Google has quietly changed the shape of the AI hardware race. On 9 November 2026, the company announced two new systems, Google TPU 8tx and TPU 8i Max, taking its eighth-generation TPU portfolio to four systems in a single year — something Google says it has never done before. One system locks your data inside hardware while the model works on it. The other doubles the memory so AI agents can hold entire codebases and hours of video in their heads at once.
Neither system is on sale yet, and Google has published no prices. Even so, the announcement matters, because it shows where Google thinks AI computing is heading: away from one giant chip for everything, and towards specialised systems built around the two problems every serious AI deployment now hits — security and memory. This guide explains what Google TPU 8tx and TPU 8i Max actually are, how they work, what they cost (or rather, what we do not know yet), and who should care.

What Google TPU 8tx and TPU 8i Max Actually Are
Google TPU 8tx is Google’s third eighth-generation TPU system. It is built for secure and confidential inference, and Google also calls it a good fit for small-scale training. Its audience is specific: regulated enterprises and governments that run frontier or proprietary models and autonomous agents alongside their most sensitive data — banks, hospitals, insurers, defence contractors and public agencies, in other words.
TPU 8i Max is the fourth system, and it solves a different problem entirely: the memory wall. When dozens of AI agents work together across an entire software codebase or hours of video at once, memory bandwidth becomes the bottleneck long before raw computing power does. Google’s answer was to take two low-latency TPU 8i chips and combine them onto a single package, doubling the compute, the SRAM and the high-bandwidth memory (HBM) throughput of a single 8i chip.
Here is the detail that surprises most people: neither system uses a brand-new chip. Google is explicit about this. A new chip takes years to design and manufacture; a new system rearranges the compute, memory, networking and security around an existing chip, and that can be done in months. Google TPU 8tx reuses the same silicon as the TPU 8t training chip, paired with new host processors and networking. TPU 8i Max reuses the TPU 8i reasoning chip, two of them in one package. That distinction — chip versus system — is the key to understanding the whole announcement.
The two systems complete a portfolio Google started at Google Cloud Next in April 2026, when it introduced TPU 8t for large-scale training and TPU 8i for real-time reasoning. At the time, Google said the 8t offered up to 2.7 times better training performance and the 8i up to 80 percent better inference price-performance than the previous TPU generation. Those April figures describe the base systems, not the two new ones — a distinction worth keeping in mind, because no equivalent benchmark numbers exist yet for Google TPU 8tx or TPU 8i Max.
Table of Contents
Google TPU 8tx: Confidential AI Inference, Explained
To understand why Google TPU 8tx exists, start with a gap in how data is normally protected. Most organisations already encrypt their data when it is stored on disk and when it travels across a network. The problem is the moment in between: while a computer is actually processing the data, it has to be unlocked in memory. Anyone — or any software — with enough privilege on that machine can, in principle, read it. For a hospital’s patient records or a bank’s trading models, that gap is the whole ball game.
Confidential computing closes that gap. It runs workloads inside a hardware-isolated trusted execution environment, a protected region of the machine whose memory the host system cannot read, even while the model is actively using the data. Google TPU 8tx brings that idea to cloud TPUs for the first time, according to AMD, which supplies the host processors for the system.
The gap in normal encryption
Think of it this way. Encryption at rest is a locked filing cabinet. Encryption in transit is an armoured van. But the moment someone opens the cabinet to read a file, the contents are visible. Traditional cloud computing has always had that moment of exposure. Confidential computing is the equivalent of reading the file inside a sealed room that nobody else can enter — not even the building’s owner.
That matters more now than it did five years ago, for two reasons. First, AI agents work with live, sensitive data constantly: customer records, internal documents, proprietary code, medical images. Second, high-profile leaks have made the risk concrete. In October 2026, a Cloudflare incident exposed cached information from earlier requests in responses delivered to other customers — exactly the kind of cross-tenant data leak that hardware isolation is designed to prevent. Google TPU 8tx is aimed squarely at organisations that cannot afford that kind of accident.
How Google TPU 8tx works: 8t silicon, AMD EPYC hosts
Google TPU 8tx is built on the same silicon as the TPU 8t training chip, but the system around it is different in two important ways. First, it is paired with 5th Generation AMD EPYC server CPUs as host processors. AMD announced the same day that its EPYC chips power the host infrastructure for Google TPU 8tx, using AMD Secure Encrypted Virtualization (SEV) capabilities to extend confidential-computing protection from the CPU to the connected accelerators, keeping data encrypted in processor memory and in transit between them.
Second, Google TPU 8tx adopts the low-latency Boardfly network that Google created for the TPU 8i reasoning chip. That matters because confidential computing has historically carried a performance tax: encrypting and isolating everything slows the work down. Google claims TPU 8tx delivers up to two times lower latency than TPU 8t, with a near-zero performance penalty from its security features. That is Google’s own claim, not an independently measured result — no third party has tested the system yet, and Google has published no absolute latency figures or test conditions.
The choice of AMD hosts is also notable. Google has its own Arm-based Axion CPUs, but for TPU 8tx it went with AMD’s x86 EPYC processors, citing their security capabilities and software compatibility. Some industry watchers have suggested the AMD partnership may be a short-term choice, with Google’s own security functions eventually moving into future TPU and Axion designs — but for now, the EPYC host is part of the product.

One system for running and customising models
Pairing training-grade compute with a fast inference network gives Google TPU 8tx a practical advantage that goes beyond security. Organisations can run a model and customise it on the same system, without moving sensitive data between separate clusters. Google says TPU 8tx enables teams to fine-tune models using their private data, distil large models into smaller ones, and train their agents through reinforcement learning — all in one place.
That combination matters because the old workflow was fragmented: train on one cluster, serve on another, and shuttle data between them, with each move creating a new point of exposure. For a bank fine-tuning a model on customer transaction data, or a hospital adapting a model to its own patient records, keeping everything inside one confidential system removes several steps where data could leak. It also removes the operational headache of maintaining two separate infrastructures.
TPU 8i Max: Built for the Memory Wall
If Google TPU 8tx is about keeping data safe, TPU 8i Max is about keeping processors fed. The problem it attacks is the memory wall: the point where a chip can compute faster than memory can supply it with data, so the expensive processor sits idle waiting. For ordinary chatbots, that is a minor issue. For agentic workloads — dozens of AI agents working across an entire codebase or hours of video simultaneously — it is the bottleneck that decides whether the system is fast or useless.
Double the compute, double the memory
Google’s fix is straightforward in concept and demanding in engineering. It took two TPU 8i chips and combined them onto a single package. The result, Google says, delivers double the compute, double the SRAM and double the HBM bandwidth of a single TPU 8i, so large workloads do not stall while processors wait for data. Reporting based on Google’s briefing puts the TPU 8i Max at 768 MB of SRAM and 17.2 TB/s of memory bandwidth — figures Google’s own announcement describes only in relative terms, as double the 8i.
Just as important is the packaging. Because TPU 8i Max fits twice the silicon within the existing rack footprint, data centres can deploy significantly more computing power per square foot and per watt. In an industry where data-centre space and electricity are the two scarcest resources, that density is a real selling point. A customer gets twice the chips without twice the floor space or twice the power infrastructure.
Why AI agents need giant context windows
The memory wall matters because of how AI agents actually work. A single prompt to a chatbot is a short, self-contained job. An agent is different: it reads a task, searches a codebase, remembers what it found, calls tools, waits for results, reasons about them, and keeps going — sometimes for hours. Every step adds to the context the model must hold in memory. Multiply that by dozens of agents working in parallel, and the memory demands explode.
This is why Google frames TPU 8i Max around “massive context windows and the highest compute density per rack.” The system is built for the moment when an agent needs to reason across a million-token context — a whole software repository, a full day of video, a thick legal file — without the hardware choking. It is the same trend driving model design: Google’s own Gemini models, including Gemini 4 Hydrogen, are built for long-horizon work over huge contexts, and the hardware has to keep up.
The same logic applies to video. An agent that watches hours of footage to find a single event needs to hold that footage in fast memory, not fetch it from slow storage frame by frame. TPU 8i Max is Google’s answer to that workload, and it is no coincidence that it arrived in the same announcement as the confidential-computing system: both are responses to what agents actually do in production.

The Full Eighth-Generation TPU Lineup
With the two new systems, Google’s eighth-generation TPU portfolio now covers every stage of the AI lifecycle. Each system has a distinct job, and understanding the four-way split is the fastest way to see what Google is doing.
| System | Announced | Built for | Key idea |
|---|---|---|---|
| TPU 8t | April 2026 | Large-scale training | Maximum “goodput” — useful intelligence per watt |
| TPU 8i | April 2026 | Real-time reasoning | Low-latency inference; cuts inter-chip delays by up to half |
| Google TPU 8tx | November 2026 | Confidential inference and small-scale training | Hardware-level confidential computing on AMD EPYC hosts |
| TPU 8i Max | November 2026 | Massive context windows | Two 8i chips in one package; double compute and memory |
Google’s own summary is simpler: TPU 8t for scale-up training, TPU 8i for low-latency reasoning, Google TPU 8tx for confidential computing, and TPU 8i Max for massive context windows and the highest compute density per rack. The strategy is specialisation. Instead of asking one chip to be good at everything, Google is building a family where each member is tuned for one job — and, crucially, where new members can be assembled from existing chips in months rather than years.
That speed matters because the AI workload keeps changing faster than silicon manufacturing can. When Google designed its first TPU more than a decade ago, the target workloads were voice recognition and language translation. Today the same lineage of chips trains and serves Gemini models, runs agentic workloads, delivers services to Google Cloud customers, and — in a recent expansion — can be deployed directly into customers’ own data centres. Google puts its TPU customer count at nearly 100, spanning frontier labs, trading firms, retailers and universities.
Why Four TPU Systems in One Year?
Google’s announcement, written by Amin Vahdat, the company’s senior vice president and chief technologist for AI and infrastructure, frames the four-systems-in-a-year milestone as a deliberate strategy, not a coincidence. The company says computing is moving “from human to agent scale,” and that shift creates demands that did not exist when the 8t and 8i were designed in April.
Two demands in particular drove the new systems. First, businesses are asking for accelerators that can reason across AI models’ massive context windows — the TPU 8i Max problem. Second, highly regulated customers are asking for specialised encryption capabilities for unique workloads — the Google TPU 8tx problem. Google says it is increasingly getting feedback from external organisations on their specific needs, not just co-designing with Google DeepMind as it did in the past.
There is also a competitive logic. Nvidia dominates AI accelerators, and its GPUs are the default choice for most AI work. Google cannot beat Nvidia by building a single faster chip — the manufacturing race is too long and too expensive. Instead, it is competing on systems: specialised configurations, open software, and tight integration with Google Cloud. The four-system portfolio is a bet that customers will choose the right tool for each job rather than one general-purpose accelerator for all of them.
The economics support that bet. Training a frontier model is a different job from serving a chatbot, which is different again from running a confidential agent on hospital data. Each job has different bottlenecks — compute, latency, memory, security — and a system tuned for the wrong bottleneck wastes money. Google’s pitch is that the eighth-generation family lets a customer match the system to the workload, instead of paying for capacity they cannot use.
The Software That Makes Google TPU 8tx Usable
Hardware announcements are easy; making hardware usable is the hard part. Google knows this, and the TPU 8tx announcement came with a package of software moves designed to remove the barriers that have historically kept developers away from TPUs. When the first TPU launched more than a decade ago, only a handful of companies ran workloads that needed it, and engineers used optimised frameworks and hand-tuned code. That is no longer acceptable: with AI adoption as broad as it is today, Google says it has been working systematically to make TPUs as easy to use as general-purpose hardware.
The biggest single change is PyTorch support. Last month, Google open-sourced TorchTPU, so PyTorch code and Hugging Face models now run natively on TPUs, alongside out-of-the-box TPU backends for leading open-source inference engines like vLLM and SGLang. For most developers, that removes the main reason to avoid TPUs: they no longer need to rewrite their code.

Google also announced a new open-source TPU agent suite, now in preview, that plugs directly into coding tools such as Claude Code, Cursor and Google Antigravity to automate TPU onboarding in minutes, diagnose bottlenecks, and autotune custom kernels using plain language. For production deployments, the new open-source XLA Serving Engine lets developers train models in Python and serve them directly in a low-latency C++ runtime across TPUs, GPUs and CPUs.
Three further software pieces target the agent workloads that Google TPU 8tx and TPU 8i Max are built for:
- Agent Compute Engine: AI agents are stateful — they frequently pause for seconds or hours while waiting on a database lookup, an external tool, or human feedback. The new Agent Compute Engine software automatically suspends and resumes idle sessions, cutting the infrastructure cost of agents that spend most of their time waiting.
- TPU Sync: Google is open-sourcing TPU Sync, the model-state layer for TPU clusters. It turns chip HBM, host DRAM, shared storage and the data-centre network into a single memory hierarchy, automatically tiering massive KV caches off the chip in milliseconds and synchronising reinforcement-learning weights. Google claims it cuts KV transfer latency and time to first token by more than 90 percent.
- Cloud Pathways and GKE dynamic slicing: Cloud Pathways, now generally available on Google Kubernetes Engine, brings the Google DeepMind runtime behind Gemini to every Google Cloud customer, orchestrating large-scale TPU systems across multiple slices as if they were one local machine. GKE dynamic slicing, also generally available, assembles TPU cubes into larger topologies or partitions them into smaller workloads on the fly — Google claims job startup times up to five times faster and fault recovery up to 4.5 times faster.
Google also points to benchmark results on industry-standard InferenceX and AgentX tests, which measure realistic agentic coding and long-context work. Its native vLLM-TorchTPU software stack on TPUs achieved what Google describes as state-of-the-art price-performance marks. As with every performance claim in this announcement, those are Google’s figures from its own testing, not independent results — but they signal where Google thinks the competitive fight will be won: not on peak chip speed, but on outcomes on real AI workloads. Developers building agents with tools like Google’s Agent Development Kit are the audience this software stack is aimed at.
Google TPU 8tx vs Nvidia: The Confidential-Computing Race
Google TPU 8tx does not enter an empty market. Nvidia, the dominant AI accelerator company, has been building confidential computing into its own GPUs, and its Confidential G4 instances (built on RTX PRO 6000 GPUs) reached general availability on Google Cloud itself on 21 October 2026, in 36 zones, using AMD SEV with Nvidia Confidential Computing. In other words, Google Cloud already offers confidential AI inference on Nvidia hardware — Google TPU 8tx adds a second, Google-built option alongside it.
The comparison is instructive. Nvidia’s confidential-computing technology has a published design, a list of adopting cloud providers, and independent security review. Google TPU 8tx, by contrast, arrives with a thinner public record. Google’s announcement names no attestation method — the mechanism by which a customer can independently verify that the system is actually running the code it claims to be running — and publishes no threat model or independent audit. AMD’s announcement describes SEV capabilities in the EPYC host CPU, but does not say which generation of SEV the system uses, and Google’s own post names no trusted-execution technology at all.
That gap matters for the exact customers Google TPU 8tx targets. A bank’s compliance team cannot approve a system on the strength of a blog post; it needs to know what is encrypted, what is attested, and how it can check. Reporting after the announcement noted that Google’s own Confidential VM documentation says TPUs cannot be attached to Confidential VM instances, and that Google has not published whether HBM memory or chip-to-chip links are encrypted on TPU 8tx. Until those details appear, the security promise is a vendor claim rather than something a buyer can verify.
There is also a caution from the security research world. In January 2026, researchers disclosed StackWarp (CVE-2025-29943), a flaw that lets an attacker with privileged control of the host break AMD’s SEV-SNP protections; AMD has released firmware updates, and its affected-product list includes the 5th Generation EPYC family that Google TPU 8tx uses. Google has not said whether TPU 8tx relies on SEV-SNP specifically, so that finding describes the host CPU family, not a confirmed design — but it is a reminder that confidential computing is a fast-moving field, and yesterday’s protection is not automatically tomorrow’s.
None of this makes Google TPU 8tx a bad product. It makes it an early one. The honest reading is that Google has announced the architecture and the intent, and the evidence — independent benchmarks, attestation design, compliance certifications — will have to follow. For now, the competitive story is that Google has a confidential TPU at all, which gives regulated customers a choice beyond Nvidia for the first time.
What Google Has Not Told Us Yet
A careful reader of the announcement will notice what is missing. There is no pricing for Google TPU 8tx or TPU 8i Max, no list of cloud regions, and no firm availability date. Google’s Amin Vahdat told Bloomberg the systems will arrive in 2027, but the announcement itself gives no date, and neither product is shipping yet. That means no enterprise can put either system into a 2026 budget, and no developer can try one today.
There are also no independent benchmarks. Every performance number in the announcement — the two-times-lower latency for Google TPU 8tx, the doubled memory figures for TPU 8i Max, the 90-percent-plus KV latency reduction from TPU Sync, the five-times-faster job startups from GKE dynamic slicing — comes from Google’s own testing, with no published test conditions. The April figures for the base 8t and 8i systems (2.7 times better training performance, 80 percent better inference price-performance) are the closest thing to a baseline, and they describe different systems.
For TPU 8i Max, one more detail is missing: memory capacity grows less than compute and bandwidth. Doubling the chips doubles the processing power and the speed at which data moves, but the total memory capacity does not double by the same factor. That is a normal engineering trade-off, but it matters for buyers doing capacity planning: TPU 8i Max is a bandwidth and density story, not a “twice the memory” story.
The practical advice is simple. If you are planning 2027 infrastructure, treat Google TPU 8tx and TPU 8i Max as real products with unproven numbers. Get pricing and availability terms in writing before committing, ask Google for the attestation and compliance documentation if you are a regulated buyer, and run your own benchmarks on the base 8t and 8i systems in the meantime — they are the closest available proxy for how the new systems will behave.

Who Should Care About Google TPU 8tx
Regulated enterprises are the obvious audience. If your organisation handles data that cannot leave a protected boundary — patient records, financial transactions, government files, trade secrets — Google TPU 8tx is the first Google-built accelerator designed for your problem. The combination of confidential inference with on-system fine-tuning and reinforcement learning means you can adapt models to your data without the data ever leaving the protected system. Just do not sign a contract until the attestation and compliance documentation exists.
AI agent builders should watch TPU 8i Max more closely. If you are building agents that work across whole codebases, long documents or hours of video, the memory wall is probably already your biggest cost and your biggest source of slowness. The long-context models now arriving — including Qwen4 Max Preview, built for million-token work — need hardware that can hold those contexts in fast memory. TPU 8i Max is designed for exactly that job, and the TorchTPU and TPU agent suite software makes it accessible without rewriting your stack.
Cloud architects and platform teams should pay attention to the systems strategy itself. Four TPU systems in one year, assembled from existing chips, suggests Google will keep shipping specialised configurations at this pace. If your infrastructure planning assumes one accelerator choice per year, that assumption is now out of date. The GKE dynamic slicing and Cloud Pathways releases matter here too: they are the tools that let you treat a fleet of different TPU systems as one programmable pool.
Everyone else can safely wait. If you are a solo developer, a small team, or a company whose AI work runs on standard cloud GPUs, neither system changes your day-to-day work today. The software around them — TorchTPU, vLLM and SGLang backends, the TPU agent suite — is the part that might reach you first, and it is free and open source. The systems themselves arrive in 2027, and by then the pricing, benchmarks and independent reviews should exist to judge them properly.
Frequently Asked Questions
What is Google TPU 8tx?
Google TPU 8tx is Google’s third eighth-generation TPU system, announced on 9 November 2026. It is built on the same silicon as the TPU 8t training chip, paired with 5th Generation AMD EPYC host processors, and designed for secure, confidential AI inference and small-scale training. Google positions it for regulated enterprises and governments running frontier or proprietary models and autonomous agents alongside sensitive data.
When will Google TPU 8tx be available?
Google has not announced a firm availability date. Amin Vahdat, Google’s senior vice president and chief technologist for AI and infrastructure, told Bloomberg the systems will be available in 2027. No pricing, cloud regions or preview programme have been published yet.
What is the difference between TPU 8tx and TPU 8i Max?
They solve different problems. Google TPU 8tx is about security: hardware-level confidential computing that keeps data and model weights protected while in use, aimed at regulated industries. TPU 8i Max is about memory: two TPU 8i chips in one package, doubling compute, SRAM and HBM bandwidth for massive context windows and the highest compute density per rack. One protects data; the other feeds processors faster.
Is Google TPU 8tx faster than Nvidia GPUs?
There is no independent evidence either way yet. Google claims up to two times lower latency than TPU 8t with a near-zero performance penalty from its security features, but it has published no absolute figures, test conditions or third-party benchmarks. Nvidia’s confidential-computing GPUs are already generally available on Google Cloud. Any honest comparison will have to wait for independent testing in 2027.
Do I need Google TPU 8tx to run AI agents?
No. Most agent workloads run fine on standard cloud GPUs or on the existing TPU 8i. Google TPU 8tx is for organisations whose data cannot be exposed during processing — regulated industries and governments. TPU 8i Max is for teams whose agents work across very large contexts. The software that arrived with the announcement, including TorchTPU and the TPU agent suite, is the part most developers can use right away.
Google’s eighth-generation TPU story is, in the end, a story about specialisation. The company spent a decade building one chip family and selling it as a general-purpose AI accelerator. Now it is splitting that family into systems tuned for training, reasoning, confidential computing and giant contexts — and shipping new members in months by rearranging existing silicon. Whether Google TPU 8tx and TPU 8i Max live up to their promises will be decided in 2027, when pricing, availability and independent benchmarks finally appear.
Until then, the announcement tells us something useful on its own: the AI hardware race is no longer about who has the fastest chip. It is about who can build the right system for the job, and build it fast enough to matter.
Sources: Google — Introducing TPU 8tx and TPU 8i Max (9 November 2026); AMD Newsroom — AMD EPYC Expands Its Leadership at the Foundation of AI (9 November 2026). Performance figures are the vendors’ own claims and have not been independently verified.
