself to scoop up assets being dumped by panicking retail funds. Morgan Stanley is simultaneously launching a new interval fund for retail private credit exposure, betting the current fear is a buying opportunity.
The AI valuations underpinning BDC lending fears are themselves built on shaky ground. ARR (annual recurring revenue) — the dominant metric for AI startup valuations — is unaudited, has no SEC definition, and is routinely inflated: subscriptions from trial customers who later cancel are counted as recurring. Stanford professor Chuck Eesley calls AI startup finance “a Wild West” with “no cop on the beat.” OpenAI’s $852 billion valuation (achieved in a $122 billion funding round) rests partly on this metric; its secondary-market shares are now almost impossible to sell as investor attention pivots to Anthropic, which recently revealed a $30 billion annualized revenue run rate.
Quotable:
“By and large, private credit does not tend to have great transparency or rigorous valuation ‘marks’ of their loans — this increases the chance that people will sell if they think the environment will get worse — even if actual realized losses barely change.” — Jamie Dimon, CEO of JPMorgan Chase, in his annual letter to shareholders
FT Opinion · email · 3 mins
Trump’s erratic threats and unpredictable policy mirror Nixon’s “Madman Theory” — the Cold War strategy of convincing adversaries you’re capable of anything to extract concessions. But the comparison breaks down: Vietnam in the early 1970s was peripheral to the global economy, while Iran is not. Trump threatening to “erase a civilization” risks widespread economic damage, making the tactic far more costly to deploy than Nixon’s Vietnam-era version ever was.
The Madman Theory contains a self-defeating paradox that renders it structurally weak: the threat must be extreme enough to be feared, but that same extremity makes it incredible. And if the threat is actually carried out — obliterating Iran, say — the strategy has by definition failed, since the goal was coercion, not destruction. A two-week Iran ceasefire (announced April 9) takes Armageddon off the table for now, but the Strait of Hormuz question remains unresolved, leaving the standoff in a limbo that neither Trump nor Iranian leadership can easily exit.
Quotable:
“The threat is too extreme to be entirely credible. On the other hand, if it is carried out, then by definition the strategy has failed.” — Janan Ganesh on the inherent contradiction of the Madman Theory
John Authers · email · 7 mins
A US-brokered ceasefire between Iran and Israel triggered a massive relief rally on April 9: world stocks rose 3.26%, emerging markets posted their best single day since November 2022 (when inflation peaked), and Brent crude fell 12.1%. Markets repriced the conflict as essentially over — stocks-vs-bonds ratios snapped back to their pre-war upward trend — even though the ceasefire terms appear to enshrine Iran’s right to charge tankers for passage through the Strait of Hormuz, something that would have been dismissed as unconscionable weeks earlier.
Rate-cut hopes were not restored alongside the equity celebration. Inflation expectations in the swaps market sit above 3% for the next year, up from below 2.25% at the start of 2026, and futures markets now price zero Fed cuts in 2025 — two had been fully priced before hostilities broke out. The rally is a “left-tail risk” reduction trade, not a growth-restoration trade; ongoing oil supply disruption and the murky Hormuz compromise keep costs elevated and profit margins under pressure.
Emerging market flows cratered during the conflict but in a notably differentiated way. Non-resident portfolio flows tracked by the IIF surged to a near-two-decade high earlier in 2026, then collapsed to the lowest since the pandemic; EM ex-China recorded net equity outflows of ~$60 billion in March alone. Crucially, Oxford Economics’ Joshua Fisher notes this was driven by energy dependency and Hormuz exposure rather than generalized risk-off — a “major departure from previous crises” — making the post-ceasefire EM rebound more durable if the waterway reopens fully.
Q1 earnings season, opening next week with the banks, begins from a surprisingly resilient base: FactSet forecasts ~13% year-on-year EPS growth, the sixth consecutive quarter of double-digit expansion, with energy (high oil prices) and tech (AI capex) as the two pillars. Alastair Pinder of HSBC warns that shipping backlogs, insurance constraints, and operational bottlenecks mean Hormuz traffic normalization could take at least three more months, capping how quickly energy-price tailwinds fade.
The Magnificent Seven — 32.3% of S&P 500 by market cap — have compressed to the same P/E multiple as Consumer Staples, a valuation level that preceded multiple-expansion both in the 2022 inflation shock and the Liberation Day selloff. Morgan Stanley’s Michael Wilson calls them “quite attractive” after six months of correction. But Nicholas Colas flags the asymmetry: AI-picks-and-shovels suppliers (Nvidia et al.) are expected to do very well, while the buyers of AI infrastructure are not — meaning blockbuster hyperscaler results will only lift broader markets if they deliver clear evidence that AI capex is converting into profits.
Quotable:
“Investors must assume that chaos, inefficiencies, and lack of transparency is not a bug but the feature of the system. Plan accordingly. Day-trade the news, but buy ongoing disruption.” — Viktor Shvets, Macquarie Group