• Favor AI inference and infrastructure as a long-term growth theme, with AWS, Crusoe, and Fireworks mentioned as potential beneficiaries; this is a sector thesis, not a specific buy recommendation.
• Look for vertical AI agents with proprietary domain data and measurable results, while recognizing that reliable products and durable advantages remain difficult to build.
• Treat customer-support AI cautiously: the sector was described as crowded, with heavy funding and entrenched incumbents.
• Avoid treating the hypothetical OpenAI and Anthropic IPO valuations as investment targets; the discussion favored OpenAI only in a forced-choice scenario and offered no valuation analysis.

AI Inference and Model-Serving Infrastructure

• Harrison Chase said he had underestimated demand for inference and now sees the market as large enough to support multiple providers. He named Fireworks, AWS, and Crusoe as businesses that could all benefit.
• He also described rising token use and growing demand from coding agents and multi-agent systems. That demand can increase costs even as the price-performance of frontier models improves.
• He identified lower-level infrastructure as an area to consider, naming Modal, Temporal, Supabase, and Convex. He highlighted Modal’s work on sandboxes as well as inference.
• Fireworks was also a podcast sponsor, so its promotional claims in the introduction should be distinguished from Chase’s interview comments.

Takeaways

• The discussion is bullish on inference and AI infrastructure overall, not a recommendation to buy any particular provider.
• For investors, the opportunity is tied to sustained growth in AI usage. A key consideration raised in the episode is that token demand and operating costs may rise even as models become more efficient.

Vertical AI Agents

• Chase was “massively bullish” on vertical agents—AI products built for specific fields such as law or finance.
• He argued that domain expertise, proprietary data, specialized tools, and the ability to measure whether an agent did a good job can help a vertical agent stand out.
• His suggested build sequence was to start with the context and harness—the information and software structure around a model—and consider post-training a model later, once the product has users and enough usage.
• He said vertical agents may be well suited to memory and learning because what an agent needs to remember differs by domain.
• He did not identify particular verticals he would overweight.

Takeaways

• The investment thesis favors companies with deep domain knowledge and a clear way to measure results, rather than generic AI features alone.
• Chase cautioned that building reliable agents remains very difficult. He also said model improvements can make parts of an existing product unnecessary, so companies need to adapt quickly.

AI Coding Tools and Enterprise AI Spending

• Chase said LangChain spends roughly 30% of engineering salary on coding tools, while token use has continued to rise. He described tracking usage, setting budgets, and improving the agent harness to manage costs.
• He noted that some coding-agent usage is highly valuable, while poorly configured tools can create large, unintended bills; one example cost $40,000 in a week.
• He said enterprise buyers still prioritize accuracy, with cost becoming a more prominent concern. Security also matters more to large enterprises than to developers generally.
• In discussing Anthropic’s potential value, he agreed that coding-agent spending is relevant, but added that agents could affect many other enterprise functions over time.

Takeaways

• AI coding adoption may create substantial demand for model providers and tools, but spend alone is not proof of productivity or durable revenue.
• The transcript points to practical areas of focus for businesses: usage visibility, spending limits, and cost controls built into the software.

OpenAI and Anthropic

• In a hypothetical choice between an OpenAI IPO at $1.5 trillion and an Anthropic IPO at $2 trillion, Chase chose OpenAI. His stated reason was that OpenAI has more diverse routes to potentially large outcomes.
• He also said Anthropic and other model providers could benefit if agents expand beyond software engineering into other parts of the enterprise.
• These were hypothetical valuations and a forced-choice discussion, not price targets or a direct investment recommendation.

Takeaways

• The conversation presents a positive long-term case for major model providers, while emphasizing uncertainty about how quickly agent adoption will spread across industries.
• The hypothetical valuation comparison does not establish that either valuation is attractive; the transcript offers no financial analysis supporting those figures.

NVIDIA (NVDA)

• Chase said he expects comparable U.S. open models within 12 months and pointed to NVIDIA’s work with the ecosystem as evidence of attention to the area.
• He described open models as useful today, particularly for narrower tasks such as routing, guardrails, extraction, and summarization. He did not think they were generally driving the core loops of major agents.

Takeaways

• The discussion suggests a potential ecosystem tailwind for NVIDIA from continued work on open models, but it does not provide a stock-specific valuation view or recommendation.
• The interview’s distinction between narrower open-model tasks and more demanding agent work is relevant to assessing how broadly open models are being adopted.

LangChain and LangSmith

• Chase described LangChain as a developer-tools company for building agents, with LangSmith providing observability and evaluation tools.
• He said LangChain had grown from two founders to about 400 employees and described Sequoia and Benchmark as important partners.
• He emphasized that agent reliability is difficult because model outputs are non-deterministic. He sees traces and domain-specific evaluations as tools for improving agents, rather than standalone sources of value.
• He said LangChain’s commercial memory product did not find product-market fit, in part because the company tried to make it too general.

Takeaways

• The episode highlights a potential market for developer tools that help companies monitor and improve AI agents, but also shows how difficult it can be to find demand for a broad, general-purpose product.
• LangChain is presented as a private company; the transcript gives no public ticker or valuation for it.

Fireworks, OpenRouter, and Model Routing

• Chase called model routing a difficult and fast-changing problem, not a commodity. He said the best placement depends on the use case:
  • For vertical agents, he favored routing within the harness, where the system has more context about the task and memory.
  • For general chat or coding, routing through a gateway may work.
• He noted that rapid model improvements can reduce the value of routing if a new model performs better across many tasks.
• OpenRouter was praised for model access, but Chase said much of its usage may come from serving as a gateway rather than from intelligent routing specifically.
• Fireworks was discussed both as a provider and in a sponsored advertisement.

Takeaways

• The investment case for routing depends on whether a provider can deliver useful, task-aware model selection as models change.
• The transcript suggests a distinction between gateway businesses that simplify access and routing products that decide which model should handle a request.

NVIDIA, Open Models, and Model Post-Training

• Chase said many companies are post-training open-weight models, including models developed in China, rather than training models from scratch. He described post-training as a later step for companies that already have users, data, and traction.
• He sees open models as especially useful for higher-volume, narrower tasks, including classification, extraction, summarization, routing, and guardrails.
• He said some enterprises are wary of Chinese models, but argued that concerns about data exposure are less relevant when models are hosted by a provider or on a company’s own infrastructure. He also said evaluations can help test whether outputs fit an organization’s needs.

Takeaways

• Open models may create opportunities in focused, high-volume tasks, but the episode does not suggest they have broadly displaced frontier models in complex agent work.
• For investors, the discussion points to a potential advantage for companies with proprietary data and enough usage to justify model specialization.

Public Software Companies Mentioned

• Salesforce (CRM) and Atlassian (TEAM) were cited as large incumbents in customer support software, alongside newer providers.
• Uber (UBER) was mentioned in the podcast’s advertisement for Fireworks, not as an investment recommendation.

Takeaways

• The episode provides no stock-specific outlook, valuation, or recommendation for these public companies.
• Salesforce and Atlassian were mentioned as examples of incumbent competition in customer support, a market Chase characterized as crowded.

Customer Support AI

• Asked where he would avoid investing, Chase called the customer-support space “wildly uninteresting”, citing around 20 providers that had each raised more than $100 million in the prior 18 months, as well as large incumbents and established software vendors.

Takeaways

• This was the clearest bearish sector view in the interview: Chase sees heavy funding and entrenched competition in customer-support AI.
• The concern is sector crowding, rather than a specific company’s financial performance.

Robotics and Humanoids

• Chase said substantial funding was going into robotics, but argued that the sector was still far from a “ChatGPT moment.” He was particularly skeptical of humanoids, saying funding and excitement appeared to exceed current capabilities.

Takeaways

• The interview expresses caution about near-term expectations for humanoid robotics, especially where investment enthusiasm is ahead of what the technology can currently deliver.
• No companies, price targets, or investment timelines were specified.

AI Infrastructure and Data Companies

• Chase named ClickHouse as a company doing very well, while describing it primarily as a traditional database provider.
• He also mentioned MongoDB and Elastic as databases suited to storing and querying software-generated data.
• He argued that the harder, still-unsolved opportunity is an intelligent system that can decide what information agents should retain—not simply the database that stores it.

Takeaways

• The discussion suggests that agent-related data infrastructure may extend beyond conventional databases to systems that organize and distill information.
• Chase’s comments distinguish existing database businesses from the less-developed “memory” layer; they do not amount to a specific recommendation on ClickHouse, MongoDB, or Elastic.

Sequoia and Benchmark

• Chase said founders should consider taking a discount to work with Sequoia or Benchmark, based on his experience with their recruiting, company-building support, and ongoing involvement.
• He described Benchmark as a close sounding board and Sequoia as especially helpful with recruiting and launches.

Takeaways

• This is fundraising advice for founders, not a public-market investment recommendation.
• The potential value discussed is the firms’ support and networks; the transcript does not compare fund performance or specify an appropriate discount.

Other Companies Mentioned

• Ramp and Harvey were cited as companies experimenting with post-training models for their domains. Chase said companies like these may move to that stage after getting substantial value from context and harness improvements.
• Modal, Temporal, Supabase, and Convex were named as examples of lower-level infrastructure businesses.
• Suno was mentioned as an example of creative AI that Chase finds interesting.
• Jev was discussed in relation to a hypothetical $9 billion round. Chase’s bull case was that the company helped define a new model category and had a talented team; he said he did not know the team well. The company name is reproduced as transcribed.

Takeaways

• These comments are examples of areas or companies the guest finds notable, not formal buy recommendations.
• The episode’s broad investment preference is for technically strong teams that execute quickly and build a recognizable brand, while recognizing that early attention alone does not ensure lasting success.