Anish Acharya on AI Software, Investment Judgment and the Limits of Automation

Harry Stebbings asks Anish Acharya, a general partner at Andreessen Horowitz, whether it is better to build an AI company outside San Francisco, where talent may cost less and stay longer. Acharya disagrees, citing the concentration of technical talent and informal knowledge in the Bay Area. He acknowledges that strong companies can be built elsewhere; his preference is a judgment, not a controlled comparison.

Acharya names Tel Aviv as another place where he sees an ambitious technology ecosystem. He argues that a smaller domestic market can push founders to think internationally early. Stebbings asks whether a $3 billion to $5 billion company is still a sufficient venture outcome, and Acharya calls it extraordinary even while noting that a trillion-dollar ambition requires different assumptions.

On the software sell-off, Acharya argues that markets have become too pessimistic about the prospect of AI replacing every existing application. He estimates enterprise IT spending at roughly 8%–12% of total spending and says companies may prefer using AI to improve their core operations rather than rebuild payroll or enterprise systems. The percentage is his estimate, not a verified measure for every organization.

Stebbings asks whether slower growth and seat contraction justify investor concern. Acharya cites his own review suggesting that 75% of public SaaS companies raised prices after ChatGPT’s launch, with a mean increase of 8%–12%. The interview does not supply the dataset, dates or methodology, and price increases alone cannot establish demand, profitability or market power.

Acharya says capable incumbents can improve existing workflows, while acknowledging that some business models and seat-based pricing may face pressure. He believes the risks of rebuilding complex systems often outweigh the savings. This is a competitive assessment from a venture investor, not proof that individual software vendors will retain customers.

He identifies a different potential disruption: AI-assisted systems integration could reduce the cost and risk of switching between established software providers. His SAP-to-Oracle example is an argument about what coding agents might enable, not evidence that such migrations are already easy, safe or inexpensive in every case.

The guests revisit the question of whether incumbents acquire innovation before startups acquire distribution. Acharya expects capable incumbents to improve existing product categories, while startups may lead in categories that did not previously exist. His example of AI filmmaking is a prediction about competitive outcomes, not a verified market result.

Acharya says he does not know whether applications will create more value than foundation models, but believes application value is underappreciated. He argues that multiple model providers and open models can offer substitutes for some tasks and specialized capabilities for others. His rough 80%–20% split is an illustration, not a measured market allocation.

He gives coding and creative tools as examples where an application could route work among models and provide a unified interface. His comparisons of model strengths reflect his experience and can change as products evolve. The argument assumes users value orchestration enough to pay for it; it does not establish that every application will be defensible.

Stebbings worries that competing coding tools could lose revenue quickly when a new product improves. Acharya responds that demand and ambition can grow as tools become more useful, and expects different developer products to find distinct users. Both are predictions about a rapidly changing market rather than evidence of a particular company’s future revenue.

Acharya compares AI infrastructure and developer tools with cloud markets, where several suppliers have coexisted, rather than assuming a two-company ride-hailing model. He thinks specialization can support multiple participants. This analogy is his view of market structure; it does not establish future margins or rule out consolidation.

On investing in competitors, Acharya says a firm that actively helps portfolio companies may face difficult conflicts over customers, staff and shared resources. He says a16z avoids directly competing investments in that context, although markets can diverge over time. The statement is his description of his firm’s approach and interests.

Stebbings asks whether model providers can absorb application features. Acharya discusses Granola, explicitly saying his firm is not an investor, and speculates that its ambitions may extend beyond meeting transcription. He argues that a specialized application could build a broader feature set and use multiple models; his description of Granola’s strategy is an assumption, not a company announcement.

He acknowledges the appeal of a bundled product with most of the needed features but thinks model providers have many competing priorities. Specialized interfaces and access to several models may help applications compete. That is an interested venture investor’s thesis, not proof that the large model companies cannot build or bundle similar products.

Acharya argues that startups may build unusual consumer experiences that large companies avoid. He cites AI companionship products and suggests they could support reflection or social connection. These are his product and social judgments; the conversation does not establish that any specific companion is safe, therapeutic or beneficial for all users.

Asked whether he would encourage a child to use such a product, Acharya imagines a contextual AI companion in Minecraft that could model prosocial behavior. Stebbings raises the risk that reliance on technology could weaken human interaction; Acharya disagrees. Their differing views are hypotheses, not evidence of developmental outcomes or child-safety guarantees.

The guests also consider companionship for older adults. Acharya imagines context-specific conversations that might be more welcome than a generic check-in. His examples are speculative design ideas, not validated care, medical monitoring or a replacement for human support.

On interfaces, he thinks voice can work well in enterprise settings but that consumer chat and fully dynamic interfaces may be overestimated. Many people browse because they want to explore or spend time rather than specify a precise intent. This is a product-design view, not a measured prediction that one interface will dominate.

Acharya believes defensibility still matters, particularly network effects. He contrasts a marketplace such as Airbnb with systems of record that may vary in how difficult they are to replace. He argues that live, proprietary data could improve an application’s usefulness, while acknowledging uncertainty over how exclusive such data can remain.

Turning to margins, Acharya compares earlier venture-backed spending on advertising with current free trials or subsidized inference. He argues that subsidized usage can help acquire paying users, though it may depress blended gross margin. Whether that spending is efficient depends on conversion, retention, cost and lifetime value; the interview offers no universal proof.

He says AI power users may pay much more than traditional consumer-subscription customers and cites several then-current product prices. Those are time-specific examples, not current prices or evidence that most users will pay the same amounts. He believes acquisition spending can be justified if high-value users remain and contribute sufficient margin.

Acharya suggests separating the cost of free trials from the margin earned on paying users when evaluating acquisition economics. He proposes treating early organic traffic differently from paid acquisition and using later cohorts to assess retention. This is an analytical framework, not a change to accounting rules or proof that organic acquisition is costless.

Stebbings questions how to estimate lifetime value when product leadership changes rapidly. Acharya emphasizes retention and suggests examining month-two users after an initial wave of trial traffic, with month-twelve retention as a further signal. His example thresholds are investor preferences and depend on cohort definitions, not universal benchmarks.

Asked about an AI bubble, Acharya says he does not believe the market is in one at the time of the interview. He cites reported demand, capacity and prices as reasons, while noting that the subject is outside his main expertise and some subsidy exists. His account of a provider’s revenue and capacity is unverified here and does not rule out overinvestment.

He argues that spending can move beyond traditional software budgets toward work previously performed by people. For voice agents, he imagines combining support, sales, operations and collections rather than automating one isolated function. His suggested productivity gains are speculative; the interview does not establish a tenfold improvement or universal job replacement.

Discussing legal and customer-support markets, Acharya distinguishes an industry from a single product market and expects room for multiple specialized companies. He and Stebbings discuss a roughly $500 billion legal-market figure and a smaller software-spending figure as broad estimates. How much labor spending AI companies can capture is an open question, not an established addressable market.

Acharya says automating an entire job can be harder than completing some of its tasks, and suggests productivity improvements may sometimes show up in working patterns rather than immediate job cuts. His reference to a four-day week is illustrative, not a measured economy-wide outcome. The eventual effects on employment and work hours are uncertain.

On market sizing, he says investors often underestimate eventual market size while overestimating how easy it is to build a working product. He prefers Series A because a shipped product and paying customers give him more evidence than a pure seed-stage idea. That preference reflects his own investment mandate and does not guarantee a better return.

Stebbings describes passing on Deel and other opportunities after misjudging market size. Acharya, who has invested in and serves on the board of Deel, says momentum and a capable founder can reveal an opportunity broader than an initial estimate. Their retrospective accounts are personal investment experiences, not proof that early growth always persists.

Acharya recalls underestimating Google’s potential scale and discusses Credit Karma, where he previously worked, as another example of behavior that surprised him. He cites user and engagement figures from memory and offers a theory that people value the feedback of checking their credit score. These are his retrospective observations and estimates, not independently verified current metrics.

He says he often uses momentum as a mental model when a founder repeatedly makes progress, while Stebbings recalls similar advice about founders meeting targets. Past execution may be useful evidence but cannot establish that growth will continue indefinitely. The speakers present this as a heuristic for judgment.

Acharya recalls asking Marc Andreessen for an investment process and receiving the advice to be right frequently. He interprets this as a focus on outcomes that can transcend a particular checklist, while describing a similar lesson from his time at Amazon. This is his recollection of mentors and his interpretation, not evidence that process is unimportant.

He says that during his time at a16z he has not lost a deal he pursued, a personal claim whose scope and definitions are not independently verified. Stebbings asks whether never losing could mean taking too little risk. Acharya says existing founder relationships can be difficult to overcome and advocates earning trust rather than playing games.

On deal terms, Acharya says he is relatively flexible on early-stage price but less flexible on ownership. He thinks a high entry price may matter especially because it raises the bar for the next round. His illustrative financing terms describe his own framework, not a recommendation or proof that price is irrelevant.

The guests discuss companies that may grow quickly enough to skip an intermediate valuation step. Acharya says this can work if core metrics justify it. Asked about the traditional triple-triple-double growth pattern, he argues that fundraising expectations should depend on the market, sales cycle and peer group rather than a single growth rule.

He emphasizes that reaching the first million dollars of revenue and then growing further is difficult. Some businesses may grow more slowly initially but create substantial value over a longer period; he calls attention to the “area under the curve.” This is a way to compare possible trajectories, not a guarantee that slower growth will eventually compound.

Stebbings argues that Series A can be particularly competitive and expensive. Acharya separates competitive, pricing, team, geographic and future-fundraising risks. He thinks winning the right deals is a core investor responsibility, while recognizing that a founder’s ability to raise later capital can affect whether a business gets time to execute.

On founders moving quickly between ideas, he says commitment to the problem or to another deeply held goal can matter when markets become difficult. He does not require every founder to share the same personal motivation, but is skeptical of an idea chosen only because a superficial analysis looks attractive. These are his founder-selection preferences.

Acharya says founders can fundraise effectively in different styles and recalls the Krea founders’ unconventional first meeting. He later notes an investment relationship with Krea. His praise is a personal account of a portfolio team, not evidence that a particular appearance or pitch style predicts success.

He sees repeat founders working in a familiar enterprise domain as potentially advantaged because they know customers and shortcuts. In consumer products, he also values a beginner’s willingness to experiment with ideas that might initially seem embarrassing. He offers these as contrasting possibilities, not demographic rules.

Acharya says he now spends more time directly using new products and models to build intuition. He believes founders and investors should test tools themselves when possible. Personal use can reveal product details, though the conversation does not claim it replaces customer research or technical evaluation.

He cautions against the most expansive claims for autonomous agents that work across long time horizons without supervision. He expects humans to remain in the loop for exceptions and to clarify vague instructions, although he also predicts agents will perform some unwanted tasks and expand what people attempt. These are judgments about current and future capability.

Asked whether agent companies must own an entire technology stack, Acharya says they can also use other providers’ tools. He distinguishes well-defined process tasks from ambiguous work that involves deciding what to build or how to respond to new information. He expects human judgment to remain useful in ambiguous settings; this is not a formal benchmark of agent capabilities.

On open and closed models, he says organizations often prioritize capability over cost at the current stage, while some open models offer distinctive characteristics or control over hosting. His comments about particular models and price reductions are time-specific personal assessments, not verified general cost comparisons. Different workloads and deployment choices can produce different economics.

Stebbings relays Jason Lemkin’s concern that a voice product was too expensive for a game. Acharya disagrees that cost optimization will necessarily dominate immediately, arguing that new capability can justify higher prices. He also says ongoing inference costs force founders to establish a business model earlier; that is his view, not a guarantee users will pay.

Acharya predicts that software could account for 80%–90% of consumer discretionary spending as it expands into entertainment, education and other services. He clarifies that he is discussing discretionary spending rather than rent and food. The interview supplies no model or time frame for that striking estimate, so it remains a speculative forecast.

On “kingmaking,” Acharya says investors can introduce startups to customers and lend credibility but cannot force customers to buy an inferior product. He gives accelerator networks and a16z’s customer introductions as examples of potential assistance. His description reflects his firm’s commercial role and is not evidence that its backing guarantees a winner.

He says effective founders know how to use their investors and cites Deel founder Alex as an example, noting how founders can seek introductions. Acharya has a board and investment relationship with Deel. He recommends that founders speak with other founders to understand what an investor actually does, then judge the product and customer response themselves.

Asked about investing mistakes, he says he was sometimes too casual in assuming product-market fit in 2021. A plausible founder theory and his own enthusiasm made it easy to treat a product as working before it had demonstrated that. He distinguishes an explicit early-stage bet from self-deception about the evidence.

Acharya says he has not found the a16z brand a disadvantage in his own experience and praises its founders. In the quick-fire round he describes memorable meetings and the different strengths he sees in colleagues. These are personal views about people and an institution with which he is professionally affiliated.

He recalls asking Ben Horowitz how the firm addresses poor investor behavior and describes a founder-feedback process conducted every two years. This is his account of internal practices, not an independently audited performance system. He says the prospect of being evaluated on responsiveness and honesty made the move from operating to investing attractive.

Acharya describes a16z’s aspiration to see all deals in its domain and win those it pursues, while saying good-faith investment decisions can still turn out wrong. The 100% figures are stated expectations, not independently verified coverage or success rates. He also explains why consistent seed investing is difficult when a product and go-to-market evidence may not yet exist.

He says early leaders in the current AI cycle have persisted longer than he expected, while predicting that new native categories could emerge. On Moltbook, he distinguishes excitement about a particular product from the broader idea of digital agents interacting. His imagined dating-agent example is a scenario, not evidence of successful matching or safe deployment.

Stebbings hopes for medical breakthroughs relevant to his mother’s multiple sclerosis, while Acharya expresses optimism about technology improving daily life and relationships. These are personal hopes, not evidence of a new treatment, clinical efficacy or guaranteed improvements in well-being.