TPU vs GPU: Specs, Prices and Which AI Uses Each

In brief

A TPU is Google's custom AI chip, and a GPU is a general purpose parallel chip. Their memory is now close: Google lists 192 GiB for each Ironwood TPU, and Google Cloud lists 180 GB for each Nvidia B200 GPU, as of November 2026. Google says Gemini 3 Pro trained on TPUs.

What is the difference between a TPU and a GPU?

A TPU does one kind of math for AI and a GPU does many kinds of parallel math. Google makes TPUs and rents them through Google Cloud. Nvidia makes the GPUs that OpenAI and Anthropic name in their announcements, and the big clouds rent them.

Two black plinths in a charcoal studio under white rim light, one holding a small square chip with a fine glass lattice and the other a larger flat board covered in a grid of identical glass cubes, with a thread running between them, in front of a terracotta Claude starburst, a brushed silver OpenAI knot and a rainbow glass Gemini star.
TPUGPUNPU
MakerGoogleNvidia, AMD and IntelApple, Qualcomm, Intel, AMD and others
DesignA chip built for neural network math (an ASIC)A parallel chip that began in graphicsA block inside a laptop or phone processor
Best atLarge, matrix heavy training and servingAlmost any parallel job, from games to AISmall, steady AI tasks at low power
Where you rent oneGoogle Cloud only on the pages readGoogle Cloud, AWS, Azure and othersYou don't; it ships inside the device
Current generationIronwood, listed as TPU7xBlackwell and Blackwell Ultra; Vera Rubin named nextSee what is an NPU

The table covers the three chips people mix up. The rest of this page backs each cell with the vendor's own page and adds what the cheap explainers skip: real per chip numbers, a rental price and which AI models run on which chip.

What is a TPU?

A TPU, or Tensor Processing Unit, is a chip Google designed for machine learning. Google's docs call TPUs "custom-developed, application-specific integrated circuits (ASICs)" used to speed up machine learning work. Google has used them for years, and its Ironwood launch post says "Gemini 2.5 and the Nobel Prize winning AlphaFold all run on TPUs today."

The core idea is a systolic array: a grid of multiply and add units that passes numbers from cell to cell instead of fetching them from memory each time. Google's architecture page says each 256 by 256 matrix unit on TPU v6e and TPU7x does 128K bfloat16 multiply and add operations per cycle.

Google lists three jobs TPUs are not suited to: programs with frequent branching or many element by element operations, work that needs high precision arithmetic, and networks with custom operations in the main training loop. That is the honest edge of the design. A TPU is fast at the thing it was built for and awkward at the rest.

What is a GPU and why did it take over AI?

A GPU is a graphics processing unit: a chip with thousands of small cores that all work at once. It began in video games. AI training turned out to be the same kind of work, huge piles of multiplication done in parallel, so GPUs became the default.

Nvidia's page says its Blackwell GPUs pack 208 billion transistors on a custom TSMC 4NP process, and its fifth generation NVLink can scale up to 576 GPUs. The bigger advantage is software. GPUs run through the tools most developers already use, which is why Google Cloud's own GPU page lists Nvidia chips from the T4 up to Blackwell and Rubin parts.

The weakness is the mirror image of the TPU's strength. A GPU can do more kinds of jobs, so it carries hardware a pure AI chip doesn't need.

Which is faster, a TPU or a GPU?

Neither wins on paper. Per chip, Google's Ironwood TPU and Nvidia's B200 land within about 15 percent on memory, bandwidth and peak FP8 compute (192 GiB is about 206 GB, so memory is the widest gap), and no neutral test in the pages read says otherwise. The table uses Google's TPU7x docs and Nvidia's DGX B200 page, as of November 2026.

Per chipGoogle Ironwood (TPU7x)Nvidia B200
Memory192 GiB HBM180 GB HBM3e (Google Cloud A4 table)
Memory bandwidth7,380 GBpsAbout 8 TB/s (64 TB/s across 8 GPUs, divided by us)
Peak FP8 compute4,614 TFLOPs (sparse or dense not stated)About 4,500 TFLOPs dense (36 PFLOPS dense across 8 GPUs, divided by us)
Chip to chip link1,200 GBps bidirectional14.4 TB/s across 8 GPUs (page figure, not per chip)
Largest linked group9,216 chips in one pod72 GPUs in one NVLink domain (GB200 NVL72)
A TPU and the newest Nvidia GPU now carry almost the same memory per chip
50 GB100 GB200 GBNvidia H100Nvidia B200Google Ironwood TPU

Two cautions. Nvidia states some of its figures as sparse, with dense at half, so we used the dense figure, and Google doesn't say which kind its 4,614 is. And a peak number isn't a result. Novita AI's popular explainer says TPUs "can achieve up to 2-3x faster training times" than a comparable GPU setup, with no benchmark, model or hardware named. Treat that as an unsupported claim.

Is a TPU or a GPU better for inference and power use?

Google calls Ironwood "the first designed specifically for inference," meaning serving answers to users rather than training. Its launch post says Ironwood performance per watt is 2x Trillium, and nearly 30x more power efficient than Google's first Cloud TPU from 2018. Those are Google's own claims about its own chips, not a test against a GPU.

Nvidia's B200 also runs inference, and any GPU can serve a model with the tools most teams already have. So the deciding facts are your software stack and the cloud you already pay, which the how to steps on this page cover. A TPU suits steady, high volume serving on Google Cloud, and a GPU suits mixed jobs or code with custom operations.

How do TPU and GPU prices compare?

Google prints TPU prices per chip hour, and they run from $1.20 for an older chip to $12.00 for Ironwood. Nvidia doesn't set a rental price because it sells chips, and each cloud prices its own GPU instances, so a true side by side needs the price of your cloud. Google's on demand rates below were read on 22 November 2026.

Google TPURegion readOn demand per chip hour
Ironwood (TPU7x)us-central1$12.00
Trillium (v6e)us-east1$2.70
TPU v5pus-east5$4.20
TPU v5eus-central1$1.20

Google also lists cheaper committed and Spot rows on the same page, and notes that billing shows VM hours while one TPU VM can hold several chips. So a bill and a chip price are different numbers.

Where can you rent each? Google Cloud's Compute Engine page lists Nvidia B200 as the A4 machine type, GB200 as A4X, GB300 as A4X Max and H100 as A3. The TPU7x page says the chip is reached through GKE, Compute Engine or reserved capacity.

Which AI models train on TPUs and which on GPUs?

Google trains Gemini on TPUs, OpenAI is signing up for Nvidia systems, and Anthropic splits across three chip platforms. The table quotes what each company says on its own page. Where a page is silent, the cell says so.

Model familyChip the maker namesWhat its own page says
Gemini (Google)TPUsThe Gemini 3 Pro model card says it "was trained using Google's Tensor Processing Units (TPUs)."
Claude (Anthropic)TPUs, Amazon Trainium and Nvidia GPUsAnthropic says its compute strategy "uses three chip platforms" and it plans up to one million TPUs. It doesn't say which chip trains which model
GPT (OpenAI)Nvidia systemsA September 2025 letter of intent covers at least 10 gigawatts of Nvidia systems to train and run next generation models
Llama (Meta)GPUs, brand not named on the pageLlama 3 trained on two custom built 24K GPU clusters

The table says what companies announced, not a full inventory of their chips. A lab can use more than one platform, and Anthropic says it does. The safe reading is that the chip is a business choice and not a quality ranking, because every family above competes at the top of its field.

How do NPUs differ from TPUs and GPUs?

An NPU is a small AI block inside a laptop or phone chip, built to run AI features at low power, while TPUs and GPUs live in data centers. You can't rent an NPU, because it ships inside the device. The TPU is a cousin of the NPU built for racks, not for your battery.

The practical difference is where the work happens. A chatbot answer comes from a TPU or GPU in a data center. Background blur on a video call runs on your NPU. What is an NPU has each maker's TOPS figure and the Copilot+ PC bar.

Do you need a TPU or a GPU to use ChatGPT, Claude or Gemini?

No. You send a question and the vendor's data center does the math on its own chips. The chip behind a reply has no bearing on which app you should pick. ChatGPT vs Claude vs Gemini compares the three on what you see, and what is an LLM explains what the chips are running.

You only choose a chip if you train or host models yourself. A local model on your own machine is a different case, and it usually means a GPU in your PC.

Three real cases where the chip choice shaped the plan

  1. Google trained Gemini 3 Pro on its own TPUs, and its model card dates the release to November 2025. Google makes the chip, the cloud and the model.
  1. Anthropic announced on 23 October 2025 that it would expand to up to one million TPUs and bring well over a gigawatt of capacity online in 2026. It also says it keeps Trainium and Nvidia GPUs in the mix, so it isn't betting on one chip.
  1. Meta wrote that Llama 3 reached a compute utilization of over 400 TFLOPS per GPU on 16K GPUs. That figure belongs to Meta's setup, and it shows that utilization, not the peak on the box, decides how fast a run goes.

What questions should you ask an AI about TPUs and GPUs?

These five prompts work in Claude, ChatGPT, Gemini or Whizi. Turn web search on, because chip lines change every few months.

  1. "Compare [TPU generation] and [GPU model] per chip. Use only Google Cloud and Nvidia pages, give memory, bandwidth and peak FP8 with links, and say not stated where a page is silent."
  1. "My model is [size] parameters, trained in [PyTorch or JAX] with [custom operations or none]. Would a TPU or a GPU fit better, and what would I have to change?"
  1. "Price one hour of [job] on [cloud] using a GPU and a TPU. Show per chip hour prices from the official pages and how many chips the job needs."
  1. "Explain peak TFLOPS versus real utilization to me in three sentences, with one example from a published training run."
  1. "List which of [Gemini, Claude, GPT, Llama] each lab says it trains on, quote each lab's own wording and mark any claim you can't find."

Where does Whizi fit?

Whizi is one app for asking Claude, GPT and Gemini the same question, so you don't need to care which chip runs behind them. Web search is on by default in each chat, which helps with a topic where specs change by the quarter.

The free plan gives you 10 messages a day on Whizi AI, Auto and Claude Haiku, with no card. Pro adds Claude, GPT and Gemini in one app, Compare puts two answers side by side, and its first week is $1.99. For a final figure, open the vendor's own page.

What can this comparison not tell you?

It can't tell you which chip is faster for your job. The numbers here are vendor peak figures, and Meta's 400 TFLOPS per GPU shows real runs sit below a ceiling. Only a test on your model settles it.

It also can't settle opinion. Investor threads argue the topic from the stock side, and a stock view is not a chip spec. One popular explainer lists TPU prices as $1.35 to $5 an hour in its table and $1.35 to $8 in its FAQ. Google's own page, not a blog, is the place to read a price.

Workflow checklist
  • Compare memory per chip before comparing peak speed
  • Check whether a vendor's figure is sparse or dense
  • Read TPU prices on Google's own pricing page, per chip hour
  • Match the chip to your framework and custom operations
  • Run a one hour test on your own model before committing
Common questions

Frequently asked questions

What does TPU stand for?

TPU stands for Tensor Processing Unit. Google's docs describe it as a custom developed application specific integrated circuit used to accelerate machine learning work. Google rents TPUs through Google Cloud, and the newest one on its docs is Ironwood, listed as TPU7x.

Is a TPU faster than a GPU?

Not by a margin the vendors' own pages show. Per chip, Google's Ironwood lists 4,614 TFLOPs in FP8 and Nvidia's DGX B200 page works out to about 4,500 TFLOPs dense per GPU, as of November 2026. Real speed depends on your model and software, so test it.

Does ChatGPT run on TPUs or GPUs?

OpenAI says it plans to train and run its next generation of models on at least 10 gigawatts of Nvidia systems, from a September 2025 letter of intent. That page doesn't list every chip OpenAI uses today, so treat it as the plan, not a full inventory.

Which is better for AI, a TPU or a GPU?

A GPU is the safer default and a TPU is a strong pick for large, matrix heavy jobs on Google Cloud. Google's docs say TPUs suit models with no custom PyTorch or JAX operations in the main training loop. If your code has custom operations, start with a GPU.