In conversation with Reid Hoffman & Robert F. Kennedy Jr.

After a round of banter about Chamath's unbuttoned shirt, Jason Calacanis introduces Reid Hoffman as a LinkedIn co-founder, Microsoft board member and former PayPal colleague of David Sacks. Hoffman credits Sacks with a fast learning curve and a "maniacal focus" on how PayPal's product cycle worked on eBay, saying he judges people he respects by what he learned from them.

Sacks returns the compliment by describing PayPal's existential threats from Visa, Mastercard, eBay and Citibank, with Hoffman acting as the emissary who kept those larger players at bay. Hoffman's most vivid memory is the X.com merger: before closing, Elon Musk praised CEO Bill Harris as a reason X.com deserved a large share; immediately afterward Musk wanted Harris fired before the first board meeting. Harris's eventual removal, planned at Antonio's Nut House in Palo Alto, became known as the "Nut House coup."

Calacanis cites Nvidia's latest results: data center revenue of $26.3 billion, or 87% of revenue, net income of $16.6 billion and a 75% gross margin, while noting that quarter-over-quarter growth is cooling and the stock fell after hours. Hoffman's answer to public-market investors has been blunt:

It's sustainable for two years, which for you guys means forever.

His reasoning is that Nvidia leads in training-cluster chips, but as demand scales, the bulk will shift to inference, where many new chips will compete. Nvidia will then face a choice between defending prices and margins or responding to a competitive market, which he expects to start within a year or two at the latest. He concludes the current heat is not sustainable, yet still advises people not to short Nvidia.

David Friedberg raises reported comments, via Gavin Baker, that leaders view the buildout as a race toward a "digital god" whose payoff would be trillions, making quarterly spending of tens of billions secondary. He asks whether Microsoft's investment is ROI-driven or purely strategic.

Hoffman stresses that board members do not speak for Microsoft, but describes Satya Nadella as blending strategy with return-on-capital discipline: treating AI as a platform shift that cannot be missed, while asking when revenue arrives and avoiding spending "like drunken sailors." He notes Nadella focuses on Office and cloud outcomes rather than AGI talk, while acknowledging that novel platforms are hard to predict, which is why every hyperscaler is fully engaged.

On open source, Hoffman argues OpenAI never promised to be open source; "open" referred to safe, broad access. He expects winners on both sides, including Meta's LLaMA, Mistral, vertical startups in areas such as coding agents and medicine, and the hyperscalers. The key unknown is when scaling hits an asymptote; the bet of OpenAI, Anthropic and the hyperscalers is that returns to scale continue, and large models also help train smaller ones.

Friedberg asks whether one "god model" will dominate. Hoffman calls the one-model view a mistake, comparing it to Sauron's ring. Because larger models cost roughly an order of magnitude more, providers already swap in GPT-3.5 instead of GPT-4 for some tasks. He predicts agents built from blends of models with routing and escalation, citing Microsoft's textbook-data paper as evidence that small, well-trained models can be effective, and leaving room for startups.