一句话总结
The Information reported late on October 27 that Nvidia's biggest customer OpenAI plans to buy Google TPU chips, terms undisclosed; it is also hiring a TPU compiler optimization engineer in San Francisco at $380K-$500K — a deeply Nvidia-tied model company shifting toward a multi-supplier matrix.
Report and job posting cross-checked: OpenAI is paving the way for frontier models on TPU
As relayed and verified by FourWeekMBA and First Squawk, The Information reporter Ryan Barwick published on October 27, 2026 at 23:59 UTC the story "OpenAI, a Longtime Nvidia Customer, Plans to Buy Google's AI Chips": OpenAI plans to buy Google's AI chips (TPU), though financial terms and other details could not be learned.
Corroboration comes from OpenAI's own careers page: as read and verified by FourWeekMBA on October 28, its Inference team is hiring a Software Engineer, TPU Compilers Workload in San Francisco, responsible for optimizing the full forward-pass performance of OpenAI's most advanced frontier models on TPU — from model code through kernels to hardware execution — at $380,000 to $500,000 plus equity. A procurement report paired with a real TPU compiler posting suggests OpenAI is not testing interest but doing engineering prep to actually run frontier models on TPU.
The binding starting point: 100,000-plus Nvidia Blackwell GPUs behind GPT-6 Astra
The flip side of the multi-supplier shift is OpenAI's deep tie to Nvidia. Per FourWeekMBA's verification of the October 27 Oracle-Nvidia joint press release, OpenAI trains and serves GPT-6 Astra on the Abilene OCI supercluster, which deploys more than 100,000 Nvidia Blackwell GPUs. OpenAI's latest flagship model, in other words, currently runs on a 100,000-card-scale Nvidia cluster.
Only against that starting point can the weight of buying TPU be understood: this is not a small customer diversifying, but one of Nvidia's largest customers reserving engineering capability for a different instruction-set architecture on top of a 100,000-card Blackwell base. In the pricing relationship, the customer having an alternative route itself changes negotiating leverage.
Three lines at once: in-house silicon, Broadcom custom chips and Google TPU
OpenAI's compute posture is actually three lines at once. Beyond continuing to load up on Nvidia (via the OCI Abilene cluster), it is pushing in-house silicon, bringing in Broadcom custom chips, and per The Information, reportedly planning to buy Google TPU. The same-day companion event: Oracle and Google jointly announced on October 27 that TPU will be available on OCI in calendar 2027 — the first time TPU is offered on a hyperscale cloud platform outside Google.
That gives OpenAI a real delivery channel for TPU: direct from Google or through OCI cloud. For Google, the earnings window is near — Alphabet is set to report Q3 2026 results on October 28; Alphabet said in July that the vast majority of TPU external-sales revenue will be recognized in 2027. A frontier model company becoming a TPU external customer is a landmark order proving Google's out-licensing chip ecosystem works.
Two-way impact: on Nvidia's pricing power and Google's external-sales ecosystem
Per FourWeekMBA's explicit flag, procurement scale, TPU generation, delivery timing and deployment mode (own data centers or via cloud) are all undisclosed, and neither OpenAI nor Google commented. Until those details fill in, interpretation should stay measured.
But the directional impact runs both ways. For Nvidia, its biggest customer publicly preparing an alternative compute route weakens, over time, unilateral leverage on pricing and allocation, even though Blackwell remains the workhorse near term. For Google, a customer like OpenAI joining the TPU external-sales roster validates the commercial loop from self-use to external sales and anchors demand for the external revenue it starts recognizing in 2027. Once a multi-supplier matrix forms, the relationship between frontier model companies and chip vendors shifts from primary-vendor lock-in to portfolio management across architectures.