Google Expands TPU Lineup With Secure AI and Larger Memory Systems to Challenge Nvidia

Two new systems target the next demands of AI

Google has expanded its eighth generation of Tensor Processing Units with two specialized systems: TPU 8tx, which protects sensitive information while AI processes it, and TPU 8i Max, which combines two chips in one package to handle much larger amounts of information. Announced on November 9, 2026, the additions bring Google's TPU lineup to four systems introduced in a single year, a first for the company.

The new systems address two practical problems facing businesses adopting AI agents: keeping private data secure and giving software enough memory to work through entire collections of documents, computer code or video. Shipments are expected in 2027. Google has not specified a delivery date or pricing for either system.

Google's TPU portfolio announcement describes both products as configurations built around existing silicon. TPU 8tx uses the same underlying chip as TPU 8t, paired with fifth generation AMD EPYC central processing units. TPU 8i Max packages two TPU 8i chips together, doubling computing capacity, certain onboard memory and memory bandwidth.

A TPU is a processor designed to carry out the mathematical operations used by AI. Google originally developed these processors for its own services, including speech recognition and translation. They now support Gemini, external cloud customers and deployments in customer data centers.

The expansion strengthens Google's challenge to Nvidia by offering customers more choices for specific jobs. Rather than asking one accelerator design to serve every requirement, Google is assembling different combinations of processors, memory, networking and security around its existing chips.

Why four systems arrived in one year

Google introduced TPU 8t for large training jobs and TPU 8i for rapid AI responses at Google Cloud Next in April 2026. Training builds or changes a model through examples. Inference is the process of using that model to produce an answer or perform a task.

Google says TPU 8t delivers up to 2.7 times the training performance of its previous generation, while TPU 8i offers 80% better inference price performance. Those comparisons are against Google's earlier TPUs. The latest additions extend the same generation into confidential computing and work requiring much larger memory capacity.

Amin Vahdat, Google's chief technologist and senior vice president for AI and infrastructure, described the priorities in the announcement:

As enterprises bring autonomous agents into production, we have created two new configurations of these foundational systems to solve two urgent new challenges: hardware-level data security and million-token context scale.

Google says designing a new chip takes years, while adjusting packaging, memory, optical networking and security around an existing chip can take months. Its modular approach lets engineers adapt systems to changing software requirements before a completely new processor is ready.

The development and delivery schedule now includes:

• April 2026: Google introduced TPU 8t and TPU 8i at Google Cloud Next.
• October 2026: Google says it released TorchTPU as open source software to support PyTorch development on TPUs.
• November 9, 2026: Google announced TPU 8tx, TPU 8i Max and additional infrastructure software.
• 2027: The new systems are expected to begin shipping, while Oracle plans to start offering TPU access to its cloud customers.

TPU 8tx protects information during processing

TPU 8tx targets governments and regulated businesses, including financial and healthcare organizations. These customers may need AI to examine highly sensitive records while keeping both the information and the model's internal parameters protected.

Confidential computing addresses the moment when information is actively being used. Encryption already protects data stored on a drive or moving across a network. During processing, information normally needs to be accessible in working memory. Google says TPU 8tx adds hardware protections that keep sensitive data and model weights isolated during that stage.

AMD's fifth generation EPYC processors provide the host computing platform. The system also uses encryption and authentication on PCIe connections, the links that move information between major computer components. Technical details accompanying the launch list 216 GB of high bandwidth memory for TPU 8tx.

Google pairs the system with its Boardfly network, which connects processors so they can exchange information quickly. The company says this configuration can reduce latency by as much as half compared with TPU 8t, while imposing a near zero performance penalty for its security protections. Latency is the delay before a system completes a step or responds.

TPU 8tx also supports smaller training jobs. Organizations can adapt a model using private information, compress a larger model into a smaller one through a process called distillation, or train agents through reinforcement learning, which uses feedback to improve their actions. Google says these tasks can take place alongside inference on one system, reducing the need to move sensitive information between separate clusters of computers.

TPU 8i Max gives large tasks more working room

TPU 8i Max addresses a different constraint: processors can perform calculations faster than memory can supply the information they need. Google calls this the memory wall. Adding computing power works best when memory capacity and the speed of data movement grow with it.

The system combines two TPU 8i chips in one package. Its specifications include 384 GB of high bandwidth memory, 17.2 terabytes per second of memory bandwidth and 768 MB of SRAM, a fast form of memory located on the chip.

Google says the package doubles computing capacity, SRAM and high bandwidth memory bandwidth compared with TPU 8i. High bandwidth memory capacity rises by 33%, so the amount of information stored grows by a smaller proportion than the speed at which information can move.

The intended workloads include whole software codebases, lengthy documents, hours of video and groups of agents working together. The system targets context windows on the scale of a million tokens. A context window is the information a model can refer to while handling a request; tokens are the small pieces of text or other content the model processes.

For a coding task, that extra working room can allow an agent to keep more related files available as it investigates a problem. Drug research and mathematical work are other intended applications. Google also says fitting twice the silicon into the existing rack footprint increases computing capacity per square foot and per watt.

AI agents create work beyond a single answer

Mark Lohmeyer, Google Cloud's vice president and general manager of AI and computing infrastructure, said an individual agent request can trigger hundreds or thousands of operations. He estimated that this shift can increase computing requirements by between 10 and 1,000 times, depending on the task.

An agent may break a request into steps, query a database, write code, call another software tool and wait for a person's approval. That pattern creates demand for quick communication between processors, persistent working memory and protection for information moving through several systems.

Google is also trying to reduce the cost of pauses between those operations. Its new Agent Compute Engine automatically suspends and resumes idle agent sessions, freeing resources while an agent waits for a database response, an external tool or human feedback.

The practical goal is to stop expensive processors from remaining occupied by sessions that have nothing to calculate. For customers running many agents at once, managing those pauses becomes part of controlling infrastructure costs.

Software aims to make switching hardware easier

Google's hardware expansion comes with a broader effort to make TPUs easier to adopt. Nvidia's position rests partly on the software that developers already use, so Google's strategy includes support for familiar development tools as well as new processors.

TorchTPU, developed with Meta and the PyTorch community, provides a native TPU backend for PyTorch, a widely used framework for building AI models. Google says existing PyTorch code and Hugging Face models can run on TPUs without rewrites. Its launch presentation described a basic setup requiring two added lines of code.

Google's Cloud infrastructure announcement also describes TPU support for vLLM and SGLang, open source engines used to serve AI models. Work with Inferact and RadixArk supports a broader approach in which customers can operate both TPUs and graphics processing units.

The new XLA Serving Engine lets developers train models in Python and serve them through a C++ runtime across TPUs, GPUs and CPUs. A TPU agent suite, available in preview, connects with coding tools including Claude Code, Cursor and Google Antigravity to help configure workloads and identify bottlenecks.

Google cites strong price performance on InferenceX and AgentX tests of coding agents and work with large contexts. Those are company performance claims; buyers still need to compare the cost, response speed and accuracy of their own applications when choosing infrastructure.

Memory management becomes part of the product

Google's new TPU Sync software moves beyond the memory physically attached to an accelerator. It coordinates chip memory, host computer memory, shared storage and the data center network so a workload can use them as different layers of a larger memory system.

A central target is the KV cache, a model's stored working information from content it has already processed. Large context windows and many simultaneous agents can make these caches very large. TPU Sync can move them off the chip in milliseconds, preserving expensive accelerator memory for active work.

Google says the software can cut KV transfer latency and the time before an initial response by more than 90%. It also supports moving context between different processing stages and synchronizing model weights during reinforcement learning.

Two other updates focus on organizing clusters. Cloud Pathways, now generally available through Google Kubernetes Engine, lets a single controller coordinate distributed TPU work. GKE dynamic slicing uses optical circuit switching to assemble larger groups of TPUs or divide them into smaller isolated groups as needs change.

Google claims dynamic slicing can make job startup up to five times faster and fault recovery up to 4.5 times faster. Together, these tools address the same commercial problem as the new hardware: turning installed computing capacity into useful work with less waiting.

A larger business beyond Google Cloud

Google says its TPU business now has nearly 100 customers, including AI laboratories, trading firms, retailers and universities. The company is extending access beyond its own cloud, including direct deployments in customer data centers. Select customers will be able to deploy TPU 8tx on their own premises.

Oracle has announced plans to make TPUs available to its cloud customers beginning in 2027. Anthropic holds options to purchase up to $150 billion in chips between 2028 and 2031, while OpenAI is also reported to be planning TPU purchases. The Anthropic figure represents a potential purchase ceiling under those options.

Jump Trading offers an example of customers choosing different hardware for different tasks. The firm says TPUs improve energy efficiency and reduce costs for some work, while Nvidia products remain its main platform for training AI models.

Analyst Mandeep Singh estimates that TPUs cost about 30% less than Nvidia GPUs while providing competitive data processing capability. A research projection puts TPU sales, including associated hardware and software, at about $12 billion in 2026 and more than $135 billion in 2028, excluding revenue from customers accessing TPUs through Google Cloud.

That forecast would mean growth to more than 11 times the 2026 figure in two years, using $135 billion divided by $12 billion. It describes a very rapid expansion whose delivery depends on manufacturing capacity, customer spending and deployment schedules.

Supply and construction limits shape the opportunity

Google said in October that customers had already reserved all the TPUs it could supply. Memory shortages and limits on manufacturing capacity remain constraints, and Vahdat's team is working to diversify supply sources.

Industry assessments place the eighth generation TPU family's advanced manufacturing and system packaging within Taiwan Semiconductor Manufacturing Company's ecosystem. Growing TPU demand therefore creates opportunities for suppliers in Taiwan as well as for Google, while making access to advanced production and packaging capacity a practical concern.

Data center construction faces a separate constraint. Singh estimates that opposition to new facilities could reduce the technology industry's projected $1.4 trillion in infrastructure spending for 2027 by roughly 5% to 7%. Applied to that projection, the reduction would amount to about $70 billion to $98 billion.

Greater computing density can help operators do more within an existing building, which gives TPU 8i Max a practical role where space is tight. Google must still obtain chips and memory, deliver systems and build a profitable business. Singh estimates that TPU profit margins remain below Nvidia's.

The competitive contest now reaches beyond peak processing speed. Google is combining confidential computing, larger memory systems, familiar software and more flexible deployment to win specific workloads. Its ability to supply those systems will determine how quickly customer reservations become revenue.

Key Points

• Google announced TPU 8tx and TPU 8i Max on November 9, 2026, with shipments expected in 2027.
• TPU 8tx combines existing TPU 8t silicon with AMD EPYC hosts and hardware protections for sensitive data in use.
• TPU 8i Max combines two TPU 8i chips, with 384 GB of high bandwidth memory and twice the computing capacity and memory bandwidth of TPU 8i.
• Google introduced four TPU systems in 2026 and says its TPU business has nearly 100 customers.
• TorchTPU, TPU Sync and other software updates aim to simplify development and reduce idle computing time.
• Oracle access and direct customer deployments widen Google's market, while memory, manufacturing and construction constraints limit expansion.