NVIDIA's Jensen Huang Almost Refused Dinner Over TPU Price Comparison
The English-speaking community recently revisited the first dinner between Jensen Huang and Anthropic CEO Dario Amodei in May 2022: Tom Brown, head of computing, presented a table claiming that Google TPU was more cost-effective than NVIDIA chips at the same dollar amount, prompting Huang to call him a "bean counter" and threaten to skip the dinner.
The dinner ultimately took place. During the event, Huang repeatedly stated that NVIDIA aims to build the world's largest data center; Amodei muttered that Huang's behavior resembled that of Trump. Brown, who came from OpenAI's large training background, later handled Anthropic's computing procurement.
Subsequently, Anthropic adopted a multi-chip approach, utilizing Google TPU, Amazon Trainium, and NVIDIA GPU in parallel. Public agreements included up to approximately 1 million TPUs and over 1 gigawatt capacity, with some leased through Google Cloud and others sold as complete systems via Broadcom.
NVIDIA announced an investment of up to $10 billion in Anthropic, linked to up to approximately 1 gigawatt of the Grace Blackwell and Vera Rubin systems; Microsoft also has parallel funding arrangements. The timeline fell within weeks to months after the public announcement of the TPU deal, creating a structure of "buy Google silicon first, then accept NVIDIA's investment".
In Silicon Valley, a "bean counter" refers to a procurement officer who only calculates unit dollar performance without securing ecosystem lock-in. Huang's counterpoint focused on cluster scale, networking, and software stack, rather than single-card efficiency metrics. Three years later, both sides included each other in their capital expenditures.
In terms of market mechanisms, what is sold is alternative silicon for training and inference, and what is bought is the uninterrupted capacity of model companies. Funds flow from Anthropic's cloud and self-built capital expenditures to TPU racks, GPU clusters, and equity investments; beneficiaries are suppliers who can simultaneously profit from leasing and complete system sales, while pressured are chip manufacturers trying to lock in customers with a single efficiency metric. The event was driven by the resurfacing of old dinner details, compounded by the established million-chip agreements.
Source: Public Information
ABAB AI Insight
The table from 2022 crystallized the conflict: one side focused on dollars per floating point, the other on who would build the largest data center. Huang elevated the negotiation from procurement to identity with his refusal to dine and insults to the bean counter; Amodei framed the same scene as a stylistic observation by likening it to Trump. The contracts three years later proved that neither side won over the other, but rather incorporated each other into their supply lists.
The capital path shows that model companies retain bargaining power through multi-cloud strategies, while chip manufacturers buy back shipping rights through equity stakes. The large TPU deal positioned Google as a commercial silicon supplier rather than just an internal infrastructure provider; NVIDIA's $10 billion investment welded customer balance sheets to its own data center strategy. The first to commit to a deal has no retreat in the next price increase.
Analogous to the 2010s, when cloud vendors simultaneously procured Intel and self-developed chips, and TSMC clients required both advanced processes and second suppliers. The industry phase transitioned from "NVIDIA is irreplaceable" to "capacity is insufficient, so alternatives must be found," and then to "even after finding alternatives, one must invest in NVIDIA to avoid being locked out of the software stack."
Structural judgment indicates a transfer of pricing power. The mechanism is that while unit dollar performance can change suppliers, cluster software, interconnectivity, and delivery times are still priced by the largest shipping entity; the efficiency metrics at the dinner table determine entry, while the gigawatt contracts determine who sets the pricing for the next generation of models.
ABAB News · Laws of Cognition
Efficiency metrics determine entry, gigawatt contracts determine pricing.
Those who insult the bean counters ultimately need to invest in the services of those same bean counters.
Multi-supplier strategies are not a belief; they are a fear of being erased by the next table of metrics.