Lux Q2 2026 Report
Nullius In Verba
Ask a general what defends the West, and you’ll hear about wings and warheads, magazines and munitions, fabs and FLOPs. All true, all necessary, but arguably all downstream. What defends the West isn’t a weapons system so much as an operating system, one in continuous deployment for roughly 350 years. It’s been constantly patched—it’s occasionally crashed—and yet, it’s never been surpassed. That operating system is how free societies find and fix their errors before their errors find and finish them.
In 1660, a dozen natural philosophers gathered in London and founded what became the Royal Society. For a motto, they chose three subversive Latin words: Nullius in verba. Take nobody's word for it. Not the king's, not Aristotle's, not the bishop's. All claims earned standing not from the rank of whomever made them but by surviving the attempts of strangers to refute them.
Jonathan Rauch recently named what they’d built the “Constitution of Knowledge.” He reduced it to two rules. First is the fallibilist rule, in which no one gets the final say. Every claim stays open to challenge, and knowledge is whatever has survived these challenges, holding its title only so long as it keeps defending it. Second is the empirical rule: no one has personal authority, and a claim counts only if anyone, anywhere, can contest it. From those two rules flow the entire epistemic order of free societies such as peer review, adversarial courts, audited accounts, investigatory journalism, contested elections and where Lux lives and invests: priced markets. From vain, petty, jealous, tribal, confirmation-hungry and status-seeking primates we get an ever-growing stock of tested knowledge via institutionally organized skepticism. Remarkably, the system doesn’t even need honest people. It manufactures honesty as an emergent property, the way markets manufacture prices from greed. Virtue from vice, voilà.
Knowledge is the ultimate source of wealth and power, and civilizations compound knowledge at the rate they correct errors. The central strategic variable of this decade isn’t compute, energy, talent, or capital. It’s error-correction rate: the speed at which an institution, a firm, a fighting force, or a model can discover that it’s wrong and stops being wrong. Karl Popper understood this first, observing that science progresses through ruthless refutation of every bold conjecture. He also added a humane corollary: criticism lets our theories die in our stead, which is why confusing a person with their ideas isn’t merely unkind but anti-epistemic. One of Popper’s intellectual heirs, David Deutsch, carried the argument further. All progress flows from good explanations that are exposed to criticism and hard to vary. Myths are easy to vary: Persephone’s descent explains winter but so would her ascent, and neither can be dispelled by evidence. Newton, on the other hand, is hard to vary: change the exponent in the inverse-square law and orbits fly apart.
Everything Lux does is a bet on error correction. Science, markets, and liberal democracy are not three independent inventions, but rather three implementations of one OS running on different hardware. Science subjects conjectures to experiments. Markets subject pitch decks to buyers and prices. Democracy subjects platforms to elections. A hypothesis dies in a peer-reviewed journal, a thesis in a stock drawdown, and a policy at the ballot box. In each, error is signal not scandal, authority is provisional, and the system's genius lies not in choosing well but in un-choosing quickly and in making mistakes cheap, visible, and reversible. The value of democracy isn’t that the majority is wise, but that a bad government can be removed without bloodshed. From Nazi Germany and Imperial Japan to today’s IRGC, impervious intransigence to truth bombs reliably summons the kinetic kind.
This formidable operating system is now being stress-tested on three fronts at once, which we’ll describe in more detail in the rest of this letter: (1) In the labs (the computational ones and increasingly the wet ones), the most consequential technology of our lives is splitting into two architectures—open and closed—a choice less about licensing than about epistemology. (2) On battlefields and in shipping lanes, hard power is deciding whether societies that allow debate can deter (and dominate) societies that don't. (3) At home, the market that converts conflict and disagreement into economic knowledge is being increasingly doubted by the very public they serve.
Nowhere is that doubt deeper than among the young, who at 5.7%, carry a higher unemployment rate than the average American. The tasks AI does best are also the ones we used to hand new grads, tasks less valuable for their output than the feedback they offer young people to advance their careers. Someone not permitted to be wrong cheaply and privately eventually gets told by a politician, publicly and with a big following, that being wrong was never their fault. Thus the dangerous drift toward socialism, communism and anti-Americanism, with serious political consequences we’re just starting to see and we’ll bemoan being blind to.
There’s a simple test we use from our friend and Morgan Stanley investment strategist Michael Mauboussin to determine whether a result owes to luck or skill: can you fail on purpose? (Try it on a run and then at roulette). There’s an equally simple test for whether progress will persist: can you fail and be wrong in public? Investigators, investors and instigators—which is to say science, markets and democracies—all depend on that answer being yes. Everything we value rests on the right mix of naive ambition and certitude, balanced against the error-correcting virtues of fallibility and humility. The West’s operating system is venerable but tried-and-true. New challenges must strengthen it rather than tear it down.
Open and Shut: The Constitution of Compute
Nowhere has the debate over open and closed systems had more purchase than in the series of “moments” that have defined AI: the DeepSeek moment, the Mythos moment, the OpenAI–Hugging Face moment, and the Kimi moment. Their pace has hastened, their independence has become chained, and together they lead to the most important question in the future evolution of AI: do we want an open bazaar or a closed cathedral? Should the most consequential technology of our lifetimes be built in the open, its weights and workings published for all, or behind the sealed doors of a handful of frontier labs whose safety teams and shareholders alike counsel discretion?
We’ve never accepted the premise of the open-closed binary, since the dichotomy between open and closed is not about AI itself but whether AI stays inside an error-correcting system or escapes one. A closed lab with red teams, external audits, adversarial evaluation, and staff free to dissent and to resign loudly (and start competitors) that also publishes enough for outsiders to check its claims, is epistemically open whatever its license says. In contrast, an open-weights model (almost never open source) released into an ecosystem with no norms of testing, institutions of review, liability or criticism is epistemically closed whatever its license says. Openness is less a legalism and more a social arrangement: where do claims get exposed to people motivated to refute them? A model is, in the Popperian sense, a conjecture, billions of parameters of compressed guesswork about language, code, chemistry and the world. Closed systems ask for trust while open systems ask for tests.
History is unsubtle about which arrangement wins at the frontier. Closed systems optimize while open systems discover. Closed systems exploit known territory well but go blind at the edge, because the edge is where maps are wrong and only critics redraw maps. The West’s advantage was never producing great first drafts but producing better second drafts, faster, with more people marking up the first. The USSR had brilliant mathematicians but closed epistemology. China's labs are formidable (whether via forging ahead or forgery), but operate inside an authoritarian state that has criminalized certain conclusions in advance. A mind, human or artificial, that’s forbidden to reach some answers can’t be fully trusted on any answer. That’s not a moral claim but an engineering one.
The warnings from the cathedral—frontier capabilities are dual-use; releasing weights is irreversible; there’s no recall button on a download; safety work benefits from staged releases—are serious arguments made in good faith. The strongest version is true: there are capabilities that should never ship as a free download, any more than the Manhattan Project should have published its schematics. The key concern is a model that designs proteins, since bits that become biology need no fab and no enrichment cascade. Yet, refusing to build these models or closing access isn’t realistic in the global bazaar, and so the only durable answer to a cheap attack is an even cheaper refutation. That’s the thesis behind Valthos, which Lux backed since inception to accelerate sequencing, identification and countermeasure design from months to days. Deterrence in biology will matter just as much as it did for Oppenheimer and his contemporaries.
Open and closed AI were never opposing bets for us, since we focus on the system, not the license. We invested in Hugging Face in 2019, and it has become part Royal Society, part Library of Alexandria for machine intelligence, the commons where models are posted, probed, benchmarked, forked, and falsified in public. Its leaderboards offer peer review at the speed of upload—and unlike the original Library—distributed, versioned, and harder to burn. Meanwhile, we have funded the closed as well, where secrecy and scale are the entire point, even as they open up new capabilities. We spun EvolutionaryScale out of Meta to build ESM-3, a frontier model that reads and writes the language of proteins, evolution compressed and then run in reverse. It’s the kind of proprietary, hard-won science whose value lies precisely in being closely held but widely available to the scientific community. The market agreed: the Chan Zuckerberg Initiative acquired it. Osmo (which we spun out of Google) is not only digitizing and tokenizing smell, it is developing the repository of all smells, a valuable proprietary dataset.
The model isn’t the moat, the data is. The richest untapped reserves of value in AI are the proprietary, longitudinal, temporal silos that exist inside one institution and nowhere else on Earth: the pharma company’s unpublished assays and failed trials; the hospital’s patient records; the defense agency’s classified imagery. For these owners of secrets, the winning architecture is the inverse of the consumer default: an open model, very nearly as performant as the frontier, fine-tuned on data no one else possesses, running on local or sovereign hardware they alone control.
Lux has funded every layer of that stack. Databricks (with MosaicML) holds data where it already lives and lets companies train on it without it leaving. Together serves the world’s best open weights as a cloud at record inference speeds. Modal turns raw GPUs into compute you summon in a line of code and pay for by the second. SkyPilot, born open-source in Berkeley’s Sky lab, is air-traffic control for AI that efficiently routes tasks to the cloud, region, or cluster with the cheapest silicon, even hardware you own. Armada deploys ruggedized, Starlink-connected data centers onto oil fields, borders and battlefields, delivering compute to where energy and data already are rather than the reverse. Majestic Labs rebuilds the server around memory, packing roughly a thousand times more memory than a GPU into one box to break the key bottleneck of long-context, agentic inference. Meanwhile and in radically different ways, Flapping Airplanes and Flourish are attacking the deepest constraint of all, data efficiency, by teaching models to wring frontier intelligence from a fraction of the data. That’s exactly the trick a bank or a hospital needs when its proprietary trove is deep but narrow, or a robot needs when it can’t quantize a data center on its back.
Then there’s the geopolitical layer, which matters because standards are destiny and defaults compound: the rail gauge you adopt, the keyboard layout you inherit, the substrate the world's developers build upon. When labs in Beijing hand the developing world capable free models, it’s not generosity but seeding and embedding their defaults into billions of downstream applications. If the West's best models retreat behind expensive APIs while China's run in every server room in Jakarta, Lagos and São Paulo on Chinese clouds, we’ll have ceded the substrate, and with it the values compiled into it. The answer is openness that is strategic, not naive. In Tokyo, Sakana AI (co-founded by an author of the original paper that invented transformers) turns evolution itself into an R&D department, merging and mutating open models and keeping what survives selection. Moats built on closed systems erode, while moats built on openness can compound.
Hard power, soft rules
Closed societies suffer a disease for which there’s no autocratic cure: power flows up, but truth doesn't. When truth can’t speak to power, power lies to itself and then lives inside the ignorance it manufactured. Chernobyl's reactor exploded and the system's reflex was secrecy, at the cost of its own citizens' thyroids and its own legitimacy. Open societies run the opposite way. We look chaotic because our errors are loud—but loud errors get fixed.
Truth wins arguments but it doesn’t, unassisted, win wars. You can’t peer-review a torpedo already in the water. Washington, in his first annual message to Congress said, “to be prepared for war is one of the most effectual means of preserving peace.” When we use the phrase "defense of the West," it’s not a hemisphere, an ethnicity, or a nostalgia that we’re describing but the set of societies (Seoul, Warsaw, Tel Aviv and Tokyo as much as Washington and London) that run the error-correcting operating system.
The post-Cold War peace ran on four assumptions: that America could impose outcomes effortlessly; escalate cheaply and win quickly; keep sea lanes open; and make resistance futile. The world benefited from and borrowed against all four. Allies underspent on defense, investors discounted geopolitical risk, and supply chains got stretched thin across oceans. Then globalization commoditized precision. The very same trillion-dollar chip ecosystem that put supercomputers in every pocket put precision munitions within reach of every militia: Shahed drones aren't remarkable for their sophistication but for their cheap cost. American security has been underpriced and underappreciated, and now we face more crises, costlier commerce, and contested chokepoints.
Add to this an underdiscussed financial risk. The United States pays more in interest (roughly $1.1 trillion annualized) than it does on defense. The Pentagon’s true peer competitor for the marginal federal dollar is not China’s PLA, but the bondholder. Interest is the perfect rival program: it requires no authorization, survives every continuing resolution, can never be sequestered, and compounds while Congress sleeps. And the creditor across the table has changed character. The patient, price-blind central banks that once held 67% of our foreign debt have receded to roughly 40%, replaced by leveraged funds who hold Treasuries only so long as the trade works. The 10-year yield has arguably become a national-security index, with Treasury auctions a readiness review. Will rates rise? A shocker: perhaps taking both the coupon cost servicing our rising debt and the stock market as Trump’s preferred barometer, the consensus of rate hikes may yet surprise to the contrary.
Modern war is the most brutal lab on Earth, with live lessons from Ukraine, the Black Sea, the Red Sea, and Hormuz. Magazines matter more than marvels: $20K drones defeat $20M armor. Iteration speed is a weapon: drone, countermeasure, counter-countermeasure, cycling first in weeks and then days. The side that error-corrects faster fields the future first and takes territory and morale.
One Ukraine lesson restored what we spent 80 years unlearning: supply is the weapon. We don’t just aim projectiles, we aim production lines. Today’s wars consume in days what factories yield in a year. We face shortages in shipyards, magazine depth, precision components and drones at a scale measured not in demos of dozens but in millions. Wisdom not new, merely forgotten: the 1915 shell crisis ended the last purely Liberal government in British history, and World War II was won as much at Ford’s Willow Run production plant as at Midway. Deterrence is measured in depth.
Our adversaries understand what they’re playing, and it’s the game theory of Colonel Blotto: spread the giant across every front—from Ukraine, the Red Sea, the Strait, and the Arctic straight into the depths of the ocean and into the heavens of space. The side that must defend everywhere risks losing to the side that may choose anywhere. The deepest constraint on American power isn’t appeasing adversaries but appeasing markets. A democracy must placate its bondholders as well as its voters and their wages before it can wage anything: every basis point conceded to deterrence is levied on a mortgage or a grocery bill. We’re left with reserve currency defending reserve stockpiles, with risk of neither replenishing.
There’s an investable thesis here: abundance. Flip the ratios: cheap electrons, cheap mass, cheap money, and cheap deterrence. We’ve long been investors in rebuilding the arsenal of democracy, without apology then and with more urgency now. Anduril puts autonomy and affordable mass back into deterrence. Saildrone gives the free world persistent eyes across oceans that are suddenly contested again. Hadrian builds software-defined factories that make precision parts at the tempo the moment demands. We funded them when defense was a moral no-fly zone for most of venture, when the consensus held that technologists building for the military were somehow betraying technology's utopian promise. We believed the opposite: the surest way to betray technology's promise is to let the societies that suppress error—our adversaries—inherit the commanding heights of it. The founders in this corner of the portfolio tend to be people with something to prove: immigrants and outsiders defending the system that let them rise. Chips on shoulders put chips in pockets, and yes, sometimes chips in fighters too.
Defense abundance isn’t limited to America. We’re increasingly convinced that the allied European rearmament cycle is real, structural, and mispriced by American investors trained by 20 years of disappointment to disbelieve it. It’s likely a decade-long capital formation and deployment opportunity. Cambridge Aerospace builds interceptor drones cheap enough to fight profitably against the ones being fired at us. Kela builds the connective software that lets commercial and military systems plug into one another at the speed of fight rather than a program office. Covenant brings that same impatience to long-range targeted autonomous mission systems. Onodrim, Traysar, Tenzai, and EnduroSat carry the thesis across Europe’s industrial base and Israel’s software depth.
Here’s the epistemic truth: militaries are knowledge systems and they fail like them. The deepest advantage of free-world defense innovation isn’t budget but a lieutenant who can tell a general he’s wrong, a startup that tells a prime it’s obsolete, and a test range that tells a program office its missiles don’t work before the enemy does. The adversary's engineers may be equally brilliant, but their systems are often not optimized to find out. Every closed system eventually fights the war its leader imagined rather than the one that exists. The invasion of Ukraine began with intelligence that told Putin what he wanted to hear (they’d fold in days) because in a system where the messenger is shot, the message comes pre-packaged. The result sat thwarted and burning on the roads to Kyiv. Meanwhile, Ukraine’s drone war iterates on cycles measured in weeks: a jammer appears, a frequency-hopping counter is invented, a fiber-optic tether answers the counter, an interceptor drone answers the tether. A learning military beats a lying one even at severe disadvantages of mass.
Our adversaries know this, which is why another key battlefront runs through all the machines we hold in our hands. The disinformation campaigns targeting open societies aim at exhaustion, with persuasion merely upside: flood the zone with lies and seeking truth seems futile. A citizenry that stops checking gets disarmed just the same as the one that stops manufacturing. Generative AI has added ammunition asymmetrically: synthetic text, voice, and video drop the cost of counterfeit conjectures to near zero while verification stays stubbornly and expensively human. This rhymes with the drone problem: lies launch for cents while refutations incur institutional expense. Verification will have to be industrialized the way interception is, with provenance, attribution, and evaluation made cheap, fast, and ubiquitous until the exchange ratio, improbable as it sounds, flips and favors the truth again.
Every major advantage America has had in war—radar, cryptography, the transistor, GPS and the internet—emerged from open societies running open science under wartime pressure, and each was unavailable to adversaries optimized for obedience. The West’s interceptor problem, cheap offense against expensive defense, is at root a problem of prediction under uncertainty. Detect better and you waste fewer interceptors on decoys; predict better and smaller, slower, and cheaper effectors can arrive on time. Do both and you have refuted the enemy’s main conjecture that volume will saturate you. The side that learns fastest fires last and least.
The Market as Truth Machine
Despite the young’s growing antipathy toward capitalism, the miracle of markets is that they run without anyone's permission and correct without anyone's apology. A price is a conjecture, published continuously, and refutable by anyone with capital and conviction. The short seller is reviled and yet is the market’s most important peer reviewer, the professional skeptic paid to hunt for the errors everyone else is paid to ignore. A drawdown is just peer review with money attached. Bankruptcy (cruel at the scale of a life, merciful at the scale of a civilization) is how markets let our companies die in our stead, releasing talent and capital from refuted conjectures to fund new ones. Command economies under socialism fail not because their planners are stupid but because their prices are lies. Show us an economy where nothing is allowed to fail, and we’ll show you an economy where nothing is allowed to be true.
Apply this framing to the present moment. The AI buildout is the largest capital conjecture of our careers: hundreds of billions of dollars wagered on the hypothesis that intelligence scales with compute and that the returns to intelligence are effectively unbounded. Parts of that conjecture will be corroborated as the transformation is real, generational, and still underpriced over the long run. Parts will be refuted with the violence markets mete out for crowded certainty. Our discipline here is the theme of this letter: hold positions as hypotheses, not identities. Ask: what would it take to vary this? Fallibilism isn’t a lack of conviction, but rather conviction anticipating in advance what could change it.
Over half a trillion dollars has flowed into private AI since ChatGPT’s launch lit the spark, while the public companies most levered to AI have added $14 trillion of market cap (that’s with capex growing nearly three times against revenue growing 60%). Debt has entered the party, powered by incremental interest payments rather than incremental belief. SpaceX, an extraordinary company, had an extraordinary IPO with extraordinary inclusion in passive indexes and yet, has shed more than a trillion dollars of market cap from its post-IPO trading peak. The broader AI edifice built on high expectations may yet be shakier than its promoters will confess.
Users can win even when investors lose. Manias are a form of price discovery from the inside. Just as the railway bust ruined the speculators, it also left us all the rails. By adoption math, AI in 2026 sits roughly where the internet sat in 1998: over one-third of American adults use it and half the country is more concerned than excited. Fear is actually a necessary precondition for platform shifts that stick: nobody feared CB radios, while over half of internet users in the early 2000s claimed zero interest in broadband right before they got it and could never give it up. In 1998, the fortunes went to the discerning and the lucky, since 1999 still lay ahead, as did 2000. As the .com domain suffix proliferated then, now more than a million .ai domains have been registered, while around 7,000 AI companies have actually been funded (a number that will prove itself too high by at least an order of magnitude). Bursting bubbles don’t forfeit the future, but they do finance its infrastructure and trim the tourists. Lux will continue to believe before others understand and to underwrite as skeptics while we do.
Venture is the strangest and most beautiful embodiment of the Western operating system. We are professional conjecturers funding guesses at the outermost edge of the possible, structured so that refutation costs one times capital while corroboration returns a hundred. Optionality is fallibilism's financial expression. Even amidst failures, every surviving conjecture becomes the floor for the next (Stuart Kauffman’s adjacent possible), widening with each solved problem. When a company dies, the capital is spent but the information is banked—by us, by the founders, and by the entire ecosystem watching. The founders who compound are not the ones who are never wrong, but the ones with the shortest half-life between being wrong and knowing it. Failure comes from a failure to imagine failure. Our own errors (too numerous to detail here) include many of both commission (none fatally costly, several embarrassing) and omission (eternally burdened by the permanence of regret). The market may be a cruel teacher, but it also offers the best feedback to the fastest learners.
The Beginning of Infinity
Error correction requires error tolerance. Societies that stigmatize failure get fewer conjectures and thus less knowledge, ultimately leading to less wealth and less power. That’s one reason why Europe has fallen behind the United States: for generations, it’s been less tolerant of failure, and thus it has been less risk-taking and less entrepreneurial. America's most underrated export has never been a technology. It is, contra Fitzgerald, the second act. Where bankruptcy is survivable and heresy is legal, we get a frontier. Epistemic arbitrage is the gap between heresy and consensus, held with discipline until evidence converts belief into knowledge, and knowledge converts into realized returns.
Popper wrote that optimism is a duty and a moral obligation because the future is open and what we do decides it. Deutsch said problems are inevitable, problems are solvable, and their solutions breed new problems, which is good because that’s the very signature of progress. We hold both without strain. Inevitable: the wars are real, the manias are real, and the fragility of the epistemic commons is real. Doubt merchants, synthetic evidence, and tribal unreason gnaw at every institution that checks power with truth. Solvable: three subversive Latin words uttered in 1660 London became a civilization’s entire immune system. None of it was inevitable and all of it began as someone’s improbable guess. Across our portfolio (we say this with fallibilist confidence), some accident is almost surely being made right now that our grandchildren might mistake for destiny when it’s just randomness and optionality.
We remain rational optimists: long the error-correctors and short the error-deniers, long empirical facts over opinions, builders over talkers, and production over demos. We will keep funding science fiction on its way to science fact, including the instruments that see, the algorithms that think, the factories that make, and the shields that defend. Which leaves us the Royal Society’s motto and its final application and confident humility: nullius in verba, take nobody's word for it, including ours. You, as partners, are part of our error-correction machinery: questioning theses, probing reasoning, telling us where you think we’re fooling ourselves. Every refutation makes us less wrong, and less wrong compounded is how we all win. Fiat Lux.