Google's new training results show speed bought with scale, and coordination bought with design, rather than a lone chip breakthrough. Ironwood reaches its headline time only when thousands of processors train together with little waiting, while a much smaller NVIDIA system on the same cloud takes far longer on the same model. The contest Google is actually winning is over keeping a vast machine busy.
| Document | Details |
|---|---|
| Issued by | |
| Published | 16 November 2026 |
| Official text | blog.google |
| Linked official pages | 2 |
| Earlier documents compared | 2 |
The document, New AI training performance benchmarks for Google TPUs, published by Google on 16 November 2026, reports 10 results in the MLPerf Training v6.1 benchmark round. MLPerf is a shared test: every entrant trains from an official reference checkpoint, a starting version of a model, until it reaches a set target for quality or accuracy. The spread matters because Google is presenting its cloud, not only its chip, as the place where either kind of hardware performs.
Pre-training speed arrives with thousands of chips working together
Pre-training is the long first phase in which a model learns from large amounts of data before later tuning. On DeepSeek-V3, a model named in the result list, 8,192 Ironwood TPUs reached the benchmark target quality in 1.70 minutes. Google also reports scaling up to 4,096 chips with 96% efficiency. Efficiency here describes how much of the expected gain from adding chips survives in practice. Adding processors brings coordination costs, because work must be split, results must be exchanged and partial results must be combined. A high figure means less of the added power is lost to that coordination.
Google attributes the outcome to its AI Hypercomputer architecture and to designing silicon, networking and open software together. That is the company's own explanation, and the scaling figure is the part that gives it weight. A cluster can be large without being useful if extra chips spend their time waiting for data or for each other. Google's claim is narrower and stronger than raw size: the machine grew while keeping most of its expected performance. For a customer training a frontier model, elapsed time on a very large job is a direct business fact, because the job occupies expensive capacity until it finishes.
Agentic learning rewards systems that keep chips fed
The round also included the first MLPerf benchmark for agentic reinforcement learning. Reinforcement learning trains a system through feedback on its actions. In an agentic version, the system does work, in this case generating code, and learns from the result while the work continues. That creates a scheduling problem as well as a computing problem: generation and learning compete for attention, and expensive training chips can sit idle if the two activities fall out of step.
Built on Google Cloud’s AI Hypercomputer architecture, these results reflect co-design across silicon, networking, and open software: Pre-training on Ironwood: 8,192 Ironwood TPUs reached the target model's benchmark target quality in 1.70 minutes, scaling up to 4,096 chips with 96% efficiency. Agentic RL on Ironwood: 640 Ironwood TPUs reached the target model’s coding benchmark target accuracy in 18.02 minutes (0.765 vs. the 0.69 goal), running code generation and learning simultaneously so training chips waited under 3% of the time.
Source: Google, 16 November 2026. Google, New AI training performance benchmarks for Google TPUsIn that test, 640 Ironwood TPUs reached the coding benchmark target accuracy in 18.02 minutes, scoring 0.765 against a goal of 0.69. The result list identifies the model as Qwen3.5-397B. Google says the system ran code generation and learning simultaneously, so training chips waited under 3% of the time. The low waiting share explains why this result belongs beside the pre-training time rather than behind it. Pre-training shows how fast a coordinated cluster can move through a known workload. Agentic learning shows whether the same kind of cluster stays productive when the workload keeps producing new material for itself during the run.
NVIDIA results make the pitch about Google Cloud, not loyalty to one chip
Google also published a pre-training result on NVIDIA hardware. On Google Cloud A4X Max instances, 256 NVIDIA GB300 GPUs reached the target model benchmark quality in 31.96 minutes for DeepSeek-V3. The company adds five A4X and A4X Max submissions on the agentic reinforcement learning benchmark, covering GB200 and GB300 systems. A reader should not treat the Ironwood and NVIDIA times as a straight contest between individual devices. The chip counts differ sharply, the instances differ, and the benchmark rewards the whole configured system that Google submitted.
Read that way, the NVIDIA entries change the meaning of the Ironwood headline. Google is not arguing that customers must choose its own silicon to use its cloud. It is arguing that its cloud is a competent home for large training systems of more than one kind, with networking and software supplied by Google in each case. Ironwood remains the flagship result and the clearest statement of what Google can build end to end. The Blackwell results broaden the commercial claim: a buyer already committed to NVIDIA hardware is still being invited to run the job on Google Cloud.
The new document changes the subject from access to capacity
Earlier Google documents in the same run of publications worked at the level of the individual user. On 13 November 2026, Google told Gemini app users in the U.S. that they could connect Google Wallet for responses informed by their spending habits, with read-only access and user control over the connection. A separate document the same day put Google Cloud credits inside Google AI Studio for Google AI Pro and Ultra subscribers, so developers could claim monthly credits and keep building without leaving their workspace. Those moves widened access and deepened personal context.
This document speaks to a different reader: a lab or company that trains models and buys computing capacity by the cluster. Its question is not whether an assistant knows more about one person. Its question is how quickly a large training job finishes, how far it scales before coordination eats the gain, and whether irregular agentic work keeps the hardware busy. The contrast is the context that makes the benchmark more than a score. Google spent the earlier announcements making Gemini and AI Studio easier to enter. It now has to show the industrial base underneath those products, and to sell that base to customers who may never use the consumer app at all.
A benchmark time is a promise with narrow terms
All reported times measure MLPerf time to train from the official reference checkpoint to the target convergence criterion, as Google states in the document. That definition makes the submitted runs comparable under the rules of the round. It also limits what a buyer learns. The published results give no price for a run, no energy figure and no statement about how much Ironwood capacity a new customer can obtain, or when. Those omissions matter because a fast benchmark system still has to be available and affordable at production scale before the time becomes a saving.
The next MLPerf round will test whether rivals close the scaling gap and whether Google can repeat the agentic result as that benchmark matures. Until then, the supported verdict is precise. Google has documented a training platform whose advantage lies in coordination at scale, across pre-training and a new agentic test, and it has placed that claim on its own chips and on NVIDIA systems sold through its cloud. The speed is real within the stated test. The commercial value will be decided by capacity and cost, which this document leaves unpriced.
Source: Google, Google, Google Cloud Google Cloud leaders share what to expect in cybersecurity for 2027., Google, Google Cloud Get certified as a Google Cloud Professional Agentic Architect. By Gary Eimerman, Google, Claim your Google Cloud Credits in Google AI Studio (13 November 2026), Google, Connect Google Wallet to Gemini for personal finance insights (13 November 2026) · Date:



