April 14, 2026
Matt Levine (Bloomberg) · rss · 17 mins
The popular belief that sportsbooks balance bets on both sides is mostly wrong. Sportsbooks — and by analogy, electronic proprietary trading firms like Citadel Securities, Hudson River Trading, Jane Street, and Virtu — routinely take directional positions when their models indicate positive expected value. The general public systematically bets irrationally (home teams, underdogs), so fading those biases is structurally profitable; staying perfectly “flat” would be leaving money on the table.
Stock market makers have a major structural advantage over sportsbooks: correlated instruments allow hedging across positions. A firm long $100 of Stock A and short $100 of correlated Stock B carries far less net risk than either position alone; options books hedge with the underlying. Sportsbooks lack this tool — the result of the Mets game tells you nothing about the Yankees game, so correlation models have very limited applicability (mostly same-game parlays).
Prediction markets sit in a regulatory gray zone that creates a market-structure identity crisis: Kalshi is a CFTC-regulated commodity futures exchange, yet 79% of Kalshi’s volume is sports bets. That means the natural liquidity providers might be either traditional prop trading firms (good at regulated financial instruments) or sports market-making specialists (good at sports bets) — but the skill sets diverge sharply in ways that matter for risk management.
An academic paper by Nick Palumbo (former DraftKings product manager), “A Microstructure Perspective on Prediction Markets,” found that passive liquidity providers in NFL prediction markets don’t primarily profit from bid/ask spread capture. Instead, they systematically absorb residual demand imbalances and retain outcome-dependent exposure through contract resolution — structurally resembling underwriting rather than inventory-neutral intermediation.
Palumbo’s NFL-season data: liquidity providers accumulated net exposure in the direction of realized outcomes in the majority of markets, generated approximately $29 million in aggregate profit, but with significant weekly volatility including pronounced drawdown weeks. Profitability depended on being directionally correct, not on staying flat — confirming that prediction market making is fundamentally a forecasting business.
Prediction market resolution — deciding whether an event “happened” for payout purposes — is the other structural problem. Kalshi’s approach is a regulated-exchange committee model: lawyers and contract writers iteratively refine resolution rules after each ambiguous case (Cardi B at the Super Bowl, US ground troops in Iran, Iranian missiles in Israel), incrementally building precedent the way credit-default swap documentation evolved.
Polymarket uses a fully decentralized crypto arbitration mechanism called UMA, where holders of UMA tokens debate on Discord and vote on outcomes. Voting power is proportional to token holdings, the largest UMA holders are anonymous, and a Substack analysis argued you could rig a disputed election market’s resolution for $10 million or less by accumulating UMA tokens — or manipulate Kalshi by bribing a committee member.
The resolution ambiguity problem may be structural rather than a maturity issue. Traditional financial contracts pay $X if a measurable price exceeds $Z — a narrow, objectively verifiable condition. Prediction market contracts pay $1 if any real-world event occurs, and the universe of possible events is unbounded. New unimagined ambiguities will continually emerge (alien landings, novel geopolitical situations), keeping resolution permanently contested regardless of how much precedent accumulates.
Anthropic’s new AI model, “Mythos,” is capable of identifying and exploiting vulnerabilities in every major operating system and web browser. Treasury Secretary Scott Bessent and Fed Chair Jerome Powell summoned Wall Street leaders for an urgent briefing on resulting cyber risk. Anthropic is limiting Mythos to a handful of firms — Amazon, Apple, and JPMorgan — under “Project Glasswing,” a program to harden critical systems before similar models proliferate more broadly.
Neobrokers built on young retail traders (Robinhood, eToro, Revolut, Public.com) are pivoting upmarket with premium services: $695 metal credit cards, airport lounges, Formula One race access, concierge services for million-dollar balances, and trust account management. Revolut is hiring multi-lingual private bankers; Public’s COO Stephen Sikes says AI-powered tools make even high-net-worth customers comfortable self-managing tens of millions — but they still want a human to pick up the phone.
Quotable:
“It is not looking to take no risk; it’s looking to take the right risk.” — On how sportsbooks and trading firms actually manage their books
Ben Thompson · rss · 13 mins
AI chips do not create marginal costs — they create opportunity costs. Unlike widget factories where raw materials scale with output, AI chips are fixed-cost infrastructure run at full capacity regardless. The real constraint is not “how much does one more inference cost?” but “what workload should this GPU serve?” This distinction refutes Doug O’Laughlin’s January 2025 claim that reasoning models killed Aggregation Theory by reintroducing marginal costs.
Microsoft’s Q2 2026 earnings are the clearest proof of opportunity cost in action. CFO Amy Hood disclosed that Azure grew below the Street’s expectations not from lack of demand, but because Microsoft deliberately allocated new GPUs to first-party products (M365 Copilot, GitHub Copilot) and R&D first. She said that if those same GPUs had gone entirely to Azure, the growth KPI “would have been over 40” — the company chose higher-margin, higher-lifetime-value internal workloads over external Azure revenue.
Anthropic’s Mythos model was announced but not broadly released, for two compounding opportunity-cost reasons. First, Anthropic is already compute-constrained serving existing models — a GitHub issue thread this weekend documented widespread user complaints about Claude being “dumbed down” over the past month. Second, broad release would accelerate distillation: DeepSeek, Moonshot, and MiniMax ran over 16 million exchanges via ~24,000 fraudulent accounts to illicitly extract Claude’s capabilities, and a more capable public Mythos would turbocharge that exploit.
Stopping distillation is strategically more valuable than it first appears: it’s not just about protecting margins, it’s about starving potential customers of a compute-free alternative. Open-source models distilled from frontier labs allow enterprises to self-host without paying frontier-lab prices or consuming frontier-lab capacity. Every successful distillation attack reduces the frontier lab’s addressable market and bargaining power for compute procurement, making the compute shortage worse for those labs specifically.
Meta’s Muse Spark debut puts the company in the best structural position in the industry for the consumer AI market. Every other major player — Microsoft, Amazon, Google, OpenAI, Anthropic — faces an opportunity cost when serving consumers because that compute could instead serve higher-paying enterprise workloads. Meta has no cloud business and no enterprise customers, so consumer serving carries zero opportunity cost. Meta also has a mature advertising business to monetize consumer usage directly, meaning they don’t need per-seat subscription revenue to justify the compute spend.
Meta should open-source Muse for the same reason they open-sourced Llama: the entities most damaged by a freely available frontier model are rival frontier labs, whose pricing power collapses and whose compute advantage narrows. Open-sourcing Muse increases competitive pressure on OpenAI and Anthropic in enterprise, making it harder for them to bear the opportunity cost of competing in consumer — clearing the field further for Meta.
OpenAI is betting compute supply beats product momentum, but the bet is contestable. OpenAI told investors in April 2026 that its infrastructure buildout gives it a decisive advantage over Anthropic, which is “rapidly gaining ground.” Thompson’s counterargument: distribution and transaction costs for AI are still effectively zero — the two preconditions for Aggregation Theory — so the winner should be whoever builds the most compelling product and attracts the most demand. Anthropic’s TPU deal with Google is an early example of demand-driven compute acquisition: enough revenue to take supply away from Google itself.
Quotable:
“My bet is that owning demand will ultimately trump owning supply, suggesting that the underlying principles of Aggregation Theory lives on.” — on whether compute constraints permanently upend internet-era platform economics
Byrne @ The Diff · email · 13 mins
The textbook model — government funds public goods because individuals free-ride — obscures what the US government actually does: mostly healthcare smoothing and retirement annuities, i.e., acting like UnitedHealth and MetLife. The more interesting modern complication is a class of extremely horizontal companies through which vast amounts of GDP flows without being fully captured, and which have begun performing government-like regulatory functions.
E-commerce concentration paradoxically improves consumer protection. A fragmented market creates a race to lax standards — a hypothetical “iBay” that allows fraudsters advertises lower prices than eBay, creating a market for lemons. Amazon Basics prices a $100 coffee pot at $20, making it economically infeasible to profit from selling counterfeits, since sourcing fakes costs more than selling the real thing.
Anthropic restricted its most powerful vulnerability-detection model to defensive users first — capturing the private upside (subscription revenue, reputational safety) while managing the social externality of handing hackers better tools. Amazon’s Ring cameras address porch piracy (disproportionately affecting Amazon-branded packages) while also tracing fugitives — not a substitute for law enforcement, but a complement that expands effective policing capacity from a fixed number of officers.
Cloudflare is the clearest case of private quasi-government: when they terminated service to 8chan in 2019, the site took several months to come back online. They’re now allowing creators to set prices on AI scrapers’ access to their content — writing intellectual property regulations themselves. The original internet was designed assuming every user was a grad student or a government employee; Cloudflare’s entire business is rebuilding it for billions of adversarial users.
All three companies face the same constitutional-style tradeoffs that legislatures and supreme courts normally resolve: where to draw the line between self-defense and enabling violence (Ring), between open and harmful speech (Cloudflare), between AI capability and AI safety (Anthropic). The opt-in private legal system they’ve created is a complement to, and increasingly in competition with, the public one — pointing toward an eventual comparison with public utilities.
AI has made it easier to fake 95% of being a good job candidate long enough to get an offer, so companies shifted to paid week-long work trials. This counterintuitively increased hiring transaction costs rather than reducing them, and introduces adverse selection: you’re drawing from unemployed candidates or people who can burn vacation days. Confident pre-AI predictions about AI reducing hiring friction were wrong in direction.
Meta committed $21bn to CoreWeave’s GPUs; OpenAI signed its first multi-year CoreWeave deal the same week. CoreWeave’s model works because frontier AI competition guarantees at least one major lab will always be relatively short on compute. Meta and Google are structurally more stable bidders because their AI infrastructure also supports ad-revenue-generating workloads; pure-play labs’ demand swings more violently.
Quotable:
“Their entire business is basically a fully-privatized, extremely technical chunk of the legal system devoted to enforcing de facto laws about how the Internet can be used.” — on Cloudflare’s de facto regulatory role
Andrew Ross Sorkin · email · 8 mins
Washington’s threat to blockade traffic from Iranian ports through the Strait of Hormuz — announced by U.S. Central Command taking effect at 10 a.m. Eastern — has pushed Brent crude above $102/barrel and West Texas Intermediate above $104, with U.S. gas averaging $4.125/gallon. The blockade is narrower than Trump’s originally threatened “complete blockade” but is a reversal from last month’s policy of allowing Iranian tankers through. JPMorgan analysts note the last tanker to clear the strait on Feb. 28 will exhaust pre-closure barrels from the global supply chain next week.
Iran’s threatened retaliation escalates the risk well beyond Iranian oil. A military spokesman said “no port in the Persian Gulf and the Sea of Oman will be safe,” while Parliament speaker Mohammad Bagher Ghalibaf warned Americans would “be nostalgic for $4–$5 gas.” Miad Maleki of the Foundation for Defense of Democracies estimates the blockade could cost Iran $13 billion a month, but the economic blowback to the U.S. is real: Trump himself acknowledged gas prices may be “a little bit higher” through the November midterms, a timing that analysts say will hurt Republicans.
Corporate earnings season opens against this backdrop with Goldman Sachs posting blowout Q1 results driven by trading and investment banking, but forward guidance is the real anxiety. Delta Air Lines pulled its full-year forecast entirely, with CEO Ed Bastian warning “our economy is going to continue to lose steam.” Citigroup’s U.S. equities strategist Scott Chronert told clients to brace for profit warnings if WTI stays above $90/barrel for six to eight weeks. Futures markets now price in zero Fed rate cuts before September 2027 as inflation climbs again.
AI is simultaneously powering cyberattacks and creating a boom for cybersecurity startups. Over 50% of cybersecurity companies funded last year were AI-native, per PitchBook, with VCs pouring $16.5 billion into the sector in 2025 — up 27% year-on-year. Large tech is following: Palo Alto Networks acquired CyberArk for ~$25 billion in February, and Alphabet closed its largest-ever acquisition with the $32 billion Wiz deal in March. The trigger: Anthropic’s latest coding model was deemed so capable at finding vulnerabilities that the company restricted its release over cybersecurity fears.
Quotable:
“Everyone’s talking about oil, but I think what the world is mainly short of is tokens.” — Ben Pouladian, engineer and tech investor, on AI compute scarcity as the deeper bottleneck
The New York Times · email · 8 mins
Trump called January’s midnight raid that captured Venezuelan president Nicolás Maduro “the perfect scenario” and repeatedly invokes it as the template for the ongoing war in Iran. Three months later the basic stability test has passed: no notable violence, no protracted U.S. ground occupation, and Maduro’s successor Delcy Rodríguez has locked in control of the government and armed forces. Venezuela has received no U.S. loans or significant aid.
Venezuela’s resource economy has visibly shifted. Chinese, Iranian, and Russian contractors are packing up; Russian and Chinese state oil firms nominally remain but the Venezuelan government now runs their fields. The free-oil pipeline to Cuba has stopped. Several hundred political prisoners have been freed, some opposition politicians have emerged from hiding, and Venezuela now sells its oil and minerals at global market prices rather than through opaque intermediary schemes — raising government revenue.
The central U.S. goal — American companies investing directly in Venezuelan oil — has not been achieved. Despite Rodríguez passing new laws with attractive investment terms, U.S. firms haven’t committed serious capital. Venezuela has burned investors before, and comprehensive sanctions remain in place. The real early winners are international commodity traders Vitol and Trafigura, both headquartered in tax havens, who buy Venezuelan oil and resell it to refiners. According to shipping analytics firm Kpler, India — not the U.S. — was the largest buyer of Venezuelan oil last month.
Venezuelan government officials privately undercut the invasion’s strategic rationale: Maduro himself had repeatedly offered U.S. companies open access to Venezuela’s natural resources and had agreed to expel Chinese, Russian, and Iranian competitors before he was captured. The economic opening now underway was available without the invasion. The operation produced a brazen display of American power, but not necessarily unique economic gains.
Quotable:
“Had he stayed, the moneymaking opportunities might well have been the same. But there wouldn’t have been the same brazen display of American power.” — Anatoly Kurmanaev, on whether the invasion of Venezuela was economically necessary
The Core · email · 6 mins
India’s economic history is a series of sector-led growth cycles, and the automotive/EV transition is the current one. Overall car retail sales grew 13% to 4.7 million units in FY26, but EV sales surged 84% to roughly 200,000 units — rising from 8% to 20% of all incremental vehicle sales in a single year. Domestic champions Mahindra & Mahindra and Tata Motors jointly hold 61% of the electric car market, meaning the coming energy transition is being captured by Indian firms, not foreign ones.
The biggest EV prize isn’t passenger cars — it’s commercial transport, which consumes 70% of India’s total transport fuel. AdvantEdge founder Kunal Khattar puts the electrification opportunity for commercial fleets (ride-hailing taxis, gig-economy two-wheelers, heavy freight) at close to $1 trillion. Delhi’s draft EV policy for 2026–2030 adds regulatory tailwind: 100% road tax and registration fee exemption for EVs priced at or below Rs 30 lakh, and a mandate that from January 2027, only electric three-wheelers will be permitted for new registrations in the capital.
The West Asia energy shock is a real drag on India — North Sea crude traded above $140/barrel, FPIs pulled Rs 48,213 crore from Indian equities in just 10 days in April, and total FPI outflows in 2026 have reached Rs 1.8 lakh crore — but New Delhi has no seat at the Hormuz negotiating table and therefore no geopolitical lever to pull. India’s only durable response is to accelerate domestic industries: since liberalisation in the 1990s, every maturing sector (telecom services → smartphones → insurance → fintech → e-commerce) has been replaced by a new growth engine, and the pattern gives grounds for confidence that EVs and AI-era services will do the same.
Quotable:
“True geopolitical leverage is rarely won by clamoring for a seat at foreign negotiating tables. It is built at home, through relentless economic focus.” — Govindraj Ethiraj, on India’s response to the US-Iran conflict and Hormuz crisis
John Authers · email · 8 mins
US-Iran talks collapsed after 21 hours in Islamabad, and the US response makes things worse in the short run. Iran refused to reopen the Strait of Hormuz — its central leverage in the conflict — so the US announced it will itself blockade the Strait. The irony, captured via a Goon Show bit, is complete: it doesn’t matter whether Iran’s mines or US warships are blocking the Strait; Brent crude jumped 8% at the Asian open Monday because the oil still isn’t coming through either way.
The blockade has a coherent strategic logic but will almost certainly escalate before it resolves. Andrew Bishop of Signum Global Advisors frames the US plan as treating the Strait as two halves — Iran’s “toll-booth” lane and the mined lane — with the US closing the former while demining and providing insurance/naval escorts for the latter, until Iran concedes politically. The problem: Iran still has its own tankers moving through; cutting that off deals a serious blow to its economy and raises the odds of retaliation, not capitulation.
March US CPI breached the Fed’s 3% upper-range target, driven by gasoline, and consumer sentiment hit an all-time record low. Headline CPI climbed to its highest in over two years; core and trimmed-mean stayed just under 3%, and tariff pass-through to goods prices has been milder than feared. But the University of Michigan’s consumer sentiment index — tracked since 1978 — has never been this negative: respondents say they feel worse now than during the 1980 hostage crisis, the 1990 Kuwait invasion, or any point in the Covid pandemic, a data point that limits Washington’s political tolerance for a prolonged supply shock.
The Hormuz disruption has broken the US and UK housing recovery. US mortgage applications fell more than 10% through March as rates re-spiked on war-driven inflation fears; Capital Economics’ Thomas Ryan sees existing home sales remaining stuck near 4.1 million annualized — roughly flat year-on-year. In the UK, RICS’s homebuyer-appetite gauge plunged to its lowest since August 2023, with lenders withdrawing mortgage products as markets speculated the Bank of England might hike to counter the supply shock. The Fed — facing a leadership transition with a new chair yet to be confirmed — is now expected to hold rates, deferring any relief to housing affordability indefinitely.
Quotable:
“It doesn’t much matter whose mines or warships are blocking the Strait; the point is that the oil isn’t coming through and supply remains interrupted.” — on why the US blockade, however strategically rational, provides no short-term market relief
Robert Armstrong · email · 6 mins
The Fed should neither raise nor cut rates in response to the Strait of Hormuz closure and the resulting energy price spike. Friday’s 3.3% headline CPI — the hottest in nearly two years — is irrelevant to monetary policy because raising rates cannot open a strait, and only rates severe enough to cause a global recession would push oil prices back down. Conversely, cutting rates is equally unjustified: before the war started, markets were already pricing in two quarter-point hikes, unemployment is well under 5%, and core PCE inflation has been above 4% for three consecutive months (December through February).
High energy prices pose a growth risk, not just an inflation risk — and the cascade is already visible in the data. Strategas economist Don Rissmiller’s framework captures it: energy prices act as a regressive tax on middle- and lower-income consumers, crimping non-energy spending, squeezing corporate margins, and ultimately triggering lay-offs that crack the job market. Q4 real GDP was just downgraded to 0.5% annualised, and the University of Michigan consumer sentiment index hit an all-time low — making the bear case harder to dismiss even though March consumer spending data still looked fine.
Any rate cuts that materialise will be the bad kind. If cuts come, they will reflect a spiking unemployment rate — which stocks hate — not falling inflation, which stocks like. The upshot: rooting for rate cuts right now is rooting for a recession. Second-quarter corporate earnings, whose reporting begins this week, will be the next key test of whether the bright spot in the economy — excellent profits — can hold.
Bonus (portfolio section, by Katie Martin): AQR’s Cliff Asness, Daniel Villalon, and Antti Ilmanen argue that the three most-cited bond replacements — crypto, private credit, and buffer funds — all carry far more equity risk than the bonds they’re meant to replace. Bitcoin has an equity beta of 2, meaning moving a dollar from bonds to Bitcoin delivers 10× the equity risk per dollar. Private credit remains too correlated with stocks; buffer funds are explicitly tied to equity market performance. AQR’s verdict: investors who think bonds are broken should apply a very high diversification bar to any replacement — and these three fail it.
Quotable:
“No one should be rooting for rate cuts at this point, because if we get them, they will probably be the unemployment-is-spiking kind which stocks hate, not the inflation-is-falling kind stocks like.” — Armstrong, on the policy trap created by war-driven energy inflation
Bloomberg Businessweek · email · 6 mins
Viktor Orbán, Hungary’s prime minister for 16 years and icon of the global illiberal right, was handed a resounding electoral defeat on April 13, 2026 by 45-year-old Peter Magyar and his Tisza party, which won more than two-thirds of Parliament’s seats — a supermajority. Orbán had entrenched himself through gerrymandering, court-packing, and near-total control of mass media, yet none of it was enough to overcome voter anger over stagnant growth, a crumbling health-care system, and Hungary’s deepening estrangement from the EU.
Magyar is a former member of Hungary’s ruling elite (as recently as 2024), not a liberal, and deliberately avoided cultural issues — his pitch was anti-corruption, EU recommitment, and ending Orbán’s alignment with Putin. Orbán’s final campaign pushed wall-to-wall messaging casting Volodymyr Zelenskiy as a warmonger who would drag Hungary into conflict; voters ignored it entirely, focused instead on decaying public services and the fact that Hungarian students were no longer welcome on exchanges across Europe.
Magyar’s supermajority gives him the power to rewrite the constitution and purge Orbán’s loyalists from state chokepoints, but the economy is deeply entangled with Orbán’s crony network and has stagnated for three years — unwinding that without tanking growth is the core challenge. An early, easy move: ending Orbán’s blockade of a €90 billion ($105 billion) EU loan to Kyiv, which Magyar pledged immediately, and which should in turn unlock EU structural funds frozen over corruption concerns. Orbán remains acting prime minister for at least a month while a Tisza government forms.
Quotable:
“Hungarians spent years under ‘Goebbelsian, North Korean-style propaganda’ — the kind of propaganda that didn’t contain an ounce of truth, that tried to keep our parents and grandparents living in terror.” — Peter Magyar, speaking to reporters the day after his election victory
The Information AM · email · 5 mins
Meta is poaching three senior OpenAI executives who helped build the Stargate data center initiative: Peter Hoeschele (who got Stargate off the ground), Shamez Hemani (compute strategy and biz dev), and Anuj Saharan (compute organization leader). They are joining Meta’s newly formed “Meta Compute” group to build capacity for TBD Lab, the internal AI unit run by former Scale AI CEO Alexandr Wang. Mark Zuckerberg has added the two heads of Meta Compute — Santosh Janardhan (global infrastructure) and Daniel Gross (strategy and supplier partnerships) — as direct reports, signaling how central the compute buildout is to Meta’s AI strategy.
Anthropic’s compute arms race is accelerating sharply: its annualized revenue jumped to $30 billion this week, up from $9 billion at end of 2025, and it signed a multi-year deal with CoreWeave (its first) to run Claude models, with servers coming online later this year. It also struck a deal with Broadcom and Google to bring more than 3.5 GW of server capacity online starting in 2027, against a longer-term target of 10 GW. OpenAI pushed back by telling investors its existing compute stockpile — worth billions of dollars — gives it a structural advantage over Anthropic.
Two adjacent signals from the newsletter: Tesla won approval to sell its “Full Self-Driving Supervised” software in the Netherlands — the first EU country — at $99/month with 1.1 million US subscribers already, though bloc-wide EU approval requires a majority vote from member states with no clear timeline. Separately, Hong Kong granted its first stablecoin issuer licenses to HSBC and a Standard Chartered joint venture (with Animoca Brands and Hong Kong Telecom), allowing HKD-pegged tokens effectively equivalent to dollar stablecoins, while mainland China continues to ban yuan-pegged stablecoin issuance.
Quotable:
“There is a need to de-escalate the rhetoric.” — Sam Altman, in a personal blog post after a man allegedly threw a Molotov cocktail at his San Francisco home at 3:45 a.m., an incident that caused no injuries