Aaron Levie - CEO of Box: Agents Will Use Enterprise Software at 100x Human Scale
Box's growth rate is accelerating because of AI. Aaron Levie has the mechanism, and it inverts the SaaS commoditization story entirely.
Agents will be the #1 user of enterprise software at 100x human scale, and Box’s accelerating growth is live evidence.
A 100x cost gap between frontier and open-source models makes the applied layer the natural orchestrator. Box, Atlassian, Salesforce benefit.
Uber’s token cuts are budget math, not demand reversal. Everything digital runs through agents on a timeline of years, not quarters.
Jevons Paradox holds: cheaper software means more demand. The most AI-forward firms, Anthropic and OpenAI, are hiring hardest.
Box Is the File System for Knowledge Work Agents
Levie’s frame is three interdependent layers moving on their own axes: chips and infrastructure, the software stack, and the AI labs. As one moves, the others adapt. As compute gets cheaper, margin shifts toward software.
“Agents are going to be the number one user of software in the future. And they might use software at like a scale of 100x more than people ever using that software.” Aaron @ 3:15
The distinction most commentary misses: coding agents need code, and that is only 5-10% of future agent work. The other 90-95% are knowledge work agents reviewing contracts, processing claims, analyzing research files. They need a file system to reach the same unstructured corporate data humans use.
Box already built that file system. Now it is building it for agents as a third constituent, alongside humans and apps.
Box’s growth rate has accelerated as agent adoption rises Aaron @ 3:15. Atlassian’s recent earnings say the same thing: the companies using agents most are growing fastest on the platform Aaron @ 7:09. The “agents kill SaaS” narrative was backwards.
Two prior Podcast Alpha pieces track Levie’s thesis - the June 20 piece and the June 1 piece. This episode sharpens it with growth data.
Model Routing Is the Underpriced Moat
Enterprises will not standardize on one model vendor. Cost volatility, performance differences, and geopolitical access risk (Levie cites governments restricting access to specific models) make multi-model flexibility a hard requirement.
Brian Armstrong’s math: a frontier model can cost 100x a near-frontier open-source model for similar tasks Aaron @ 7:09. Even at 10x, the incentive to route is severe.
“The value accrues to the layer that is completely agnostic to which model you want to deploy.” Aaron @ 7:09
The applied layer becomes the orchestrator: Box for unstructured data, Snowflake and Palantir for structured, Atlassian for engineering. That routing layer does not exist at scale yet. Whoever builds it first holds a position that is hard to attack from either the model side or the enterprise side.
Token Budget Headlines Are Noise
Uber capped token budgets. Microsoft pulled Claw Code. Meta stepped back from token-maxing leaderboards. The headlines read as demand reversal. Levie reads them as EPS math.
“You should assume that anything that is digital in the world, we are going to be running through agents in some capacity.” Aaron @ 15:13
A public company cannot let any employee trigger unlimited compute spend before a budget governance layer exists. Once line-of-business owners learn to price the ROI - is a half-week task at a premium model worth it - tokens flow Aaron @ 18:54.
That handoff is the timeline risk. Coding adoption was a vertical takeoff because developers own their toolchain and the ROI is immediate. Marketing, sales, legal, and HR have never owned compute budgets.
The leading indicator: when line-of-business teams start owning AI budget lines independent of IT. Until then, the enterprise AI revenue ramp is a multi-year event, not a one or two quarter cycle.
Levie is also blunt on the infrastructure under all of this: demand for compute, memory, storage, and CPU only goes up. CPU is the underappreciated one, because inference needs a compute sandbox Aaron @ 12:01.
Jevons Paradox and the Last Mile
The labor question resolves through two arguments at once.
Jevons Paradox: make software cheaper and the world wants far more of it, more than 2x by what Levie sees across customers Aaron @ 22:41. Companies automating workflows are doing more work, not cutting headcount.
The last mile is the other half. Even the best models get 3-10% of tasks wrong. Every AI project ends with a human reviewing, re-prompting, and making the judgment call.
“I just worked for two and a half hours on something that I actually wouldn’t have even started if AI hadn’t existed.” Aaron @ 26:53
The proof is in hiring. Anthropic and OpenAI have hundreds of open roles across nearly every function Aaron @ 22:41. The firms furthest ahead in AI deployment need more people, not fewer.
Alpha Takeaway
Box’s growth acceleration is the most tradeable signal here - a public-company CEO with quarterly results staked his thesis on a specific, falsifiable mechanism.
Implication: Track Box’s next two earnings reports against the 100x agent-usage claim.
Model routing is the underpriced catalyst - the 100x frontier-vs-open-source cost gap builds a real economic case for an orchestration layer that does not exist at scale yet.
Implication: Watch which applied platforms ship model-agnostic agent orchestration in H2 2026.
The LOB budget handoff is the timeline risk - the thesis is directionally strong, but pace depends on IT ceding AI budgets to line-of-business owners.
Implication: Avoid pricing a fast enterprise-AI ramp until that handoff shows up in adoption data.
Bottom line: The applied software layer is core agentic infrastructure, not a disruption casualty, and Box’s accelerating growth is the first hard data point.
This newsletter is for informational and educational purposes only. It is not investment advice, not a research report, and not an offer or solicitation to buy or sell any security.
Any valuations, price scenarios, projections, and forward-looking statements are hypothetical illustrations based on assumptions that can change without notice. They are not price targets, forecasts, or guarantees of future results. Model outputs depend heavily on their inputs, and small changes in assumptions can produce large changes in the results. Actual outcomes will differ, often materially. Do not rely on any figure here as a prediction or a basis for an investment decision.
Information is drawn from third-party sources we believe to be reliable, including public filings and podcast interviews, but we do not guarantee its accuracy or completeness. All views attributed to named speakers reflect their own opinions, not Podcast Alpha’s.
This content is general in nature and does not account for your financial situation, objectives, or risk tolerance. The author may hold positions in securities mentioned. Consult a licensed financial, legal, or tax advisor, and do your own research, before making any investment decision.