·
Article
For three years, building serious AI has meant waiting in one line: Nvidia's. This week, more doors opened. The line didn't get any shorter.
Door one: Google's chips leave home
At Oracle AI World on October 27, Oracle and Google announced that Google's Tensor Processing Units (TPUs) will be offered on Oracle Cloud Infrastructure starting in calendar 2027. Both companies describe it as the first time TPUs will be available from a hyperscale cloud other than Google's own. An Oracle executive called it a multi-year, multi-gigawatt partnership and said new TPU generations will arrive on Oracle as soon as Google launches them. Oracle also plans to offer Cerebras inference and OpenAI's in-house Jalapeño chip, alongside Nvidia's latest systems.
Door two: OpenAI shops around
Hours later, The Information reported that OpenAI plans to buy Google's TPUs. Terms, volume and timing weren't disclosed, and neither company has confirmed it. OpenAI began renting TPUs in 2025. Analysts read the report less as a near-term hit to Nvidia's revenue and more as evidence that its biggest customers are building real alternatives, which weakens Nvidia's pricing power over time.
The line: still long
Then came the reminder that supply is tight everywhere. On October 28, Microsoft CFO Amy Hood said delays at third-party GPU providers will push some of Microsoft's available capacity out to the first half of 2027. Alphabet's finance chief said outside computing power came online only at the very end of the third quarter. Both companies reported capital spending below analyst estimates, not because they want less compute, but because they couldn't get it fast enough.
Why the chip choice reaches software
Switching chips used to mean rewriting code, which kept most teams locked in. That's starting to loosen. Google says TorchTPU, its PyTorch stack, works with common serving tools such as vLLM, SGLang and Ray, so a model can move between TPUs on Google Cloud and Oracle without being rewritten. If that holds up in practice, compute starts to behave more like a commodity you shop for on price and availability.
What it means for builders
For startups that use AI through APIs, the effect is indirect but real. More chip options should pressure inference prices over time. At the same time, shortages through mid-2027 can mean rate limits, waitlists for new models and price changes with little warning. For teams running their own models, portability is becoming a strategic asset, because the ability to move workloads between providers is leverage when capacity is scarce.
There's a quieter signal too. The biggest buyers diversifying suggests the industry expects demand to stay high for years, which keeps pressure on every provider to lock in capacity early, and tends to favor customers who commit early as well.
Build for more chips and less certainty
The practical lesson for founders is to stop treating any single model provider or chip vendor as permanent. Keep your prompts, evaluation tests and deployment setup portable enough that switching providers takes a week of work rather than a quarter, and track your cost per completed task so you can tell when a cheaper option is actually better. If AI sits at the core of your product, ask your providers directly about capacity commitments through mid-2027, and build a fallback model into your architecture now, while switching is still a choice rather than an emergency.
Leave a comment