Point the Innovation Bazooka at the Other 90 Percent

20VC with Harry Stebbings · Anish Acharya, General Partner at Andreessen Horowitz · Video ID Aq0JSbuIppQ · February 9, 2026 · 80:02

Bloomberg is trying to make “SaaS apocalypse” stick. Public-market investors have decided enterprise revenue is no longer sticky, and the folk theory underneath that panic is that anyone can vibe-code an ERP, a payroll system, or a CRM. Anish Acharya’s objection is arithmetic, not nostalgia. IT and SaaS are 8 to 12 percent of enterprise spend. Even if you rebuilt those systems with all the operational risk that entails, you would save 8 to 12 percent. The models are an innovation bazooka. Pointing them at payroll is a misuse of the weapon.

Acharya leads consumer and fintech investing at Series A for a16z. He has founded and sold companies (Snowball to Credit Karma, SocialDeck to Google), scaled Credit Karma’s U.S. card business, and sat on boards including Deel, Mosaic, Clutch, Titan, and HappyRobot. He is not arguing that every incumbent is safe. He is arguing that the market is oversold on software dying, and undersold on where the bazooka actually belongs: the other 90 percent, plus categories that did not exist before this product cycle.

Cities are still the original network effect

Stebbings opens with a London pitch. Talent is cheaper. It stays. It does not hop roles with San Francisco promiscuity. Acharya has built in Canada and in SF. He disagrees, and says he wishes he did not. Talent may be equally distributed; opportunity is not. Cities are the original network effect, and for this moment in technology — secrets whispered down shadowy hallways — the builder network in San Francisco is enormous. You can make it happen in New York, London, Toronto, or Tel Aviv. Moving everything to SF is a selection-bias test: do you care enough to be singular?

The one other city he blesses on those terms is Tel Aviv. Ambition can be uncompromising there, and the country is 10 million people, which makes it impossible to fool yourself that the domestic market is enough. London has 60 million. In fintech, where LTVs are high, that might be sufficient. For most mass-market products it is not, and companies that start domestic often fail to leave. He names ElevenLabs as a counterexample and still says SF is easier. A $3 to $5 billion outcome, he adds, is extraordinary and stacks into meaningful funds. It is not uninteresting. It is also not how you underwrite a trillion-dollar intention.

Prices went up. That is not how a massacre behaves.

Stebbings cites seat contraction at CRMs and “the Mondays of the world.” Acharya looked at the data that morning: 75 percent of public SaaS companies have raised prices since ChatGPT launched, a mean of 8 to 12 percent, with a large group at 25 percent or more. If they were under enormous competitive pressure, they would typically be cutting. Price, to him, is a measure of product-market fit. ServiceNow is not IBM; it is a capable incumbent that raised guidance. Incumbents have a right to win inside existing workflows. There will be disruption — especially companies once priced on seats and now priced on outcomes, which he calls a real drag — but for the majority of SaaS there is little upside and a lot of downside in being rewritten.

The interesting enterprise effect of coding agents is not replacement. It is switching costs. Alex Rampell’s line, which both men love, is that some vendors have hostages, not customers. An SAP shop is a hostage. Switching to Oracle is a multi-year, high-risk, career-ending project that usually does not happen, so SAP can do the bare minimum after the win. Coding agents collapse the complexity, speed, and risk of that transition. More customers, fewer hostages. That, Acharya says, is a positive incentive for the whole ecosystem — and the actual way agents show up in public enterprise names.

Incumbents keep Word. Startups get the category that did not exist

Rampell’s other question: does the incumbent acquire innovation before the startup acquires distribution? History, Acharya says, is unkind to people who skip it. In a product cycle, a capable incumbent usually makes a better version of what it already sells. Microsoft will make a better word processor. Google will make a better search engine. Adobe will make a better Photoshop and Illustrator. The native categories that did not exist before the cycle go to startups. AI-assisted movie-making is not Adobe’s category, and he is betting a native company wins it.

He will not claim the application layer creates more value than foundation models. He will claim it is under-discussed. In late 2022 and early 2023, a single supplier a generation ahead — OpenAI — looked like the Beatles on one label: charge 99 percent of customer gross margin, or 100, or 110. That did not happen. Foundation-model providers now innovate roughly in lockstep. About 80 percent of what they do, he treats as substitutes, plus open-source models doing the same things. The remaining 20 percent is where they specialize. That split is why an aggregation layer is valuable.

Coding is the clean example. Gemini is great for front end, Codex for backend; a vibe-coded project wants both, and nobody wants two CLIs. Cursor as an orchestrator is the point. Creative tools fragment the same way: Midjourney and Krea’s Krea1 are aesthetically opinionated; Ideogram, used by graphic designers, is intentionally not. A creative at a big company needs both, which means an apps company. Stebbings worries Cursor loses half its revenue to Claude Code. Acharya’s reply is that the error is holding ambition and customer count fixed while efficiency rises. Demand for making and consuming software already exceeds supply. Some people will want a rich IDE. Some will want to be closer to the metal. That composition looks like AWS and Google Cloud — an oligopoly with reasonable margins and rough substitutability — not Uber and Lyft, the extreme case of pure substitutes with price competed away.

Model companies invading the apps layer is a real copy threat, not a feature-surface threat. Granola, which a16z did not invest in and which he admires, was first to live meeting recording and transcription and has been “copied to the moon,” including inside ChatGPT. He assumes Granola’s vision is a productivity suite — Word, Docs, spreadsheets — around that primitive. Recreating the primitive is what model companies do, including product-marketing moves like Anthropic attaching Claude to legal. Building opinionated UI for the legal community, with rich feature surface, while remaining multi-model, is not how those companies are set up to prioritize. OpenAI will only ever give you OpenAI models.

Weird wins where committees cannot go

Stebbings reminds him he once said “boring wins.” Acharya reverses it on the spot: weird wins. Previous technology was quantitative, clinical, bounded in the feelings it could capture. These models are wild, non-predictable, emotional, and human. They can be pointed at disagreement, persuasion, sexuality. Google and Apple have a thousand committees designed to ensure none of that appears. Startups can live in the pocket the labs find uncomfortable. Companionship is the example: Character, Janitor, Replika — products customers receive well and big tech would rather not build, “perhaps modulo Grok.”

He would encourage his children to use them. His request for startup is a contextual companion for his son in Minecraft: pro-social, still cool, a better influence than whoever is on the server. He does not buy the withdrawal critique. Most people do not have therapy, an embarrassment of dinners, or a spiritual practice as an outlet. Technology can be one. For seniors, the companion should not call to check in — that attacks self-respect — but to ask about medicine, the day, World War II, maybe lightly flirt. Indirection delivers the nourishment.

The same human grain shows up in UI. Voice, he says, is amazing for enterprise. Dynamic UIs and chat UIs are overstated in consumer. Eugenia Kuyda, who founded Replika, is the thinker he cites: most people do not want to save time, they want to spend it. Sam Altman and Elon Musk are the highest-agency people in the world; for them the optimal UI is a chat box. Everyone else is often not sure what they want and cannot articulate it. Browse-based interfaces largely stay. He is still skeptical that the future of intent is chat.

The moats that survived, and the ones that were always fake

Defensibility still exists. Networks remain the gold standard. Airbnb does not lose because someone can vibe-code a listing site. He flags Moltbook as a possible new synthetic network that might weaken some network types, without throwing the category out. Systems of record split: an on-prem database with no engagement layer and few human workflows is at risk; a bank core with thousands of transactions per second, hundreds of humans, and a brutal accuracy demand is “as good as gold.”

The fake moat of the last decade was the “data network effect,” the phrase you used when you could not think of a moat. What he now takes seriously is proprietary and live data — health data that changes, live data about a running product. Put a commodity model in front of that and you beat a cutting-edge model without it.

Margins still matter. Blend them and you will lie to yourself

Stebbings walks his mother around London saying “margins matter.” Acharya’s nuance is about the form of distortion, not the death of gross margin. He does not believe this is a bubble period, but superheated markets always distort. In 2021 the distortion was an indirect Google/Facebook subsidy: give a fintech $10 million, watch $8 million become ads — empty calories. Today’s subsidy is zero- or negative-gross-margin credits so users can try the product. Those are a drag and also healthy calories, because they convert into high-paying power users. Blended margins for AI-native companies look worse. Unbundle trial CAC from the durable margin of converters.

Treat month-one traffic as organic, not acquired. Treat M2 as the new M1, because M1 is tourists you did not pay for. Then apply the old retention bar. An M12 of 50 percent is solid; 60 to 70 percent makes him very happy. Power users broke the old consumer ceiling. Andrew Chen used to say power users are just users, because even 100x value did not mean 100x price: the richest Spotify SKU still sat at $20 to $25 a month. Grok Heavy is $300, ChatGPT $200, Gemini Ultra $250, plus consumption on top. Sales and marketing spend against those users is no longer a tax on a $25 ceiling. Jason Lemkin’s line, which Acharya endorses at 100 percent: for the best companies, influence is the new sales and marketing.

Not a bubble, and it is good that the spend is loud

His quip: it is not a bubble, and it is good that it is. He marks the caveat that this is not his expertise, then offers three facts. OpenAI’s recent announcement put it at $20 billion of topline; they 3x’d capacity and 3x’d topline, so new inference supply is 100 percent spoken for — the opposite of a classic overbuild. Customer prices are going up, not compressing. Whatever subsidization exists is mostly paid by big tech and the labs and benefits consumers and startups. “God bless.”

Rory O’Driscoll’s test, via Stebbings: this works if spend migrates from the 12 percent SaaS budget to the human-labor budget. Acharya says it is already happening. He points at CH Robinson as something David George discussed on the show. Voice is the wedge into the enterprise, but the near-term story is not “take cost out of support.” Support, sales, operations, and collections were separate because they required different human archetypes — the empathic listener versus the high-energy yapper. Models can be either at any time. Sophisticated companies bundle them toward a broad goal like CAC improvement. That is the 10x, not a cheaper helpdesk.

Stebbings cannot square 50 customer-support vendors over $50 million in funding and 10 over $100 million with Peter Thiel’s monopoly gospel (Decagon, Sierra, Intercom, and the rest). Acharya’s correction: you are calling an industry a market. Legal is infrastructure for capitalism, on the order of $500 billion, not one TAM for one winner. There is room for another dozen specialists, as there are dozens of legal specializations today. Traditional legal software at $50 billion is the old 8-to-12-percent world. He thinks AI-native legal lands somewhere between $50 and $500 billion, closer to 500. Getting to 60, 70, 80 percent of a job is easy; 100 percent is not. A 20 percent productivity gain is showing up so far as a four-day week more than 20 percent fewer jobs, because jobs are bundles of tasks, and someone still has to take the customer to a steak dinner.

Underestimate the market. Overestimate zero-to-one. Then shut up and watch inertia

He does not worship TAM models. Venture consistently underestimates how big markets are and overestimates how easy zero-to-one is. That is why he focuses on Series A: shipped and sold is a dramatic signal, the better point on information versus entry ownership versus price. Seed can be anything. Once something works, companies tend to become greater versions of themselves for a long time. The common VC mistake is not seeing how big “working” can get.

He has made the mistake himself. After the Google acquisition he told his co-founder the stock might go up 10, 15, 20, 30 percent — how much bigger could it get? Credit Karma looked, on a back of the envelope, like a torso product: people check a score once or twice a year, the excellent do not need to look, the terrible do not want to. Then: over 100 million Americans, 50 million quarterly actives, logging in on average four times a month, because the score is a mirror of how you are doing as an adult. He would not have predicted the feedback loop. The underwrite, once a formidable founder is showing nonlinear progress, is inertia. Everything happening today defaults to happening forever. Tie-break in the founder’s direction.

Marc Andreessen’s process, which Acharya found maddening as a new partner: just be right a lot. Process does not matter if you are consistently winning. Amazon, where he was an engineer in 2003, had the same leadership-principle irritant for a 23-year-old: be consistently right, never mind the why. The quality of being right supersedes the mental models.

Win the deal. Do not skip the physics of the market

Asked for his most painful loss: he has not lost a deal in six and a half years at a16z. Stebbings wonders whether that means the risk aperture is too tight. Acharya’s answer is process, not cowardice — plus the pre-existing condition you cannot overcome, a healthy decade-long relationship with a prior investor. In those cases there are no games. Respect it, earn the next round, or both. He is not very elastic on ownership; that is the whole “all the chips in” model. He is elastic on price below a threshold that starts to matter in the hundreds of millions. At early stage — $50, $70, $100, even $120 million — price mostly shows up as next-round impairment. He will tell a founder that 12-on-60 versus 15-on-75 is a wash on check size and a question of what expectations they want to sign up for later. At $200 versus $300, he thinks more about the $500 million growth round than the absolute dollars in.

Triple-triple-double-double is not dead, and you do not have to be Lovable or ElevenLabs to get funded. Calibrate to the market. You have to be top quartile versus your peer set. Bottoms-up consumer can explode. ERP and payroll have slower physics, even if Rampell’s team at Deel turned a slow-boil sale into a fast one. Some new primitives go 10-to-100 or 10-to-200. Getting anyone who is not a family member to pay you is hard; one to five or ten is hard; ten to 100 is tremendously hard. He hates investors who are flip about that. Heuristics are for reading the area under the curve, not for shaming a million of revenue.

Some of the most significant companies are area-under-the-curve companies: 20-year overnight successes that venture lore underestimates while lionizing one-to-100 slope. Figma spent years in the build. What it has now is an n-of-one network-effects product, positioned as work moves from execution (being subsumed by coding agents) toward thinking. He is not sure Dylan Field saw that ten years ago. Those companies still have to show enough momentum to keep raising and enough substance that customers pay upfront and expand.

Series A is supposed to be hard. The risks he will take, in order of what the job actually is: competitive risk (can I win the process), pricing risk (did I make the next round harder), then team, geography, fundraising. Winning deals — trust, market intelligence, first to conviction — is the number-one job. Promiscuous founders are a setup problem, not a 2026 novelty: you need irrational optimism and irrational interest in the domain, or at least outlier commitment, including pure capitalism. A case-study tourist is a bad setup. Repeat founders in the same enterprise domain are alpha — the Clutch team sold to Carvana, then built another auto company and took the shortcuts. In consumer, beginner’s mind and a high willingness to be embarrassed are the advantage. After an exit, dinner-party cool can kill the idea that looks silly.

Use the products. Do not wait for the fully autonomous day

Free advice to investors and founders: use the products, more than ever. Five or seven years ago, a small-business factoring tool was hard to get intuition for. Today, if three new models ship, try three, and make something. Most people still do not.

Agent maximalism — autonomous agents doing everything over long horizons while you chill — is ahead of the present. Humans remain in the loop for exception handling. Instructions are as vague as the way you manage a team. Agents will take low-NPS work you have to do, and they will expand the circumference of ambition until execution and expertise are no longer the constraint. They will not, in his view, break you out of local maxima into global ones. That still takes human intuition. BPOs, with well-defined tasks pulled off a queue in an offshore call center, are set up for replacement. Software development’s hard part is not the coding. He declines to opine on UiPath; he does not know the company well enough, and he thinks vision models have not kept up with the way people talk about them.

Open versus closed is not, yet, a cost-optimization cycle. Closed is still a bit advantaged for maximizing ambition. Open has idiosyncratic product qualities: Kimi K2, in his telling, was not post-trained to restrain what it could say, which made the text more interesting, especially for companion companies. Closed has been cutting costs too — a GPT-4o token down 100x since release. He does not buy Jason Lemkin’s “this is the year we substitute on price.” Capability keeps outrunning cost. Nobody is going back to Sonnet 3.7 because it is cheaper when Opus and Codex keep sparking “what more can we do.” Costs in these products are a feature: they kill field-of-dreams investing. You cannot ship free and invent a business model later. That hygiene did not exist across the board ten years ago.

Would he tell ElevenLabs to subsidize, cut price, and land-grab after a $500 million raise? No. There is more to do at the frontier. Software is 8 to 12 percent of enterprise spend and a few hundred dollars a month of consumer spend. He believes it asymptotes toward 80 to 90 percent of discretionary spend — companionship, entertainment, therapy, healthcare, professional tools, education — not by taking cost out of the current slice. Rent and food stay. Stebbings, European, adds fashion. Acharya adds that nobody buys wine anymore.

You cannot anoint a winner. You can refuse to believe in luck

Kingmaking is overstated. YC is a real catalyst for enterprise startups selling to other YC companies. a16z can connect a small company to the Fortune 500 and 2000; it cannot force the buy. Buyers have perfect information. The right investor is a catalyst, not an anointer. The best founders know how to maximally leverage VCs — Rampell runs what Acharya describes as a sales quota on him, on David George, even on Ben Horowitz. Could Deel do it without them? Of course. Advice to founders: pick an investor that does stuff, verify with other founders, and understand the early gift is a loaned brand. You are an Andreessen Horowitz company until your brand is bigger than that. You still have to win on product.

The most memorable first meeting: Krea, after nine months of not being able to get hold of them, Thanksgiving week, Marc in the room, two people in matching kimonos with Celsius drinks, holding the room as technologists rather than MBA polish. On the partnership: Marc paints the future and has read everything outside technology; Ben’s The Hard Thing About Hard Things is the first honest business book, wartime stories unmatched; David George is a pure-play investor with growth-stage clarity the way Chris Dixon is at early stage. Acharya’s own hardest decision was leaving hands-on building. What sold him was Horowitz’s answer on bad-investor behavior: they do not measure you on near-term returns. Every two years they 360 you with every founder. If founders say you tell the truth, show up, and stay responsive, you are doing the job regardless of company performance. If they say anything else, you are looking for work, also regardless of performance.

The operating rule is not “losing is unacceptable but missing is fine.” He rejects that. They are not allowed to believe in luck. See 100 percent of the deals in the domain. Winning 100 percent of the ones they go after is the expectation. Being wrong after seeing the information is the business. Missing the meeting is not.

The last-twelve-months update is the one that should scare copycats. In mobile, 2008–2009 anointed winners were not the eventual winners — Friendster, then Facebook. In this cycle, early leaders from 2023 and 2024 have kept the lead: Harvey, Gamma, and others he calls impressive. Late 2022 was ChatGPT; 2023 was the obviously good ideas; late 2024 reasoning models (o1, DeepSeek) made some of those ideas actually work; 2025 they scaled. Existing markets — support, chat, creative tools, code — now have early leaders and are hard to enter as another coding or support tool. 2026, he thinks, is native categories that were inconceivable two years ago. OpenClaw and Moltbook are the beginning, not the map. Moltbook-as-point is probably overhyped — “robot dogs barking at each other,” with any humanity borrowed from owners. The slope is underhyped: digital twins having pseudo-dates and coming back to say the humans should meet. Look at the slope, not the point.

What he wants personally is not another agent demo. He has meditated, transcendental style, for 25 or 30 years. The idea that more people could have a slice of that peace — fewer rote parts of life, more of the relationships they find fulfilling — is the thing he will say out loud. The NPS of the human experience, for lack of a better phrase, is on the way up. That is where he wants the bazooka pointed, not at a cheaper payroll rebuild.