Inside Thrive Capital: Investing in OpenAI, Wiz, Cursor, Nudge, Physical Intelligence

Summary

• Molly places Thrive’s OpenAI entry in the January 2023 round at $29 billion; Clark says Thrive was one of only two firms to give the company a term sheet — “an investment that required belief,” not a competitive edge. Clark saw GPT-4 in a private demo in November 2022 and says it passed the Turing test (“we don’t talk about it anymore, but that was a really important moment in technological history”); Molly’s reported valuation path is $29B → $86B → $157B → $500B, and the product now has more than a billion users as “the front door to AI.”
• Clark’s core research claim on OpenAI: “every single major paradigm in large models… has come from OpenAI, or at least originated at OpenAI,” to his knowledge — and paradigm generation, not just the current product, is the lab’s durable moat. He maps pretraining → post-training (supervised fine-tuning and RLHF) → reinforcement learning, and guesses “there’s going to be a couple more of these” — “we are basically figuring out in real time how to do the alchemy of turning pieces of metal into thinking machines.”
• Cursor went from tens of thousands of users and low-single-digit millions of ARR at Thrive’s May 2024 investment to “many, many hundreds of millions of run-rate revenue” and millions of developers in 18 months. Clark met founder Michael the day he pivoted away from AI-for-CAD, and cites Michael Dell’s framing: the internet was chess, “the AI era is like playing speed chess.”
• Wiz’s diligence signature was the very best deal-size-to-implementation-speed ratio of almost any company Clark had seen — healthy six-figure deals closing in weeks or months — plus 50% of users being developers, not security people, and only one customer having churned in all the time since the company started. Thrive flew into Tel Aviv in December 2023 or early January 2024 during the Israel–Gaza war to handshake the deal; Google announced the acquisition earlier this year, which Clark thought was the largest sum ever paid for a startup. The house philosophy: “the people who win deals are the people who want to win deals the most.”
• The hardware opportunity set is now far greater than at any time in the last 20 years: lidar fell from ~$75,000–80,000 per sensor in 2014 to a couple hundred dollars, and SpaceX/Anduril have trained a generation of scale-ready hardware founders. Some hardware businesses can show accelerated growth at scale — Anduril won a program of record within two years, “fastest timeline since the Korean War,” and Starlink now drives a very large majority of SpaceX’s free cash flow and revenue — while five of the ten most valuable companies are hardware companies (Apple, NVIDIA, Broadcom, Tesla, and TSMC).
• On the AI bubble, Clark’s honest answer is that concentration substitutes for market timing: “a lot of the rounds we participated in felt eminently rational to us,” and the Google entry-point analogy says variability can wash out if you hold enduring companies for decades. Venture is “the asset class with the worst beta and the best alpha,” the Magnificent 7 are 50%+ of Nasdaq market cap, and the worst investments are ones where one bad quarter or year changes your mind — that’s momentum, not fundamentals.
• On jobs, his bottom-up evidence cuts against the layoff narrative: he cannot name a single portfolio company that has laid off engineers because of AI tools — “much more of an augmenting technology than a substituting technology.” The next 12 months to watch: RL entering production software (still scarce beyond lab work, with some coding-company efforts only rumored), intelligent hardware addressing “well north of 50%” of global GDP, and AI turning science problems into engineering problems — Isomorphic Labs (Thrive-backed) and Periodic (not an investor).

Deep dive

1. From physics Substack to Thrive — via semiconductor reshoring

• Clark’s self-description: “a technologist and optimist” whose childhood heroes were Oppenheimer, Claude Shannon, and Elon Musk. He studied physics, concluded he “probably wasn’t quite good enough,” switched to CS, and saw two paths into “the technological engine of history”: start a generational company (without a great idea) or partner with them — hence Thrive.
• The origin story is serendipity-by-obsession: he maintained a semiconductor Substack, and Josh — exploring reshoring in 2021–22 — found it via a former Bridgewater colleague. The stats Clark cites: ~40% of semiconductor manufacturing was U.S.-based in the 1990s versus ~10% today. Reshoring turned out to be difficult and would require “a little bit more money than Thrive had available,” but Clark’s reaction was that Josh’s ambition was unique: “ask for advice, get a job.”
• No chip companies yet, but Thrive led the seed at the beginning of this year for Mesh Optical — data-center interconnect built by Travis and Cameron, former SpaceX Starlink photonics engineers whose work used lasers to pass data between satellites — riding the hundreds of billions of dollars spent each year on AI data-center capacity.

2. Cursor: back the person on pivot day

• Clark met Michael in late 2022 — the day he decided to abandon an AI tool for mechanical engineers using CAD. The whole meeting was a list of possible pivots; Clark’s one contribution was steering him away from customer support toward software. “With great founders, there’s this almost electric energy you can sense in the first meeting.”
• Michael and Aman merged with another company, bringing in co-founders Swala and Arvid, and started working on the first AI IDE. Clark tried the product early: “one of those magical one-way-door moments. Once you walk through a door like that, you can’t not walk through a door like that.” He and partner Miles spent the next year trying to partner before investing in May 2024.
• The growth arc: a couple tens of thousands of users and low-single-digit millions of ARR at investment; today, many hundreds of millions of run-rate revenue and millions of developers — in 18 months. Michael Dell’s chess/speed-chess line frames the compression.

3. Wiz: the war-zone handshake and the metric behind it

• The origin was a Thrive ritual — everyone explores a city outside the venture core — that put Clark in Tel Aviv in 2022, where Wiz was “the talk of the town.” The first call was a three-way Zoom spanning New York, Tel Aviv, and Tokyo; the shock stat was that 50% of Wiz’s users were software developers, not security people — “not your mom’s security company,” but a developer platform rather than “check-the-box compliance.”
• In December 2023 or early January 2024, during the Israel–Gaza war, Assaf called with new numbers; Thrive got to Tel Aviv “by hook or by crook” within 12–24 hours for a four-hour dinner and a handshake. Clark reframes the Colossus fly-into-a-war-zone story: it was not about deal competition — “you play against yourself, you don’t play against others” — but showing founders “we’re willing to do anything.”
• The diligence logic: Wiz had almost never churned — only one customer had churned between the company’s start and that round — and had the very best deal-size-to-speed ratio of almost any company Clark had seen. A rep could obtain a customer’s cloud API keys, scan the environment, and surface unknown vulnerabilities within minutes; six-figure deals closed in weeks or months. “Normally you have small deals and fast cycles, or big deals and slow cycles. Wiz had more or less cut the Gordian knot and got both.”
• Third leg: “most products like to say they’re platforms. Very few products are actually platforms.” Wiz had healthy adoption across code security, cloud-infrastructure security, and runtime security at barely four years old — “the mark of a lot of great companies is that their first product gives you the right to build some really exciting second and third products.” Google announced the acquisition earlier this year, which Clark thought was the largest sum ever paid for a startup.

4. Why hardware’s moment is now — and why it stays hard

• The barriers-down case: lidar fell from ~$75,000–80,000 per sensor in 2014 to a couple hundred dollars now; software intelligence enables more interesting hardware, including self-driving cars that Clark thinks increasingly outnumber human-driven cars on San Francisco streets; and — unlike Elon in the early 2000s or the Anduril founders in the late 2010s — a trained talent pool from that first wave now exists to draw on.
• Clark concedes hardware is genuinely harder — atoms cost money, and you need manufacturing and deployment lines, not a “ship the binary” button — but once over the hump it can be uniquely durable: “it’s really hard to imagine who is possibly going to disrupt SpaceX at this point, knock on wood.”
• The distinctive return shape is that some hardware businesses can have accelerated, “geometric” growth at scale. Anduril won a program of record within two years — fastest since the Korean War — on tens-to-hundreds-of-millions-dollar contracts, and is now being tapped across air, space, and the Navy. SpaceX’s launch business enabled Starlink, which now drives a very large majority of the company’s overall free cash flow and revenue. “The bigger you get, the bigger markets you can access.”
• On whether hardware funding is a bubble: returns accrue to N-of-1 founders solving unusually hard-to-replicate problems — venture has “the worst beta and the best alpha of any financial asset class” — and even now five of the ten most valuable companies are hardware: Apple, NVIDIA, Broadcom, Tesla, and TSMC.

5. Nudge: the brain becomes an engineering problem

• The thesis: the human brain is “basically one of the last frontiers of engineering” — long treated as a science problem, without products being built for it in the way they are for other parts of the economy or human health. It demands “life’s-work founders”: Jeremy, a former Neuralink engineering leader, and Fred Ehrsam, a Coinbase co-founder and early Neuralink investor — people who “no matter what island you put them on, only have one missionary purpose in life.”
• Mechanism as told: using ultrasound waves, which are also used to image pregnant women, Nudge targets very specific brain points associated with neurological problems and stimulates activity. The goal is “a treatment easier than taking a pill” for illnesses such as depression and addiction: a headset on your temples.
• His positioning map: Neuralink starts on the read side — an invasive implant, “Telepathy,” and a roadmap including Blindsight and attempts to restore limb movement that Clark calls “basically biblical miracles” — while Nudge is mostly on the stimulation side. Today’s closest comparison, TMS, is “pretty effective” but unpleasant and clinic-bound, so second-line adoption is “very, very low.” The ceiling is consumer: “why shouldn’t I be able to stimulate my mood or my energy levels or my focus?” — in 10 years, “we’re both wearing headphones on our temples.”

6. Concentration as alignment, not just math

• Partner Kareem’s line: “mathematically, there are probably a lot of ways we can make money as a venture firm. The question is, what feels really aligned to the founders we partner with?” A life’s-work founder holds a portfolio of one, so Thrive’s portfolio should not be too large: “all the wins should really feel great and all the losses should really hurt.”
• The financial half is the power law: Molly cites Carta research showing concentrated funds, typically with about 20 companies, outperforming spray-and-pray funds; the Magnificent 7 are 50%-plus of Nasdaq market cap. Once a company is judged to be in the right tail, the strategy is “concentrating as much capital into that company as possible,” against the backdrop of Molly’s rumored — and perhaps reported — $25B AUM figure.

7. OpenAI: belief at $29B, homegrown talent, paradigm machine

• Molly’s reported timeline: $29B in January 2023, $86B in February 2024, $157B in October 2024, and $500B in October 2025. Clark’s correction to the popularity narrative: “people probably overestimate how popular the company was in its early days.” Thrive was one of two firms to give the company a term sheet in 2023; Clark saw GPT-4 in November 2022 before its public release and says the model had passed the Turing test, even though $29B “is undoubtedly a high price for a company when ChatGPT hadn’t even launched yet.” The original memo never mentioned ChatGPT; the product now has more than a billion users.
• On talent density: beyond stars like Ilya, who Clark says invented AlexNet, OpenAI grows its own — Alex Radford “famously did not even really study AI” and became one of its most prolific researchers — while also betting on young people who “almost hold the model weights in their brains.” For Sora, Clark estimates the team’s average or median age was in the low twenties. The strategy: “find people who could be great versus who are already great.”
• The competitive-advantage claim: labs win by “constantly figuring out what the next paradigm is even as you’re productizing the last” — pretraining through GPT-4, post-training (supervised fine-tuning and RLHF) that made models “less alien intelligences and more helpful assistants,” and now reinforcement learning. “Every single major paradigm in large models” that has been successful at the macro level has come from OpenAI, or at least originated there, “to my knowledge.”

8. After SaaS, not after OpenAI — jobs, bubble, and the next 12 months

• Asked “what comes after OpenAI,” Clark rejects the premise — “is there an after Meta or an after Google?” — and reframes: “it’s an after SaaS that we’re going to see.” Productized intelligence called via API or fine-tuned into models creates a new breed of software: Cursor as collaborator, Harvey doing legal work, and even Wiz shipping a security assistant.
• On layoffs — Molly cites a Calshe chart showing up to 86% expecting more tech layoffs than in 2024 — Clark cannot name a portfolio company that has cut engineers because of AI tools. He sees AI as “much more of an augmenting technology than a substituting technology,” allowing companies to pursue more problems without adding as much headcount, and ultimately redirecting talent toward oncology, sustainable mining, and spacefaring. He expects most great technology companies to become public companies over time, which he calls morally right.
• Process over outcomes, via Dalio’s “focus on the swing, not where the ball goes” and Miles’s “paranoid but patient”: an investor in OpenAI in 2017–18 might have wondered whether the nonprofit would go anywhere, while other companies that looked hot early later weakened. The worst investments are ones “where, if the numbers change one quarter or one year, you would all of a sudden start to feel bad,” because that signals momentum rather than fundamentals.
• The forward watchlist: Clark is excited for reinforcement learning to enter production software; currently, there are relatively few such products, with some lab projects and rumored coding-company work. He also expects an intelligent-hardware renaissance — Physical Intelligence, Nudge, Mach, and Anduril — addressing “well north of 50%” of global GDP where affecting the physical world matters, plus AI-for-science through Isomorphic Labs (a Thrive investor) and Periodic (not an investor), as “science problems become engineering problems.”