Google TPUv7 posts strong agentic inference results against Nvidia GB300

Google’s TPUv7, the chip also known as Ironwood, has delivered early benchmark results that put it ahead of Nvidia’s GB300 NVL72 on cost efficiency for agentic AI workloads, according to SemiAnalysis.

On the AgentX benchmark, TPUv7 submissions posted performance per dollar between 1.42 and 1.67 times better than Nvidia’s rack-scale system.

The timing matters because the AI industry’s spending focus is shifting. Training giant models grabbed the headlines for years. Running those models for millions of users, and increasingly for autonomous agents, is where the recurring bills now pile up.

What the benchmark actually measured

AgentX tests inference for agentic workloads. Inference is the stage where a trained model answers questions or carries out tasks, and agentic workloads involve AI systems working through multi-step jobs instead of single replies.

Those jobs tend to involve long context and multi-turn exchanges, meaning the model has to remember a lot and talk back and forth repeatedly. SemiAnalysis’s research notes that TPU architecture now supports exactly that kind of long-context, multi-turn inference.

The test model was Qwen3.5 with 397 billion parameters, an external model and not one built by Google.

Results were checked across several interactivity levels. Interactivity, roughly speaking, describes how fast a single user gets responses versus how many users a system can serve at once.

Then there is the precision question. Google’s submissions ran in FP8, while Nvidia’s ran in FP4. These are number formats that determine how many bits a chip uses to represent each value in the model.

Fewer bits generally means less memory and faster math, which is why FP4 is described as the more aggressive option. Google, in other words, still came out ahead on cost efficiency while playing with the less aggressive format.

The hardware footprint was also lopsided. SemiAnalysis says the TPUv7 benchmarks used only four chips, compared with full Nvidia racks of 72 GPUs.

From internal tool to external product

Google has been building TPUs since 2013, and for most of that history they were mostly a house specialty powering Google’s own services.

In November 2025, SemiAnalysis reported that Google had begun selling TPUs to outside customers. The firm now says that call has held up, naming Oracle, OpenAI, Anthropic and Meta as TPU customers in line with its Accelerator Model tracking.

The research outlines how some of those relationships look in practice:

• Anthropic has committed to more than one million TPU units, making it the largest external customer.
• Oracle plans to resell TPUs through Oracle Cloud Infrastructure starting in 2027 under a multi-gigawatt-scale contract. SemiAnalysis describes this as a first for a hyperscaler outside Google.
• OpenAI is reportedly preparing to commit to direct TPU purchases.

The software gap is narrowing

SemiAnalysis argues Google is chipping away at that advantage too. Google has partnered with upstream maintainers across the open-source AI stack, including the Meta PyTorch team, vLLM (via Inferact and RedHat), SGLang (via RadixArk), llm-d and Ray.

Those names matter because they are the frameworks and serving engines developers already reach for. If TPU support lives inside the tools people use daily, switching hardware starts to look like a configuration change instead of a rewrite.

What this means for Nvidia, Google and AI buyers

The most direct stakeholder here is Nvidia. Performance per dollar is the metric that large buyers obsess over once inference bills scale up. SemiAnalysis’s research suggests results like these may put upward pressure on Nvidia to compete harder on cost efficiency and performance.

The customer roster tells its own story. When Anthropic, Oracle, Meta and reportedly OpenAI are all routing at least part of their compute toward TPUs, the market is no longer a one-vendor conversation.

The items worth tracking from here are concrete. Watch whether OpenAI’s reported TPU commitment gets formalized, how Oracle’s 2027 rollout on OCI proceeds, and whether independent benchmarks across other models confirm the AgentX results.