The $13B phone call

The Threads

The $13B phone call

The deal that crystallised a decade of work at Groq closed faster than most term sheets. Chamath Palihapitiya traces the origin to 2015, when Sundar Pichai mentioned the TPU on CNBC and Chamath became “really shocked — it was not something I expected a search or a software company to either have the expertise or the wherewithal to really do.” He hunted down Jonathan Ross — the Google engineer who’d built the TPU in his 20% time — wrote a $10m check for roughly a third of the company, and then watched seven years pass without product-market fit.

What changed wasn’t the chip. It was the workload. LLM inference arrived and Groq’s SRAM-heavy architecture turned out to be accidentally perfect for the decode phase. In May 2025 Nvidia shipped NVLink Fusion. Groq’s team experimented with it, the early data looked compelling, Chamath called Jensen Huang on a Saturday. A few weeks later the message came back: “There’ll be 13 billion in the account on Tuesday. Get it done.”

The lesson Chamath draws: the business wouldn’t exist but for Jonathan Ross’s technical brilliance. When the founding team fractured, the heuristic was simple — back the technical person the company wouldn’t exist without. Builders over managers. Non-negotiable.

Back the builder, not the org-builder

This is where Chamath turns the episode into a manifesto. Modern org design — VP of Engineering, CTO, CFO with VPs of Audit, Tax, FP&A — is “a vestige of some crappy piece of enterprise software that was sold to somebody.” He points to counter-examples: Jensen Huang runs with 60-80 direct reports, and Elon Musk ping-pongs across 5-7 companies simultaneously. Both create trillions in enterprise value.

Organizational charts accelerate and celebrate the politician or the maneuverer or the kisser, but very rarely do they accelerate and highlight the technical genius.

His corollary on talent: A-players are meaningless in isolation. What you want is an A-team — high-agency, self-directed people who take motivation from the work itself rather than validation from peers. The insecure politicker with one foot out the door is the tax orgs impose on themselves.

Inference silicon as the asymmetric bet

The technical framework worth noting: training and inference are architecturally different problems. Training is systems-heavy. Inference is north-south traffic — pre-fill and decode as distinct phases with distinct bottlenecks. Groq’s SRAM bet was the wrong choice for training but the right choice for decode. This asymmetry is why a startup could get paid $13B by Nvidia for what is, in the end, a different answer to a different question.

Seven years of darkness is the job

We didn’t have product market fit for seven years… most of these businesses will go through many years of darkness — five, six, seven years where nothing is working until it all of a sudden works.

Chamath’s warning to anyone considering a startup: if you can’t find motivation in technical progress through those years, the path isn’t for you. He cites Replit’s revenue chart — years of flat nothingness, then an exponential hockey stick — as the emotional shape of the thing.

This Week’s Episode

Chamath (solo) — “What You Can Learn From Our $20B Deal” · YouTube · 12 min

A compact memoir of the Groq investment from 2015 through the 2025 Nvidia deal, with the conclusions Chamath draws about org design, technical founders, and the nature of breakthrough bets. Worth the 12 minutes.

Issue 01 is a smoke test — one episode while the rest of the pipeline came online. From next week: the full 12-show feed. This is the foundation.