Google today added two systems to the eighth generation of its Tensor Processing Units, the custom AI chips it designs in-house, bringing the number of TPU systems it has introduced this year to four. The new TPU 8tx is built for secure and confidential workloads, and the new TPU 8i Max packs two chips into a single package for AI models that must hold very large amounts of context, according to a post by Amin Vahdat, Senior Vice President and Chief Technologist for AI and Infrastructure at Google.

The post also puts a size on a business Google rarely quantifies. TPUs are now "a rapidly growing business with nearly 100 customers," according to Vahdat, a group that includes frontier labs, trading firms, retailers and universities. What the announcement does not contain is a price, a release date or a list of regions for either new system.

In Short

Google announced two more versions of its own AI chips today: one keeps data and AI models sealed off even while they are being processed, and the other combines two chips in one package to handle AI tasks that need huge amounts of memory. The chips matter beyond engineering because they set the cost and capacity behind Google's Gemini models, its cloud business and the nearly 100 organizations that now pay to use TPUs. What changes for now is the menu rather than the bill, since Google gave no prices, release dates or regions for either system.

Four systems in one year

Google first designed TPUs more than a decade ago for internal needs such as voice recognition and language translation. According to the post, the chips now serve four purposes: training and serving Google's own Gemini models, running agentic workloads, delivering services to Google Cloud customers, and, "most recently," letting customers deploy TPUs "directly into their own data centers."

That last use is the newest commercial line. When Alphabet raised its 2026 capital expenditure guidance to $180 billion to $190 billion in June, PPC Land reported that the company was moving to deliver TPUs directly into enterprise customers' own data centers, alongside the TPU 8t and TPU 8i chips announced in April.

Those two chips set the baseline for the generation. At Google Cloud Next in April, Google introduced TPU 8t for large-scale training and TPU 8i for real-time reasoning. According to the post, TPU 8t offers up to 2.7 times better training performance than the previous TPU generation, and TPU 8i offers 80% better price performance for inference, the work of running a trained model to produce answers. Today's additions extend that pair rather than replace it. "This is the first time we have introduced four TPU systems in a single year," Vahdat writes.

Why two more, and why now? The post attributes the expansion to customer demand from outside Google. "While we've always co-designed our TPUs in partnership with Google DeepMind, we're now increasingly getting feedback from external organizations on their specific needs," Vahdat writes. Two requests are named. "Businesses are asking for accelerators that can reason across AI models' massive context windows. In addition, some highly regulated customers are asking for specialized encryption capabilities for unique workloads."

A new system is not a new chip

The distinction matters for reading the announcement. Neither TPU 8tx nor TPU 8i Max is new silicon. TPU 8tx is "built on the same silicon as TPU 8t," and TPU 8i Max combines two existing TPU 8i chips on a single package. Four systems in a year, in other words, rest on two chip designs.

Vahdat makes the point himself. "There's an important difference between a new chip and a new system," he writes. Designing a chip is "a multi-year process bound by the laws of physics and a complex, long-duration manufacturing lifecycle." A system, by contrast, "is about how we bring together the compute, networking, and storage around the chip."

Google presents this as a deliberate choice made when the first TPU was built. "We knew AI software would evolve much faster than silicon manufacturing ever could," the post states, so the company "made a foundational bet on system balance and modularity," treating TPUs as building blocks within what it calls its AI Hypercomputer. Engineers can adjust packaging, high-speed memory, optical networking and security around the chips, an approach that, according to Google, "allows us to deliver specialized supercomputers in a matter of months rather than years."

What a TPU system looks like in practice was spelled out in Alphabet's third-quarter materials. The earnings presentation described trays carrying four proprietary chips on one liquid-cooled board, racks of 16 TPU machines and pods of 144 TPU racks, connected by optical circuit switches and supported by software including JAX, PyTorch, the XLA compiler and the vLLM and SGLang serving engines. Today's post does not say how TPU 8tx or TPU 8i Max map onto those trays, racks and pods.

TPU 8tx: protecting data while it is processed

TPU 8tx is described as Google's third eighth-generation system, "optimized for secure and confidential inference and is also great for small-scale training." Google says it is built for regulated enterprises and governments that run frontier models or proprietary models, and autonomous agents, "alongside their most sensitive data."

The problem it addresses is a gap in how data is normally protected. "Normally, data is encrypted when stored or sent over a network, but has to be unlocked in memory while a computer processes it," the post explains. This is the territory of confidential computing: keeping data shielded during processing, not only at rest or in transit.

Google's answer pairs TPU 8t silicon with 5th Generation AMD EPYC CPU hosts. According to the post, TPU 8tx "builds security directly into the hardware layer, keeping sensitive data and model weights protected and isolated even while in active use." Model weights are the learned parameters that make up a trained AI model, and for companies that have spent heavily on training their own models, they are among the most valuable assets they hold.

The system also borrows its network from TPU 8i. TPU 8tx adopts the Boardfly network that Google created for real-time reasoning, and according to Google, that gives it up to 2x lower latency than TPU 8t, "with a near-zero performance penalty."

Two details in that sentence deserve attention. The comparison is with TPU 8t, a system built for training, not with TPU 8i, the reasoning system whose network TPU 8tx borrows. And the "near-zero performance penalty," presumably the cost of the security features, is not quantified anywhere in the post.

The second half of the pitch is consolidation. "Pairing training-grade compute with a fast inference network also means organizations can simultaneously run and customize their models on a single system," the post states. According to Google, TPU 8tx lets teams fine-tune models on their private data, distill large models into smaller ones and train agents through reinforcement learning, "all without needing to migrate sensitive data between separate clusters." A caption in the post sums up the intended use: "hardware-level confidential compute for private fine-tuning and secure agent execution."

What the security claim does not explain

The post describes outcomes, not mechanisms. It does not say which component holds the encryption keys, how a customer can verify that the isolation is in place, whether the protection covers the TPU's own memory as well as the AMD host, or whether any outside party has assessed the design. For a product aimed at regulated enterprises and governments, those are the questions a compliance team would ask first, and the announcement leaves them for later documentation.

Confidential computing is already part of ad tech

For marketers, the idea behind TPU 8tx is not new, even if the hardware is. In September 2024, Google introduced Confidential Matching, which uses confidential computing and trusted execution environments to process first-party data for Customer Match and Google Ads Data Manager, where it became the default for data connections. In April 2025, Google extended the approach to enhanced conversions for web and said it would open-source the technology.

Meta has taken a similar route for AI. Its WhatsApp Incognito Chat, announced in May 2026, runs on a server-side system called Private Processing that uses AMD SEV-SNP CPUs and Nvidia H100 GPUs with confidential computing features. TPU 8tx applies the same principle to Google's own accelerators and, notably, to training as well as serving.

The commercial worry behind such products has also been aired at the top of the industry. In July, Microsoft chief executive Satya Nadella argued that companies using AI services must reveal their own know-how to get useful results, and that enterprises will demand the right to use model outputs to fine-tune or train their own models. Hardware isolation does not settle who owns the knowledge a model absorbs, but it is aimed at the related fear that a cloud provider, or anyone else with access to the machine, can see the data while it is in use.

The post itself makes no reference to advertising. Its named audiences are regulated enterprises and governments.

TPU 8i Max: two chips, one package

The fourth system targets a different constraint. "Our fourth system, TPU 8i Max, is built for a different problem: the memory wall," Vahdat writes.

The memory wall describes the point at which a processor sits idle because it cannot get data from memory fast enough. TPU 8i was already designed to tackle it, "cutting inter-chip communication delays by up to half," according to the post. The workloads Google now describes are larger still. "When dozens of AI agents work together across an entire software codebase or hours of video at once, memory bandwidth quickly becomes the bottleneck," the post states. "So we took two low-latency TPU 8i chips and combined them onto a single package."

The result, according to Google, scales "compute and memory in lockstep, delivering double the compute, double the SRAM and double the HBM bandwidth of TPU 8i." SRAM is the small, very fast memory built into the chip itself; HBM, or high bandwidth memory, is the larger pool of stacked memory placed beside the processor. Keeping both well supplied is meant to ensure, in the post's words, that "processors remain equipped with essential data and so that large workloads do not stall."

The doubling follows from the design. Two chips on one package deliver twice the resources of one chip, so the per-chip figures are unchanged. The gain Google emphasizes is density: "because TPU 8i Max fits twice the silicon within the existing rack footprint, data centers can deploy significantly more computing power per square foot and per watt." The post gives no figure for that improvement.

Nor does it give absolute specifications. There is no HBM capacity, no bandwidth in terabytes per second, no power draw and no indication of whether software sees the two chips as one device or two. Each of those determines what "massive context windows" means in practice.

Memory is also a cost line, not only an engineering limit. In Anthropic's IPO exhibits, the lease under which Broadcom rents TPU8i pods to an Anthropic subsidiary allows memory-related increases in TPU prices to pass through to pods not yet delivered. A system that doubles HBM per package doubles exposure to that component as well.

Google summarizes the four systems as "a purpose-built accelerator for every stage of the agentic lifecycle": TPU 8t for scale-up training, TPU 8i for low-latency reasoning, TPU 8tx for confidential computing and TPU 8i Max for "massive context windows and the highest compute density per rack."

The software around the hardware

A large part of the announcement concerns software, and some of it is available now. The stated aim is to make TPUs "as easy-to-use as general-purpose hardware." When the first TPU launched, according to the post, "there were only a handful of companies running workloads that needed the processing power of a TPU," and engineers relied on "optimized frameworks and hand-tuned code." Google lists six areas of work.

PyTorch support. Last month, Google open-sourced TorchTPU, so that, according to the post, "PyTorch code and Hugging Face models now run natively on TPUs," with ready-made TPU backends for the vLLM and SGLang inference engines. PyTorch is one of the most widely used frameworks for building AI models, and the post's heading for this item promises "no code rewrites needed." A new open-source TPU agent suite, in preview, plugs into coding tools including Claude Code, Cursor and Google Antigravity to "automate TPU onboarding in minutes, diagnose bottlenecks, and autotune custom kernels using plain language." Claude Code is made by Anthropic, whose IPO filings show large commitments to TPU capacity. A new open-source XLA Serving Engine lets developers train models in Python and serve them from a low-latency C++ runtime across TPUs, GPUs and CPUs.

Benchmarks. Google says its vLLM-TorchTPU software stack "achieved state-of-the-art price-performance marks" on the InferenceX and AgentX benchmarks, which test agentic coding and long-context work. The post contains no figures, names no competing hardware and does not identify the models tested. Its supporting links point to two posts on the SemiAnalysis newsletter. "Upcoming software optimizations on this same hardware will drive those results even further," the post adds, without a date.

Idle agents. "Because AI agents are stateful, they frequently pause for seconds or hours while waiting on a database lookup, an external software tool, or human feedback," the post observes. A new Agent Compute Engine "automatically suspends and resumes idle sessions, cutting operational infrastructure costs." The size of the saving is not stated.

TPU Sync. Google is open-sourcing what it calls "the model-state layer for TPU clusters." TPU Sync treats chip HBM, host DRAM, shared storage and the data center network as one memory hierarchy. It moves KV caches, the stored intermediate results a model keeps for text it has already processed, off the chip "in milliseconds," streams context between separate prefill and decode nodes, and synchronizes reinforcement learning weights. According to Google, the result is to "cut KV transfer latency and Time to First Token (TTFT) by more than 90%." Time to first token is the delay before a model starts producing its answer. The baseline for the 90% figure is not given.

Cloud Pathways. Now generally available on Google Kubernetes Engine (GKE), Cloud Pathways "brings the Google DeepMind runtime behind Gemini to every Google Cloud customer." It uses a single-controller design to orchestrate large TPU deployments across multiple slices "as if they were one local machine," with sub-minute slice hot-swapping and elastic resharding to keep long-running jobs going through hardware failures.

GKE dynamic slicing. Also generally available, this feature uses optical circuit switching to combine TPU cubes into larger topologies or split them into smaller isolated ones on demand. According to Google, it "slashes job startup times by up to 5x and accelerates fault recovery by up to 4.5x."

The post closes with a claim about how the chips themselves are made. Google uses Google DeepMind's AlphaChip reinforcement learning system and Gemini "to deliver chip layouts in hours rather than weeks," which Vahdat calls "true co-design and a unique Google differentiator." Google scheduled a broadcast on the new systems for 9 a.m. PT today.

What the post leaves out

For an announcement of this scale, the gaps are notable:

  • No prices for TPU 8tx or TPU 8i Max, and no indication of how they will be sold: cloud rental, on-premises sale or both.
  • No availability date, preview status or general availability timeline for either system, and no regions.
  • No absolute specifications: compute throughput, HBM capacity, bandwidth, power or pod size.
  • No named customers for either system, and no breakdown of the "nearly 100 customers" between cloud users and on-premises buyers.
  • Every performance comparison is against Google's own hardware. There is no comparison with Nvidia or AMD GPUs.
  • Benchmark claims without numbers, and percentage improvements ("up to 2x," "more than 90%," "up to 5x") without baselines.

A business that now shows up in the accounts

The customer figure arrives as TPUs become visible in Alphabet's financial statements. In the third quarter, Google Cloud revenue rose 103% to $30,709 million, and Alphabet said TPU systems sales were a meaningful contributor to that growth, without disclosing a figure.

The quarterly filing added detail. Alphabet said it had signed a limited number of agreements to supply TPU systems for customers' on-premises infrastructure, began recognizing revenue from them in the second quarter of 2026, and expects the significant majority in 2027. Inventory rose from $2.4 billion at the end of 2025 to $13.3 billion at September 30, consisting mainly of TPU hardware for enterprise customers and consumer devices, and Alphabet's financial guarantees include $6.6 billion tied to repurchasing TPU systems.

The build-out is expensive. Third-quarter capital expenditure reached $51.4 billion and free cash flow turned negative at $9.1 billion. Demand, by Google's own account, still exceeds supply: after Gemini 4 Argon processed almost 25 trillion tokens in its first 72 hours, Sundar Pichai said in an October 21 interview that Google remained capacity constrained.

One buyer dominates the public record. Anthropic's IPO documents show TPU8i pods that Google sells to Broadcom and Broadcom leases to an Anthropic subsidiary, and its November 2 amendment estimated additional TPU compute from Google at $125 billion to $150 billion, for delivery between 2028 and 2031. That concentration is a recurring concern. UBS estimates cited by Ed Zitron in July projected that OpenAI and Anthropic would account for more than 48% of Google Cloud revenue in 2027, a projection rather than a reported figure. A list of "nearly 100 customers" that includes retailers and universities suggests a broader base, though the post gives no revenue split.

Google is not the only hyperscaler building its own silicon. Amazon activated Project Rainier with nearly 500,000 Trainium2 chips in October 2025, with Anthropic as its user. Anthropic's IPO filing describes a compute pool spanning AWS Trainium, Google TPUs, Nvidia GPUs and AMD GPUs, so the rival chips compete for the same workloads even inside a single customer.

Why this matters for the marketing industry

Advertisers do not buy TPUs. They buy the products that run on them. According to the post, TPUs train and serve Google's Gemini models, and Gemini now sits inside AI Mode in Search and the Gemini app. When Google launched Gemini 4 Argon in October, it set an introductory price of $2 per million input tokens and limited consumer access to paid subscribers. How cheaply Google can run inference shapes both how widely it can deploy AI features in Search and what it charges others to use its models.

The agent framing is also directly relevant. Google's post describes computing "moving from human to agent scale" and builds products around agents that pause, wait and resume. The Agent Compute Engine addresses the cost of agents that sit idle while waiting for a tool or a human, the same pattern found in agents that shop or manage campaigns on someone's behalf. Lower infrastructure costs for idle agents would make long-running agents cheaper to operate, though the post offers no figures to measure by how much.

Then there is data. Brands, agencies and retailers hold customer lists, conversion data and purchase histories that could feed fine-tuned models or agents. Ad tech already relies on confidential computing to match that data without exposing it. TPU 8tx extends the same promise to model training and serving on Google's own chips, but whether it reaches advertisers at a price and on a timeline that matters is a question the announcement does not answer.

Timeline

Summary

Who: Google, through a post by Amin Vahdat, Senior Vice President and Chief Technologist for AI and Infrastructure. The systems target regulated enterprises, governments, AI labs and Google Cloud customers, and the nearly 100 organizations that already use TPUs, including frontier labs, trading firms, retailers and universities.

What: Google added two systems to its eighth TPU generation. TPU 8tx, built on TPU 8t silicon with AMD EPYC hosts, keeps data and model weights protected while in use and claims up to 2x lower latency than TPU 8t. TPU 8i Max combines two TPU 8i chips on one package, doubling compute, SRAM and HBM bandwidth per package. Google also open-sourced TPU Sync, made Cloud Pathways and GKE dynamic slicing generally available, and introduced an Agent Compute Engine for idle agent sessions. No prices, availability dates or regions were given.

When: Today, November 9, 2026, following the April introduction of TPU 8t and TPU 8i at Google Cloud Next. It is the first time Google has introduced four TPU systems in a single year.

Where: The systems are offered through Google Cloud and, according to the post, TPUs can now also be deployed directly in customers' own data centers. No regions were specified.

Why: Google says outside customers asked for accelerators that can handle very large context windows and for specialized encryption for regulated workloads. For the marketing industry, TPU cost and capacity sit behind Gemini, AI Mode and Google's agent products, and confidential computing already underpins how Google processes advertisers' first-party data.