Aaron Levie believes in AI value but ignores corporate IT reality
By Imani Sutton · Reporting from Atlanta ·
Aaron Levie's vision of AI value overlooks the crumbling hardware and financial burden facing corporate IT.
On the Sequoia Capital podcast, Box CEO Aaron Levie said the AI boom is no longer about raw intelligence. The fight has moved to the application layer, where software hits the actual workflow of a company. Companies do not want a raw API; they need an "agent harness"—the digital glue that connects the AI to their specific file folders, user permissions, and search tools. Levie believes this shift will create trillions in value, similar to how Snowflake and Databricks grew by building on the foundation of early cloud infrastructure.
He noted the friction between the labs building foundation models and the developers building apps. Once public companies face the actual cost of training runs, he said, they will stop subsidizing tokens. Open-weight models are already taking over routine tasks, preventing one or two labs from owning the entire market. But deploying these tools to office workers is slower than giving them to coders. Old servers in basement closets, rigid login permissions, and a fear of hallucinated legal citations hold the process back.
Why adoption stalls
Levie highlights the gap between a model's intelligence and the reality of corporate IT, but he ignores who pays for the transition. Engineering teams adopted coding agents quickly because their work lived in GitHub. This reveals the wall facing everyone else. Southern law firms, rural hospitals, and city halls do not use modular clouds. They rely on twenty-year-old servers and fragmented databases—often just spreadsheets saved on local drives—that crash when an agent tries to read them. The bottleneck is not a lack of imagination. It is a lack of working hardware, a cost borne by workers whose roles shrink to fund a speculative tech cycle.
The hidden cost of the token
Levie expects market discipline to return to the foundation labs, but this overlooks the forces keeping the industry afloat. He says OpenAI and Anthropic must eventually stop subsidizing tokens. Yet he ignores state-funded compute clusters, massive data centers, and the grip of chip monopolies that distort the cost of running a model. Executives often describe AI deployment as a neutral shift in efficiency. This erases the physical toll: power draws that brown out city grids and the low-wage laborers in overseas hubs who spend hours scrubbing training data to remove bias and gore.
Enterprise AI is not about sliding an agent into a law firm's digital archives. It is about who pays to replace the old cables and servers while the software companies collect the fees. To unlock that massive sum in value, the application layer must stop treating workflows as abstract software problems. It must deal with the broken hardware, the outdated operating systems, and the chaotic org charts the algorithm leaves behind. The real victory will not be a smarter model, but a world where the plumbing actually works.