a16z GP, Martin Casado: Anthropic vs OpenAI & Why Open Source is a National Security Risk with China

Summary

• Casado’s core call: the model question has been around for three years, and “there’s only been one sin and that one sin is zero sum thinking.” The hosting providers and model companies investors wrote off as non-defensible — alongside Nvidia — have kept growing in value, so the mistake is not playing. But he’s explicit about the paradox: the game is capital-intensive and non-leaders are already getting wiped out — “you kind of have to play, but it’s very very high risk.”
• On the vibe-coding stack’s dependence on Anthropic, he bets oligopoly over monopoly, on the cloud analog: AWS held 70–80% share early — “way more dominance than Anthropic has now” — and still ended in an oligopoly once Microsoft and Google spun their way in. Models distill easily (he cites recent launches including likely Qwen and likely Kimi), Gemini 2.5 is his personal standard model on price-performance, and Google can subsidize arbitrarily — so an independent consumption layer stays a “very healthy layer.”
• Brand effects are back for the first time since the internet: markets are expanding so fast that the household name wins the frontier by default (“My mom knows ChatGPT”), and the brand leader “tends to get 80% of the market” until growth slows and share disperses — exactly what happened to AWS. Investing implication, stated flat: “you just try to invest in the leader and it’s worth paying up for the leader.”
• “Models” are not one asset class: diffusion businesses (ElevenLabs, Midjourney, Black Forest Labs, Ideogram) have great economics because the models are smaller and Google doesn’t subsidize speech — while frontier language is a subsidized, high-stakes game where “if you’re not in one of the leaders, that capital is forfeit.” Markets also fracture as they expand: OpenAI was first to code, image, and video, lost all three, and was still rational to keep language — “by far the largest market.”
• His contrarian call on national security: “open source is most dangerous because China is better at it than we are” — the answer to Chinese model proliferation is to fund US open source, national labs, and academia “like crazy,” the same posture that worked when he simulated nuclear weapons at Livermore in 1999. The worst VC take he’s heard: “open source is bad for national security.”
• The productivity call that cuts against the hype: coding models make “10x engineers 2x,” not 100x — every company he works with uses Cursor, yet product velocity hasn’t visibly jumped because “the things that are hard remain really hard.” The average production PR is two lines, and those two lines encode field learning no model has; apps were always copyable (“it’s fucking CRUD, man”), so AI doesn’t change the defensibility paradigm.
• His one sin in investing proper: missing the winner — losing money on a dead category is forgivable, picking the wrong horse in a live one is not, because “there’s basically no amount of work you can do to determine if a space is going to work” but you can diligence which of ten companies is best. Harry’s open pushback — he’d fire the person who picked the wrong market — is left standing as a genuine philosophical split.
• Quickfire tells: most overhyped category is “ASI”; the founder he’d wire money to blind is likely Michael Truell; his most impactful trait is “deep-seated anxiety from being poor,” from food stamps on a Montana dirt road to running a16z’s $1.2B infrastructure fund while working 80–100-hour weeks.

Deep dive

1. The one sin is zero-sum thinking — every layer has winners

• Casado opens with an admission of epistemic humility: “my intuition doesn’t really work like it has the last 20 years,” because this is the first time software creation itself is being disrupted. But observationally the record is one-sided: “there’s only been one sin and that one sin is zero sum thinking. We always worry about like oh is this defensible?… And the answer has kind of been unilaterally yes. The answer has been every layer has gotten value.”
• On Harry’s 2021 trauma about “playing the game on the field”: Casado’s distinction is that behavior should follow business, not marks. In 2021 behavior followed marks — Tiger deploying into prices the public market decided. Today it follows “some of the fastest growing companies we’ve ever seen by users, by revenue… If not, I mean, what are we doing?”
• The honest dissonance he insists on keeping: Nvidia, the written-off hosting providers, and the written-off model companies all keep growing in value — yet “we’ve seen tons of wipeouts already for the non-leaders.” His summary of the regime: “it’s almost this bipolar or paradoxical situation where you kind of have to play, but it’s very very high risk.”

2. Vibe coding on Anthropic: two futures, and he bets oligopoly

• Harry’s question — vibe-coding players are predominantly sitting on Anthropic, which “could eventually shut them off.” Casado frames two futures for code: an Anthropic monopoly, or an oligopoly of coding models. In the multi-model world, an independent consumption layer (for non-technical users, Python users, professional coders) is “a very healthy layer.” In the monopoly world, Anthropic “will put pressure on anybody that they view in their core focus” and capture margin or share — though even monopolists draw lines: “we’re never going to be like an app dev tool company.”
• His timing caveat is worth keeping verbatim: the conversation lands “pretty soon after Claude 4 launched. And these models are so episodic. Every time one launches, everybody’s like, ‘It’s the future’” — citing the OpenAI Ghibli image moment that came and passed. “We always assume every time a model launches it’s going to be a monopoly and it just really hasn’t been the case.”
• Pressed for the bet: “Oligopoly. This is how the cloud played out.” AWS had 70–80% early share — “way more dominance than Anthropic has now” — and Microsoft and Google spun their way in anyway. Models distill too easily to hold an edge (he cites recent launches including likely Qwen and likely Kimi), Gemini 2.5 is “the one that I actually use as my standard model” on price-performance, Google can subsidize arbitrarily, and “never count out OpenAI… the rumor is GPT-5 coding is going to be great.”
• On whether the 10-year model leaders exist yet: RL-era scaling “doesn’t generalize nearly as much,” so models fragment by flavor — Mira and Ilya are out building, science models are starting up. And it’s early by precedent: “Google was third generation search, Facebook was third generation social networking” (Friendster, MySpace first). Still, Anthropic and OpenAI have done “a remarkable job” on brand and share and should remain stalwarts. Disclosed: a16z is an OpenAI investor.

3. “Model” is not one business — tease the asset class apart

• Harry’s challenge — are models fundamentally good venture investments given stock comp and dilution? Casado’s headline lesson: “there is no one way to think about models and the models themselves are entirely different businesses.” Diffusion models — ElevenLabs, Midjourney, Black Forest Labs, likely Ideogram — are “wonderful businesses that have great economics because the models are smaller” and the ecosystem isn’t subsidized: “Google subsidizes language and code and video but not speech.”
• Frontier language is the opposite: Meta, Google, and Chinese entrants subsidize it, so it’s “a high stakes game where the winners really win, but it requires a lot of capital to enter… and if you’re not in one of the leaders, that capital is forfeit.” Even three years in, “there’ve already been a number of companies that have had to exit early.”
• On Rory O’Driscoll’s line that AI investors have simply accepted going massively up the risk curve, Casado reframes: it’s not willingness but “the requirement of the game… we’d be very happy not to.”

4. Brand effects are back — invest in the leader and pay up

• His attached thesis (offered with the hedge “so early we don’t know and maybe in a month all this gets proven wrong”): markets are so large and growing so fast that “we’re actually seeing brand effects take place. And we haven’t seen that since the internet.” The household name wins because the expanding frontier of new users hears one name first — “people know ChatGPT… My mom knows ChatGPT.” Harry’s confession folded in: he backed Lovable “for the exact same reason” — consumer brand wins.
• The specimen case is Midjourney: “first that got above the quality bar. It’s taken zero investment from institutions. It’s still the market leader” despite a crowded field. Historically “the brand leader tends to get 80% of the market. It just tends to break out Pareto for a while.”
• The mechanism has an expiry: brand dominance holds while markets expand and loses its edge when growth slows and the frontier “hears both names.” AWS again — 70–80% share until cloud growth decelerated, “and then all of a sudden market share started to shift dramatically” to GCP and Azure — less because Google and Microsoft entered, more because expansion itself slowed. Portfolio translation: two questions — is it the leader in its area? “If it is, it’s definitely worth paying up.”

5. Markets fracture as they expand — OpenAI’s rational losses

• His second structural observation: expanding markets fracture into white space. OpenAI was first to code with GitHub Copilot (he says they provided the weights, as far as he knows), and lost that; it was first to image with DALL-E, and as far as he can tell lost that; it was first to video with Sora, and as far as he can tell lost that — yet remains dominant in language, “by far the largest market by far. And so OpenAI acted totally rationally.” That retreat is what let Midjourney and BFL take image, Google seems to have grabbed video with V3 (likely Veo 3), and Anthropic turn code into “this wonderful business.”
• Even submarkets sustain multiple viable winners on positioning: likely Ideogram for professional designers, BFL for the open-source/developer crowd, Midjourney for the stylized fantasy aesthetic. His call: “we’re going to see fragmentation for quite a while before we see consolidation.”
• On Harry’s live deal — a European medical-transcription player at 1→8 million in a year, up against Abridge: AI has revived regional advantages (“regulatory environments are quite balkanized,” so are language and culture). “The thesis cannot be European company X wins the American market, but I promise when it comes to AI, the European market is large enough. I promise that.”

6. Thin margins are a board-level choice, not an endemic flaw

• Against the “pass-through funnels to the LLMs” denigration: “I just don’t buy that these are endemic to the business model.” With cheap private capital in land-grab time, “the rational business decision is to sacrifice margin for distribution” — the incremental user can be monetized forever; miss them now and never. Precedent: “the web wasn’t even monetized… literally massively negative because we didn’t even have a business model until advertisements come up.”
• When it’s time to turn margins on: classic moats (marketplace, brand), long-tail domain and regulatory integration — a healthcare company that has “cracked the European market and understands all the regulation… Anthropic is not going to take the time to do that” — or technical differentiation, because “in this phase of model scaling a lot of the approaches to scaling don’t generalize. This gives a ton of room for the application developers to build their own models.”
• From inside his portfolio (numbers withheld): break-even margins are usually “a board level specific choice to prioritize distribution, not just because this is systemically something they have to do.”

7. Open source and China: the national-security argument, inverted

• On Vinod Khosla’s “nuclear secrets” framing and Founders Fund’s anti-open-source turn: “it’s crazy to have VCs talking against open source… we were in very bizarro land for a while.” His credentials for the safety question are unusual — intelligence community, Lawrence Livermore, a PhD half in security. The internet had concrete proof points (the Morris worm, critical infrastructure going down) that forced a real doctrine shift to defense asymmetry. AI has none yet: “I still have yet to see the dramatic new attack. It’s going to come for sure, but we haven’t seen it yet” — so the discourse “is not in line with the reality.”
• The novel dynamic he flags: historically the builder was the booster and the firewall vendor was the fearmonger — “the same voices but in two different bodies based on interests. The interesting thing this time is they’re in the same body” — the people creating the thing also call it dangerous, which “has been very confusing for everyone.”
• The China turn is his sharpest line: “open source is most dangerous because China is better at it than we are” — Chinese open-source models are proliferating everywhere and the US can’t regulate Beijing. “The answer is yes [it’s a risk], because of China, not because of us.” The response: fuel American open source, “fund this stuff like crazy,” national labs, academia, national priority. His 1999 analog — working on nuclear-weapons simulation at Livermore, PlayStation export controls on Saddam Hussein — ended with the US choosing to lead everything and “we won and were able to control the technical discourse of the planet. This time instead we want to put our head in the sand.”
• On Trump’s university cuts, he refuses both easy readings: indirect costs “have gotten way out of hand” and everybody in academia agreed “until it was Trump doing it” (Obama tried too) — but yes, fund as much or more. And on Harry’s point that Llama’s going closed: he agrees the ecosystem is moving away from open even as rhetoric (the American AI policy) endorses it — though open-source AI is a misnomer anyway: you release the small model, keep the capable one closed, and get brand without eroding the business, since weights alone don’t let anyone replicate the data and training pipeline. Historically open source captured ~20% of software market value; “I would say it’s much higher than that for AI.”

8. Coding models make 10x engineers 2x — and the hard things stay hard

• What he’s consistently gotten wrong: “just how fast these coding models advance.” A programmer since the ’90s who still codes most evenings, he “couldn’t work back to working without them” — they offload “all the shit I didn’t want to learn,” the framework and environment sludge that by 2015 consumed “90% of your time.” Now old systems programmers are back at it: “the modern version of the old man making the train set is these old systems programmers vibe coding at night just because it’s become pleasant again.”
• But on the productivity multiple, he deflates Harry’s premise: not 1x→10x, not 10x→100x — “I think they make 10x engineers 2x.” Every company he works with uses Cursor, yet “if I actually look at, has that increased the velocity of the products coming out — I don’t think that much.” Creating a frontier model is still loss curves, data pipelines, and experimentation no coding model does; the tools shine at tests, docs, visualization. His long-run bet: “more robust maintainable code bases with less bugs is just as likely to be the impact as feature velocity.”
• On time-to-copy collapsing (Harry quoting Misha’s line that it’s “basically been reduced to nothing”): split apps from infra. Apps never had defensible technology — “every time I look at vertical SaaS… it’s fucking CRUD, man” — the business was always the long tail of domain understanding. Infra has real computer-science tradeoffs learned through market exploration that models can’t shortcut. Aaron Levie’s stat carries the argument: the average accepted PR on a production codebase is two lines — “that two or twelve lines signify some learning in the field… The hard thing isn’t the two lines of code.” AI “is getting rid of the middle” — net accretive.
• The philosophical coda: research had become “sweeping the dust around the room” — too much literature to know if anything was new, the important problems stuck between disciplines. AI “knows all the literature… and is very good at tying different disciplines” — “a bit of a liberator” from the morass. On jobs, his translator cousins are the microcosm: work shifted to spot-checking AI output they can only fix by rewriting, “but they won’t pay for that… I can’t work on something without a soul.” AI’s oddity versus electricity: “it actually requires today a human handler” — every monetized use case has a human on the other side.

9. The one sin in investing is missing the winner — Harry disagrees

• The a16z doctrine: “The only sin in investing — and I’ve sinned so much — is missing the winner. It’s fine to invest in a category that doesn’t work… but if you choose the wrong company, that’s not okay.” Rationale: “there’s basically no amount of work you can do to determine if a space is going to work — that’s like weather prediction. But given a set of companies you can actually do the work to understand which one is the best.” Can that strategy beat the market? “Yes. Can you specifically pick the winner every time? Absolutely not.” A scaling note: a firm needs a strategy “you can articulate and teach other people.”
• Harry’s pushback, unsmoothed: “some points need to be given for the insightfulness to pick the right market… It’s fucking hard to pick the horse. Almost I’d fire the one who picked the wrong market entirely.” Casado doesn’t concede — you can’t predict technology adoption, big companies and innovations wipe out whole markets (“you could argue AI is invalidating tons of markets”), but ten companies with traction can be diligenced concretely.
• His confessed graveyard: the data streaming market — “just turned out to be a subset of the analytics batch market,” tough at every layer (dashboards, transformation, feature stores), with ClickHouse under Aaron Katz (likely) “maybe the one breakout since Confluent.” Many spaces where “we played multiple bets where it just didn’t work out.” And on AGI-versus-SaaS as a contradiction: “humans are AGI and we still invest in enterprise SaaS” — AGI doesn’t mean anything he wants to disappear in the future disappears.

10. Ownership over price, one mortal enemy, and dirt-road anxiety

• Fund mechanics, stated cleanly: “the market sets the price… Everything comes down to ownership for us, not price” — walk away on ownership, not price, because the median outcome has to return a fifth to half of the fund. Two legitimate games now exist: the specialist (Harry’s — deep network, early, concentrated) and the full-product-suite adaptive platform (a16z’s — “the strategy that works has shifted this entire time”), and he runs the $1.2B infrastructure fund inside the latter. On conflicts, the a16z bugbear: companies “pivot all the time after you invest,” and conflicts are among the top reasons they pass — including a recent deal killed when a portfolio founder called about something merely on his roadmap. Harry’s pushback: “if it’s not on the road map… I’m not having you tell me how to do my job.” Casado’s talk track, stolen from Chris Dixon: “you have one mortal enemy, and whoever it is, I’m with you… but you get one.”
• On Elad Gil, a genuine read: not market-first as advertised — “very good with founder market fit, maybe the best in the industry… this kind of boyband construction based on the market,” but the founder is the primary decision once the market’s chosen. Quickfire: most overhyped category “ASI”; worst VC take “open source is bad for national security”; the founder he’d wire money to sight unseen is likely Michael Truell: “he knows what he wants, he’s got an intuition that’s impeccable, and he listens incredibly well” — a “very potent combination.”
• The personal engine: his most impactful trait is “deep-seated anxiety from being poor” — food stamps, a Montana dirt road, school closed for hunting season. After selling his company he drove toward Hollywood to make movies with his cousin Vincenzo Natali, turned around two hours out: “don’t use those dreams that you concocted when you were in the pressure cooker… you’re probably where you are for the love of it.” Casado still works 80–100 hours a week. Harry, separately, describes driving “a fucking Volkswagen” with three dogs and three chickens; Casado calls family support and grounding “the full catastrophe” (Zorba the Greek) and says they’re the only way he can do what he does. And on a16z in 2035: Marc and Ben at the top is “a feature, not a bug” — the VC partnership model suits “a dentist office or a law firm,” but for decision velocity and disruptive change “it’s death.”