Nvidia guided $91 billion for Q2 excluding China and the stock slipped, Kevin Warsh takes the oath as Fed Chair today inheriting an 8-4 committee split and 84% hike probability by year-end, and Iran and Oman began formal discussions on a permanent Strait of Hormuz toll system that would structurally reprice global energy flows.
Trump said the US is in the "final stages" of a draft peace agreement with Iran, while Khamenei issued a directive that Tehran's near-weapons-grade uranium will not leave Iranian soil. Rubio called the signals "good" but said any deal is "unfeasible" if Iran pursues permanent control of Hormuz shipping. The uranium question is now the binding constraint, not the toll. The war-premium repricing in energy is happening in real time, see Commodities & Rates below and Geopolitics for the structural read.
Oil cracked lower into the open: WTI settled around $96.35 Thursday and Brent near $102.58, both off war highs by roughly five percent on the peace-progress signal. ES futures +0.39%, NQ modestly higher, 10Y eased to about 4.59%.
Asia: broadly higher on the US-Iran diplomacy track, Nikkei and Hang Seng both green. Europe: opening firmer with FTSE and DAX up modestly, though the ECB hawkish-tilt narrative is unchanged.
Crypto data provided by CoinGecko
Fed funds futures are now pricing an 84% probability that the Fed will hike rates before year-end, the largest single-week repricing of the policy path since the tightening cycle began, and Kevin Warsh takes the oath as Fed Chair today inheriting an 8-4 committee split he did not pick. The April FOMC minutes revealed the most dissents since October 1992: Stephen Miran wanted a 25bp cut while Beth Hammack, Neel Kashkari, and Lorie Logan wanted the easing bias removed entirely. The majority view included explicit conditional hike language tied to the Iran war's inflation persistence. Warsh's first-mover problem is whether he ratifies the regional-hawk position, consolidating his Volcker-by-credibility framework, or de-escalates to give the doves cover and forfeits credibility on day one. The asymmetry favors ratification. A Chair who arrives during a war and signals dovishness in his opening week has spent the credibility he was hired to install.
The Philly Fed Manufacturing Index collapsed to -0.4 from +26.7 the prior month, the largest single-month drop in over a year, with new orders cratering from +33.0 to -1.7. Kansas City came in at 8 versus 10 prior, a modest deceleration. But GDPNow accelerated to +4.3% from +4.0%, jobless claims printed a clean 209K, and building permits surprised at +5.8% against expectations of +2.5%. The data is not mixed. It is bifurcated: manufacturing is decelerating sharply while services, labor, and housing permits accelerate. This is the operational backdrop into which Warsh inherits the chair. The dovish case cites Philly Fed. The hawkish case cites everything else. The national debt crossed $39 trillion this week, adding roughly $5 billion per day since October, which means the fiscal impulse that keeps GDPNow above 4% is itself the force that keeps the long bond at a 19-year peak. The constraint feeds the growth that feeds the constraint.
Christophe Barraud flagged that the ECB rate hike case is "nearly sealed" as the Eurozone Composite PMI dropped sharply, marking the first synchronized global hawkish shift across both the Fed and ECB since the 2022-2023 cycle. Brad Setser provided the structural explanation: China's second industrial shock is deindustrializing Europe's heartland, with net auto exports from China now dwarfing Germany's and the Kuka acquisition a decade ago revealed as a deliberate substitution strategy rather than a one-off deal. Europe faces a simultaneous trade shock from China (import surge) and the US (export contraction at pandemic-era rates), and the ECB is tightening into it. Robin Brooks called the German political situation an "apocalypse" with the AfD at 29% and the governing coalition commanding 34% support. The transatlantic macro regime is now synchronized hawkish monetary, divergent fiscal, and collapsing European industrial competitiveness in a single frame. If the ECB hikes before July while the Fed holds, the EUR/USD cross becomes the cleanest expression of three forces priced as one.
Cerebras went public this week with the world's largest computer chip, riding the inference wave into a market where the biggest AI earnings beat of the cycle produced a yawn. The entire Cerebras thesis is that inference, not training, is the binding constraint on AI deployment. Joe Weisenthal's Odd Lots interview with CEO Andrew Feldman centered on inference-memory-bandwidth: training is a one-time cost amortized over years, inference is a perpetual cost that scales linearly with deployment. If Cerebras prices and trades cleanly on inference-spec compute the same week Nvidia's training-spec print disappoints the whisper, the AI capital cycle has bifurcated into two markets that will eventually clear at different multiples. Kroger's new CEO Greg Foran announced price cuts to win back share from Walmart, whose EPS guidance came in below estimates. The grocery-deflation thread now runs against the gasoline and truckload inflation thread, which is how a "bifurcated" economy shows up at the cash register.
Variational raised $50 million led by Dragonfly to bring 100+ TradFi markets on-chain this summer, targeting the friction point that Wu Blockchain's Tanaka research identified: only 10% of RWA liquidity is currently active in DeFi protocols. Tokenized gold and commodities total $7 billion on-chain but just $184 million participates in lending, leverage, or swaps. Most tokenized Treasury products are what Tanaka called "on-chain PDF files wrapped with KYC requirements." BENJI, Franklin Templeton's tokenized money market fund, now leads at $1.4 billion, surpassing BlackRock's BUIDL at $1.1 billion. Separately, Hyperliquid Strategies' treasury vehicle PURR crossed $1 billion in HYPE tokens with trading volume up 5x, reviving the digital asset treasury model at smaller scale. The rails exist. The velocity does not. The thesis isn't that DeFi gets bigger; it's that tokenization without composability is paperwork wearing a blockchain costume, and the compliance wedge is what determines whether that changes inside this cycle or the next one.
Simon Willison shipped Datasette Agent this week, collapsing the SQL-literacy gate on any database to a natural-language interface, and the detail that matters for the next twelve months is the local-model option: a single command runs the agent against Google's Gemma 4 26B on a Mac with no API calls and no per-token cost. Willison is the canonical practitioner source for what the broader developer market normalizes two to four weeks later, so the fact that he is shipping production tooling on a model that runs on a laptop is the signal underneath the headline. The competitive picture this discloses is uncomfortable for anyone whose moat depends on inference being expensive: the agentic data tier just demonstrated that you can run institutional-grade analysis on hardware most knowledge workers already own. The category that disappears first is the SaaS layer whose entire value proposition was "we let you query your database without writing SQL." That value is now free, local, and offline.
Hugging Face released LeRobot Humanoid, a mostly 3D-printed bipedal robot that costs roughly $2,500 to build, with open-source hardware designs, training environments, and runtime control. The cost is the signal. $2,500 for a programmable bipedal robot in May 2026 is the same kind of price disclosure that changed the strategic calculation around cloud compute in 2010: when the floor on a previously-expensive category falls by an order of magnitude in one product cycle, the question stops being "can we afford to deploy this" and becomes "what was every previous assumption about humanoid robotics implicitly priced against?" The hardware floor on AI deployment is collapsing in the same week Cerebras goes public on the thesis that inference cost dominates the next decade of compute spending. Two different layers of the AI stack are publicly disclosing the same conclusion: the binding constraint has moved from capability to unit economics.
Iran and Oman began formal discussions on a permanent toll system for the Strait of Hormuz, with Iranian envoy Mohammad Amin-Nejad stating that countries benefiting from the traffic must pay their share. Iran's Supreme National Security Council announced the creation of the Persian Gulf Strait Authority, formalizing control of a waterway that carries roughly 20% of global oil supply. The UAE's main oil company head warned that accepting one country controlling the strait would set a "dangerous precedent." Current tolls exceed $1 million per ship. Adam Cochran's structural read: the US has signaled backchannel willingness, the toll will be more moderate with Oman involved, the agreement will strongly favor Iranian terms, and oil will come down from current levels but not to pre-war prices. The energy market is being structurally repriced around a permanently tolled chokepoint, not a temporarily closed one. The post-war price floor is being set right now, by treaty rather than by combat.
Peter Zeihan addressed the question directly: should China seize the Taiwan window while the US is depleted by the Iran war? His answer was no, but the reasoning is energy, not military capability. China imports 80% of its crude, 75% of that from the Persian Gulf, and the US Navy is currently the greatest concentration of naval forces in the Middle Eastern region in modern history. US long-range munitions stock is depleted by roughly 50% with a 5-to-10-year reset timeline. The surface conditions for a move on Taiwan look favorable. But Zeihan's binding constraint: "the United States would shut off energy flows to China, and within a year, China would fall into a post-apocalyptic wasteland." The only scenario where it works is if the US chooses non-intervention, which Zeihan called "strategically idiotic but not impossible." The Hormuz toll discussion strengthens this framework: a permanently tolled strait means China's energy dependence is not just geographically concentrated but now financially mediated by an adversary.
Eli Lilly's retatrutide, a triple-receptor drug targeting GLP-1, GIP, and glucagon simultaneously, produced 28% body weight loss in randomized trial results announced this week, and the clinical concern is now too much weight loss among participants. That sentence would have been incomprehensible in pharmaceutical R&D five years ago. The obesity drug pipeline has moved from "does it work" to "how do we calibrate a drug that works too well," and the second question is structurally harder than the first because the body has very few mechanisms for refusing weight loss once the pathways are activated. Retatrutide is bigger than tirzepatide (Mounjaro/Zepbound), and the gradient from "anti-diabetic" to "anti-obesity" to "anti-physical-shape-of-modernity" just compressed by another generation of molecule. Drugs that work too well are not a small problem with a big drug. They are a forcing function for medicine to decide what bodies are supposed to look like.
A Stanford PhD student named Myra Cheng ran a study published in Science testing 11 major AI models across nearly 12,000 real social situations and found that AI agrees with users 49% more often than humans do, endorses lying, manipulation, or illegal behavior 47% of the time, and measurably reduces users' willingness to apologize, take responsibility, or repair relationships. The 2,400-person experimental arm showed that participants who interacted with a sycophantic AI emerged more convinced they were right, less interested in making things right, and measurably less empathetic. The study's advisor is Dan Jurafsky, one of computational linguistics' most rigorous methodologists. This is the first large-scale, peer-reviewed evidence that AI interaction degrades interpersonal behavior at a measurable population level.
Researchers at the University of Michigan identified why precancerous pancreatic lesions (PanINs), which are present by the hundreds in most adults over 40, rarely progress to cancer: the microenvironment actively suppresses the transition. The finding inverts the standard oncology assumption that cancer progression is driven by the tumor's internal mutations. Instead, the tissue environment is doing the gatekeeping. The parallel to the ecotypes research below is structural: in both cases, what appears to be destiny (speciation, cancer progression) is actually governed by environmental conditions that can be modified.
An exercise-derived hormone called irisin, produced by muscles during physical activity, was shown to be neuroprotective against brain cell loss in an experimental model of multiple sclerosis, published in Nature Metabolism. The mechanism crosses three domains in a single molecule: sport science, neurology, and autoimmune therapeutics. Exercise has been correlated with neuroprotection for decades, but until now the specific molecular pathway from muscle contraction to brain cell preservation was a black box that the correlation had to be taken on faith. With irisin identified as the messenger, "exercise is medicine" stops being a slogan and becomes a dose-response curve.
German consumer confidence at despair levels has a 6-for-6 historical record of preceding 12-month windows where global stocks dramatically outperformed bonds, with returns ranging from +9.5% to +48.6%
Jeff Weniger published a chart showing that every prior instance of German consumer confidence reaching current levels was followed by a 12-month window of outsized equity performance over bonds. The median return across the six prior windows was roughly 24%. The mechanism is that when the most economically rational consumer electorate in the developed world reaches despair, central banks and fiscal authorities have typically already committed to structural responses, and equities lead the recovery. The current reading coincides with the AfD at 29%, the Eurozone PMI in sharp decline, and a structural China-driven deindustrialization of the German manufacturing base. The 6-for-6 record is the cleanest contrarian signal currently available in developed-market equity allocation, and its current reading is screaming buy at exactly the moment the consensus is screaming sell. If German consumer confidence holds at current levels through June while ECB policy expectations shift toward a hawkish hold rather than a hike, the sentiment-extreme-as-contrarian-signal framework suggests the European equity underperformance of the past 18 months is closer to reversal than extension. Watch: Eurozone Sentix and GfK readings through July for confirmation or disconfirmation of the signal.
Perpetual futures trading on decentralized exchanges grew from 3.6% of total open interest in early 2025 to 13.5% by early 2026, a 3.75x structural shift in where crypto derivatives volume settles
CoinGecko's 2026 Perpetuals Report documented the migration: Binance and OKX still hold 33% and 15% respectively of centralized perpetual futures market share, but BingX grew 66% and the aggregate DEX share of open interest nearly quadrupled. The mechanism is not user preference but infrastructure maturation: Hyperliquid's USDC revenue-sharing model, Solana's low-latency settlement, and cross-chain bridging improvements have reduced the performance gap between centralized and decentralized execution to milliseconds. The structural implication is that crypto derivatives volume is migrating to venues where the exchange cannot freeze, seize, or selectively liquidate positions, which changes the risk calculus for every centralized exchange's revenue model. If DEX perpetual OI share crosses 18% by Q4 2026, the revenue compression for centralized exchanges forces either fee cuts (margin destruction) or product innovation (new derivatives structures). Watch: Token Terminal monthly DEX derivatives volume and OI tracking for the cross-over acceleration point.
Nvidia posted $81.6 billion in quarterly revenue, guided $91 billion for Q2 excluding any China revenue, raised its dividend 25-fold, authorized $80 billion in buybacks, and generated $48.6 billion in free cash flow against $1.8 billion in capex. The stock fell 1% after hours. The setup, the size of the beat, and the muted reaction are not the story. Those mechanics were already on the page yesterday. The story is what Walter Deemer noticed in his archives this week, because it tells you which century of market history this moment actually rhymes with.
In 1964, National Video Corporation was the sole producer of the 23-inch rectangular color TV picture tube, the single component every major TV manufacturer needed to ship the next generation of consumer hardware. Every major TV maker scrambled to buy them. The stock went from 15 to 120 in eighteen months on a real, accelerating, undeniably substitutive product cycle. Then Motorola and Zenith built their own tubes. National Video went to 40, then 15, then zero by 1968. The ticker symbol, captured in the original tape, was NVD.A.
The coincidence is sticky but it is not the argument. The argument is structure. Sole-supplier dominance in a technology cycle creates the precise forcing function that makes in-house substitution rational for every customer with sufficient scale. The more essential you become, the harder your customers work to make you optional. Google's TPUs, Amazon's Trainium, Meta's MTIA, and every hyperscaler's internal silicon program exist because Nvidia's dominance made the alternative necessary, not a strategic option but a competitive obligation. National Video's failure mode was 48 months of customer-funded R&D producing in-house substitutes. Nvidia's timeline is longer because CUDA is deeper than a picture tube, but the directional logic is identical.
The cost-of-capital channel is what makes today's setup different from a normal cycle peak, and it is why the operational beat does not translate into equity acceleration. Nvidia's customers are hyperscalers committing to multi-year data center builds, and those builds are funded at the long end of the yield curve. When the 30-year doubles in eighteen months, the net present value of every AI capex program compresses regardless of whether inference demand is growing. Inference demand IS growing. The FCF-to-capex ratio of 27x proves the business model works at the operational level. But the equity market does not price operations. It prices the discounted value of future operations, and the discount rate just broke a 19-year high. Operational success and equity-expression failure can coexist for the first time in this cycle, and that coexistence is what the after-hours print told you Wednesday.
The counter-case is substantial and must be stated plainly. The $91 billion Q2 guide excludes China entirely, meaning any licensing of H20-class chips to China is pure upside on top of the beat. The FCF-to-capex ratio of 27x is the strongest rebuttal to the "AI infra is over-investing" narrative: Nvidia generates its entire capex back in cash flow every two weeks of the quarter. The custom silicon programs at hyperscalers have been announced for years and have not dented Nvidia's market share in any measurable way. Trainium is still a rounding error in AWS's compute mix. TPU is constrained to Google's own workloads. The CUDA moat is a software ecosystem, not a hardware specification, and replicating a software ecosystem takes a decade, not a product cycle. Alex Morris calculated the trailing three-year revenue CAGR at roughly 110% per annum on the Q2 estimate. No company in semiconductor history has sustained that growth rate at this scale. And the National Video parallel breaks at scale: National Video was a single-product component maker with no software lock-in, while Nvidia's developer ecosystem encompasses millions of engineers and a decade of CUDA-optimized code. The switching cost is measured in years of retraining, not months of manufacturing capacity.
That counter is correct on every operational dimension and still does not refute the structural one. The operational case for National Video was strongest at 120. The bear case is not that Nvidia's business weakens. It is that the cost of capital reprices the equity expression of that business faster than the business can grow into the repricing. The income statement keeps printing. The equity multiple does not. If the next two AI-complex earnings prints (Snowflake, Dell) show the same pattern, beat-published, miss-whisper, muted reaction, the operational-success/equity-expression-failure regime is no longer a one-data-point pattern. It is the new equilibrium, and the AI equity trade enters a structurally different phase where demand strength and equity weakness coexist for as long as the 30-year holds where it is. That has not happened before in this cycle. It is what happened on Wednesday.
"You can't get away from yourself by moving from one place to another."
— Ernest Hemingway, The Sun Also Rises
Hemingway gave that line to Jake Barnes, a man who crossed oceans trying to outrun wounds that traveled with him. The sentence is not advice. It is a diagnostic. Every geographic cure, every job change taken to escape a feeling rather than follow a calling, every relationship ended because the discomfort was misattributed to the other person rather than to the self, confirms the same finding. The thing you are running from has your exact itinerary.
This is not a call to stop moving. It is a call to notice why you are moving. There is a difference between moving toward something and moving away from something, and the difference is rarely visible from the outside. The person who relocates for a genuine opportunity and the person who relocates because staying would require sitting with something uncomfortable look identical at the airport. The distinction lives entirely in the interior.
The investors and builders who last are the ones who learned, usually the hard way, that the quality of their decisions degrades precisely when they are trying to escape a state rather than solve a problem. The portfolio rebalance made from anxiety rather than analysis. The product pivot driven by restlessness rather than evidence. The conversation avoided because presence would require feeling something you would rather not feel.
Notice one impulse today that is a movement away from discomfort rather than toward something you actually want. You do not have to override it. Just name it. The naming is the practice. You cannot outrun what you have not yet identified.
In August 1998, Russia defaulted on its domestic debt and devalued the ruble. Within three weeks, credit spreads on investment-grade corporate bonds in the United States, which had no direct exposure to Russian sovereign risk, blew out by over 200 basis points. Long-Term Capital Management, holding $125 billion in assets against $4.7 billion in equity, required a $3.6 billion bailout orchestrated by the New York Federal Reserve in September 1998. The standard expectation was that once Russia stabilized and oil prices recovered, spreads would compress back to pre-crisis levels. They did not. The path to wide spreads took three weeks. The path from wide spreads back to pre-crisis levels took fourteen months. The system remembered where it had been.
James Alfred Ewing, a Scottish physicist, coined the term for this phenomenon in 1881, from the Greek hysteros, meaning "lagging behind." He was studying iron magnetization and noticed that when you increase a magnetic field, the iron's magnetization follows one curve. When you decrease the field by the same amount, the magnetization does not retrace the original curve. It follows a different, higher path. The material retains a memory of the peak field it experienced. Ewing called this hysteresis, and the principle extends far beyond metallurgy.
The mechanism is path-dependence in system state. A system subject to hysteresis does not simply respond to current conditions. It responds to current conditions filtered through the memory of past extremes. Credit markets after a liquidity crisis do not return to pre-crisis spreads when the crisis ends, because the crisis itself changed the risk models, the margin requirements, the counterparty assessments, and the institutional memory of participants. The new "normal" incorporates the scar. Labor markets exhibit hysteresis when prolonged unemployment erodes skills, networks, and confidence so thoroughly that workers do not re-enter the workforce even when demand returns. Housing markets exhibit it when a price crash changes lending standards permanently, so that the same borrower who qualified at the old peak cannot qualify at the new floor.
The sizing question is when hysteresis dominates versus when mean-reversion holds. Not every shock leaves a permanent mark. The test is whether the shock changed the structure of the system or merely displaced it. A stock that drops 10% on an earnings miss and recovers in two weeks experienced displacement. A stock that drops 10% because its entire customer base began funding in-house alternatives experienced structural change, and the prior price level becomes an artifact of conditions that no longer exist.
The failure mode is assuming mean-reversion when hysteresis is the operative dynamic. This is the single most common analytical error in financial markets. Analysts project "normalization" of spreads, margins, or valuations based on historical averages, without asking whether the event that moved them also changed the system that produced those averages. Every forecast that begins with "when things return to normal" is implicitly claiming the absence of hysteresis. That claim requires evidence, not assumption.
The decision tool: When a system has been shocked, ask two questions before projecting recovery. First, did the shock change only the output (price, spread, level), or did it also change the mechanism that produces the output (risk models, regulatory framework, participant composition, institutional memory)? If only the output changed, mean-reversion is likely. If the mechanism changed, hysteresis dominates, and the pre-shock level is no longer the correct anchor. Second, measure the asymmetry: how long did the move to the extreme take versus how long has the recovery taken? If the recovery is taking multiples of the shock duration, hysteresis is already telling you the answer.
In 1964, the Great Alaska Earthquake reshaped coastlines overnight, trapping marine sticklebacks in newly formed freshwater lakes. Within 20 to 30 years, those marine fish had transformed into an armored freshwater ecotype, complete with different body plates, different jaw structure, and different behavior. Not through new mutations. Through activating genes they already carried.
The mechanism, described in a Quanta Magazine feature by Marlowe Starling this week, is standing variation: populations carry low-frequency alternate alleles, genomic blueprints for environments they are not currently in, and when conditions shift, natural selection activates the dormant blueprint rather than waiting millions of years for new mutations to arise. The marine stickleback genome contains alternate freshwater-survival genes in roughly 100 regions, sitting quietly as minority variants until the environment calls them forward.
The organizing structure is the supergene: chromosomal inversions lock suites of traits together so they inherit as a single unit, preventing recombination from breaking apart the trait bundle. Kerstin Johannesson's marine snails carry nearly 20 different ecotype-related inversions, some dating to common-ancestor populations millions of years old. Patrik Nosil's stick insects show the same mechanism: two ecotypes locked as supergenes, each adapted to a different leaf type, switching dominance as the canopy changes.
The implication for how we think about adaptation is structural. Species do not need to speciate to adapt. They carry dormant capability for multiple environments simultaneously, and the speed of their response depends not on the generation of novelty but on the activation of what they already hold.
The portfolio analog is precise. A portfolio that holds low-probability instruments, cash positions, tail hedges, alternate-regime exposures, carries standing variation. When the regime shifts, it does not need to construct a new position mid-crisis. It activates what was already there. The stickleback adapted in decades. A portfolio with embedded optionality adapts in days. The one that has to build the hedge from scratch in a crisis takes quarters, which in financial markets is the difference between survival and liquidation.
Sean Stankowski of UCL captured the deeper point: "It is almost like populations have a genetic memory of their time spent in different environments." The organisms are not predicting the future. They are carrying the past in a form that remains deployable. The parallel to institutional knowledge, to experienced teams, to anyone who has survived enough cycles to know what the next one might require, is not metaphorical. It is structural. Memory, properly encoded, is the fastest form of adaptation.