Alpha in the AI startup rat race
Every venture-backed AI company is now its own chaebol
After years of guessing whether the best AI investments would be in frontier labs, data centers, semis, high-bandwidth memory chips or data center REITs, an investor recently posited to me that concrete was where the real action is. After all, you can’t build fabs or compute cathedrals without it.
The data seems to back him up: the price of cement in the United States has swollen 46% since Covid, almost triple the growth rate of the previous decade. CRH, one of the largest cement producers in the world, is up 110% over the past five years, while Dan Sundheim at D1 made bank on building materials conglomerate James Hardie International.
We’re all trying to find alpha, but I don’t think a trip to World of Concrete is in order. Yes, VC is a total frenzy right now. According to Crunchbase News, roughly half a trillion dollars of venture capital was deployed in the first half of this year. That figure compares to $26.8 billion for the entire mobile-social-cloud-year-of-our-lord 2012, when I first entered VC.
With so much capital flooding the market and inflating even the first draft of a pitch deck to a billion-dollar-plus valuation, I understand the urge to find alternatives to tech companies (here’s one: premium kimchi refrigerators for all the SK Hynix workers loaded with bonuses).
Yet, even the most “picks and shovels”–pilled VC understands you don’t invest in the hot dog stand in front of the data center (after all, no one works there). Yes, you can potentially invest in a tungsten mine in Gangwon, South Korea, and you might even make money blasting rock. But the big returns that will make or break VC funds are going to come the way they always have: investing in the best technology companies.
What’s changed in this new era is the definition of “best.” In the 15 years between the Global Financial Crisis and Covid, the best startups featured a great core product that generated momentum toward monopolization. In social and marketplaces, it was network effects. In cloud and APIs, it was developer relations and platform lock-in. In applications, it was self-service and product-led growth that drove word-of-mouth. In consumer fintech, it was a viral loop in a primary beachhead with many cross-promotes to reduce marketing costs. In crypto, it was a mix of many of these plus the shameless pursuit of fraud (I only half joke).
AI is an entirely different beast, since the source of profits will constantly shift as users (and their agents!) seek the most efficient intelligence in a hyper-competitive marketplace. Sometimes profits will come from the frontier model, sometimes from the compute infrastructure, and sometimes from applications. Unlike the monopolization effects of yesteryear’s startup winners, none of the leading players in any layer of AI seems to have locked in its status. The closest is maybe Nvidia and its CUDA platform, but even here, a number of competitors are working hard to break its lead.
Why can’t AI companies become entrenched? Part of the answer is a function of their own products: Claude Code, OpenAI’s Codex and similar products can make shifting to a new API or platform nearly effortless. Rapid innovation of new techniques around model training ensures that the leading options are always churning. The lavish funding of semis has given chip designers better tools to go from conception to tapeout (Etched being the latest example). Even brand affinity seems less powerful than it once was. I thought OpenAI had the game won given the near-universal global awareness of ChatGPT, only to find that Anthropic has completely upended the race.
DeepSeek shows how different the best companies of this generation are going to operate.
The best companies of this generation will thus compete through rapid, multidimensional execution that constantly reinvents their business. These companies will be organized more like a hedge fund than a traditional tech company, less about monopolization and much more about constantly rebalancing a portfolio of products, selling off one category that’s underperforming in order to invest like hell into another with brighter prospects. If a company and its engineering team have capacity to work on only one core product at a time, that company will be subsumed by the improvisation of competitors who can readily chase wherever the money will come from next.
That’s very different from, say, Uber as it grew up. Its engineers honed product loops to drive its marketplace for taxis and later food delivery, building up lock-in via network effects that no AI company is going to have this generation. Here’s the easiest way to see the difference: if Uber launched no new features this year, would anyone care? Would its market share plummet? (In fact, given that all of their new features seem to be some form of enshittification, I as a user would probably encourage a company-wide engineering holiday). That’s absolutely not the case for leading AI companies.
I was thinking about this new model of startup while reading Gary Sernovitz’s profile in The New Yorker of Ken Griffin, CEO of Citadel, one of the most successful hedge funds of the last few decades. It was a boring profile, and it was clear that the author struggled to turn immense access and assiduous research into any sort of propulsive narrative. What’s the secret to Citadel’s success? Well, after twenty pages, it’s pretty much as trite as “survival of the fittest.” Sernovitz sums up:
Citadel has never publicly codified how it achieves “alpha”—outperforming the market. But my conversations with Griffin and twenty-eight of his current and former employees revealed that they had consistent tenets. The output of Citadel’s factory is alpha. Every aspect of the factory must be constantly improved, and to accomplish this everyone must forage for information. The purpose of every improvement is to minimize unknown risk so that Citadel can maximize intentional risk. The biggest risk of all is complacency. The objective is not just profit but victory over rivals.
Of course, that describes everyone in finance, and so the piece keeps searching for a better answer. But there isn’t one, since what’s unique about Citadel is precisely what’s quoted above. Griffin just keeps relentlessly reinventing year after year, decade after decade and apparently never gets bored or complacent.
That’s just not normal for mortal humans, which is one reason why Citadel has massive employee turnover. It reminds me of the hard-charging and adaptive cultures at Apple and Huawei I talked about two weeks ago. Like in finance, in the Hunger Games of an extremely competitive tech market, only the most adaptive and fastest executors are left standing at the end of the tournament.
Given how different “best” looks today, what should VCs be looking for in a company? Small, technical and highly competent teams that rapidly seek out new profit pools and exploit them before they are competed away. There’s a reason international olympiad winners are so prized right now by VCs and founders alike: their acumen plus aggressive competitiveness is the magic sauce. Who is out? The visionary product founder who is fastidiously obsessed with the details of the button interaction on their app’s signup page.
We see the outcome of this thesis in one of the biggest AI funding rounds of the past few weeks into DeepSeek, the company behind the most popular open-source model from China. DeepSeek was spun out of a high-performance hedge fund, High-Flyer, and carries with it finance’s competitive DNA. Despite its popularity, the company has only 300 employees today, compared to several thousand at Anthropic and upwards of 10,000 at OpenAI. Yet that small squad has produced one of the most competitive models in the world right now.
That model, DeepSeek V4 Pro, is crazily efficient, both in terms of its training (although that’s heavily disputed) and the cost of its inference. It’s an order of magnitude cheaper than Anthropic’s Claude Fable 5, with DeepSeek V4 Pro coming in at $0.87 per 1 million tokens out versus Claude Fable 5’s price of $50 on OpenRouter. Yes, Fable is tremendously capable and outperforms DeepSeek, but is it 57 times better? What percentage of queries demand better intelligence rather than cost savings? This is the competition that will constantly shift from week to week and year to year.
Anthropic’s launch of Mythos spooked DeepSeek, according to The Information, and triggered its $7 billion fundraise, the first institutional round in the company. The deal structure is highly unusual compared to Silicon Valley standards. Up to now, DeepSeek has been funded by its founder Liang Wenfeng from the profits of High-Flyer, which he also founded. Liang’s net wealth is unknown, but he must have done well: he is funding 40% of DeepSeek’s round himself. He’s also keeping complete control of the company by forcing VC investors to fund DeepSeek through a limited partnership with no voting or governance control.
With greater resources at hand, the company is now expanding beyond just its best-known open-source models. This week, Reuters revealed that the company is developing its own AI chips, and it has a multitude of other ambitions far outside its frontier lab reputation.
DeepSeek shows how different the best companies of this generation are going to operate. As VCs fund a different kind of company, they will need to think differently about portfolio construction. Using the previous era’s model, investors would select a bet in a chip company, an infrastructure platform, a harness, a frontier lab and on and on, putting checks in each of the boxes on the industry market map. This approach assumed every company did one thing, and one thing well, and every logo could be placed in exactly one spot on the map.
Great execution is the dumbest blandishment about what makes a great company, but it’s also the hardest goal to sustain forever.
That’s not how AI is shaking out. Even early-stage companies are taking on multiple product categories simultaneously that would be more expected of a Mag 7-scale company. Everyone is a model builder. Seemingly everyone is working on some sort of chip and infrastructure play. Every company is considering applications and how to own the full profit cycle of a user. Hell, a lot of companies seem to be buying nuclear power plants or investing in fusion. While an exaggeration, it’s almost as if every startup and big tech logo can be placed in every box, or at least, certainly more than one. In short, every venture-backed AI company is now its own chaebol.
VC portfolios today are less about placing a bet-per-box built upon a bunch of careful market analysis and much more about investing in the most competitive talent who can sustain their game for a decade or more. That’s why even eye-watering valuations can seem reasonable: if the right team is in charge, they will conquer the market and investors will ultimately be rewarded.
The magisterial and zany magazine profiles of yesteryear’s founders will (unfortunately) read more like the tedium of Griffin’s. Great execution is the dumbest blandishment about what makes a great company, but it’s also the hardest goal to sustain forever. There’s nothing concrete about success anymore.
Also check out
What would happen if China hacked the U.S. water supply.
What it’s like waiting for electricity in Monrovia.
Tick hunting and alpha-gal.
The Apple-Bananna paradox.
Why grocery-store ice cream is so bad.
The Searchlight Institute looks at public sentiment on U.S. national security and foreign policy.
America’s drone supply, mapped.
And why is there a Palantir chore coat?