Bitter Lessons in Venture vs Growth: Anthropic vs OpenAI, Noam Shazeer, World Labs, Thinking Machines, Cursor, ASIC Economics — Martin Casado & Sarah Wang of a16z

Episode Summary

Executive Summary: A16Z partners Martin Casado and Sarah Wang discuss how AI investing has blurred traditional lines between venture and growth, infrastructure and apps, and even product and research. They argue the field is now driven by capital flywheels, compute contracts, and founder talent wars, while also highlighting underinvested areas like boring enterprise software and the still-uncertain robotics/hardware stack. They end on how AI apps like Cursor and Anthropic are reshaping model economics.

Main Topics: Venture vs. growth blur in AI (Priority: 5/5): The speakers argue that frontier AI rounds are no longer clean venture or growth deals; they combine early-stage founder bets, large capital needs, BD, compute procurement, and strategic investors. Compute, capital flywheels, and market demand (Priority: 5/5): They debate whether circular financing is a problem, concluding the key question is real demand for compute and whether capital can be translated into capability gains and revenue. Infrastructure vs. apps and the rise of platform companies (Priority: 5/5): Model companies are described as simultaneously infrastructure and application businesses, with blurred lines between API revenue, vertically integrated apps, and ecosystem/platform effects. Founder talent wars and personnel churn (Priority: 4/5): They emphasize unprecedented founder movement, poaching, and compensation pressure in AI, driven by AGI ambitions, strategic money, and intense visibility on social media. Underinvested areas: boring software and robotics (Priority: 4/5): They suggest traditional enterprise software is underappreciated by investors, while robotics/hardware remains hard to diligence because it is highly vertical and lacks a clear ChatGPT-like breakthrough. Model economics, specialization, and the 'token path' (Priority: 5/5): The discussion explores whether general frontier models will dominate everything or whether specialized layers/apps can still win by building on top of model APIs and extracting margins. World Labs, spatial intelligence, and multimodal models (Priority: 3/5): Sarah explains her coding work on Gaussian splats and 3D scene generation, using it as a case study for how generative AI can drastically lower the cost of creating valuable content.

Key Arguments: AI rounds now often mix venture and growth because companies need early growth-scale resources, BD, and compute contracts almost immediately. Compute spending is not inherently bad if there is real demand; unlike the dot-com era, there is no obvious supply overhang of GPUs. Capital can now be traced more directly to capability improvements, making R&D spend feel more like a measurable growth investment. The core open question is whether frontier model layers can outspend and subsume the app ecosystem built on top of them. Founder motivations are more unified around AGI than in past waves, which increases churn, poaching, and anxiety. Boring enterprise software is likely underfunded because investors over-index on extremely fast growth, despite attractive returns in large markets. Robotics and hardware are important but hard for generalist investors because they are vertical and require deep market-specific diligence. Cursor is cited as proof that an app-layer company can build its own model and win by focusing on a specific workflow and user base. Agentic/app companies may enjoy better margins than model labs because they price against human labor rather than per-token commodity pricing. Model companies may still win if they can keep raising more capital than the aggregate spend of all downstream users of their models. Spatial intelligence likely requires different representations than language, even if language models can help with coding and tooling around it. 3D generation has huge economic upside because it can reduce the cost of creating valuable 3D content by several orders of magnitude.

Data Points: Time working together: 7 years - Sarah and Martin said they have worked together on AI infrastructure for seven years. Years doing startups: 20 years - Martin said he has been doing startups for two decades. Years doing this investing work: 10 years - He said he has been doing this kind of investing for a decade. A16Z team size: 600 people - Sarah mentioned A16Z has around 600 people. Open recruiting roles: 3 - Sarah said she currently has three open recs on the investing team. Reported offers in market: $10 million/year - Sarah said some people they speak with have active offers around this level. Reported L5 offers: tens of millions - They said even L5-level AI talent can get offers in the tens of millions. Character investment timing: January 2023 - Sarah said A16Z invested in Character in January 2023. Character Google IP licensing deal: August 2024 - She noted Character signed the Google licensing deal in August 2024. Typical revenue scale after GA: tens of millions - A company was described as reaching tens of millions in revenue within weeks of going GA. Model cost vs. image generation: 100x cheaper / a hundredth of a penny inference - Martin described image generation economics as dramatically cheaper than human-created equivalents. 3D content cost: $4,000 to $10,000 amateur; $30,000 professional - Sarah estimated the cost to recreate a room in 3D at different quality levels. Potential 3D generation cost: less than $1 - They argued generative 3D could slash marginal cost by 4–5 orders of magnitude. Global market growth example: 5x - Sarah argued a company growing 5x in a large market can be a strong investment even if not hypergrowth by VC Twitter standards. Anthropic model market share assumption: large percentage - Used as a hypothetical to discuss whether a frontier model provider could outspend downstream apps.

Pivotal Quotes: "There are these kind of large compute contracts, which can take months to do. And so it's very different times." — Martin Casado: On why AI fundraising and commercialization now resemble a hybrid of venture and growth with strategic procurement complexity. "If I can raise more money than the aggregate of everybody that's using it, I will consume them whether I'm a GI or not." — Martin Casado: On the possibility that frontier model companies can outspend downstream app ecosystems even without full AGI. "The lines are blurring even more today, where everyone is... but of course, these companies all have API businesses." — Sarah Wang: On how model, infrastructure, and application categories are increasingly intertwined in AI.

Implications: AI investing is shifting toward capital-intensive, high-speed flywheels where compute, talent, and product launch speed matter as much as code. Expect continued concentration at the frontier, but also room for specialized apps, enterprise software, and 3D/spatial tools that translate model gains into real products.