Context

Box co-founder and CEO Aaron Levie joins hosts Theo Jaffee and Sofia Puccini to make the case for open-weight AI, prompted by an open-weights letter that Nvidia's Jensen Huang wrote and Box signed. Levie unpacks why he thinks open and closed models aren't actually in competition, how distillation and the US-China AI race should really be framed, and what the latest frontier models mean for knowledge work. He also speaks from Box's own vantage point as a company running AI against real corporate documents, contracts, and regulated industry data, on how AI has changed software engineering hiring and where he thinks enterprise value is heading next.

The Big Idea

Open-weight AI is not zero-sum with closed models: it expands the total number of AI use cases rather than stealing revenue from frontier labs, because the real economics of AI ultimately settle around the cost of inference infrastructure, not who holds the weights.

Because token costs converge toward infrastructure cost over time regardless of open or closed licensing, Levie argues the more urgent strategic question for enterprises isn't which model to bet on, it's who controls the layer that routes work across all of them.

Key Insights

Open and closed models are not zero-sum

Levie's core claim is that open-weight models do not cannibalize closed frontier labs. Open weights let more people build on top of AI, train models for narrow use cases, and create new applications that would not exist otherwise, which grows the total pie rather than splitting it. He argues this actually pushes closed labs to innovate faster rather than undermining them, since more experimentation across the ecosystem raises the bar everywhere.

Distillation is hard to draw an ethical line around

Levie says he cannot credibly argue that training a model on another model's outputs (distillation) crosses some ethical line that training on the public internet does not.

• His reasoning: almost nobody opted in to having their data used to train the frontier models in the first place, so treating a model's outputs as uniquely off-limits is inconsistent.
• The economic angle: distillation via API access still means the lab gets paid, often more than what it was ever paid for the original human-generated training data it used.
• The comparison: he sees little difference between someone distilling a model directly through repeated API calls and code generated by an AI model ending up in public repositories, where it later gets scraped into the next model's training data anyway.
• He still says that if he ran a frontier lab, he would try to defensively block distillation as much as possible, even while doubting there's a clean ethical argument against it.

Blocking China won't stop Chinese AI progress

Levie sides with Nvidia's Jensen Huang in the debate (referencing the widely discussed Huang interview with Dwarkesh Patel) that restricting US AI exports to China will not meaningfully slow China down, because China treats AI as strategically important and has the raw talent, data-generation capacity, and industrial capacity to compete regardless. He argues the practical effect of cutting China off is that Chinese labs simply train on a different hardware stack, one that could then get built out and exported globally through China's own infrastructure, which he sees as a worse long-term outcome for US influence than staying engaged.

Inference, not model secrecy, is where AI money flows

Levie's economic argument for openness rests on a market-structure observation: with three to five US labs all treating frontier capability as existential, competition pushes the price of tokens down toward the cost of the underlying infrastructure over time, landing somewhere in a 20 to 40 percent margin range above infrastructure cost rather than the 70 to 90 percent margin a monopoly could sustain. If that's true, a lab that powers inference for an open model, or builds the preferred post-training environment for it, can capture nearly as much revenue as keeping the model closed, without giving up the ecosystem-growth benefits of openness.

Anthropic's closed stance is a safety belief, not a market one

Levie contrasts Anthropic's refusal to open-source anything against Google's partial move with Gemma. He believes Anthropic's position is not primarily commercial. Rather, the company believes a single entity needs to control the flow of tokens to prevent things like prompt injection and to route queries safely across models, something that is not possible to enforce in an open-weights environment. He expects Anthropic to stay closed for reasons tied to its founding mission rather than short-term market share.

AI expanded Box's engineering ambition, not shrank it

At Box, Levie says AI has not reduced headcount needs; it has expanded what the company is willing to attempt. Before AI, most software projects clustered in a "one to six month" middle band: ideas too small to bother with got shelved, and ideas too large to justify got rejected. AI collapses both ends of that range: multi-year projects become tractable, and week-long low-priority items become two-hour jobs worth just doing. The result is Box has taken on multiple dozens of projects it would never have greenlit before AI, and Levie says the binding constraint now is budget for other parts of the business, not a lack of engineering work to do.

Frontier models are trained to be overly conservative on timelines

Levie and the hosts observe that coding agents tend to badly overestimate how long a task will take, guessing weeks or months for work that finishes in hours. Levie's explanation is that these models were trained on historical human communication (Slack messages, planning docs) written before AI existed, where engineers habitually padded estimates, so the model inherited that conservative bias rather than learning the AI-accelerated reality.

Mental Models & Frameworks

The applied AI layer captures router value

Levie's framework for where enterprise AI value concentrates: the more a task requires coordinating across multiple models (a cost-efficient workhorse model for routine work, a frontier model for hard reasoning, different vendors for different strengths), the more value shifts to whichever layer understands the task, holds the data, and manages the workflow, the layer that decides which model handles which piece. This is the layer companies like Box, Cognition, Factory, and Cursor are positioning to own. It also solves a real enterprise problem: analysis paralysis from constantly leapfrogging model releases, since a routing layer lets a company build its workflows and data pipeline once and swap the underlying model freely as better ones ship.

Two-axis model quality for enterprise knowledge work

Levie evaluates frontier models for Box's use case on two separate axes rather than one general score:

• Deep domain understanding: whether the model actually understands the specifics of a regulated industry, such as life sciences, law, or banking.
• Horizontal capability: general intelligence for working with large volumes of data, using tools, and doing analysis across any domain.

He says Claude's Opus 5 showed meaningful gains over its prior generation on both axes at once, which is why he considers it a broadly compelling upgrade for knowledge work rather than a narrow one.

Trade-offs & Nuance

Open source dollars still flow to infrastructure

Levie is careful to note that open-weight AI does not change AI's underlying economic structure, only who can build on top of it. Running an open model still requires large GPU clusters; nobody is going to circumvent the need for cloud infrastructure by running a frontier-class model on a laptop. The dollars still flow to whoever provides that infrastructure, which is why he thinks even a lab like OpenAI could benefit from open-sourcing a prior-generation model on a faster cadence: it keeps more use cases inside that lab's own ecosystem while still capturing the inference revenue.

The national security argument against Chinese open models

Levie acknowledges a real counter-argument: if the US startup ecosystem becomes dependent on Chinese open-weight models and China later stops releasing them, or if cheap Chinese models undercut American companies' margins, that is a legitimate strategic risk. His response is not to dismiss the concern but to question the proposed fix, comparing it to arguing the iPhone should cost five thousand dollars instead of one thousand to avoid relying on Chinese manufacturing. He believes the realistic alternative to engaging with open-weight competition is that China keeps building capable models anyway, just without any American influence over the ecosystem that adopts them.

Safety guardrails carry a real usability cost

Levie raises a concrete downside of current safety tuning: he says people in biotech have told him they effectively cannot use certain advanced models because the systems push queries down to a weaker, less capable version too often when something looks sensitive, such as code that touches permissions or access control. He supports Anthropic's proposed multi-point framework for agreeing on capability risk levels, but flags an open question: who decides which risks in that framework are real, versus overcautious defaults that make people abandon the tool. He argues that today's level of restriction would be untenable if it became the permanent norm as AI usage scales.

Common Mistakes

Mistake: assuming AI eliminates the need for engineers

Levie pushes back directly on companies cutting engineering headcount because AI can now do more of the work. He argues that if a company believes it has eliminated its need for software engineers, that is a sign its product roadmap isn't ambitious enough, not evidence that engineering work has run out. The right response to AI-driven productivity gains, in his view, is to take on more of the previously-impossible multi-year or previously-not-worth-it small projects, which increases rather than decreases the need for engineering talent, constrained mainly by budget elsewhere in the business.

Practical Application

Reassess your team's stalled backlog

Revisit any project that was shelved as "not worth it" because it was too small to justify engineering time, or as "too big" because it looked like a multi-year effort. Levie's experience at Box is that AI collapses both ends of that range: week-long tasks become two-hour jobs, and multi-year builds become tractable within a normal roadmap cycle. A backlog audit specifically for these two categories can surface real opportunities that were dismissed under old cost assumptions.

Evaluate a multi-model routing layer instead of a single vendor

If your product or internal tooling depends on AI, consider building or buying a layer that can route different tasks to different models rather than standardizing on one vendor. Levie's argument is that with several credible frontier labs all racing on cost and capability at once, locking into a single model family creates switching risk and forces you into analysis paralysis every time a new model ships.

Score any frontier model on domain fit and general capability separately

When evaluating a model for a specific regulated or specialized use case, test it on two distinct axes rather than one aggregate benchmark: how well it understands the specific domain (legal, medical, financial), and how strong its general, horizontal capabilities are (tool use, data handling, multi-step reasoning). A model can improve meaningfully on one axis without the other, and Levie's evaluation process at Box treats both as necessary, not interchangeable.

Push back on overcautious AI safety defaults that block real work

If a model is silently downgrading your queries to a weaker version because of overcautious safety triggers, especially in legitimate, security-adjacent work like access-control code, treat that as a usability defect worth escalating to the vendor, not something to quietly route around. Levie's Box customers in biotech have hit this problem directly.

Questions to Consider

• If your team assumes it needs fewer engineers because of AI, is that a real capacity signal, or a sign your product roadmap has not scaled to match what AI now makes possible?
• Which of your team's currently-shelved projects were rejected as "too small to be worth an engineer's time," and would that math change now that AI can shrink a week of work into a couple of hours?
• If you rely on a single AI model vendor for a core workflow, what would it cost you in migration effort if that vendor's pricing, capability, or safety behavior changed tomorrow?
• Where in your product does an AI safety guardrail meaningfully reduce the quality of output for legitimate use cases, and have you measured how many users that guardrail is quietly pushing away?

Bottom Line

Aaron Levie argues that open versus closed AI is a false fight, because inference cost, not model secrecy, is what actually determines where AI revenue ends up. For a PM, the more consequential strategic bet is not which single model to standardize on, but whether to build the routing layer that can move work across whichever model is best, and cheapest, for each task at any given moment.

Case Studies Mentioned

Box's multi-dozen AI-unlocked project backlog

Box currently has multiple dozens of active projects that Levie says the company would never have greenlit before AI existed, because they previously fell into one of two categories: too complex to justify (multi-year efforts) or too small to bother with (week-long tasks with limited payoff). AI collapsed both categories, making the multi-year projects achievable in a normal roadmap cycle and the small tasks fast enough to just do rather than defer indefinitely. The lesson for a PM is that AI's biggest roadmap impact may not be making existing plans faster, but unlocking a different, larger set of projects that were never on the plan in the first place.

People to Follow

Aaron Levie

Co-founder and CEO of Box, the cloud content management company that handles enterprise documents, permissions, and collaboration across regulated industries like life sciences, law, and banking. Levie is a longtime, outspoken commentator on AI and technology strategy on social media, and speaks in this episode from the vantage point of a company running frontier models directly against real corporate and regulated data at scale.

Notable Quotes

"If you think that you've eliminated the need for software engineers, there's just no chance you're being ambitious enough with your product roadmap." (Aaron Levie)

"Open weights AI does not dramatically change the economic structure of AI, other than to just provide even more avenues to innovation and more avenues of use cases that begin to emerge." (Aaron Levie)