a16z, Anish Acharya: Is SaaS Dead? Do Margins Still Matter? Why We Are Not in an AI Bubble?

Summary

• Software is oversold. Anish Acharya’s answer to the public-market “SaaSacre”: IT is only 8-12% of enterprise spend, so “you have this innovation bazooka with these models — why would you point it at rebuilding payroll or ERP or CRM?” 75% of public SaaS companies have raised prices since ChatGPT (mean 8-12%, a large group at 25%+), and “price is a measure of product market fit.” The vibe-code-everything story is “flat wrong.”
• Coding agents’ real effect on enterprise software is that switching costs collapse — “some companies have hostages, not customers” (Alex Rampell), and the SAP-to-Oracle migration that used to be a multiyear, high-risk project that would probably fail and get you fired gets dramatically easier. Fewer hostages, more customers — a positive incentive for the ecosystem, not a death sentence.
• The apps layer aggregates the models: foundation models innovate “roughly in lockstep,” 80% substitutes / 20% specialists, so orchestration (Cursor across Gemini front-end and Codex back-end; creatives across Midjourney/Krea vs Ideogram) is where much of the value is — a cloud-style oligopoly, not Uber-vs-Lyft. Harry’s pushback: “there’s a chance Cursor loses half of their revenue this year” to likely Claude Code; Anish’s reply — demand isn’t fixed: “our ability to be ambitious… always grows so much faster than our means.”
• Not a bubble: OpenAI hit $20B topline by 3x-ing capacity and 3x-ing revenue — inference supply is “100% spoken for” — and customer prices are rising, not compressing. Today’s subsidy (free-trial credits) is “healthy calories” converting into power users who pay $200-300/month (likely Grok Heavy $300, ChatGPT $200, Gemini Ultra $250) versus the old $20-25 Spotify ceiling.
• The SaaS-to-labor-budget shift is already underway: voice is “the wedge into the enterprise,” and the 10x is bundling support, sales, collections, and operations under one goal like CAC improvement. Legal software is $50B; after Anish calls legal “$500 million” infrastructure for capitalism, Harry calls it a “$500 billion market” — AI’s capture lands “closer to the 500 than the 50.”
• Weird wins: these models are emotional, human technology, and Google/Apple have “a thousand committees explicitly designed to ensure there’s never any persuasion, disagreement, or sexuality” — leaving companionship and other uncomfortable categories to startups. Traditional moats otherwise hold: networks are still “the gold standard,” and live proprietary data (likely OpenEvidence) beats a frontier model without it.
• The a16z operating bar, stated flat: “I’ve never lost a deal” in six and a half years — “we’re not allowed to believe in luck… we have to see 100% of the deals in our domain and win 100% of the deals we go after.” Elastic on price (sub-$100M entry “is a little bit of a wash”), never on ownership.
• The 2026 call: unlike mobile, where the Friendsters lost to later Facebooks, the 2023-24 early leaders (Harvey, Gamma) have kept their lead — and 2026 brings new AI-native categories, with Open Claw and Moltbook “just the beginning.” Moltbook “as an individual data point is probably overhyped right now, but what it points at directionally is underhyped.”

Deep dive

1. Cities are the original network effect — SF still wins

• Harry opens with his London pitch — cheaper talent that retains longer, none of the role-hopping promiscuity. Anish doesn’t soften it: “I disagree with you. I wish it was true.” Cities are the original network effect, and at a moment when “so many of the secrets are these sort of things whispered down shadowy hallways, the benefit of being in SF is enormous” — plus the selection bias of giving everything else up to move there.
• The one other geography with positive signal: Tel Aviv — at 10 million people “you can’t possibly fool yourself into thinking that the domestic market is going to be big enough,” so you go global immediately, whereas London’s 60 million is just big enough to trap you domestically (high-LTV fintech excepted). A $3-5B outcome is “extraordinary,” but trillion-dollar companies require “a set of assumptions that can lead to that” from day one.

2. The “SaaSacre” is oversold — point the bazooka at the other 90%

• His core call on the public-market sell-off (Bloomberg’s “SaaS apocalypse”): “software is completely oversold… a silly story.” IT is 8-12% of enterprise spend, so even a fully vibe-coded ERP and payroll saves 8-12%: “You have this innovation bazooka with these models. Why would you point it at rebuilding payroll or ERP or CRM?” You’d point it at extending your core advantage, or at optimizing the other ~90% of spend.
• The dissonant fact against the seat-contraction bears: 75% of public SaaS companies have raised prices since ChatGPT — mean 8-12%, a large group at 25%+. Harry: isn’t that forced, because seats aren’t growing? Anish: “price is a measure of product market fit” — under real competitive pressure you cut prices, not raise them. And incumbents aren’t relics: ServiceNow “is not IBM” — they raised guidance.
• The hedge stays in: “of course there will be secular losers” — models priced on seats being forced onto outcome pricing face “a big drag” — but for the majority of SaaS, being rewritten has “so little upside… and so much downside.”
• The under-discussed mechanism: coding agents crush switching costs. Rampell’s line — “some companies have hostages, not customers” — described SAP, where moving to Oracle was a multi-year, probably-failing, get-you-fired project. Now that transition is dramatically cheaper and faster: “decrease switching costs, more customers, less hostages — a positive incentive for the entire ecosystem.”

3. Apps aggregate the models — cloud oligopoly, not Uber vs Lyft

• On Rampell’s race — “will the incumbent acquire innovation before the startup acquires distribution?” — history says capable incumbents make their existing products better (“Microsoft will make a better word processor than they’ve ever made”) while native categories go to startups: AI moviemaking has no incumbent, and “it probably won’t be Adobe.”
• Why apps-layer value is under-discussed: the 2022 nightmare was a single foundation model as unique supplier — like being the only label with the Beatles, “you can charge 99% of your customer’s gross margin, and you actually tend to charge 100 or 110%.” Instead, providers innovate “roughly in lockstep”: 80% substitutes (plus open source doing the same things), and in the 20% “where a lot of the value is,” they’re specialists.
• That makes aggregation valuable: Gemini is great for front-end, Codex for back-end — Cursor becomes “a single way to orchestrate all the models”; in creative, Midjourney and Krea (Krea 1) are the aesthetically opinionated models while Ideogram is intentionally unopinionated for graphic designers, and a working creative needs both. Market structure: closer to AWS/Google Cloud than Uber/Lyft — rough substitutes with real specializations and reasonable margins, not price competed away.
• Harry’s pushback, worth keeping: “I think there’s a chance that Cursor loses half of their revenue this year” to likely Claude Code — “I don’t know anyone who’s not moved.” Anish: the error is assuming efficiency rises while ambition and customer count stay fixed — “our ability to be ambitious for wanting more things always grows so much faster than our means.” Cursor, Codex as app and CLI, and Claude Code “are going to find market fit and all grow.”

4. Startups’ pockets: the feature surface labs won’t build — and “weird wins”

• Granola (not a portfolio company) has been “copied to the moon” — OpenAI shipped meeting transcription inside ChatGPT — but its presumed vision is a productivity suite around that primitive. The question is whether OpenAI has “the prioritization, the resources and the ambition” to build all that feature surface. Models “will often recreate the primitive and even do product marketing — which I think the Claude legal stuff was” — and they’re single-model by construction; “multimodel rich feature surface” favors apps.
• Asked about “boring wins,” he flips his own line: “I think weird wins.” These models are “wild, non-predictable, emotional, very human” — pointed at disagreement, persuasion, sexuality. “If you’re Google or Apple, you have a thousand committees that are explicitly designed to ensure there’s never any persuasion, disagreement, or sexuality expressed in your products.” That’s the startup pocket.
• Companionship is the proof: likely Character.AI, likely Janitor AI, and likely Replika (“probably one of the most healthy and nourishing forms”) — well-received by customers, uncomfortable for the labs, “perhaps even Grok.” Would he let his kids use them? “Absolutely” — his request-for-startup is a contextual companion that plays Minecraft with his son and “models pro-social behaviors and is still cool and chill”; for seniors, an AI that calls to check medicine, “maybe lightly flirt with them, talk about World War II” delivers spiritual nourishment without feeling like babysitting.
• Harry’s pushback — doesn’t this make people more withdrawn? “The exact opposite”: the wealthy have therapy, an “embarrassment of social riches,” or religion, but “for the majority of our society today they just don’t have an outlet, and I do think technology can be that outlet.” His closing optimism runs through the same idea: “the NPS of the human experience, for lack of a better phrase, is on the way up.”

5. People want to spend time, not save it — and the old moats hold

• Against the everything-becomes-voice consensus: voice is amazing for enterprise, but chat and dynamic UIs are “overstated in consumer.” Channeling Eugenia (Replika’s founder, now likely Wabi): “most people don’t want to save time, they want to spend time” — products get designed by “the most high-agency people in the world like Sam and Elon,” for whom a chat box is optimal. Browse-based interfaces largely stay; chat as the future of intent, “I’m still a little skeptical.”
• Defensibility survives: “networks are the gold standard” — all the vibe coding in the world doesn’t touch Airbnb — though Moltbook-style synthetic networks may make certain network types less defensible than they once were.
• The moat he was always skeptical of — the “data network effect,” the thing “that got thrown out a lot when you couldn’t think of what moat to say” — is now real in one form: live proprietary data (likely OpenEvidence): “you can put a relatively commodity model in front of it and get much better results than the most cutting-edge model” without it. Systems of record split the same way: an on-prem database with no engagement layer is “at some risk”; a bank’s core system — thousands of transactions per second, hundreds of humans, extreme accuracy demands — is “as good as gold.”

6. Margins: healthy calories now, and power users finally pay

• The nuance he insists on: 2021’s distortion was an indirect subsidy of Google and Facebook — raise $10M, spend $8M on ads — “empty calories.” Today’s distortion — zero or negative-gross-margin credits for trials — is “very healthy calories,” because it converts into high-paying power users. Blended margins of AI natives look worse, but the form of distortion is much better than five years ago.
• Andrew Chen’s pre-AI line “power users are just users” has broken: Spotify’s maxed-out plan set a $20-25/month consumer ceiling; now likely Grok Heavy is $300, ChatGPT $200, Gemini Ultra $250 — 10x prices plus consumption revenue on top, making the S&M behind those users “very wisely invested.” He endorses Jason Lemkin’s adjacent line as “100% correct”: “for the best companies, influence is the new sales and marketing.”
• The operating framework for Harry’s team: treat month one as free organic traffic rather than acquired users, book trial-margin cost as CAC, and read durable margin off converts. M2 is the new M1 — then apply the old retention bar: an M12 of 50% is solid, “60-70%, we’re very very happy.”

7. Not a bubble — supply is 100% spoken for, and prices are rising

• His quip: “it’s not a bubble, and it’s good that it is.” OpenAI is at $20B topline and got there by 3x-ing capacity and 3x-ing revenue — inference supply is “100% spoken for,” versus prior bubbles’ buildout of supply far ahead of demand. Customer prices are going up, not compressing. And what subsidization exists is “intelligent subsidization… mostly being paid for by big tech and the labs, and it benefits consumers and startups. God bless.”
• Rory O’Driscoll’s test (via Harry): this all works if spend migrates from the 12% SaaS budget into the human-labor budget. Anish: “we’re already seeing it” (the CH Robinson story). “Voice is the wedge into the enterprise” — and the near-term prize isn’t cheaper support: models can be the empathetic listener or the charismatic “yapper” at will, so sophisticated companies bundle support, sales, collections, and operations under one goal like CAC improvement — “that is going to be the 10x on productivity.”
• Lemkin’s counter: this is the year of substitution on price — ElevenLabs is “amazing… it’s too expensive.” Anish disagrees: productized capability has outstripped cost — “is anybody saying I should go back and use Sonnet 3.7 because it’s cheaper… or use something other than Opus 4.5 or Codex 5.2? Nobody” — and GPT-4o token costs have fallen 100x since release anyway.
• Costs are a feature: they force “business model hygiene” that field-of-dreams free products never had. And rather than ElevenLabs subsidizing a land grab off its $500M raise, push the frontier: software should asymptote to “80 to 90%” of consumer discretionary and enterprise spend — companionship, entertainment, therapy, healthcare, education.

8. Legal isn’t a $50B market — it’s $500B infrastructure for capitalism

• Harry can’t square his Thiel-school “competition is for losers” with ~50 customer-support startups over $50M raised (10 over $100M). Anish: “what you’re calling a market is actually an industry.” Anish initially says legal is “$500 million” infrastructure for capitalism; Harry calls it a “$500 billion market,” and Anish says no single company wins that — AI’s capture lands “somewhere between the 50 and the 500 — closer to the 500.”
• What capture looks like: dramatic productivity gains, not headcount elimination — jobs are bundles of tasks that resist 100% automation, so a 20% productivity gain shows up “more as a 4-day work week than 20% less jobs.” “You can do all the customer support you want, but sometimes you got to take the customer out for a steak dinner.”
• Agent maximalism (“you just chill out for the day and your AI does everything”) is “probably a little bit ahead of where we actually are”: humans stay in a tight loop for exception handling, our instructions are “frustratingly vague,” and models drive into local maxima — “often it takes human intuition to break from the local to the global.” BPOs — well-defined tasks pulled off a queue — automate first; on UiPath he demurs, noting only that “vision models haven’t nearly kept up.”

9. The a16z bar: see 100% of deals, win 100% — luck not allowed

• Asked his most painful loss: “I haven’t lost a deal” — in six and a half years. Harry, gently: doesn’t that mean risk aperture is too low? Anish holds the line: “I don’t think we’re allowed to believe in luck at Andreessen. We have to see 100% of the deals in our domain, and we win 100% of the deals that we go after” — wrong decisions on good information are fine; not seeing the company is not. It echoes Marc’s maddening onboarding advice: “just be right a lot.”
• On Harry’s claim that Series A is the hardest place to invest ($1M revenue, 100-200x multiples, price-to-progress mismatched against $25M seeds): “I disagree.” Investors choose their risks — competitive, pricing, team, geographic, fundraising — and “having shipped something and having sold something is such a dramatic signal”; the A is his optimal point of information versus entry ownership. “It’s supposed to be hard, and you should be winning anyway.”
• Very elastic on price, “not very elastic” on ownership: below ~$100M, “12 on 60, 15 on 75 — it’s a little bit of a wash”; price’s real cost is next-round expectations, and in growth “the difference between having to raise at 300, 500, and 700 is pretty significant.” Ownership is the whole all-chips-behind-you model.
• Triple-triple-double-double isn’t dead — the bar is “calibrated to your part of the market,” top-quartile versus your peer set. “There is just physics to some of these markets” (ERP’s cautious buyer; payroll’s slow-boil sale that Deel turned fast-boil), while new primitives allow 10→100 or 10→200. And he defends “area under the curve” companies — Figma’s three-to-four quiet build years yielding an N-of-1 network-effect product now positioned for the shift from execution work to thinking work — as underestimated next to lionized 1→100 stories.

10. Changed mind: early leaders kept their lead — 2026 births the native categories

• The self-confessed mistake: being “a bit too casual about product market fit” in 2021 — funding a credible founder theory that matched his own theory, instead of asking “is this actually working… investing with this sort of self-deception of like, well, let’s just assume it’s working when it’s not quite working is a mistake.”
• The TAM lesson: “we consistently underestimate how big the markets are and consistently overestimate how easy it is to go from zero to one.” His specimen is Credit Karma: back-of-envelope says a torso of people who need a credit score twice a year; reality is 100M+ Americans, 50M quarterly actives, logging in four times a month — the score is “a mirror that people like to look in and see how they’re doing objectively as an adult.” Corollary: with a formidable founder making nonlinear progress, “inertia is the most powerful force in the universe… you have to tie-break in the direction of them doing it forever.”
• What surprised him this cycle: in mobile, the 2008-09 anointed winners (the Friendsters) lost to later Facebooks; this time the 2023-24 early leaders — Harvey and Gamma — have kept their lead. His map: Nov-22 ChatGPT; ‘23 the obviously good ideas get started; end-‘24 reasoning models (o1, DeepSeek) make the not-yet-working ones work; ‘25 they scale. “In 26 we’re going to see a whole new set of categories… Knowing what we all know now, what company would you build? That is the operative question.” Open Claw and Moltbook are just the beginning.
• On Moltbook itself: “just so damn cool,” even granting the critique (likely Balaji) that it’s “robot dogs barking at each other.” The direction — digital twins going on pseudo-dates and matchmaking their owners, the UGC story that just knocked Match’s stock — is what matters: “Moltbook as an individual data point is probably overhyped right now, but what it points at directionally is underhyped.”