IRR, Dollars of Alpha, or Dollars of Fees
In this issue: IRR, Dollars of Alpha, or Dollars of Fees—It's often possible to earn eye-popping returns on small amounts of capital, and many investors have shown that they can fundraise past the point where they can deploy more capital effectively. Which means that every asset manager is in one sense actively looking for the next opportunity and in another is passively harvesting alpha the market makes available to them. Intelligence—AI has gotten so important that its national security implications are in the hands of people who don't really get it. Nearly-Public—OpenAI is going to act more like a public company. Insiders—Entertainment is a great prediction market honeypot, since gamblers' interest is so high relative to its economic consequences, and there are so many people who get exposed to market-moving information. The Micropolitics of AI—AI polls badly in the abstract, but a tax cut polls well for the people it affects. Factors of Production—Agents make programming a different kind of exhausting. Talk to this post on Read.Haus . IRR, Dollars of Alpha, or Dollars of Fees
Last week was a big week for asset managers retrenching: several hedge funds jettisoning their long/short equity strategies ($, Diff ), and Sequoia wound down its London office, scout program, accelerator, and weekly partner meeting . In one sense, asset managers can expand indefinitely: the business is fragmented, and there's always something adjacent to what they're already doing: an early-stage investor is already thinking about later-stage investors, because the extent to which their portfolio grows into something appealing to those investors is an approximation of how well those portfolio companies will turn out and how diluted their earlier investors will be by the time that happens. Systematic investing strategies require data and infrastructure that a discretionary strategy can use, so that tends to be a natural outgrowth. It's very hard to launch a fund that's going to run many strategies at once, and relatively more straightforward to start with one strategy and hop to adjacent ones. All else being equal, new strategies increase a firm's risk-adjusted returns by adding a stream of uncorrelated profits, and also increase its capacity.
But that's the firm-level perspective. The outside world doesn't care at all if a merger arbitrage expert dreams of some day expanding into distressed debt, or if a biotech investor would love to do something in robotics. When they try to expand, they're generalists competing with specialists; they need some mix of a unique insight into the new strategy, a unique one-way complementarity (i.e. some reason to think that doing X and Y is easier if you start with X and then expand to Y), or the ability to time the market in investment talent. [1] Absent that, the first payoff from diversification is higher overhead, more distraction, and lower average returns. The good part comes later, if it comes at all. And adjacency doesn't save you, either; sometimes, there's just a big leap from one strategy to another. When Citadel entered the discretionary long/short business, people were skeptical , but it's worked out well. But some of the high-profile closures this week followed exactly that pattern, where a quant shop diversified into discretionary long/short without having much experience. It's not clear whether the lesson is "this approach works if you do it early when everyone thinks you're crazy" or "this approach works if you're Ken Griffin."
So asset managers tend to be stuck between the imperative to grow and the need to specialize. Or, in terms of payoff, the choice between unbounded growth with more risk of total failure or a relatively modest business that's also subject to forces beyond its control: if you happened to be good at something adjacent to SPAC arbitrage, 2020-21 was a windfall; if you specialized in picking small-cap value stocks, the period since the financial crisis has been one where you lagged during rallies but participated in market declines much like everyone else.
Multi-manager hedge funds and multi-stage venture-plus-holding-past-IPO are two models that try to take something simple and scale up what works. A venture investor who starts out early-stage has a natural way to extend their strategy: just keep backing the same companies as they raise later rounds. And the more that venture investor is truly value-added for the world, the more likely it is that the companies they back will still have good returns well after IPO. [2] Overall GDP growth sets an implicit ceiling on the number and size of growth companies: there's about $4tr of incremental GDP growth to go around, which sounds like a lot until you remember that eight billion people are fighting for their piece of it. In that model, if there's a company that can keep expanding its topline at a mid-teens rate while maintaining margins and being able to distribute capital to shareholders, it's a pretty special business, and someone who has unique knowledge of the company and a cost basis that rounds to zero should delay selling for as long as they can get away with it. So the natural extension is for an early-stage investor to become a multi-stage investor, or, if they don't keep refreshing the portfolio, to evolve into a concentrated long-only manager with a focus on large-cap growth.
As venture funds scale, it gets harder for them to keep going after the lower end of the market. Their most successful investors are also the ones who have the most responsibility with current portfolio companies. There's also a tendency for investors to thoroughly figure out one paradigm, and then get wiped out when the next one shows up. Venture is a heavily network-driven business: at any given time, there are lots of obscure but promising ideas, and any one person won't be personally aware of most of them. But if that person has friends who are, and positions themselves to be a very useful person to pitch ideas to, they'll pick up on them. This tends to make VCs specialize, sometimes in opaque ways (if you were very good at assessing growth hacks on mobile, the 2010s were a wonderful time, but every platform's incentive is to make ads the best growth hack). In some ways, this means that VCs follow a definite lifecycle: you might spend your thirties underwriting the next triple-triple-double-double in SaaS, and your forties figuring out leveraged recapitalizations of those same companies as revenue growth slows to single-digits but EBITDA margins stay high. To stay multi-stage, a firm needs to keep hiring new people with fewer preconceptions, and basically continuously rotate through a series of temporarily-useful preconceptions: there was a time when "founder-friendly" meant low liquidation preferences and taking less than half of the equity in the first round; there was a more recent time when the most founder-friendly thing you could do was tell someone that they should wait six months and raise at a lower valuation rather than trying to grow into an impossibly generous number. To bring on those people, VCs have to be able to pick good companies and good themselves--twenty-years-ago, but different.
On the multi-manager side, the evolution is more deliberate: these companies are in the business of manufacturing and selling alpha, and they often start out with a single strategy, find ways to subdivide it into more strategies, and ultimately end up doing basically everything that a) requires a mix of talent, capital, and infrastructure as an input, and b) can be done in a way that doesn't systematically correlate with some return stream that's available more cheaply. Adding new strategies helps their topline, by diversifying their returns and increasing their capacity to raise fee-paying assets. It also helps to solve another problem they face: returns are skewed, and people have egos, so a fund that's in the business of nurturing great investment talent is also providing seed capital and a track record to future competitors. Whoever is the most profitable manager at a given fund, every year they're more experienced and better-capitalized to start their own competitor. And there's just something appealing about having your name on the door, or at least having the door feature a novel combination of words you'd find in physics textbooks, Herodotus, or brochures advertising a new subdivision.
The upside case for the hedge funds is that if someone can generate $100m of alpha in a typical year, Citadel can pay them more for it than going directly to limited partners would, and that the going-directly-to-LPs strategy involves a lot of repeating yourself in conference rooms rather than doing whatever it is that you do to make all of that money, and that other details—which meetings you get into, the price impact of being 20% of a stock's volume for a day, how much you pay to borrow when you short and whether or not you get to stay short when supply is tight, etc., are all better back at the mothership. On the other hand, maybe your $100m/year strategy pays off for long enough that you can find an adjacent strategy and build a mothership of your own.
Meanwhile, back in the world of excess return-seekers, different dynamics apply. The big pod shops are a pretty easy decision for a typical allocator: get a stream of returns that's mostly uncorrelated to the broader market (a little more correlated on bad days, but not so much in bad years—2022 was awful for growth stocks, and thus for the funds that loaded up on them, but, for related reasons, it was a great year for commodities). On the other hand, they're bad for allocators' job security: the market-clearing wage for "just put as much as you can into the following five funds" is pretty low. And those funds are, partly by design, hard to underwrite: there are constantly-evolving cross-subsidies, and competitors are always trying to gather intelligence, so there's a bigger incentive for funds to keep secrets (or share them retrospectively) than for allocators to dig for them.
The asset management business is ultimately defined by the relative contribution of capital or everything else. In the everything-else bucket, running passive strategies, or index-tracking strategies that try to be more transaction cost- or tax-efficient, can work. But it's a business where companies compete on price and the model involves multiplying trillions of dollars by basis points of fees. At the other end, there's an ROI at which the market for asset-gathering simply doesn't clear: all the people with a credible story about how they can repeatedly earn 500% annualized on a million dollars already have the million dollars. [3] In the middle, there are strategies where fixed costs are high, variable compensation costs are higher, but there really is a way for capital to add value. But the clock is always ticking for outside investors: unless the pool of available alpha grows as fast as assets under management, the fund grows to the point that it becomes a fight to divvy up a fixed reward between passive outside investors and the firm itself, with the firm's manager getting a fee-free return plus a cut of LPs' upside. The best investors are in a continuous process of buying out their outside investors to keep more for themselves, and the world doesn't produce enough available alpha to match everyone's ambitions. Alpha is always an error term: it's what you get when you run a regression that predicts returns and track what it doesn't cover. Which means that alpha is always being either pushed to zero or packaged as a cheap product; if you can describe it systematically, you're describing how to systematize it out of existence. So the firms that grow do so until they've reached the limit of their own internal complexity and the aggregate size of their competitors' mistakes.
There are some strategies that are feast-or-famine—sometimes, there isn't a lot of distressed debt, sometimes there is, sometimes it's concentrated in one industry, etc.; strategies like index rebalancing go through booms and busts; they work when index funds and quasi-indexers outbuy and outsell the people running the strategy, but if index funds need to buy 10m shares of something, and the people running the index rebalancing trade have 12m to sell them, the result is not fun. But the next rebalancing, when there are fewer participants and they're feeling a little more shy, might be better. ↩︎
Perhaps one of the most fundamental beliefs underpinning venture capital is that technology creates net new economic value. And that financing its development can create incremental opportunities for investment at magnitudes that are very hard to anticipate: there’s just no way early backers of OpenAI could have anticipated that one day they’d raise $122B of equity capital in a single round. ↩︎
Maybe not completely true, but if someone is pitching you on such an investment, your first pass should be to ctrl-F for words like "prison," "rehab," or "nasty divorce." But even this demonstrates that they have some deficiencies in risk management. ↩︎
You're on the free list for The Diff . Last week, paying subscribers read about how the deficit will matter more soon ($), and how this will affect which politicians have an easier time at the national level; voters mostly cared about things other than datacenters, and that debate is more local than national ($); and what a pod shop unwind looks like ($). Upgrade today for full access.
Upgrade Today
Diff Jobs Companies in the  Diff  network are actively looking for talent. See a sampling of current open roles below: A leading AI transformation & PE investment firm (think private equity meets Palantir) that’s been focused on investing in and transforming businesses with AI long before ChatGPT (100+ successful portfolio company AI transformations since 2019) is hiring experienced private equity investors to help lead leveraged buyout co-investments in healthcare services, financial services, and business services. If you’re excited about underwriting AI value creation and have experience leading traditional LBOs (financial,  commercial diligence, etc.) please reach out. (Remote) High-growth startup building dev tools for wrangling and debugging complex codebases is looking for someone who can personally execute the SaaS bear case: review the third-party software they use and figure out what to keep, what to drop, and what to implement in-house. (SF, DC) Series A startup that powers 2 of the 3 frontier labs’ coding agents with the highest quality SFT and RLVR data pipelines is looking for growth/ops folks to help customers improve the underlying intelligence and usefulness of their models by scaling data quality and quantity. If you read axRiv, but also love playing strategy games, this one is for you. (SF) Ex-Bridgewater, Worldcoin founders using LLMs to generate investment signals, systematize fundamental analysis, and power the superintelligence for investing are looking for machine learning and full-stack software engineers (Typescript/React + Python) who want to build highly-scalable infrastructure that enables previously impossible machine learning results. Experience with large scale data pipelines, applied machine learning, etc. preferred. If you’re a sharp generalist with strong technical skills, please reach out (Remote) Ex-Citadel/D.E. Shaw team building AI-native infrastructure that turns lots of insurance data—structured and unstructured—into decision-grade plumbing that helps casualty risk and insurance liabilities move is looking for forward deployed data scientists to help clients optimize/underwrite/price their portfolios. Experience in consulting, banking, PE, etc. with a technical academic background (CS, Applied Math, Statistics) a plus. Traditional data scientists with a commercial bent also encouraged. (NYC) Even if you don't see an exact match for your skills and interests right now, we're happy to talk early so we can let you know if a good opportunity comes up. Find a Role If you’re at a company that's looking for talent, we should talk! Diff Jobs works with companies across fintech, hard tech, consumer software, enterprise software, and other areas—any company where finding unusually effective people is a top priority. Find Talent with Diff Jobs And: we're now actively deploying capital into early-stage companies through  Anomaly . Our focus is on defense, logistics, robotics, and energy. If you'd like to chat, please reach out. Contact Anomaly Elsewhere
Intelligence
Last week, Jensen Huang was accused of having an extramarital relationship with a Chinese spy, Megaspeed CEO Alice Huang. Nvidia has announced an investigation , and Huang has written an apology letter to employees . (Never one to miss an opportunity to promote his business, some of this letter appears to be AI-generated.) One odd detail on timing is that Marco Rubio informed Donald Trump of this earlier this year, but apparently neither of them told Huang . Around that time, a short seller, Culper Research, put out a tweetstorm about Nvidia's relationship with Megaspeed, and Megaspeed's relationship with Alibaba .
Just from the perspective of policymakers, there are gains from trade to be had: Xi Jinping takes AI literally, and wants China to have access to chips and to information about what the US is up to, while Trump doesn't take it very seriously except as a driver of the stock market. (Adding Nvidia to the Dow turns out to be a big deal—Trump likes to quote changes in the Dow , probably because it involves bigger numbers than the S&P.) But AI is also getting more important. So we're at a point where what to do with this intelligence is a Presidential-level decision, and Trump sees AI proliferation as less of a risk than a bad Nvidia quarter. This makes Nvidia a much harder company to analyze; you might pencil in a much higher cost of capital, or lower revenue, starting around January 2029.
Disclosure: long NVDA.
Nearly-Public
Anthropic will be public in a few weeks, barring an unfortunate sequence of bad headlines or a big divergence in investor sentiment. (If consensus is that Anthropic is worth $1.25tr, it's probably still worth doing an IPO and raising less than planned. But if public markets say $1tr and private ones say $1.5tr, there's a case for raising one more round and delaying the offering.) OpenAI would have liked to be the first lab to IPO, but the next best thing is to act more like a public company, so they're doing a quarterly earnings call as part of finishing a fundraise ($, The Information ). But that same kind of valuation bifurcation makes it hard for them to go public. They'd like a market where transactions generally set valuation records. That's easier privately, because when everyone's pessimistic, no one wants to have to change where a position is marked. Once they're public, they get a daily scorecard on how they're doing, whether they like it or not.
Insiders
Prediction markets continue their streak of red-teaming every possible situation where people are entrusted with more information than they should be: Hollywood studios are starting to ban employees from trading on information they learn on the job . Things like casting decisions get public attention vastly disproportionate to their economic impact, which makes them a great venue for insiders to monetize their knowledge. There's $2.3m riding on who's the next James Bond , compared to $1.0m on the next CEO of Apple . Sloppiness about secrets is proportionate to the risk and cost of their getting released. And prediction markets essentially guarantee the release of secrets relevant to the resolution of a market through someone betting on it. In one sense, this is the efficient market at work: the American public happens to be disproportionately interested in personnel choices in one specific industry, and, through prediction markets, their demand for it can be met by trading against them. But it's also redistributing some of a studio's intellectual property to whoever's unscrupulous enough to trade on it, so of course they need to shut that down.
The Micropolitics of AI
Datacenters were on the ballot in various place last week, and the Washington Post visited Harwood, ND, population 845, which was the only one to explicitly vote in favor (or, at least, to vote against a stricter review process). Their concerns are classic local politics: it adds jobs, and it generates tax revenue, so even if people are skeptical of AI, they're fine with the tradeoff. This is roughly the pattern to expect: small towns with low populations are a great place for a building that takes up space and doesn't need that many people to operate. In this case, the property taxes are about $2k per person per year, and there are a lot of places where $2k goes pretty far.
Factors of Production
Something I've noticed, too: vibecoding is more mentally tiring than regular programming . When you're writing code by hand, you can get into a flow state where, even if you don't know exactly what you're doing next, you can basically follow an implicit flowchart that's almost certain to get you to the output that you wanted. But describing that process is roughly how we prompt coding agents, who then get to keep the flow state to themselves. What we get instead is that we're waiting for one big chunk of code to get written, while coming up with the next one. So writing code with LLMs means putting yourself in a position to maximize decision fatigue, all the time. This might be one reason the forward-deployed engineer model is getting more popular: it's a way to maximize developer productivity by outsourcing some of the task of what to write next to someone else; it gives developers a flow chart again, just with human and agent interaction rather than toggling back and forth between an IDE, prototype, documentation, and Stack Overflow.