Handy Robots (The Diff) — Byrne Hobart makes the case that a dexterous robotic hand is to the physical world what ChatGPT was to abstract information. The unlock for the world of atoms.
Will the Ayatollahs Chicken Out (Bloomberg Opinion) — TACO is dead, WACO is born. Iran is charging ships $2M/passage through Hormuz, Houthis just joined in, and markets are still underpricing how long this lasts.
AI Agenda: Anthropic’s Success Sparks Server Crunch (The Information) — Anthropic doubled ARR to $19B but is running out of servers, accidentally leaked that its next flagship model is “too expensive to release,” and its IPO timeline is now at risk.
Iran War, Oil & Geomacro
Will the Ayatollahs Chicken Out ⚡
John Authers · Bloomberg Opinion · 8 min
Why read this: TACO (Trump Always Chickens Out) made fortunes on Liberation Day—but Iran’s new supreme leader has nothing left to lose after a month of military pounding, and that changes every incentive calculus. Authers maps why “WACO” is the new trade framework.
Sharp takes:
Iran has discovered it can selectively tax global oil shipping without mining a single channel—charging $2M/passage turns the Strait of Hormuz into a toll road, not a blocked artery, which is far more durable as a pressure instrument.
The 1973 oil embargo parallel isn’t just vibes: Iran has recreated Sheik Yamani’s leverage over tanker operators by making the risk/reward calculation for any “unfriendly” cargo no longer worth it, without firing a shot at every ship.
The WACO framework is already breaking down equity assumptions built around a short, contained conflict—Houthis joining on Sunday was the moment “prolonged war risk” stopped being a tail scenario.
Quotable:
“The ayatollahs have very little left to lose after already suffering a military pounding. That reduces their incentive to quit.” — Authers on why Iran’s game theory is structurally different from Trump’s tariff retreats
Growth Jitters 📌
Bloomberg · 5 min
Why read this: The cleanest single-brief summary of how markets are processing the war’s second month: Houthis in, bonds rallying, Brent heading for its biggest-ever monthly gain, and Trump threatening to seize Kharg Island in the same breath as claiming “good negotiations.”
Sharp takes:
The bond rally is paradoxical—Treasuries are rising alongside crude, which signals markets fear a growth shock more than an inflation spiral, inverting the standard oil-war playbook.
Trump’s threat to “take the oil in Iran” and seize Kharg Island is being read less as strategy and more as a negotiating ceiling-raiser; but the more times you say it, the more it prices itself in.
Aluminum surged 6% after Iran attacked Middle East production sites—the war’s collateral damage is spreading fast beyond energy into industrial metals.
Quotable:
“Markets spent a month pricing a short, contained conflict. That wishful optimism has now broken with the Houthis’ entry over the weekend.” — Hebe Chen, Vantage Global Prime, on the inflection point
Unhedged: Short Rates Have Overshot 📌
Robert Armstrong · FT · 6 min
Why read this: Armstrong argues the market has overcorrected—erasing two expected Fed cuts and pricing a one-in-five chance of a hike this year—despite the textbook answer to an energy supply shock being: do nothing, the oil prices will dampen growth for you.
Sharp takes:
Central banking 101 says leave rates alone in a supply shock; the Fed can’t print oil, and tightening just “kicks the economy while it’s down”—but the pre-war inflation overshoot of ~1pp above target complicates the clean version of that rule.
Five-year forward inflation breakevens are actually falling, not rising—the market is signaling growth worry, not an inflationary spiral, which is the real-time refutation of the “hike” case.
The irony: official Fed hawkish posturing (designed to anchor expectations) may be causing the market to over-tighten via the two-year yield, achieving tightening without the Fed doing anything.
Quotable:
“Responding to this official posturing by taking out the two expected rate cuts would have been overkill — but for the fact that US inflation was about a percentage point above target, and stuck there, before the war began.” — Armstrong on why this supply shock is messier than the textbook
India Brief: Rupee Falls Despite Intervention 📌
Veena Venugopal · FT India Brief · 6 min
Why read this: The RBI’s emergency order forcing banks to cap FX exposure at $100M/day (down from billions) created a bizarre two-act Monday: the rupee surged 1.4% at open, then gave back almost everything to close at a fresh record low—textbook futility of fighting a structural shock with position limits.
Sharp takes:
Requiring banks to comply within 10 days forces a fire-sale of positions measured in billions—the cure may be worse than the disease if disorderly unwinding spooks the market further.
India’s vulnerability is structural: the war is hitting both oil import costs and the rupee simultaneously, a double bind that neither monetary nor fiscal policy can cleanly resolve.
Trump’s stated preference to “take the oil in Iran” is landing in Indian markets as a signal of indefinite disruption, not negotiated resolution.
Quotable:
“To be honest with you, my favourite thing is to take the oil in Iran but some stupid people back in the US say: ‘Why are you doing that?’ But they’re stupid people.” — Donald Trump to the FT, as quoted by Venugopal, on what Indian policymakers are actually managing against
Asian Countries Defend Struggling Currencies 📎
FirstFT Asia · 6 min
Why read this: Quick-read on the coordinated Asian FX stress—yen at ¥160.30 (Japan’s prior intervention threshold), rupee at a record, and the threat of direct yen-buying intervention “imminent” per Nomura’s chief FX strategist.
Sharp takes:
Japan’s top currency official Atsushi Mimura’s “decisive action” warning was read by Tokyo traders as a “final warning”—intervention could come any session this week.
The rupee’s RBI order forces banks to unwind positions worth billions in 10 days; the White House is simultaneously floating asking Gulf states to help fund US/Israeli war costs—a new category of macro uncertainty.
Yen and rupee are the two most exposed Asian currencies to the Hormuz closure; their simultaneous weakness signals this is a systematic oil-shock effect, not country-specific stories.
When Commodity Prices Become ‘Academic’ 📌
Tracy Alloway · Bloomberg Odd Lots · 4 min
Why read this: Middle East urea is now quoted at $800/ton (up from $300 a year ago) but those prices are essentially theoretical—no one can actually source the product. The physical/financial disconnect is a leading indicator of how commodity markets seize up before formal shortages register.
Sharp takes:
“The real issue on pricing,” per Green Markets, is that sellers and buyers can quote almost any price but “nothing is getting done”—a phantom market where the bid/ask spread is measured in availability, not dollars.
Urea hit $1,000/ton after Russia’s Ukraine invasion in 2022; the current move toward $800 suggests this shock is already approaching that severity for agricultural inputs.
The breakdown between paper prices and physical delivery is the mechanism through which commodity disruptions turn into food price crises—this is worth watching well before it registers in inflation data.
AI: Infrastructure, Business & Strategy
Handy Robots ⚡
Byrne Hobart · The Diff · 15 min
Why read this: The best framing I’ve read for why physical robotics is at its ChatGPT moment: a robotic hand that can turn a doorknob, dice an onion, and operate a yo-yo would unlock the world of atoms the same way LLMs unlocked the world of bits. Byrne traces this from 16th-century prosthetics to today’s dexterous robot race.
Sharp takes:
Wait, really? The human hand has coevolved with tools—fingernails (not claws) exist specifically to enable object manipulation, and the entire built world (blister packs, Lego, cellophane) is engineered around human hand geometry, meaning any robot hand gets this “universal adapter” for free if it can match the form factor.
The path to general robotic hands isn’t necessarily biomimicry—a hand that’s much slower but still capable may unlock vast swaths of non-speed-sensitive work (assembly, logistics, care) long before a humanoid robot can juggle.
The tension worth holding: Byrne notes that personal animosity (Altman vs. Amodei) has shaped the structure of the AI industry, and wonders if a more fragmented robotics landscape—driven by similar dynamics—might actually be better for society than a monopoly on physical-world automation.
Quotable:
“A robotic hand that can turn a doorknob, dice an onion, and operate a yo-yo is an unlock for the world of atoms in the same way that a program that can finish a dialogue or turn a natural-language description into a working computer program is for the world of bits.” — Byrne Hobart, on what the ChatGPT moment for robotics actually looks like
AI Agenda: Anthropic’s Success Sparks Server Crunch ⚡
Stephanie Palazzolo · The Information · 8 min
Why read this: Anthropic more than doubled ARR to $19B in just two months on the strength of coding tools—but that explosive growth is now creating the exact problems it was designed to avoid: capacity failures, an accidental leak that Claude Mythos is “too expensive to serve,” and gross margin pressure right before a potential fall IPO.
Sharp takes:
The leaked internal blog post is the most interesting detail: Claude Mythos is described as so compute-intensive it needs to be made “much more efficient before any general release”—meaning Anthropic’s most capable model is currently unusable at scale, a constraint its competitors won’t face if they’re willing to burn cash.
Dario Amodei’s “cone of uncertainty” thesis (secure too little compute = can’t serve demand; too much = bankruptcy) is now being tested in real time, and the answer appears to be: they under-secured, but deliberately.
The IPO timeline tension is sharp: Anthropic is trying to groom margins for a public market debut even as spot server costs eat into gross profits and demand keeps accelerating—a classic “good problem to have” that is genuinely hard to solve.
Quotable:
“Claude Mythos is ‘a large, compute-intensive model’ that is ‘very expensive…to serve, and will be very expensive for…customers to use.’ In fact, it is so expensive that the company said it’ll have to make it ‘much more efficient before any general release.’” — Stephanie Palazzolo summarizing Anthropic’s accidentally-public internal blog post
AI Infrastructure Roadmap: Five Frontiers for 2026 📌
Janelle Teng et al. · Bessemer Venture Partners · 10 min
Why read this: Bessemer’s investing thesis articulated as a roadmap: the first AI infrastructure wave (compute, training, data) is mature; what’s needed now are five categories of infrastructure that ground AI in operational reality rather than benchmark performance.
Sharp takes:
The five frontiers are: (1) harness infrastructure for memory/context management, (2) real-world grounding and evaluation, (3) agentic coordination systems, (4) continuous learning from deployment, (5) human-AI interface layers—all address the gap between “impressive demo” and “production system.”
The “organizational amnesia” framing for enterprise AI is sharp: basic RAG solved the retrieval problem but not the context continuity problem—enterprises keep rebuilding context from scratch, paying the inference tax every session.
The implicit investment thesis: companies that won the first wave (Anthropic, Cursor, VAPI) were picked when benchmarks mattered; the next wave winners will be picked on production reliability metrics no one is yet tracking well.
Quotable:
“The infrastructure that got us here — which was optimized for scale and efficiency — won’t get us to the next phase. What’s needed now is infrastructure for grounding AI in operational contexts, real-world experience, and continuous learning.” — Bessemer team, on why their 2024 roadmap is already obsolete
The Cost of Computing 📌
The Core (Govindraj Ethiraj) · 8 min
Why read this: An India-angle entry point into the Sora economics story: Sora’s $15M/day inference cost vs. $2.1M lifetime revenue is the brutally clean case study for why AI energy economics matter—and India’s $200B data center bet is colliding with a power grid that can’t handle it.
Sharp takes:
Sora’s unit economics weren’t just bad—they were structurally impossible: lifetime revenue of $2.1M against a single day’s operating cost of $15M means the product could have run for about 3.4 hours before the revenue ran out.
India’s AI ambitions are being built on a grid already stretched to breaking: reports of kerosene resurgence in some areas as AI-driven power demand competes with household electricity illustrate how the digital superpower narrative is colliding with analog infrastructure limits.
The broader claim—that even as models get more efficient, demand will always outpace efficiency gains—is the AI version of Jevons’ Paradox applied to energy, and it points to a structural energy bottleneck no efficiency improvement can solve.
Quotable:
“Sora’s daily inference costs were estimated at a staggering $15 million, driven overwhelmingly by energy consumption. Its lifetime revenue? A meager $2.1 million.” — The Core, on the unit economics that killed OpenAI’s video product
The Briefing: AI Wariness Syndrome 📌
Martin Peers · The Information · 4 min
Why read this: Amazon is trading cheaper than Walmart for the first time ever. Nvidia is at a seven-year earnings multiple low. Microsoft and Oracle have converged in valuation. Peers asks if this is rational war-driven selling or a market that has suddenly decided to punish AI spending.
Sharp takes:
Nvidia at 19.9x forward earnings with 71% projected revenue growth vs. Apple at 28.7x with 12% growth is the clearest expression of the market pricing AI capex risk over AI revenue reality—it’s not valuation, it’s sentiment.
The Microsoft/Oracle convergence is the sharpest tell: two years ago Microsoft traded at 34x to Oracle’s 20x; now they’re roughly equal. Oracle is borrowing heavily and growing faster—but it’s also a fraction of Microsoft’s size. This compression doesn’t make fundamental sense.
The IPO question buried in the piece is the right one: if public market investors are punishing AI-exposed companies, what does that mean for Anthropic and OpenAI’s fall IPO plans?
Quotable:
“The last time Amazon shares traded as cheaply as they are now, as a multiple of earnings, was during the 2008 financial crisis.” — Peers, on how deeply the market has discounted AI infrastructure spending
Big Tech Schools Big Energy on Powering AI 📌
Ann Davis Vaughan · The Information · 5 min
Why read this: At CERAWeek in Houston, it was Big Tech—not energy companies—calling the shots on where capital flows for AI power infrastructure. The ideological fight between gas-only vs. renewable-plus-storage-plus-nuclear is now a billion-dollar procurement decision.
Sharp takes:
Google’s Amanda Peterson Corio shut down the “gas or nuclear” binary at a public panel: “If we’re just relying on gas, we’re also screwed”—because supply chain bottlenecks mean gas plants take years too, and nuclear even longer.
The shift is structural: Microsoft and Google are now sponsoring dinners at CERAWeek alongside Chevron—the energy industry’s power conference is being colonized by the entities that are actually writing the checks.
The 100-hour battery storage system Google is deploying in Minnesota is the most interesting technical detail—it’s a direct challenge to gas peaker plants that only exist because storage couldn’t bridge multi-day weather gaps.
How Stripe Built “Minions”—AI Coding Agents Shipping 1,300 PRs/Week 📌
Steve Kaliski (Lenny’s Newsletter) · 5 min
Why read this: Stripe’s AI agent system kicks off from a single Slack emoji reaction and ships ~1,300 pull requests weekly—but the deeper insight is that good developer experience and good AI agent performance are the same investment, not two separate things.
Sharp takes:
Wait, really? The “activation energy” bottleneck, not coding speed, is what Stripe is actually solving. Steve Kaliski hasn’t opened a text editor to start work in months—work now begins in Slack threads, Google Docs, and support tickets.
Stripe’s years of DX investment (comprehensive docs, blessed paths, robust CI/CD) are directly translating to higher AI agent success rates—good docs written for human engineers are the training data that makes the agent work.
The virtuous cycle is the key mechanism: investments in DX improve agent performance, and investments in agent infrastructure benefit human developers—making the AI-native and human-native codebases converge rather than diverge.
Quotable:
“What’s good for human developers is good for AI agents. Stripe’s years of investment in developer experience—comprehensive documentation, blessed paths for common tasks, robust CI/CD—directly translates to higher AI agent success rates.” — Steve Kaliski, on why DX investment pays double dividends
Space Is the Place 📎
FT Newswrap · 5 min
Why read this: Startups are raising hundreds of millions to put AI compute in orbit—the pitch is unlimited solar power with zero land/water constraints. Critics say it’s detached from reality and will clutter an already congested sky.
Sharp takes:
The elevator pitch is genuinely novel: “By moving AI compute to space, we unlock access to unlimited solar power and completely remove the energy bottleneck”—a direct response to the grid capacity crisis hitting every terrestrial data center build.
The counterargument from former NASA associate director Rebekah Reed points to satellite congestion already threatening weather monitoring, GPS, and climate observation—the externalities of space debris are being socialized while the benefits are privatized.
Google’s announcement of a space-based prototype solar data center is cited as evidence the race has “become detached from reality,” but Google’s track record of moonshots with eventual commercial payoff makes that verdict premature.
Google in Talks to Finance Multibillion-Dollar Data Center for Anthropic 📎
The Information AM · 5 min
Why read this: Google would offer construction loans for Anthropic’s 2,800-acre Texas campus (500MW capacity, enough for 500,000 homes)—while Microsoft simultaneously absorbed the Oracle/OpenAI campus in Abilene that both walked away from. The hyperscaler infrastructure wars are getting strange.
Sharp takes:
Google financing a competitor’s data center while also being its cloud provider and $3B investor is a novel relationship structure—the distinction between “investor,” “customer,” “supplier,” and “financier” has fully collapsed in AI infrastructure.
Oracle and OpenAI walking away from a 900MW Abilene campus only for Microsoft to take it the same week is either elegant market clearing or a sign that the original specs were wrong for current needs.
Physical Intelligence’s $11B valuation discussion (from $5.5B) is the third data point in this briefing—the robotics-plus-AI infra investment story is accelerating alongside the software layer.
Trump’s Old-School Tech Advisers 📎
Bloomberg Technology · 4 min
Why read this: Trump’s new PCAST is dominated by semiconductor and infrastructure stalwarts—Jensen Huang, Lisa Su, Larry Ellison, Michael Dell—while Musk, Altman, and Amodei are conspicuously absent. The “no Elon?” reaction was immediate.
Sharp takes:
The semiconductor/infra dominance over software/AI-native companies reflects a deliberate framing: this is a “national industrial capacity” council, not an AI policy council—a meaningful distinction for regulation and export control debates ahead.
Anthropic’s pending IPO and its simultaneous battle against the Pentagon’s attempts to ban it as a supply chain risk are the backstory to Amodei’s absence—a company can’t easily advise the government while suing part of it.
Sony hiking PS5 prices $100 to $650 (buried in the same briefing) is a clean data point on how tariff/inflation pressure is flowing through to consumer hardware.
Finance & Markets
Money Stuff: Private Assets Are Coming to 401(k)s 📌
Matt Levine · Bloomberg · 15 min
Why read this: The Trump Labor Department just proposed giving private equity, private credit, crypto, and real estate legal safe harbor in 401(k) plans—and the timing could not be more ironic: the current headline is that retail investors are desperately trying to exit private credit funds and finding they can’t.
Sharp takes:
The safe harbor mechanism is the actual lever: 401(k) plans have avoided alternatives not because fiduciaries think they’re bad, but because class-action litigation risk makes offering them career-ending for plan administrators—removing that liability shifts the equilibrium.
The “people are trying to get out of private credit” irony is real and documented: illiquidity gates are being hit, and the political optics of expanding access to the same structures that are currently trapping retail money in are genuinely bad.
The LME nickel NDA story (buried in the same edition) is another Levine specialty: the London Metal Exchange apparently asked dealers to sign confidentiality agreements about its nickel market dysfunction, which is itself a signal that the dysfunction is ongoing.
Quotable:
“It is not the most auspicious possible moment to ask people to put more private credit in their 401(k)s. But financial-market memories are short, and regulatory timelines are long.” — Matt Levine, with characteristic restraint, on the timing problem
Will Lawmakers Ban Teens From Social Media? 📌
Kurt Wagner · Bloomberg Businessweek · 5 min
Why read this: Back-to-back jury verdicts against Meta and Google—one finding they created addictive products for children, one finding Meta misled teens about sexual predator risks ($375M penalty)—may be the tobacco moment that finally cracks legislative gridlock on social media regulation.
Sharp takes:
The “Big Tobacco” comparison is now being used by the New Mexico Attorney General who won one of the cases—not by critics, but by the prosecutor who extracted the judgment, which changes its rhetorical weight significantly.
The mechanism for legislative tipping points is usually liability, not outrage: the tobacco settlement in the 1990s moved policy because it moved money, and $375M in New Mexico penalties plus thousands of pending cases could do the same.
The irony: these verdicts come as the FTC under Trump is less likely to pursue tech regulation, creating a split where state AGs and plaintiffs’ lawyers are doing the work federal regulators aren’t.
Quotable:
“This is a real watershed in terms of holding social media companies accountable.” — New Mexico AG Raúl Torrez, comparing the verdicts explicitly to Big Tobacco’s courtroom reckoning
The Electric: Battery Recycling Goes Small and Snags a $1 Billion Contract 📎
Steve LeVine · The Information · 5 min
Why read this: Nth Cycle won a $1.1B, 10-year contract with Trafigura by doing the opposite of the 2020 EV mania playbook: processing a tenth the volume of competitors, using far less energy, and targeting the Chinese-dominated MHP supply chain gap.
Sharp takes:
The EV recycling graveyard is getting crowded: Li-Cycle declared bankruptcy and was absorbed by Glencore; Redwood Materials pivoted to powering AI data centers with old EV batteries—an irony that the EV mania’s recycling infrastructure is now being repurposed for the AI infrastructure boom.
Nth Cycle’s bet on modular scale—small, low-energy equipment vs. gigafactory-style plants—is the battery supply chain equivalent of the “boring company” thesis: consistent unit economics beat moonshot capex.
The nickel, cobalt, and lithium supply chain dependence on Chinese-controlled Indonesian MHP facilities is the geopolitical bet underneath the $1.1B contract—Western buyers are paying a premium to reduce that exposure.
Briefs
In 1977, After Citi Spent $50 Million to Install ATMs… 📎
Alex Johnson · X (Twitter) · 2 min
The bank teller parable for anyone making confident AI-jobs predictions: every generation was wrong—ATMs didn’t kill tellers, then they did, but only after deregulation drove branch expansion, then smartphones changed the game again, and JPMorgan Chase just opened 500 new branches. One data point: “Nobody knows anything.”
Veblen & Jevon Walk Into a Data Center 📎
Tomasz Tunguz · Theory Ventures · 2 min
Token prices fell 10-20x in 18 months and demand exploded (Jevons’ Paradox)—but Anthropic’s surge past $19B ARR may now be introducing Veblen dynamics: as the best models become status signals for enterprises, price sensitivity declines and premium pricing sticks.
America Is Relying on the Gig Economy 📎
FT Opinion · 3 min
Weak payroll growth + flat unemployment claims + record new business applications = possible structural shift toward freelance/gig work at both ends of the income spectrum. The precariousness is real; the consumer spending resilience it may be generating is the puzzle.
The Science of Quitting 📎
FT Edit · 2 min
Most people are one “jolt”—a divorce, a big birthday, an uninspiring new hire’s week—away from resigning, even when they shouldn’t. Pilita Clark’s column on resignation triggers is the behavioral economics angle most managers ignore.
Seven Things I’ve Learned Getting Companies to Use AI 📎
Mike Taylor · Every / Also True for Humans · 1 min
Top lesson from a head of AI consulting: buy models directly and skip the third-party tools. Resistance to AI adoption is almost always an activation energy problem, not a skepticism problem.
Hegseth’s Broker Looked to Buy Defence Fund Before Iran Attack 📎
FT Exclusive · 1 min
Pete Hegseth’s broker attempted a large purchase of major defense company stocks in the weeks before the US-Israeli attack on Iran, per three sources. The timing question is now formally on the record.
Embittered and Emboldened 📎
FT One Must-Read · 2 min
Iran is reportedly charging ships $2M each for safe passage through the Strait of Hormuz—potentially billions in revenue for the regime. Gideon Rachman’s analysis: the US and Gulf allies have very few tools to stop it because interdicting the “fee collection” risks escalating every transit into a military incident.
Apple Escalates Crackdown on Vibe Coding Apps 📎
The Information · 1 min
Apple removed the app “Anything” from the App Store last Thursday, the clearest signal yet of a systematic crackdown on AI coding apps Apple views as competitive threats. The Diff’s newsletter simultaneously notes that app store submissions are up 55% YoY from AI-generated code.
Chart of the Day: The Hateful Eight Is 85% of S&P 500 Decline 📎
Paul Kedrosky · 1 min
S&P 500 down 7% YTD, but the Mag 7 + Oracle (the “Hateful Eight”) account for 85% of the drawdown—roughly 576 points. The other ~490 companies are net positive. The index concentration that drove gains is now driving the pain symmetrically.
Westerners Are Fleeing Their Countries in Record Numbers 📎
The Economist · 1 min
Record emigration from Western countries, with economic consequences for both origin and destination nations. The Economist teases this as part of a broader package on the “expat economy”—no specific data point surfaced in preview, but directionally consistent with polling showing declining institutional trust in the US and UK.
TotalEnergies Nets Windfall on Middle East Oil Bet 📎
FT Exclusive · 1 min
TotalEnergies cornered the physical UAE and Oman crude market this month, buying every May-loading cargo and making more than $1B as oil prices surged. Dominant physical positioning in a war-disrupted market: the trade of the month.
Crypto’s Nasty Downturn Is Getting Worse 📎
The Information · 1 min
Six months into a crypto bear market with no bottom in sight—companies are cutting staff and pivoting to stock trading and prediction markets. The pivot-to-prediction-markets detail is the tell: the bull case has shifted from “crypto as asset” to “crypto as gambling infrastructure.”
Hungary Exploits EU Fraud Loophole 📎
FT Exclusive · 1 min
Hungary recovered and handed over less than a fifth of the €1.39B flagged by EU anti-fraud watchdog OLAF between 2015–2024. Viktor Orbán is simultaneously trailing in polls ahead of next month’s election—the fraud recovery failure is a political liability he can’t easily explain away.