Connected Stack 2026: What Has to Be Built Next

What's changed over the past year 

I opened this year's Connected Stack by acknowledging how much we had underestimated since the year prior. We were still saying “copilot” a lot last year. We understood that AI would rewrite every level of the technology stack, but we were still wildly underestimating the scale.

One example: Dell increased its estimate of inference tokens by 2028 from 1 quadrillion to 57 quadrillion — a 57x revision. Even the companies building the infrastructure have struggled to forecast how much will be needed. Dell’s investor presentation captures just how dramatically those expectations changed.

The infrastructure is being built at extraordinary speed. Enterprise demand has to grow into it. That was the question we came together to discuss: what has to be built in the enterprise to justify all of this investment?

Perspectives in the room

Anthropic’s Kate Jensen, in conversation with my partner Mike Montano, brought that question down to the level of an actual deployment. A company needs a committed business owner, a clear outcome, and a willingness to change the underlying process. Buying access to the technology is one part of a much larger undertaking.

HappyRobot Co-founder and CTO Luis Paarup discussed what this requires from founders. An agent operates inside a real business, with exceptions, institutional knowledge, and experienced employees making judgment calls throughout the day. Understanding those details is essential to building a product that works for the customer.

Context came up repeatedly. Glean Founder and CEO Arvind Jain discussed making organizational knowledge accessible to agents. Sailplane’s Sam Ramji raised a related problem: when experienced people leave, their knowledge often leaves with them.

How do we preserve that knowledge and make it available to the next employee — or an agent? There is a substantial company-building opportunity in that question alone.

The new stack

We heard from teams building different parts of this new enterprise stack. Corridor addressed security as code is generated, while Resolve AI focused on operating software in production. LlamaIndex discussed the documents that underpin knowledge work, and Valar examined the infrastructure required for long-running agents.

As agents produce more code and take on more responsibility, these requirements become increasingly important. We need middleware to govern and deploy that code, systems to manage agent identity and authorization, and ways to understand what agents are doing in production. The applications above the models also have to provide the context that makes them useful in a particular business.

Naveen Rao of Unconventional AI discussed how compute and energy brought us back to the physical constraints beneath the software. Even with the scale of investment underway, there is room to rethink how the infrastructure itself gets built.

A collective effort

We are grateful to our co-hosts at Accel, Addition, a16z, Ballistic Ventures, Battery, Conviction, Felicis, Forerunner, Greylock, GTM Capital, GV, Kleiner Perkins, Lightspeed, Menlo Ventures, MTF.VC, Notable Capital, NVIDIA Inception Program, Redpoint, Sequoia, and Sound. We also would like to thank our partners at Lowenstein, SVB, Aduro, Attivo, Goodwin, and PWC for making this all possible. 

And to every founder who took the stage: you’re doing the real work on the field day in and day out. Thank you for sharing your vision with us.