The world's largest startup: Jeetu Patel on running Cisco at founder speed

Key points

Key takeaways from the Source interview with Cisco President and Chief Product Officer Jeetu Patel (October 2026):

Picks and shovels for the AI era. Patel frames Cisco as the AI infrastructure company — networking for GPU clusters, security for agents and machine-scale data via Splunk — with its hyperscaler AI business reaching $9.3 billion in orders within two years, a position nobody would have credited the company three or four years ago.

Run like the world's largest startup. Cisco operates 32,000 engineers in what Patel calls wartime mode, mixing speed with scale: speed alone is a startup, scale alone is a big company, and he tells his team to be both — calling himself a late-stage co-founder who arrived 36 years late.

The rule of thirds. His leadership team is split into a third of Cisco insiders, a third of outside systems thinkers who have operated at scale, and a third of founders from acquired companies, who are given charters larger than the businesses they sold.

AI is not about efficiency. Patel's stated goal is using AI to generate original insights that solve problems Cisco could not previously imagine, and he expects the company to need more engineers, not fewer, because every automation step creates a new human bottleneck — coding gives way to code review, then to judgement.

Use AI or lose your job. Cisco told employees the opposite of the usual fear framing: those who do not use AI will lose their jobs, because attacks now happen at machine scale and need machine-scale defences — and the company initially gave everyone unlimited tokens to learn, on the logic of getting familiar before becoming efficient.

Proof in the code. One sceptical engineering leader was converted after Cisco refactored roughly half a million lines of code into about 120,000 in two weeks, and the company became an early design partner with OpenAI on Codex.

Planning cycles across the stack. Cisco builds silicon, photonics, hardware and software: silicon runs on a five-year planning cycle, hardware 18–30 months, software 12–18 months and models and agents on a three-month or even weekly cycle — co-designing all layers is the hard-to-replicate moat.

Infrastructure is the constraint. With fewer than 2% of people using agents in any power-user capacity, all infrastructure produced today is already consumed; Patel argues that reaching 25–50% usage demands a step-change in capacity, and that AI safety — including agents that go rogue or are poisoned by adversaries — is the second great problem to solve.

Read more: Y Combinator on YouTube