What this covers

Eric Vishria, a General Partner at Benchmark, talks with host Patrick about what a decade of software and hardware investing, including Fireworks, Sierra, and Cerebras, teaches about the present. The history he leans on most is AWS. When S3 and EC2 launched, investors reacted badly, and he guesses almost none of thirty smart investors would have called it a durable high-margin business. Cloud adoption was led by digital natives, with Snapchat once a large share of GCP, while banks and insurers waited until about 2014-2016. The market ended as an oligopoly, with Datadog and Cloudflare growing large outside the big three.

The software half is about operating models breaking. Public SaaS multiples fell from around 30x to around 6x despite strong growth, and Vishria argues a company hitting its traditional plan is destroying equity value. He says the quota capacity model assumes pushed demand, which fails when customers pull. He claims AI agents make database migration trivial, undercutting a classic moat. Product managers now need to understand the jagged edge of AI capability, not just customer problems.

The hardware half turns to Cerebras: the three levers of hardware speedup, roofline performance, a supply chain of roughly thirty vendors beyond TSMC, and an IPO attempt that failed over CFIUS and timing. Energy is framed as the binding constraint on intelligence. The most contested stretch applies the same logic to Geoffrey Hinton's radiology prediction and to forecasts of mass unemployment. Vishria says both ignore data, liability, and deployment friction, and he expects radiologist demand to rise.

Sharpest takeaway

Vishria argues that AI, like cloud before it, will produce an oligopoly of several large winners rather than one company consuming the market, and that the criteria for winning in software and hardware have shifted, so founders and investors must discard legacy playbooks and reason from first principles.

  • AWS was widely written off in 2007 and feared as an everything-eater in 2014, yet the market ended up with multiple $100B+ winners
  • Database and SaaS moats built on migration cost and specification are eroding because AI agents make migration cheap and well-specified work easy
  • Hardware success depends on physics, supply chains, timing, and geopolitics that software founders do not control

The argument · threads1 threads · 71 claims
0.85

Cursor's large share of its usage pool mirrors Snapchat's early share of GCP.

1 pointcentrality 5/5
  • Cursor represents a large share of a usage pool, similar to how Snapchat was a large share of GCP early on, which Patrick compared to Cursor's share today.

    “That's like Cursor being 30% of all these things. Same thing.”