TPU vs GPU in 2026: Prices, Frameworks, When Each Wins

What a Google TPU is, what each generation costs per chip-hour in September 2026, how that sits beside live GPU rental prices, and when to pick which.

By Faiz Ahmed
12 min read

The short answer

A TPU is Google's own machine learning accelerator, rented by the chip-hour inside Google Cloud and programmed through the XLA compiler from JAX or PyTorch. It is a real option if your code is already JAX or plain PyTorch with no custom CUDA kernels, you are content to run on Google Cloud, and your job is large or long enough to earn a commitment discount. For everything else, rent a GPU: it runs every framework as it is, and you can buy it from many providers instead of one.

CPU, GPU, TPU in one table

CPUGPUTPU
What it isGeneral-purpose processor. Runs the operating system, data loading and any branching logicParallel processor with its own high-bandwidth memory. Runs graphics, scientific code and machine learningGoogle's accelerator built for the matrix arithmetic in neural networks. It does nothing else
Who rents it outEvery cloudEvery cloud, plus specialist GPU providersGoogle Cloud only, for self-serve buyers
How you program itAny languageCUDA on NVIDIA cards, under every major frameworkThe XLA compiler, reached from JAX or PyTorch. No CUDA
Memory per chipSystem RAM, shared24 GB on an L4 up to 180 to 192 GB on a B20016 GB on a v5e up to 192 GiB on Ironwood
How it scales upSockets and networkNVLink inside a server, InfiniBand between serversGoogle's ICI links, up to 9,216 chips in one Ironwood pod
Weak atLarge matrix arithmeticNothing specific to machine learningFrequent branching, custom operations, high-precision arithmetic (Google's own list)

Two notes on the table. Every TPU still needs a CPU next to it: Google's documentation shows a four-chip Ironwood virtual machine with 224 vCPUs and 960 GB of RAM. And the GPU memory figures come from this site's datasheet-verified spec table:

SpecL4H100H200B200
VRAM24 GB80 to 94 GB141 GB180 to 192 GB
Memory bandwidth300 GB/s3,350 GB/s4,800 GB/s8,000 GB/s
ArchitectureAda LovelaceHopperHopperBlackwell
Launch year2023202220242024
Figures from the vendor datasheets: L4, H100, H200, B200, checked 13 Sep 2026. "Not published" means the vendor gives no figure.

Google says TPUs suit models "dominated by matrix computations", models "that train for weeks or months", and large models "with large effective batch sizes". It describes the wrong fit as programs "that require frequent branching", workloads "that require high-precision arithmetic", and networks "that contain custom operations in the main training loop". Both lists are from Google's TPU introduction page on 21 September 2026.

The buyers are large. Google Cloud announced on 23 October 2025 that Anthropic would have access to up to one million TPU chips. Apple's 2024 foundation model paper says its server model was trained on 8,192 TPU v4 chips. Neither fact tells you how a TPU will run your job.

What TPUs cost in September 2026

Google publishes TPU prices "per chip-hour in USD". These are list prices read from the Cloud TPU pricing page on 21 September 2026. Many older articles quote different numbers, and those numbers are out of date.

GenerationHBM per chipRegionOn demand1-year commitment3-year commitmentDWS Flex-start
Ironwood (TPU7x)192 GiBus-central1 (Iowa)$12.00$8.40$5.40$6.00
Ironwood (TPU7x)192 GiBeurope-west2 (London)$13.20$9.24$5.94$6.00
Trillium (v6e)32 GBus-east1, us-east5$2.70$1.89$1.22$1.35
Trillium (v6e)32 GBeurope-west4 (Amsterdam)$2.97$2.08$1.34$1.35
Trillium (v6e)32 GBasia-northeast1 (Tokyo)$3.24$2.27$1.46$1.35
v5p95 GiBus-east5, us-east1$4.20$2.94$1.89$2.10
v5e16 GBus-central1, us-east5, us-west1, us-west4$1.20$0.84$0.54$0.60

The v5e costs more in some other regions: $1.416 on demand in us-south1 (Dallas) and $1.56 in europe-west4 and asia-southeast1. Three older generations are still priced. A v4 pod in us-central2 lists at $3.22 on demand, a v3 pod in europe-west4 at $2.00, and a v2 pod at $1.50. Google's documentation says the v2 runs on an API that is "no longer under active development" and tells users to migrate, and it gives similar advice for v3 and v4. Do not start new work on them.

How to read the table:

  • The unit is a chip, not a machine. You rent chips in fixed groups attached to a host. Google's example: a single v4 host, with four chips and one VM, shows in the console at $12.88 per hour. Billing runs while the TPU node is in a READY state, whether or not your code is doing anything.
  • Commitments cut the price by a fixed share. On Google's numbers, the one-year price is 30% below on demand and the three-year price is about 55% below, for both Ironwood and Trillium in their US regions. Google describes the one-year commitment as aimed at "32-512 cores" and the three-year at "512+ cores", billed monthly on reserved quota. These are not plans for someone renting one chip.
  • Flex-start is half price with a catch. DWS Flex-start lets you queue for TPUs "for up to 7 days" without a reservation. Every Flex-start price in the table is exactly half the US on-demand price. You wait for capacity, and the run is capped at a week.
  • Calendar mode lets you check availability up to 120 days ahead and book a slot. Its listed price equals the US one-year commitment price for each generation above.
  • Spot has no list price. Google says only that "Spot prices are dynamic and can change up to once every 30 days".
  • Regions are not settled. Google's pricing page and its zones page disagree about where Ironwood and v5p are offered. Check both before you plan around a region. Google also warns that larger TPU types "are available only in limited quantities".

The 2026 facts older articles miss

  • Ironwood is generally available. TPU7x entered preview on 24 November 2025 and reached general availability on 31 March 2026, per Google's release notes.
  • TensorFlow does not run on Ironwood. Google's TPU7x page says: "You can use the JAX and PyTorch frameworks on TPU7x. TensorFlow is not supported." The newest TensorFlow release Google lists as supported on any Cloud TPU is 2.18.0, from October 2024. TensorFlow tutorials for older TPUs do not carry over to the newest hardware.
  • TPUs are leaving Google's buildings, for a few. Yahoo Finance reported that Sundar Pichai said on 29 April 2026: "we'll begin to deliver TPUs to a select group of customers in their own data centers." For self-serve buyers nothing has changed.
  • An eighth generation is announced but has no price. Google presented two chips in April 2026, TPU 8t for training and TPU 8i for inference, and said both "will be generally available later this year". Neither appeared on the pricing page on 21 September 2026. Ironwood is the newest TPU you can price today.

Where you can get one

Google Cloud, in the regions above, is the only self-serve source. Colab's free tier offers a single-chip TPU v5e-1, which replaced the old v2-8 according to a Colab team member's announcement of 30 August 2025. That is useful for learning JAX and not for much else. None of the RunPod, Lambda or Vast.ai pricing pages reviewed on 21 September 2026 lists a TPU. The live table in the next section covers GPUs, not TPUs.

What GPUs cost right now

These are live rental prices, the cheapest current listing for each GPU across the providers this site tracks.

GPUCheapest $/GPU-hrProviderProviders in stock
H100$2.50Hyperstack11
H200$3.50Ori5
B200$3.75Packet.ai2
L4$0.33Vast.ai4
Cheapest in-stock on-demand price per GPU-hour, from providers with live stock tracking. Latest stock observation: .

Read this before you compare the two tables. A TPU chip and a GPU are not equivalent units. They differ in memory per chip, in how many chips come attached to one host, and in what software runs on them. A lower price per chip-hour does not mean a lower cost per trained model or per generated token. Google's prices above are list prices from one vendor. The GPU prices are the cheapest live listing across many vendors. We have not benchmarked TPUs against GPUs, and this page gives no performance ratio because we do not have one we trust.

What the two tables can tell you is narrower and still useful:

  • Memory per dollar of rental. The cheap TPU generations have little memory per chip: 16 GB on a v5e and 32 GB on a v6e, against 80 GB on the standard H100. A model that fits on one H100 must be split across several v5e or v6e chips. Count the chips you need for the memory, then multiply by the chip price. Our LLM VRAM calculator gives the memory figure.
  • The price of commitment. Google's large discounts need a one-year or three-year term. The GPU prices in the live table are on demand with no term. If you want to see what commitment does on the GPU side, the GPU cluster page covers reserved and multi-node pricing.
  • Direction of travel. GPU rental prices move daily and are tracked in the GPU price index. TPU list prices change when Google edits a page.

For the GPUs themselves, see H100 vs H200 and B200 vs H200, or the best GPU for LLM work guide.

What changes when you leave CUDA

This is the real cost of a TPU, and it does not appear on any price page.

Your code goes through a compiler. Google's product page says code written in the supported frameworks "is compiled by the Accelerated Linear Algebra (XLA) compiler". JAX and PyTorch are the two frameworks Google names as native on TPUs, and both go through XLA. Neither reaches a TPU through CUDA.

PyTorch support is improving and not finished. Google announced TorchTPU on 7 April 2026 as a native PyTorch stack for TPUs. Its claim is that developers can "take an existing PyTorch script, change their initialization to 'tpu', and run their training loop without modifying a single line of core logic". That is a vendor claim. The same post describes a public GitHub repository as a 2026 roadmap item, so check what has shipped before you plan around it.

Custom kernels do not come with you. CUDA kernels run on NVIDIA hardware only. Google lists "custom operations in the main training loop" as a poor fit for TPUs. If your model depends on a hand-written attention kernel or a CUDA extension from a research repository, someone has to rewrite it. NVIDIA's own speed-up libraries stay behind too: Transformer Engine, for example, describes itself as a library "for accelerating Transformer models on NVIDIA GPUs".

Inference has a supported path. vLLM runs on TPUs through the tpu-inference plugin, which its documentation describes as "unifying JAX and PyTorch under a single lowering path within the vLLM project". If you serve standard open models with vLLM, this is the shortest route to trying a TPU.

You take on one supplier. On GPUs you can move between many providers in an afternoon with the same container. On TPUs the supplier, the regions, the quota process and the price list all belong to Google. If your training data sits in another cloud, add the cost of moving it, which our data egress reference covers.

Budget time for the move. We have no measured figure for how long a port takes. A test is cheap: run your model on a single v5e or v6e chip on demand for a day and see what breaks.

Decision rule: TPU or GPU

Choose a TPU if all of these are true:

  • Your code is JAX, or PyTorch with no custom CUDA kernels in the training loop.
  • You are already on Google Cloud, or willing to move your data there.
  • The job is dominated by large matrix arithmetic, such as transformer training or high-volume serving of a standard model.
  • You will run for long enough to sign a one-year or three-year commitment, or your jobs fit inside the seven-day Flex-start window.
  • You do not need TensorFlow on the newest hardware.

Choose a GPU if any of these is true:

  • You depend on CUDA kernels, NVIDIA libraries or a research codebase you do not want to port.
  • You need a single accelerator with a lot of memory. One H100 holds what several v5e or v6e chips hold between them.
  • You want to rent by the hour with no term and leave when a cheaper provider appears.
  • You need to run outside Google Cloud, in a specific country, or on your own hardware later.
  • Your team is small and time spent porting costs more than the hardware saving could return.

The rule: default to a GPU. Move to a TPU only when your code already compiles through XLA and your usage is steady enough to commit for a year, and then test on one chip before you sign anything.

Sources

Source pages were retrieved with automated research tools on the dates shown and each figure was traced back to its source before publishing. Rental prices on this page are not typed: they are read live from GPUperhour's own data.

All pages accessed 21 September 2026.

Frequently asked questions

What is a TPU?

A TPU (Tensor Processing Unit) is Google's own accelerator chip for machine learning. You rent it by the chip-hour inside Google Cloud and program it through the XLA compiler, normally from JAX or PyTorch.

Is a TPU faster than a GPU?

There is no general answer, and this page does not give a ratio. A TPU chip and a GPU are different units with different memory sizes, so compare the cost of finishing your own job on each.

Can I run PyTorch on a TPU?

Yes. Google lists PyTorch and JAX as supported on its newest TPU generation, and vLLM runs on TPUs through a plugin. Code that relies on custom CUDA kernels has to be rewritten first.

Does TensorFlow still run on TPUs?

On older generations, yes. Google's documentation for the newest generation, Ironwood, says TensorFlow is not supported there.

Where can I rent a TPU?

For self-serve buyers, Google Cloud is the only place, with a single small TPU also offered in Colab's free tier. None of the RunPod, Lambda or Vast.ai pricing pages we reviewed lists a TPU.

Can I buy a TPU for my own data centre?

Not as a normal customer. Google said in April 2026 that it would begin delivering TPUs to a select group of customers in their own data centres, which does not cover self-serve buyers.

Related Posts