E167: Google's Woke AI disaster, Nvidia smashes earnings (again), Groq's LPU breakthrough & more

Nvidia’s AI boom is already a cash machine: Q4 revenue hit $22.1 billion, up 265% year over year, with $12.3 billion net income and 76% gross margins. Data-center demand drove the surge, and management was already guiding toward about $110 billion revenue for the new fiscal year, nearly double the prior year.

The strongest challenge to the Nvidia story is not demand today, but whether AI applications can earn enough to justify the hardware binge. Chamath argued the buyers are mostly cash-rich incumbents pre-purchasing capacity, while most current AI products remain demos or proofs of concept rather than large, proven profit engines.

Big Tech can accelerate AI infrastructure spending because GPU purchases are capitalized and depreciated, rather than hitting earnings immediately. Friedberg added that cash-rich platforms also face antitrust limits on acquisitions, making data-center buildout one of the few large-scale ways to deploy balance-sheet cash toward future growth.

Groq went from essentially no customers two months earlier to a sudden rush of Fortune 500 companies and developers testing its inference hardware. Chamath said users were finding Groq’s LPUs meaningfully faster and cheaper than Nvidia-based inference, after an eight-year build that survived failed pitches to Tesla, traders, agencies, and earlier vision workloads.

Chamath’s deep-tech filter is simple: fund hard engineering, not businesses that require a new law of physics to be true. He said Groq’s first roughly $7–10 million was aimed at a bounded chip-and-compiler problem on mature 14-nanometer technology, unlike fusion concepts whose feasibility may depend on unavailable fuels or breakthroughs.

Two famous companies were nearly steered into safer, smaller businesses by investors: Tesla into selling drivetrains and Uber into selling software to taxi fleets. Jason Calacanis said a prominent VC offered to preempt Uber’s Series A only if Travis Kalanick abandoned the consumer marketplace; Calacanis never passed the investor along.

Gemini’s image failures exposed a harder problem than one bad launch: AI turns Google from an information-retrieval company into an information-interpretation company. The panel’s sharpest point was that chatbots must choose a single answer, making tuning, citations, uncertainty, and user controls far more consequential than search’s familiar list of links.

If handed Google, Chamath said he would cut 50–60% of the workforce and redirect savings into roughly $100 billion a year of proprietary data licensing. His proposed moat was a product marketed around fewer errors and “defensible truth,” using exclusive data access to make AI answers more trustworthy than competing models.