Lux Q3 2026 Report
Anomalous Instruments
Karl Popper described how knowledge grows: a bold guess met by a hard test. Thomas Kuhn described how knowledge moves: long stretches of "normal science" until anomalies pile up so high that a theory breaks. Kuhn noticed who does the killing: the inventors of new systems are almost always "either very young or very new to the field whose paradigm they change." Max Planck, who launched quantum theory and then watched younger physicists carry it forth, put it darkly: new scientific truths don’t triumph through persuasion, but because a new generation rebels against the status quo while old proponents eventually die. Knowledge advances one funeral at a time.
Where do these anomalies come from? Rarely from just thinking harder. Instead, they come from new instruments that open the world to fresh observations. Freeman Dyson inspired this idea: a new instrument reveals an anomaly, an outsider takes it seriously, an incumbent paradigm cracks under testing, and then a novel market emerges. Van Leeuwenhoek didn’t out-argue Aristotle, he made a better lens. Galileo didn’t reason his way to the moons of Jupiter, he pointed a tube at the planet and saw three “stars” that moved the wrong way. The telescope made the anomaly, the anomaly made the crisis, and the crisis made the revolution. Richard Feynman told Caltech's graduates in 1974 that the first principle is that you must not fool yourself, and you’re the easiest person to fool. An instrument is a device for making it harder to fool yourself.
From science to financial markets, paradigms don’t fall from better arguments but through better instruments, held by people who’d rather be refuted by reality than flattered by consensus. Last quarter, we wrote about the method: conjecture and refutation, the Royal Society's three subversive words (nullius in verba), and the error-correcting operating system that defends the West as importantly as any army. This quarter, we write about the instruments, because the method without instruments is just philosophy, and instruments without method are just toys.
Here lies everything Lux does: backing rebellious founders who uncover anomalies by embedding their unique know-how into instruments that act on the world. Our instrument makers tracked single molecules in living cells toward a cancer drug, kept a pig's kidney working inside a human for nine months, grew crystals in orbit, lasered data down from space, and rolled defensive interceptors off new production lines while Iranian missiles kept falling on our allies. They’re scattered from Sofia and Lagos to Tel Aviv and Tokyo, many reaching gold in math or physics olympiads—even as many haven’t reached the age to drink alcohol. All are driven by an outsider’s perspective into markets where anomalies are piling up faster than the old paradigm can patch them. The instruments change, the method doesn't. It's the method free societies invented, the one their institutions, markets and militaries still run on, and the one on which we're building Lux’s next chapter, the tech of science and the science of tech.
The Instruments of Revolution
Our knowledge is embodied in our objects: the angle of a blade, the curve of a lens, the bend of a flask, the throat of a rocket nozzle, the weights of an AI model. Our friend Cesar Hidalgo calls these crystals of imagination. We've long called it matter that matters. A page stores knowledge, but an object puts it to work. Most technologies are tools, know-how built to act on the world. A rarer and more subversive kind are instruments, know-how built to read the world and report back what we don't yet know. Every instrument is a kind of heresy machine: it doesn’t just answer our questions but questions our answers, manufacturing anomalies our existing paradigms can’t accommodate.
It’s not just mechanical, but biological as well. Domesticating the wolf and later the boar, we wrote the traits we wanted into their genomes by trial and error over millennia. We bred the dog to assist us and the pig to feed us. Now, Lux companies are editing the dog to defend us and the pig to heal us. This quarter, eGenesis reported that a patient lived a record 271 days without dialysis on a gene-edited pig kidney before bridging to a human one. An FDA-cleared trial for 33 patients is set to begin early next year. Warhound emerged from stealth to clone and genetically edit military working dogs, ethically addressing a quiet geopolitical vulnerability: more than 80% of America's hounds are bred in Europe. Worse, after three years of acquisition and training, fewer than half are fit for service. Warhound will preserve the biology of the best performers, clone it, edit out the hip dysplasia and cancers that end careers early, and deliver the dogs in Anduril-built goggles and sensors. The first are expected to be born early next year and ready for service by 2028, with the American Humane Society certifying welfare. Osmo, which we spun out of Google, is giving machines a sense they never had by digitizing smell and building the only repository of it. Atrandi, in Vilnius, Lithuania, is building instruments for single-cell biology. Valthos, our Lux Labs biodefense company, exists because a cheap biological attack can only be answered by a cheaper refutation: sequence, identify and counter in days rather than months.
As for inner space, so too for outer space. In January 1920, the New York Times ran an editorial mocking Robert Goddard, a physics professor, for believing a rocket could fly in the vacuum of space. Everyone knew a rocket needed air to push against, and Goddard seemed to lack the knowledge "ladled out daily in high schools." Six years later, Goddard launched the world's first liquid-fueled rocket from his Aunt Effie's farm in Massachusetts. It flew for 2.5 seconds and landed in a cabbage patch. On July 17, 1969, with Goddard already dead for 24 years and Apollo 11 en route to the Moon, the Times printed a correction: a rocket can, in fact, fly in a vacuum. "The Times regrets the error." Error correction in this case took half a century and a funeral just as Planck predicted. Markets rarely grant 50 days.
Today, the descendants of Goddard's cabbage-patch rocket haul factories and fleets (and serious funds). This summer, Impulse Space raised $500 million and became the first upper-stage provider selected for the Space Force's national security launch program; Varda raised $250 million, returning capsules from orbit that double as hypersonic test beds; and EnduroSat in Sofia, Bulgaria raised more than $200 million to build satellite buses for allied constellations. Observable Space, whose $90 million Series A we led in May alongside a $94 million sole-source Space Force award, is building laser links that rapidly move data from orbit to ground. Reflect Orbital received FCC authorization in July for its first demonstration satellite and has secured launches to sell sunlight after dark. Invert the parabola, as we wrote earlier this year, and the same physics defends. Cambridge Aerospace raised $300 million at a valuation that nearly tripled in four months. The money will fund Skyhammer, an interceptor Britain is supplying to its own forces and its Gulf partners against Iranian Shahed-style drones, with production targeted at 2,500 a month by next spring. Covenant emerged from stealth with a long-range cruise missile that has already successfully passed 200 test flights while building a new factory in Texas. Finally, Hadrian raised $1.37 billion to machine the precision parts all of this requires at the tempo deterrence now demands, while securing $900 million of U.S. Navy backing for high stakes submarine components.
As for outer space, so too for latent space. We impress our know-how into the weights of today’s AI models, which have become the first human-made objects that also generate hypotheses of their own. When the instrument starts posing its own conjectures, our revolution compounds. It also decays and evolves faster than any instrument before it, which makes Lux’s work an endless frontier. A teardrop hand axe kept its design for a million years. A frontier model keeps its edge for maybe two quarters. Each model is wildly profitable to serve for six to twelve months, then must be replaced by a successor that costs more to train, all while open-weight copies currently trail by about three to six months at a fraction of the cost. Whether that half-life is a moat only the richest can afford or a capex tax that never ends may be the most expensive open question in markets, and will be answered by how quickly the cheap copies continue to arrive. Sakana AI in Tokyo is building an AI scientist that designs and runs its own experiments. CuspAI in Cambridge searches chemical space for materials no chemist has imagined. And the protein models of EvolutionaryScale, now part of the Chan Zuckerberg Initiative, read and write the grammar of biology.
Here's the key investment lesson from all this history: the instrument comes first, the anomaly second, the paradigm shift third, and the market nearly always comes last. A new instrument is a short position on an old consensus.
Instrumental Companies
When new partners join Lux, their first task is learning what a Lux company is. As generalists investing at the edge of the known, we don’t screen by sector but by method. Does this founder build instruments—competitively advantaged—that let them see reality more clearly and run experiments faster than the incumbent?
Consider five Lux companies in fintech, defense, software, autonomy and open-source AI. Each has different products, different customers, different supply chains, and different networks, but the founders of all five share the same method: hypothesize, experiment, measure and iterate, all at a tempo incumbents can’t match. Science is a method, not a sector.
First, Ramp entered a category that already existed against a well-funded competitor. It won by running more experiments, faster, and by treating its own data stream as an instrument: its AI Index reads actual transactions rather than surveys measuring how American businesses adopt AI. This summer, Ramp passed $2 billion in annualized revenue while generating cash. Second, Anduril, which we backed from the beginning, inverted the defense procurement paradigm by building and testing with its own capital before a program of record exists. Its Fury autonomous fighter was flying before any production decisions were made. This month, an F-22 pilot gave Fury, flying as the Air Force’s YFQ-44A, its tasks midflight and let its software fly the maneuvers. Here is man and machine, a fifth-generation fighter with a robotic wingman. The company’s method compounded this quarter: a Navy contract worth up to $2.9 billion for Virginia-class submarine components, an Army command-and-control award worth up to $1.8 billion, and Arsenal-2, a $3.7 billion shipyard at Sparrows Point in Maryland. The test range has become Anduril's peer review, and the adversary is arguably the final referee. Third, Cognition is run by Scott Wu, a three-time gold medalist at the International Olympiad in Informatics who posted a perfect score in 2014 at age 17. The company turned a hypothesis, "can AI code?", into repeatable experiments measured against real-world software problems. Fourth, Applied Intuition makes simulation an experiment that can afford to fail: a million failures in silico before the code goes inside an actual vehicle. Last month, Honda and Nissan both adopted its vehicle operating system, a serious vote of confidence from the heart of Japan's automotive establishment. Meanwhile, its defense work extends the same simulate-measure-retrain loop to counter-drone systems for the Pentagon. Fifth and finally, Hugging Face turned AI research papers into reproducible infrastructure: models posted in open to be tested, compared and improved. We led its Series A in 2019, shortly after it pivoted from a chatbot for teenagers to an open-source library that few outside research labs recognized as an instrument at all. It agreed this quarter to be acquired by Nvidia for nearly $13 billion.
The same method powering these five also runs through Hadrian (the factory as experiment), Nominal (test-and-evaluation software, literally the instrumentation of the method), Saildrone (the ocean as a sensor field) and Varda (orbit as a laboratory).
Where some focus on ‘frontier’ or ‘deeptech’ investing, we’ve stayed true by pursuing method over sector. Across 39 incubations at Lux Labs as well as our f(Lux) Fellows, Riskgaming, Sci-Tech Diplomacy (global and geopolitical), Civic-Tech (domestic and emerging) and other strategies, we find founders forging ahead on the experimental edge, empowered by compounding tools that make each generation more natively capable than the last.
High Potential-Energy Youth
Consider some of history’s most influential inventors of instruments. Isaac Newton was 23 when the plague closed Cambridge and he went home to lay the foundations of calculus, optics and universal gravitation. Einstein was 26 when he unlocked relativity. A century ago, quantum mechanics was so disproportionately dominated by men in their twenties (Heisenberg at 23, Pauli at 24, Dirac at 25) that the Germans called it Knabenphysik: boys' physics. Kuhn's observation that paradigm-breakers are very young or very new wasn't just a sociological curiosity, it described a mechanism. The young and new haven't spent decades defending the old ideas, so they have less to lose by abandoning them.
Venture is going through its own Knabenphysik. Studies once showed the average founder of America's fastest-growing startups was 45, and a 50-year-old founder was nearly twice as likely to build one as a 30-year-old. Within AI, however, the founder profile has shifted: the average age of founders of companies valued at a billion dollars or more fell from 40 in 2020 to 29 by 2024. Founders barely old enough to rent a car are running companies with real revenue run-rates in the billions. Coincidentally over the past few years, a new generation of partners has dropped the median age of the Lux investment team from 43 to 30.
Many of our top founders share a pedigree: the International Olympiads in mathematics, informatics and physics. Cognition counts on its staff at least one competitor from every International Olympiad in Informatics held between 2006 and 2025. Why? Because these olympiads train a specific muscle: speed and rigor on hard problems under the pressure of a clock. Flapping Airplanes, an AI lab barely a year old studded with math, physics and informatics medalists, raised a $1 billion Series A led by Lux, nine months after our first check. We believe before others understand, and some of the people we back were solving the problem before others had even finished reading about it.
Here's the twist: this summer, an AI scored a perfect paper at the International Mathematical Olympiad. When a model can win gold, the medal may no longer be the signal. What’s scarce is what’s valuable, and the scarce skill moves from solving well-posed problems to choosing which problems are worth solving. That leap is why we built f(Lux), a fellowship for undergraduates in computational biology, AI infrastructure, advanced materials, aerospace and defense. Of our inaugural 13 fellows, four have already founded companies and four more joined startups as early employees. This year's class of 16—selected from nearly 800 applicants—gathered this past quarter to spend time with one another as newly designated f(Lux) Fellows.
A caution against ageism is in order. Kuhn's phrase was very young or very new. Young helps, new helps, but what matters most is being unencumbered by what’s come before. Paradigms don't change minds, they change generations and sometimes a generation of one: a single person willing to think differently and start over. As William Blake put it: I must create my own system or be enslaved by another man’s.
Larger Checks, Faster Clocks
As the quip goes, time exists so everything doesn’t happen at once. Dimson said risk is more things can happen than will. Revolutions compress time, and so does the capital that funds them. Through the first three quarters of 2026, U.S. venture investment surpassed half a trillion dollars—44% more than the entire record year of 2021—with more than 80% flowing to AI. Later-stage prices rose fast as capital concentrated in perceived winners, with Series C median pre-money valuations up over 60% from last year and doubled at Series D and beyond. Check sizes rose too, with median Series A rounds growing by more than a quarter. Seed was the only stage where the typical check shrank. In Q3 alone, a record 27 rounds of $1 billion or more absorbed a third of all venture dollars globally.
Behind those valuations sits a genuine acceleration of fundamentals. Zoom took seven years to reach $100 million in revenue, Snowflake six and a half, and Salesforce five. Our own companies are getting there in two years or less: Ramp in about two years from launch, Cognition in under two years from founding (including Windsurf), and Together AI in about a year and a half from launching its neocloud and AI infrastructure services. Cognition then nearly doubled its run-rate revenue in four months, to roughly $900 million. When the clock runs faster, the cost of being late compounds as quickly as the revenue does. Errors of omission (and the burden of regret) are more costly than errors of commission, and we’ve had our share of meaningful misses.
Our checks have grown as our companies have. Our seed checks remain consistent between $250,000 to $1 million, because a first check is still an option on a conjecture. In core investments, though, our median first check has risen from roughly $2 million a decade ago to over $10 million today. Meanwhile, the median post-money valuation at our entry has gone from $85 million to over $250 million. Our checks are getting bigger because the clocks are getting faster. The test is whether the business has advanced faster than the price. Across portfolio companies reporting revenue, median year-over-year growth was 76%, and over a dozen Lux companies are above $100 million in annualized revenue. Of the roughly 25 Lux companies that raised in the third quarter, 24 appreciated at a median step up of 1.7x. Among them: Hadrian at nearly $8 billion, Cambridge Aerospace at $3.4 billion, Varda at $1.6 billion, EnduroSat past $1 billion and Cognition at $48 billion and growing.
Bigger checks demand stricter epistemology. A valuation is a conjecture on the future. Technology that works, customers who pay, and revenue that compounds are the tests. AI startups currently command median step-ups roughly half again higher than everyone else's, reason enough to both believe earlier and underwrite as skeptics even while believing before others understand.
We also observe that capital itself is concentrating like never before. Established firms took nearly 90% of the dollars raised by U.S. venture funds this year, and one firm raised more than 20% of the total. As we’ve repeatedly hypothesized and now observably verified in our view of the Minnows and the Megas: the number of emerging managers able to close a fund collapsed from roughly 930 in 2022 to about 200 this year. The Minnows and Megas look less like a spectrum than a speciation and extinction event.
The West’s Global Instrument
When we write "defense of the West," we’ve never meant a hemisphere. As we put it last quarter, we mean the societies that run the error-correcting operating system of Popper, Kuhn and Planck: Seoul, Warsaw, Tel Aviv and Tokyo as much as Washington and London. The West isn’t so much a place as a process, one that builds strong companies and strong nations alike. Our portfolio increasingly reflects that: over 40 active companies are headquartered outside the United States across 17 countries on five continents. Europe and the United Kingdom are the largest cluster, with our reach extending through Israel, Japan, the United Arab Emirates and even Nigeria.
The allied rearmament cycle we called real, structural and mispriced last quarter is now visibly underway. America’s 80-year role as guarantor of world security has changed as the current administration pulls back, prompting others to step up. Saudi Arabia, Turkey, Pakistan, Qatar and Indonesia are seeking to develop weapon systems without Western strings attached, while South Korea, Japan, Singapore and others are all increasing their investment in a future of autonomous warfare. The upshot is that when the U.S. steps back, every ally also steps up to be a customer. NATO members are spending more than $1.8 trillion this year, up 11%, with Lithuania leading at over 5% of GDP. Europe’s €150 billion joint defense loan program is oversubscribed, and Japan’s leadership just requested a record defense budget for the fifteenth straight year.
Our companies are building inside this historical cycle. In Britain, that includes Cambridge Aerospace as well as Agon, which is building synthetic battlefields to train AI for European defense and whose seed round we co-led. In Europe, we’ve funded EnduroSat in Bulgaria as well as Onodrim, which is rebuilding a continental defense manufacturing base from the Netherlands and Lithuania. In Israel, Kela connects command-and-control systems across domains, Covenant builds long-range, low-cost defensive interceptors, and Tenzai’s AI agents hack their own customers before adversaries can. Several new AI cyber security investments with elite teams will be announced soon. Terra Industries, in Nigeria, is building autonomous drones and towers to guard critical infrastructure, with a drone factory rising in Ghana. And 1001, founded in London and Dubai, is applying AI to Persian Gulf aviation and ports, in a round we led with Saudi partners. Saildrone opened its European headquarters in Copenhagen, extended its Navy counter-narcotics work in August, won an Atlantic seafloor-mapping contract, and keeps persistent eyes on oceans that are suddenly contested again. Finally, in a perfect expression of cross-border Western error correction, two Lux companies, Revenant (founded by elite U.S. special operations veterans) and PDW, announced last month that they'll team up to bring Ukrainian counter-drone and electronic-warfare technology into production in Huntsville, Alabama. None of these companies were chosen because of their flag, but rather because their founders iterate experiments faster than the incumbents, and the incumbents are everywhere.
While we wish it were otherwise, conflict is reality. The war with Iran that began (kinetically) in February sent Brent crude from roughly $70 at the start of the quarter to more than $130 at its peak. Attacks on tankers recently reached a wartime high and, at time of writing, rumors and reports of renewed strike plans loiter around the U.S. midterms. Iran is striking the Gulf’s past, present and future at once: respectively, the oil and gas that built it, the airports and hotels and events it lives on now, and the connectivity and compute that it’s betting the next economy on. Because Iran is China’s gas station, every barrel the war takes off the market is also a tax on Beijing and a small check on its Taiwan ambitions. In Ukraine, talk of local ceasefires coexists with the most innovative drone war in history as well as experimental proof of subterranean warfare.
Unearthing underground was a hypothesis we first heard from Palmer Luckey (founder of multiple Lux companies including Anduril, Erebor, ModRetro, Warhound and more). It was made manifest by Traysar and its growing set of cutting-edge subterranean technologies. Meanwhile, the Pentagon is moving from boutique to bulk: its drone-dominance plan demands some 30,000 one-way attack drones costing around $5,000 each, with an ultimate target of $2,000. Mass on target has become a mathematical reality and requirement—and as we’ve written before, mass like math is truth.
The other truth is that hard power is only effective with the soft touch of careful operators and investors. We and our partner and veteran White House Mideast diplomat Brett McGurk have long argued that sci-tech diplomacy is the new frontier of statecraft, and we witnessed a vivid demonstration this summer: Washington eased export rules on advanced chips for the UAE, proving semis are the new alliance currency. Lux venture partners including Brett, retired four-star general Tony Thomas and former Commissioner for NYC Emergency Management Zach Iscol, anchor an increasingly international practice spanning diplomacy, defense, intelligence and the civic infrastructure that determines whether a city withstands the shocks geopolitics sends downstream.
This quarter, we also welcomed three new partners. On our investment team, we added Sarah Hess, who helped launch the NATO Innovation Fund and worked on U.S.-China technology competition at the CIA, as well as Mia Krishnamurthy, who joined us from CRV and General Atlantic. Meanwhile, Mahi Joshi joined our capital formation team from Evercore, working alongside Emily Haight (from Coatue) and Adam Kalish.
Our Riskgaming program applies the experimental method to policy itself. Riskgames are controlled experiments for decisions too dangerous to test in reality. Simulation is to statecraft what Applied Intuition is to autonomy: a chance to encounter failure before paying its price. In our scenario Dead Reckoning, grad students played best; London’s civil servants and military officers played worse, driving a majority of the game’s startups into the ground (a quite accurate if unpardonable simulation of procurement).
As we wrote last quarter, standards are destiny and defaults compound. The open bazaar of models, the inference platforms that serve them and the neoclouds that rent GPUs by the hour are becoming sovereign infrastructure, tested by the cost of capital as well as political and physical threats. Our investments in this layer include Together AI, Modal, SkyPilot, Armada and Majestic Labs: more intelligence per watt, per dollar and per square foot, deployed where data and energy already are. The pressure on price comes from abroad, too. Cheap Chinese tokens are the newest export (or dumping) of Chinese overcapacity, after steel, solar panels and EVs. Beijing is running races Washington mostly isn’t, for diffusion, adoption, price and industrial application, betting that a frontier lead of a few months matters less than putting good-enough intelligence into the most factories. The inference layer is delivering more intelligence per dollar while the infrastructure beneath it demands more dollars and more debt.
We back the first trend while watching for any exposure to the second. Alliances written in treaties are increasingly written in supply chains. Anthropic’s IPO filing notes that its chips are made in Taiwan by one company and no one else while warning that blockade, cyberattack and the island’s power and water are risks to its business. Once those shares land in American retirement accounts, millions of households will own a stake in the Taiwan Strait whether they know it or not. Beijing’s leverage always rested on the idea that Taiwan was too small to fight for. Geopolitically, the AI boom is making it too big to ignore.
Six Anomalies, or the Normal Science of Markets
The market has its own Kuhnian normal science, and for 15 years, its paradigm worked so well that most participants forgot it was one at all. Accepted axioms: rates fall when trouble arrives, bonds hedge stocks, debt is cheap, passive investing is free, index rules are fixed, the marginal buyer of Treasuries is patient, and the largest capex boom in history is indifferent to the price of money.
Three things are now moving at once: the need for capital is rising, the cost of it is increasing, and confidence in the return on it is falling. We are tracking six anomalies that separately or together could presage a change in the market regime. The first and largest anomaly: time itself got expensive. Rates are up. The 10-year Treasury yield rose 87 basis points in the third quarter—its worst quarter since 1994—and touched 5.36% in early October, its highest since 2002. A 30-year auction cleared at 5.62%, the highest since 2000. Notably, this wasn’t an inflation scare. The most persuasive explanation is that governments and AI companies spending at sovereign levels are pushing real rates up and squeezing more rate-sensitive sectors. This repricing is global: Japan's 30-year bond hit a record and its 10-year touched 3% for the first time since 1996; British 30-year gilts crossed 6% for the first time since 1998; and French spreads over German bonds hit their widest since 2011.
ZIRP wasn’t a law. The U.S. 10-year has averaged about 5.4% since the 1950s, roughly the economy’s own growth rate in dollars, which is where the price of money tends to settle. Borrow below it and debt is free money. Borrow above it and debt exceeds the income that pays it back. Unusually low rates came from a cascade of temporary accidents: Chinese trade and Gulf oil surpluses got recycled into Treasuries, boomers were reaching peak saving years, inflation lulled after 2008, and central banks bought up trillions of bonds through quantitative easing. By 2020, money was cheaper than at any point in 5,000 years of recorded interest rates. Today: every one of those forces is running in reverse, with retirees spending instead of saving, supply chains coming home, and governments and AI companies bidding against each other for the same dollars. What looked like the new normal was the anomaly, and what looks like the anomaly is the return to normal.
This quarter, the bond market went looking for the developed world’s weakest link and found it in Paris. Rioters torched the town hall in Limoges, where citizens marry, vote and honor their dead, while taking hammers to Louis Vuitton’s windows, where the glass held better than Macron’s polls did. The bond market voted before the voters could: France’s 10-year crossed 5% for the first time since 2002. The state takes in 52% of GDP and spends 57%, and it now borrows at 5% while growing the French economy at 0.8%, all in a currency it can’t print. In an aging democracy, the past outvotes the future. A state that pays for its past (pensions, social security) by starving its future (education, tech, defense and innovation) isn’t running a budget, but running down its global competitiveness.
Before we Americans feel smug: among major economies, only one will run a bigger deficit than France this year, and it’s ours. U.S. debt held by the public crossed 100% of GDP in September, a level last seen in 1946 to pay for a world war. Interest now costs more than our military. In 2001, Greenspan worried about what would happen to markets if America paid off its debt—that’s how fast a paradigm turns. The patient buyers are fading: Japan and China cut their Treasury holdings by roughly 10% and 16% this year, while holdings booked in the Caymans, home of the leveraged basis trade, keep rising. Last quarter, we called the bondholder the Pentagon’s true peer competitor. Increasingly, the risk is that the bondholder may be a hedge fund with a margin call.
The second anomaly: risky credit. High-yield spreads remain unreflective of true risk, at about 3.4 percentage points over Treasuries. But spreads on the lowest-rated CCC bonds widened from 9.7 to 13.4 percentage points during the quarter, wider than at the peak of 2025's tariff panic. Private credit, which has been absorbing risk, seems to be accumulating it. Its default rate hit a record above 6%—more than double the rate of public high yield—while maturity extensions, payment-in-kind amendments, and big funds gating redemptions for many quarters appear to be obscuring more losses. Stability breeds instability and opacity breeds it even faster.
The third anomaly: capex. The four largest hyperscalers will spend roughly $750 billion on capital expenditures this year. In the third quarter, Meta's capex exceeded its operating cash flow and it repurchased no shares. In a dramatic inversion, Meta plans to sell close to $100 billion of new stock, with chip suppliers Nvidia and AMD among the buyers. If free cash flow was the moat, equity issuance might now be the drawbridge dropping down. In Q2, Alphabet had its first negative free-cash-flow quarter in its history and also launched a sale of up to $85 billion of stock. At the time of writing, it just lost half a trillion dollars of value in a single day—its worst on record—after reporting a second straight quarter of negative free cash flow in Q3 and admitting that its AI answers earn significantly less than the search links they replace. Hyperscaler bond issuance passed $220 billion by late summer, more than double last year's. The debt-funded share of hyperscaler capex has risen from under 10% to about 33%, on top of roughly $820 billion in off-balance-sheet lease commitments.
All those debt guarantees deserve attention. Nvidia now guarantees as much as $105 billion of residual value for a single customer's data center campus. Blackstone and others announced a $63 billion structure (at higher rates) to lease Broadcom equipment to Anthropic, with Broadcom providing residual-value support for an amount equal to that financing. The vendor financing the customer is an old anomaly in finance. In 2000, when Lux was starting and the dot-com boom was ending, it was Lucent and Nortel lending to the carriers that bought their gear. When the carriers flopped, so did the receivables. Today, Broadcom is guaranteeing chip debt so a frontier lab can buy its chips, basically running two businesses: selling chips and underwriting credit. The market may be mispricing one of them. One hyperscaler stretched the assumed useful life of its data centers from 15 years to 25, flattering earnings with a keystroke without swapping a single server. Dubious depreciation schedules and accounting adjustments are not the things we see at market bottoms.
Follow the debt to its final resting place, and you won’t find a bank, but rather insurance companies and retirees. Less than half of the hundreds of billions of debt is held by the AAA hyperscalers. The rest is the long, complex, BBB-ish paper behind levered neoclouds, data-center SPVs and leases. The risky cascade is that the natural owner is the life insurer, whose liabilities run as long as a data center lease. More and more of these insurers are owned by the same private equity firms that make the loans. It works like this: sell annuities to retirees, invest the proceeds in your sponsor’s private credit, get it rated BBB by some agency and reinsure it in Bermuda. We’ve seen the film before. In the 1980s, the biggest buyer of Drexel’s junk was Executive Life. When junk broke, California seized the insurer in 1991, and retirees who thought they owned the safest product in finance were left holding the bag. Insurance is supposed to be boring, but when boring becomes levered, pay attention. One concern: AI hits loans the insurer already owns, having underwritten existing software businesses from buyout funds (written at zero rates but which will start refinancing at higher rates next year). PE built a wall of debt in 2021, and it is now running into it with roughly $300 billion of buyout loans due in 2028 and 2029, against about $60 billion a year previously. A third is software, and a quarter is already rated CCC-plus or worse. Only 5% of the software debt due in 2028 has been refinanced, against 42% in other industries. Debt doesn’t matter until it’s due to mature.
Importantly, none of this requires AI to fail. GPU rents have climbed more than 40% this year, which might make GPU-backed loans look safer than they are. What if GPU prices shocked and fell while rates refused to? We’d see debt priced for an appreciating asset become a more expensive loan against a depreciating one. An AI bust would likely look more like dot-com than housing: investors lose their capital, and the world gains the infrastructure and cheaper intelligence. The detritus becomes combinatorial fodder for the next wave.
The fourth anomaly lives in market structure: concentration, composition and share counts. The ten largest companies make up roughly 40% of the S&P 500, a single chipmaker is worth more than $5 trillion, and more than half of U.S. fund assets are passively managed. Passive flows aren’t price setters, but price takers. Roughly two-thirds of next year’s expected earnings growth come from chip stocks, a hugely concentrated bet. Meanwhile, some index providers rewrote their rules this year to admit SpaceX within days of its IPO. Even so, SpaceX now trades about 35% below its summer peak. At quarter-end, the S&P 500's forward earnings yield roughly equaled the 10-year Treasury yield, an equity risk premium of about zero, last seen around 2002, while the cyclically adjusted P/E sits near 41, around the December 1999 peak of 44. The question for equity valuations is what happens if debt dries up, capex slows, revenues fall and earnings are battered.
For 32 straight years, from 1994 through 2025, American companies retired more stock than they issued, about $8 trillion in all, while the number of listed companies was cut in half. A shrinking supply of shares chased by automatic retirement savings has been the quietest tailwind in finance making geniuses of tech bulls. In the first half of 2026, it reversed for the first time since 1993, led by the giants that used to be the biggest buyers of their own stock. SpaceX floated about 5% of itself, the same sliver Palm floated days before the 2000 peak. Those shares unlock next summer. As insiders are freed to sell, index weights rise and 401(k)s buy up while selling other assets, what happens? Manias used to end with a glut of the thing. Might this one end with a glut of shares?
The fifth anomaly runs through retail participation and pledges from margin. American households hold a record 48% of their financial assets in stocks, versus less than 39% at the peak of the dot-com bubble. Margin debt is near a record $1.5 trillion, up a quarter in a year. Same-day expiring options compose almost a third of all options volume. More leveraged ETFs launched by mid-August than in all of 2025. Beneath a calm VIX, pod-shops have grown from about $300 billion in 2020 to over $1 trillion, and they all crowd into the same stocks. When one misses a quarter, they all run for the same exit at once, so the index looks calm while single names drop 30%. As we go to print, history has begun to rhyme: hedge funds that had crowded into the Mag 7 unwound them violently into earnings. The Nasdaq 100 fell 3.7% in a day, and Intel—a tenth of which is owned by the U.S. government—fell 30%, its worst day since 1974. Underneath sits leverage most investors never see: some estimates have over $5 trillion of credit lines secured by investment portfolios, much of it borrowed against the same stocks that make up the index. When the collateral is the market, a selloff becomes a much wider margin call.
The sixth and final anomaly is slower but larger: the flows. Retirement plans are being opened to private assets, with a federal safe harbor proposed in March and some target-date funds already allocating up to 10% of their assets to private markets. This is happening right as venture secondary discounts are widening from roughly a quarter to about a third. As companies are staying private longer, the risk is retail savers become the bid for illiquidity at the very moment institutions are trying to sell it. More than four million Americans turn 65 every year through 2027, shifting from accumulation to distribution. A new tax on the largest university endowments begins this year, squeezing the very LPs who have long been venture's most patient capital.
None of this is a crash call. Channeling Kuhn: the old paradigm still explains most of what happens, but the anomalies are starting to pile up faster than the patches. Our long-held conjecture that the natural cost of capital is higher and rising has now survived its harshest test in 30 years. For venture, a higher cost of capital punishes narratives and rewards fundamentals and competitively-advantaged companies. That's why we're writing larger, higher-conviction checks into faster-compounding companies with discipline not exuberance, and why we ensure Lux has dry powder ready for the dislocations that crises always produce.
We’d put greater than 50% odds of the following: AI erodes the very cash engines that fund it. As internal cash shrinks, capex leans harder on debt. Then depreciation lands in 2027 and payroll pays it with layoffs. Then the revenue and ROIC tests come. Warsh and the Fed may look hawkish and talk tough on inflation, right up until they reverse swiftly and move fast on a credit crisis. The moment losses travel from equity holders to debt holders, expect rapid intervention. Warsh will get blamed for taking the punch bowl away, though the problem will have been who spiked it. A hawk on inflation becomes a dove on credit, and the bowl gets quietly refilled.
Lighting up the Anomalies
Lux friend and Columbia neuroscientist Stuart Firestein calls science the practice of uncertainty, and argues that uncertainty is possibility and possibility is optimistic. That is the temperament we look for in founders and aim to keep in ourselves: optimistic about what is possible, sober about what is priced. Intelligence is getting cheaper. Capital is getting dearer. The token is deflating, the watt is inflating, the coupon is rising, and a great deal of margin is caught in the middle. Overbuilds are how societies ultimately get infrastructure at a discount. The railroads bankrupted many but built the American economy anyway. The dark fiber of 1999 lit up the decade that followed. Society kept both; investors kept the losses.
When the cost of thinking collapses, the scarce thing becomes the ability to make instruments that open whole new questions for investigation and innovation. Surplus intelligence goes looking for atoms: materials, chemistry, biology, energy, defense and so much more. Matter that matters is where Lux has always been. Our job is to see and think clearly and distinguish what’s compounding intrinsically from what’s just merely being repriced extrinsically.
We named the firm for light, looking where others weren’t, long ago without knowing how literal it would become. As we look toward our next fund, Lux X, in the quarters ahead, we’re grateful for the trust and partnership that let us keep ambitiously pursuing our method: finding and funding the ambitious founders who are scientists upturning systems, and turning speculative hypotheses into observable returns. Our job is to keep earning the right to do that, again and again. Fiat Lux.