Jack Altman and Ev Randle on AI markets, security and the physical world

Benchmark partners Jack Altman and Ev Randle introduce Benchmark Breakfast as a public version of their firm’s open-ended Monday conversations. Rather than following a fixed pipeline agenda, they want to discuss the questions and companies that came up during the week.

Randle explains that Benchmark partners spend hours together each Monday without a set agenda. They see the meeting as an ongoing conversation across weeks, rather than a checklist of prospective investments. The podcast is intended to share that style of discussion.

The first topic is Anthropic’s reported S-1 filing. The hosts say assembling the financial information in one place made the company’s growth and operating leverage easier to assess. Their discussion is an investor interpretation, and the episode does not independently audit the numbers.

They explain operating leverage as substantial upfront spending followed by revenue growth that can support profitability. Randle points to a chart comparing quarterly revenue with total compute and infrastructure expense, including both training and serving costs.

The figures he cites suggest the relationship changed substantially between early 2025 and the second quarter of 2026. The hosts see that as evidence that a frontier-model business can work economically at scale. They distinguish that interpretation from a prediction about how its shares will perform.

The hosts say the reported revenue growth appeared sufficient to offset major fixed training costs in the period they are discussing. Randle also notes a reported operating-margin figure. These observations relate to a specific filing period and do not answer every question about future costs.

Randle says growth volatility and consumption-based revenue remain questions for public investors. He contrasts unusually rapid sequential growth with a later, slower rate and asks how much future revenue visibility can be inferred when customers pay for usage rather than only long-term commitments.

Altman is surprised by how much customer momentum seems to follow the best available model, even when the previous model can perform many tasks. The hosts suggest that users may pay more for a marginal capability gain when the possible upside from a task is very large.

They distinguish tasks where success has a relatively clear ceiling from work such as coding or discovery where additional capability may create more value. At the same time, Altman cautions that buyers have not demonstrated productivity gains in every domain simply because token sellers are growing.

Altman says that in some tasks being correct is enough, while in others a better result can create additional value. In his view, that difference helps explain why buyers may still prefer a stronger model even if an older one is already competent.

The hosts compare code generation with the rapid adoption of video meetings during the shift to remote work. A market can experience a pull-forward and eventual saturation, but they are unsure how much of the wider economy has adopted AI coding or how many agents a team might ultimately use.

Randle describes startups in which founders run several coding agents while doing other work. He sees software factories as a different organizational pattern from traditional development. That observation comes from their startup network and does not establish that every company will work this way.

The hosts debate whether coding-agent usage could continue to expand even among early adopters. More agents per person and new ways of organizing work might increase demand, but neither treats a particular growth rate as assured.

The discussion turns to the “AI bottleneck trade”: investors followed demand from GPUs to servers, networking, cooling, power, memory and other components. The hosts cite historical share-price moves as examples of how market attention shifted; those figures are time-specific observations, not investment recommendations.

They draw a parallel in venture markets around code generation. Databases, deployment, inference, data, browser tools, code review and security can all benefit if software creation expands. They name several companies and disclose connections to their portfolio, so the examples are not an independent sector ranking.

They observe that several companies in the same category can grow quickly at once when the overall market expands. A product tailored to a different user or workflow can coexist with a general coding tool, in their view, though future competitive dynamics remain uncertain.

Randle calls AI code generation a historically consequential product-market-fit moment. The hosts argue that digital distribution may allow adoption and related demand to spread faster than technologies constrained by physical-device rollout. That is their interpretation of the current cycle.

Rapid growth in one category changes opportunity costs for founders and investors elsewhere. The hosts note that a software growth trajectory once considered exceptional can look less impressive next to unusually fast AI companies. They recognize the danger of treating today’s comparison as a permanent standard.

They consider a trade-off between exposure to a fast-growing AI trend and older signs of durability, such as long contracts, switching costs and enterprise stickiness. Altman says durability is harder to judge now because perceptions of application software can change quickly as model capabilities improve.

A conventional enterprise investment may look durable because customers take time to choose a product and then keep it for years. Altman questions whether that intuition transfers neatly to a period when models and application interfaces are changing quickly.

The hosts describe two possible investment mindsets: participate in a trend already pulling demand, or try to identify the next one before it becomes obvious. They mention robotics and biology as areas they believe could become important, while acknowledging that timing and outcomes are uncertain.

Randle notes that once an investment theme is widely understood, the obvious companies may already have expensive valuations. Investors can still believe the market underestimates the size of an opportunity, but the price paid changes the risk.

On “circular” AI revenue, one concern is that companies within the AI ecosystem buy from one another while outside customers have not yet proved lasting value. The hosts say spending by banks or health systems with independent budgets could be a stronger signal of broad demand.

The circularity concern is about dependence: if spending only circulates among AI companies, weakness in one part of the ecosystem could affect the others. The hosts therefore ask whether customers outside the sector are obtaining enough value to sustain demand.

They also offer the opposing view: young AI-native companies may become large customers in their own right. The comparison to internet-native businesses illustrates why early demand from a new ecosystem is not automatically empty, although end users still need to receive real value.

Altman and Randle then try to make the strongest case for closed and open models rather than treating the debate as binary. They disclose exposure to open-model infrastructure through investments including Fireworks. Their economic arguments depend on assumptions about model quality, costs and customer needs.

For closed models, Randle describes a low-cost-producer argument. A large provider might spread infrastructure and research expenses across aggregated demand, co-design models and serving systems, and use internal models to improve efficiency. If those assumptions hold, closed inference could be cheaper at comparable capability.

Altman counters that open-weight users do not directly pay the original model’s training cost in the same way, though they still pay for inference. He also asks whether frontier labs will choose to lower margins for routine tasks or instead focus on more demanding work. The actual comparison depends on prices, utilization and costs.

The open-model argument also depends on what frontier providers choose to optimize. Altman imagines a division where older, sufficient models handle routine work while leading labs concentrate on the frontier. He presents that division as a possibility, not an established market outcome.

Beyond price, he emphasizes control over data, models and dependence on a supplier. Some organizations may accept extra expense to retain that control, especially if an outside provider could later compete with them. The hosts present this as a customer trade-off rather than a universal reason to choose one approach.

They also discuss tailoring smaller open models to specific tasks with good proprietary data. They say some founders report strong performance at lower cost, but the result depends on the task and the quality of the data; the interview does not provide a controlled benchmark.

The hosts use inference-cloud revenue relative to frontier-lab revenue as an imperfect proxy for open-model adoption. They acknowledge that the estimate excludes some cloud activity. Their illustrative future calculations depend on assumptions about growth and market share, not a verified forecast.

Randle illustrates how even a minority share of a much larger future market could support large inference companies. He explicitly describes the calculation as conditional on growth continuing. Neither the projected market size nor the resulting company values are guaranteed.

Altman says he feels more optimistic about a middle path in which frontier labs matter greatly without making every other business irrelevant. That optimism is conditional: he also recognizes uncertainty about future model capabilities and their broader effects.

Cybersecurity is a more immediate concern for the hosts. They recount what company leaders told them about internal AI-assisted security tests finding serious issues. They also note that an internal test can have access and advantages unavailable to an outside attacker, so the results are not directly equivalent.

Randle cites reported examples of software and connected-device security failures as reasons the risk feels more tangible. The hosts worry that vulnerabilities may be difficult to remediate across old systems and physical devices. Their examples are reported in conversation rather than independently investigated in the episode.

They discuss modified open-weight models with safeguards removed and the possibility that both attackers and defenders could use them. The question they raise is how organizations can improve defensive testing while addressing the wider risks created by more capable tools.

The pair discuss the pace of capability advances relative to defensive modernization. They hope organizations respond faster than in past software transitions, but acknowledge that hardening a wide range of systems could take much longer than a few months.

The hosts hope organizations prioritize defenses because they perceive the risk more clearly. They note that adopting and modernizing systems across institutions such as hospitals can be slow, even if the technology needed to help is available.

Altman says the most meaningful benefits may ultimately be physical: better health, housing, food and transport, not only faster software production. He is interested in biology and robotics because both could change those outcomes if the technology works.

The hosts observe that AI research organizations are investing in data and teams for these fields. They caution that predicting timing is difficult. Examples from autonomous driving suggest that an apparently visible destination can still take years to reach.

Randle contrasts digital code with physical experimentation. Software can often generate large quantities of relatively cheap feedback, while robots cost money to deploy and maintain and biological experiments can answer only part of a question after a delay.

The hosts speculate that running more carefully designed experiments in parallel and investing in data infrastructure could accelerate progress. Physical time and experimental quality remain constraints. They describe the necessary effort as potentially very large, not as a solved engineering problem.

Randle invokes the idea that the internet provided a large initial stock of digital data for language models. Biology and robotics have data too, but collecting the varied physical feedback needed for new systems may require dedicated infrastructure and substantial investment.

Randle shares a family experience with cancer to explain why better treatments matter to him. He describes a prognosis improved by newer therapies and expresses hope that future research can bring similar benefits to more people. His account is personal, not evidence for a specific medical outcome.

The conversation closes with the hosts’ hope that AI investment eventually improves life beyond the screen. Their comparisons between software, biology and robotics are forecasts about possible impact, not guarantees that future discoveries will occur.