TPU vs GPU in 2026: Which AI Chip Should Your Business Use?

Last Updated: November 12, 2026

Summary: A TPU is Google's custom chip built only for the matrix math behind AI, while a GPU (mostly from Nvidia, increasingly from AMD) is a more general parallel processor that runs AI and much else. TPUs can be cheaper per unit of work for large, steady jobs on Google Cloud, and businesses rent them through Google Cloud. GPUs are available on every major cloud, run almost any AI software through Nvidia's CUDA, and are the safer default for most businesses.

In 2026 this is a budget question, not just an engineering one. Anthropic has announced plans to use up to one million Google TPUs, and Nvidia and AMD have both moved to new rack-scale GPU systems.

This guide compares the current chips on official specs, lists cloud prices per chip-hour, and ends with a verdict. Prices and specs were checked against official Google, Nvidia, AMD, AWS and CoreWeave pages on November 12, 2026.

The difference between a TPU and a GPU is specialization. A TPU (Tensor Processing Unit) is an application-specific chip that Google designed to do one job: multiply large grids of numbers, the core operation in neural networks. A GPU (Graphics Processing Unit) is a programmable parallel processor that runs AI, graphics and scientific computing.

Google's own Introduction to Cloud TPU describes TPUs as "custom-developed, application-specific integrated circuits (ASICs)" with hardware designed for "large matrix operations often found in machine learning algorithms." Every TPU program has to pass through Google's XLA compiler before it runs.

A GPU takes the opposite approach. Nvidia and AMD data center GPUs pack thousands of programmable cores plus dedicated tensor units, so the same GPU can train a language model and run a physics simulation.

Where do CPUs fit? Google's documentation recommends CPUs for quick prototyping and small models, GPUs for medium to large models, and TPUs for "models dominated by matrix computations."

A TPU works by streaming numbers through a systolic array, a grid of multiply-and-add units that pass results directly to their neighbors instead of writing them back to memory after every step. Google's documentation describes each matrix unit (MXU) as "a 128x128 systolic array." Cutting memory trips is where TPUs get their efficiency on AI math.

The trade-off is flexibility. Google lists the workloads TPUs handle poorly: programs that branch often, work that needs high-precision arithmetic, and neural networks with custom operations inside the main training loop. Standard transformer models rarely hit these limits.

TPUs also scale differently. Google links chips with its own inter-chip interconnect, and an Ironwood pod holds up to 9,216 chips, according to Google's TPU7x (Ironwood) documentation. Nvidia's GB300 NVL72 rack connects 72 GPUs over NVLink, and larger clusters link racks over networking.

The newest chips with published rental prices are Google's Ironwood TPU (TPU7x) and Nvidia's Blackwell and Blackwell Ultra GPUs, alongside AMD's Instinct MI355X. Newer generations from all three are ramping now.

Ironwood is Google's seventh-generation TPU and the newest one on Google's public price list. On April 22, 2026, Google announced TPU 8t for training and TPU 8i for inference. Google's announcement said both would be generally available "later this year." On November 9, 2026, Google added two variants: TPU 8tx, built on TPU 8t silicon for confidential inference, and TPU 8i Max, which puts two TPU 8i chips in one package. As of November 12, the Cloud TPU pricing page lists no TPU 8 prices.

Nvidia's shipping flagship is Blackwell Ultra, sold as the HGX B300 eight-GPU board and the GB300 NVL72 rack with 72 GPUs and 36 Grace CPUs. Its successor, Vera Rubin, was announced in January 2026, when Nvidia said Rubin-based products would be available from partners in the second half of 2026. On May 31, Nvidia said Vera Rubin was ramping into full production, with production shipments "set to begin starting this fall."

AMD's MI355X is shipping, and its successor, the MI455X, powers AMD's 72-GPU Helios rack. On AMD's third-quarter 2026 call, CEO Lisa Su said, "We started shipping MI455X in volume in the quarter," and named Anthropic, Meta, Microsoft, OpenAI and Oracle as lead Helios customers, according to Constellation Research's coverage.

On paper, Google's Ironwood and Nvidia's Blackwell Ultra sit in the same performance class per chip, while the newest GPUs from Nvidia and AMD carry far more memory bandwidth. All figures below are the vendors' own peak numbers, not measured results.

Sources: Google's Ironwood documentation, Nvidia's GB300 NVL72 and HGX pages, and AMD's MI350 Series and MI400 Series pages.

Read spec sheets with care: "sparsity" can double a headline FP8 figure. Nvidia's HGX table notes its dense FP8 value is half the sparse one, which works out to about 4.5 petaflops dense per Blackwell Ultra GPU. That is close to Ironwood's 4.6 petaflops. Real speed depends on software and model shape.

No independent benchmark settles whether TPUs beat GPUs in 2026, because Google has left its newest TPU out of recent rounds of the main public benchmark. Most claims come from the chipmakers themselves.

What each vendor claims:

MLPerf, run by the industry group MLCommons, is the closest thing to a neutral scoreboard. Nvidia, AMD and cloud providers submitted to MLPerf Inference v6.1 in September 2026, where Vera Rubin posted its first peer-reviewed numbers, according to StorageReview's report, which says Google's v6.1 submission focused on DeepSeek-R1 without naming the hardware. Google did not submit Ironwood to MLPerf Inference v6.0 in April 2026, having last submitted a TPU (Trillium) in April 2025, according to journalist Tae Kim.

The practical takeaway: treat every "x times faster" claim as marketing until you test your own model on both.

Ironwood costs $12.00 per chip-hour on demand in Google Cloud's Iowa region, falling to $5.40 with a three-year commitment. Nvidia GPU prices range from about $6 per GPU-hour for an older H100 to more than $16 for a Blackwell Ultra, depending on the cloud and the contract.

Sources: Cloud TPU pricing, Google Cloud accelerator-optimized pricing, AWS EC2 Capacity Blocks pricing and CoreWeave pricing, all checked November 12, 2026. Per-GPU figures for CoreWeave and Google's A3 VM are instance prices divided by GPU count.

AWS Capacity Blocks are short-term reservations, not on-demand rates, and AWS plans its next price update for January 2027. CoreWeave lists GB300 NVL72 and HGX B300 as "contact sales." Microsoft Azure sells GPUs through ND-series virtual machines with region-based rates we could not confirm as a fixed public price.

Eight accelerators running for a 30-day month use 5,760 chip-hours (8 x 720 hours). At list prices:

The catch: a chip-hour is not a fixed unit of work, so an Ironwood chip and a B200 finish different amounts of your job in that hour. The number that matters is cost per finished task, such as cost per million tokens. If you don't need to manage chips at all, buying models by the token through the Gemini API, Amazon Bedrock or Azure OpenAI is usually simpler.

Google uses TPUs to power its own Gemini models. Outside Google, the best-known TPU user is Anthropic, the maker of Claude, and Google said on November 9, 2026, that TPUs have "nearly 100 customers," including frontier labs, trading firms, retailers and universities.

Anthropic's case shows that TPUs and GPUs are not either-or. Anthropic says it trains and runs Claude on "AWS Trainium, Google TPUs, and NVIDIA GPUs," with Amazon as its primary cloud and training partner. Our guide to Nvidia's competitors covers the other chips.

Software is the biggest practical difference between TPUs and GPUs. GPUs run almost any AI framework through Nvidia's CUDA platform. TPUs run JAX and PyTorch through Google's XLA compiler, and Ironwood does not support TensorFlow, according to Google's Ironwood documentation.

CUDA's head start is large. Nvidia's NVIDIA in Brief 2026 overview counts more than 7.5 million developers in the NVIDIA Developer Program, and most AI tools and tutorials assume an Nvidia GPU. AMD offers its own ROCm software stack.

Google is narrowing that gap for PyTorch. In October 2026, Google and Meta released TorchTPU, an open-source PyTorch backend for TPUs. The TorchTPU team says porting should be as simple as changing model.to("cuda") to model.to("tpu"), while noting that peak TPU performance takes more work. Google also supports the popular serving engines vLLM and SGLang on TPUs.

The rule of thumb: code built on custom CUDA kernels stays cheapest on GPUs, while JAX or standard PyTorch models can now move to TPUs realistically.

Most businesses should start on GPUs and test TPUs once they run large, predictable AI workloads on Google Cloud. TPUs can pay off at scale for teams willing to commit to Google.

Choose TPUs if:

Choose GPUs if:

Choose neither directly if you mainly use AI through ChatGPT, Claude or Gemini. The model provider picks the chips, as with the products in our guide to Google AI agents.

Sameer's take: I think Nvidia's AI chip lead is safe for years, because of CUDA and scale.

TPUs win on specialization and committed pricing inside Google Cloud. GPUs win on flexibility and software.

Our Nvidia statistics page tracks Nvidia's revenue and data center growth.

What is the main difference between a TPU and a GPU?
A TPU is Google's custom chip built only for the matrix math in neural networks, while a GPU is a programmable parallel processor that runs AI and many other workloads. TPUs trade flexibility for efficiency and are rented through Google Cloud.

Is a TPU better than a GPU for AI?
A TPU can be cheaper per unit of work for large, standard AI models on Google Cloud, but no independent benchmark shows TPUs beating GPUs across the board in 2026. GPUs support more software and run on every major cloud, so they are the safer default.

How much does a Google TPU cost per hour?
Google's Ironwood TPU costs $12.00 per chip-hour on demand in us-central1 (Iowa), $8.40 with a one-year commitment and $5.40 with a three-year commitment. Trillium costs $2.70 and TPU v5e $1.20 per chip-hour on demand, per Google Cloud's pricing page in November 2026.

Can you buy a Google TPU?
Businesses rent TPUs through Google Cloud by the chip-hour, with on-demand or committed-use pricing. Google's Cloud TPU pages do not offer TPUs for sale, unlike Nvidia and AMD GPUs, which are sold through server makers and offered by many clouds.

Does PyTorch run on TPUs?
Yes. Google and Meta released TorchTPU in October 2026, an open-source PyTorch backend that lets most models run on TPUs by changing the device from cuda to tpu. Peak TPU performance still takes extra tuning.

TPUs are Google's specialized chips, rented through Google Cloud, and with Ironwood at $5.40 per chip-hour on a three-year term they can be a strong deal for large, steady JAX or PyTorch workloads. GPUs from Nvidia and AMD run everywhere and support almost any software.

For most businesses, GPUs remain the default. If your AI spend is large and you already use Google Cloud, test your own model on both and compare cost per finished task, not price per chip-hour.

⭐ Get more AI Business Weekly in Google: add us as a preferred source on Google.

Written by Sameer Khan with AI assistance, fact-checked against the sources linked above.