Technology
TPUs
Tensor Processing Units designed specifically for machine learning matrix operations and neural network training.
5 companies
Explainer
TPUs explained
A TPU, or tensor processing unit, is a chip designed by Google to run machine learning. A tensor is a grid of numbers, and neural networks are mostly built from multiplying such grids. A TPU is organised around doing that one operation very efficiently.
The core of a TPU is a systolic array, a grid of multiply-accumulate units where data flows from one unit to the next in a regular rhythm. This reuses each value many times without going back to memory. Google began using TPUs internally in 2015 and now offers them to customers through Google Cloud.
How it differs from the usual approach
A GPU is a general parallel processor that has been extended for AI. A TPU was designed from the start for neural network maths, and it gives up some flexibility for efficiency on those workloads. TPUs are mainly rented through Google Cloud rather than bought as chips, though Google has said it will supply a few large customers directly, and they are programmed through frameworks such as JAX and TensorFlow, with PyTorch support also available.
Where it is used
- Training and serving Google's own models and products, such as search and translation.
- Renting TPU capacity on Google Cloud for model training and inference.
- Large-scale training jobs spread across many TPUs joined in pods.
- Research workloads using the JAX framework.
- Third-party AI developers who choose TPUs as an alternative to GPUs.
What to watch
- Google publishes new TPU generations on a regular cycle, with later ones focused on inference as well as training.
- Other companies use "TPU-like" designs, so the term is sometimes applied to any matrix-focused AI chip.
- Demand for non-NVIDIA capacity makes TPUs a visible alternative for large AI developers.
- Software support and portability remain the practical question for teams moving between TPUs and GPUs.
Frequently asked questions
What does TPU stand for?
TPU stands for tensor processing unit. It is a chip family designed by Google for machine learning workloads.
Can I buy a TPU?
Mostly no. Google uses TPUs in its own data centers and rents them through Google Cloud. It has said it will also sell them to a small number of large customers, but they are not sold as general-purpose chips the way GPUs are.
Is a TPU better than a GPU?
It depends on the workload and software. TPUs can be efficient for large matrix-heavy models. GPUs are more flexible and have wider software support.
- AWS Trainium/Inferentia Trainium and Inferentia are custom AI accelerators designed by Amazon Web Services for training and inference, offered as cloud instances and supported by the AWS Neuron software stack. They come from AWS's in-house silicon effort, which grew from its 2015 acquisition of the Israeli chip designer Annapurna Labs and also produced the Nitro and Graviton chips. Trainium2 powers Project Rainier, which AWS built for Anthropic. AWS positions the chips as an alternative to merchant GPUs. AI Accelerators · ASICs · TPUs —
- SambaNova SambaNova Systems builds AI inference hardware and software based on its Reconfigurable Dataflow Unit, or RDU, a chip architecture that maps neural network operations directly onto the processor. Its products include the SN40L, announced in 2023 and made by TSMC, and the SN50, with a three-tier memory design. The company sells systems and offers cloud access through SambaCloud. SambaNova is headquartered in Palo Alto, California, and announced a multi-year collaboration and strategic investment with Intel in 2026 after earlier acquisition reports. AI Accelerators · TPUs · HPC & AI Systems $350M (Intel, Feb 2026)
- Google DeepMind TPU The Tensor Processing Unit is a family of custom AI accelerators designed by Google and offered to customers through Google Cloud. Ironwood, the seventh generation, was unveiled in April 2025 and is aimed at inference, with up to 4,614 TFLOPs and 192 GB of HBM per chip and pods of up to 9,216 chips. TPUs run Google's own models and services as well as outside workloads. The project belongs to Google rather than to DeepMind, which is a sister unit within Alphabet. TPUs · AI Accelerators —
- CSCerebras Systems Cerebras Systems builds AI processors around the Wafer-Scale Engine, a single chip made from an entire silicon wafer rather than a cut die. The third generation, WSE-3, powers the company's CS-3 systems, which it sells for on-premises use and also offers as cloud inference and training services. Cerebras is headquartered in Sunnyvale, California. It listed on Nasdaq under the ticker CBRS on May 14, 2026, after pricing its initial public offering at $185 per share. TPUs · AI Accelerators · HPC & AI Systems IPO May 14, 2026 ($5.55B raised, ~$95B market cap on Nasdaq)
- Google (TPU) The Tensor Processing Unit is a family of chips that Google designed for neural network workloads, with Broadcom helping turn the designs into manufacturable silicon. It is an ASIC, meaning a chip built for one job rather than general use, so it trades flexibility for efficiency on matrix maths. Google began using TPUs internally in 2015 and announced them in 2016. It now rents them to outside customers through Google Cloud. The seventh generation, Ironwood, was unveiled in April 2025 and can be linked in pods of up to 9,216 chips. TPUs · Custom Silicon —
