The Twenty Minute VCa16z, Anish Acharya: Is SaaS Dead? Do Margins Still Matter? Why We Are Not in an AI Bubble?
CHAPTERS
Why SF still matters for AI founders: cities as the original network effect
Anish pushes back on the idea that it’s better to build outside San Francisco due to cheaper, stickier talent. He argues that in this AI moment, proximity to builders, information flow, and the cultural “selection” of committing to SF create compounding advantages.
- •SF’s density creates a builder network effect and faster access to tacit knowledge (“whispered down shadowy hallways”)
- •Founder commitment signal: relocating to SF can reflect singular focus
- •You can win from anywhere, but probability and pace can differ
- •Tel Aviv as another ambition-dense hub with strong global-first mindset
The “SaaS Apocalypse” is overstated: why vibe-coding payroll/ERP is the wrong target
They tackle public-market pessimism around SaaS durability. Anish argues enterprise software is “oversold” as a narrative: even full rewrites only touch a small slice of enterprise spend, and AI ROI is bigger when applied to core advantage and the other 90% of costs.
- •IT/SaaS is ~8–12% of enterprise spend; savings from rewriting systems are capped
- •AI should optimize the larger non-software spend and expand business advantage
- •Some SaaS models will lose (e.g., seat-based pricing), but broad collapse is unlikely
- •Public-market negativity doesn’t equal fundamental obsolescence
SaaS pricing power and the real AI impact: lower switching costs, fewer “hostages”
Anish cites evidence that many SaaS companies have raised prices since ChatGPT, which he treats as a PMF signal rather than desperation. The more structural AI impact he highlights is reduced integration and migration complexity, which weakens legacy lock-in.
- •~75% of public SaaS companies have raised prices post-ChatGPT; many by 25%+
- •Capable incumbents aren’t ‘Sears’—examples like ServiceNow can still guide and ship
- •Agents reduce SAP/Oracle-style switching friction, time, and risk
- •Result: more competition, fewer hostage customers, healthier incentives
Incumbents vs startups: incumbents improve categories; startups create new ones
Using historical pattern-matching, Anish argues capable incumbents typically use new tech to make existing products better, while startups capture the “native categories” enabled by the new cycle. The battle is less about feature parity and more about category creation.
- •Incumbents tend to reinforce existing workflows (better Word, better Search)
- •Startups win net-new categories without established incumbents
- •Examples: AI-native movie-making likely won’t be led by Adobe, though Adobe improves Photoshop
- •Distribution matters, but category definition often matters more
Why apps still win in a multi-model world: aggregation and orchestration as value
Anish explains that the risk of one foundation-model monopolist has eased because models are now substitutes for many tasks and specialists for others. This fragmentation creates room for application companies to orchestrate multiple models and deliver integrated workflows.
- •Competition among model providers limits extreme supplier power
- •80% substitution + 20% specialization creates a need for an “aggregation layer”
- •Coding example: different models excel at front-end vs back-end; orchestration reduces friction
- •Creative example: opinionated vs non-opinionated models (Midjourney/Krea vs Ideogram) suit different jobs
Who wins dev tools: Cursor vs Claude Code and why the market looks like Cloud, not Uber
Harry argues developer tool revenue may be fragile (e.g., churn from Cursor to Claude Code). Anish believes the market will fragment by archetype and workflow preferences—more like cloud oligopoly dynamics than pure-substitute price wars.
- •Demand for building/consuming software expands with AI; ambition doesn’t stay fixed
- •Multiple tools can coexist because developers differ (IDE-centric vs CLI/“closer to metal”)
- •Uber/Lyft is the extreme substitute model; cloud shows multiple winners with margins
- •Competitive investing: hard for “service-heavy” firms to back direct competitors, but product divergence is fast
Models invading the app layer: primitives get copied; feature surface and multi-model win
They discuss whether foundation model labs will vertically integrate into apps (e.g., legal assistants, meeting transcription). Anish argues labs can replicate primitives and market them, but many domains require deep feature surface, prioritization, and multi-model flexibility that labs may not pursue.
- •Example: meeting transcription is commoditized, but productivity suites require broad product build-out
- •Labs may do “primitive + marketing,” not full workflow products
- •Vertical UIs (e.g., legal) demand sustained prioritization and opinionation
- •Multi-model support is a durable differentiator vs single-lab offerings
“Weird wins”: companionship, contextual agents, and products big tech won’t build
Anish flips “boring wins” into “weird wins,” arguing AI’s human, emotional capabilities unlock categories that large companies avoid due to brand constraints. They explore companionship products and contextual companions that improve social outcomes rather than replacing them.
- •AI enables emotionally resonant products (persuasion, disagreement, sexuality) that big tech avoids
- •Companionship products: Character/Janitor/Replica as examples of demand and discomfort
- •Contextual companion idea: an AI that plays Minecraft with a child and models pro-social behavior
- •For seniors, “indirection” (shared context) can preserve dignity while delivering care and connection
The “death of the chatbox” (in consumer): browse-based UIs still dominate
They debate the future interface paradigm (voice, chat, dynamic UIs). Anish argues chat/voice is powerful in enterprise and intent-based flows, but consumers often prefer browsing and “spending time,” not maximizing efficiency.
- •Voice is a strong wedge in enterprise; consumer UI change is overstated
- •High-agency builders over-index on chat as the ‘optimal’ UI
- •Many consumers can’t articulate intent; browsing helps discovery
- •Future likely mixes intent-based (maybe chat) with persistent browse-based experiences
Moats and defensibility in AI: networks endure; “proprietary + live data” strengthens
They address whether AI destroys moats and switching costs. Anish argues classic moats—especially network effects—still matter, while certain systems-of-record may be more disruptable depending on workflow entrenchment; he also elevates live proprietary data as increasingly powerful.
- •Network effects remain the gold standard (e.g., Airbnb)
- •Some systems-of-record are riskier if they lack human workflow entanglement
- •Banks’ core systems remain defensible due to accuracy demands and deep integration
- •“Proprietary + live data” can beat frontier models without that data access (e.g., health data)
Do margins matter? Inference as the new S&M, power users as the new profit engine
Anish reframes margin analysis: free/negative-margin usage can be treated like CAC if it converts to high-paying power users. They discuss pricing ceilings breaking in consumer AI (hundreds/month) and the importance of separating trial costs from durable unit economics.
- •AI companies may show worse blended margins due to subsidized trials/credits
- •Treat inference spend for trials as CAC; analyze post-conversion margins separately
- •Power users now pay far more (e.g., $200–$300/mo tiers + consumption)
- •Retention remains key; consider “M2 as the new M1” due to tourist top-of-funnel traffic
Not an AI bubble (yet): demand meets supply, prices rise, and spend shifts from labor
Anish argues this cycle doesn’t match classic bubble dynamics because added compute capacity gets absorbed quickly and pricing is not collapsing. They also discuss AI driving a shift from SaaS budgets into labor budgets via productivity gains, especially through voice and function consolidation.
- •Bubble test: supply isn’t outrunning demand—capacity increases are quickly consumed
- •Prices customers pay are going up, not compressing as in overbuilt markets
- •Subsidization exists, but is “healthier” than 2021 ad-spend subsidy
- •Big productivity comes from bundling functions (sales/support/collections/ops) into unified goals like CAC improvement
Why legal and support will have many winners: industries vs markets and specialization
Harry questions why customer support has so many funded competitors; Anish responds by distinguishing “industry” from “market.” In huge industries like legal, specialization ensures multiple large winners, and software spend can expand far beyond legacy legal-software budgets.
- •Many ‘markets’ are actually broad industries with multiple durable sub-markets
- •Legal as “infrastructure for capitalism” supports many specialized winners
- •AI spend likely expands legal software beyond ~$50B toward a much larger share of ~$500B activity
- •Job automation is partial; productivity gains may show up as shorter weeks before mass job removal
a16z operating philosophy: being right beats process, and winning deals is non-negotiable
Anish shares Marc Andreessen’s maxim (“just be right a lot”) and discusses how performance can supersede formal process. He describes a16z’s internal expectations: see every deal in-domain, win the ones you pursue, and use brand and services as leverage for founders.
- •Marc’s view: outcomes matter more than an elegant process if you consistently win
- •a16z expectation: “see 100% of deals in our domain” and win pursued deals
- •Founder/investor fit: no games—respect prior relationships, earn the right for the next round
- •How VCs help: lend brand early, connect to Fortune 500, but can’t ‘force’ a product to win
Rapid-fire: open vs closed source, kingmaking limits, and what’s next in AI categories
In quick-fire and closing segments, they cover open vs closed model adoption, skepticism about price-driven substitution, and whether “kingmaking” is real. Anish predicts early leaders from 2023–2024 may remain dominant, while truly new AI-native categories emerge in 2026.
- •Open vs closed: open isn’t mainly about cost yet; sometimes it wins on idiosyncratic product qualities
- •Price substitution is less likely while capability improvements unlock bigger value than cost concerns
- •Kingmaking: investors can catalyze (e.g., intros, credibility) but can’t anoint a non-winner
- •Market cycle view: 2023 obvious ideas, 2024 reasoning unlocks, 2025 scaling, 2026 new AI-native categories (e.g., ‘digital twins’/agent-to-agent like “Maltbook”)
