Back to Explore
NewsVibeFix editorialUpdated Nov 13, 2026

Four TPU Systems in One Year: Google's 8tx and 8i Max Bet on Assembly Speed, Not New Silicon

On November 9 Google announced TPU 8tx and 8i Max: one for hardware-level confidential computing, one packing two chips together to break the memory wall. But with no price and no ship date, this reads as a roadmap announcement.

Google's official render: the four eighth-generation TPU systems — 8t, 8i, 8tx, and 8i Max — shown side by side.

On November 9, Google announced TPU 8tx and TPU 8i Max. They are the third and fourth TPU systems the company has introduced in a single year, following the TPU 8t and 8i unveiled at Cloud Next in April. Amin Vahdat, Google's Senior Vice President and Chief Technologist for AI and Infrastructure, wrote a sentence worth pausing on: this is the first time Google has introduced four TPU systems in a single year.

Here is the judgement that matters. The point of this launch is not a new chip, but a way of working — the "system" approach. Neither 8tx nor 8i Max uses new silicon: the first reuses 8t silicon, and the second packages two existing 8i chips together. Google's logic is that a chip takes years to build, while recombining existing chips, networking, memory, and security into a new system takes months. When AI demand changes direction every few months, being able to assemble a new system quickly matters more than being able to fabricate a new chip. What this announcement is really about is how Google is turning the TPU from a chip product into compute building blocks that can be assembled on demand.

Google's official render: the four eighth-generation TPU systems — 8t, 8i, 8tx, and 8i Max — shown side by side.

1. Two new systems, each solving one new problem

TPU 8tx targets confidential computing. Normally, data is encrypted at rest and in transit, but has to be decrypted in memory while a computer processes it — which is exactly why regulated industries hesitate to hand their most sensitive data to the cloud. 8tx builds security into the hardware layer: built on the same silicon as 8t and paired with 5th Generation AMD EPYC processors as hosts, it keeps sensitive data and model weights protected and isolated even while in active use. AMD's own announcement puts it more bluntly: this is the first cloud TPU system with hardware-enforced confidential computing.

8tx is not just an encrypted edition. It also adopts the low-latency Boardfly network built for real-time reasoning, delivering up to 2x lower latency than TPU 8t with what Google calls a near-zero performance penalty. Training-grade compute plus an inference-grade network means an organization can run and customize models on a single system: fine-tune on private data, distill large models into smaller ones, and train agents through reinforcement learning, all without migrating sensitive data between separate clusters. For banks and governments, "the data never has to move" is a selling point in itself — every move is a chance to leak.

TPU 8i Max solves a different problem: the memory wall. When dozens of AI agents work together across an entire software codebase or hours of video at once, memory bandwidth quickly becomes the bottleneck. Google's fix is direct: take two low-latency TPU 8i chips and combine them onto a single package, doubling the compute, the on-chip SRAM, and the HBM bandwidth, while fitting twice the silicon within the existing rack footprint. A data center can therefore deploy significantly more computing power per square foot and per watt. Vahdat sums up the four-way division of labor in one line: 8t for scale-up training, 8i for low-latency reasoning, 8tx for confidential computing, and 8i Max for massive context windows and the highest compute density per rack.

2. The numbers look strong, but the two most important columns are missing

The figures in the announcement are not small. TPU 8t and 8i offer up to 2.7 times better training performance and 80% better inference price-performance than the previous generation. The TPU business now has nearly 100 customers, including frontier labs, trading firms, retailers, and universities, and some have begun deploying TPUs directly in their own data centers rather than only renting cloud capacity. Google is also using DeepMind's AlphaChip reinforcement learning system and Gemini to design next-generation chip layouts, compressing work that took weeks into hours.

But the announcement is missing the two most important pieces of information: price and availability. It says nothing about what 8tx or 8i Max will cost, gives no general-availability date, and lists no launch regions. Some press coverage quotes Vahdat telling Bloomberg that the two systems will arrive in 2027, but that is not stated in the announcement itself — until Google confirms it, it counts only as second-hand reporting. For anyone planning procurement around this launch, the correct reading is "roadmap announcement," not "product you can order."

The performance claims deserve the same discount. 8tx's "up to 2x lower latency with near-zero performance penalty," 8i Max's "doubled compute and bandwidth," and the vLLM-TorchTPU stack's "state-of-the-art price-performance" on the InferenceX and AgentX benchmarks are all Google's own measurements, with no independent testing attached. The benchmark links point to a SemiAnalysis newsletter, not a publicly reproducible test report. A vendor's numbers are a reference, not a basis for a purchasing decision — something the announcement does not deny, but also does not volunteer.

3. Half of this launch is actually software

Easy to miss is that the second half of the announcement is almost entirely software. Last month, Google open-sourced TorchTPU, letting PyTorch code and Hugging Face models run natively on TPUs with no rewrites, alongside out-of-the-box backends for leading open-source inference engines like vLLM and SGLang. That was already announced separately on October 20, and covered at the time; its reappearance inside the TPU launch narrative shows that Google now considers "it runs on TPUs" as important as "how fast the TPU is."

There is new software too. The open-source TPU agent suite, now in preview, plugs into coding tools like Claude Code, Cursor, and Google Antigravity, automating TPU onboarding in minutes, diagnosing bottlenecks, and autotuning custom kernels using plain language. The 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. The Agent Compute Engine automatically suspends and resumes idle agent sessions — agents frequently pause for seconds or hours waiting on a database lookup, an external tool, or human feedback, and that idle compute used to burn money. The open-source TPU Sync turns chip HBM, host DRAM, shared storage, and the data center network into a single co-designed memory hierarchy, which Google says cuts KV transfer latency and time to first token by more than 90%.

On the cloud side, two items reached general availability: Cloud Pathways on Google Kubernetes Engine, which brings the Google DeepMind runtime behind Gemini to every Google Cloud customer; and GKE dynamic slicing, which uses optical circuit switching to assemble TPU topologies on demand, claimed to slash job startup times by up to 5x and accelerate fault recovery by up to 4.5x. Taken together, Google's intent is clear: the TPU moat is no longer the chip alone, but the full stack from silicon to runtime, from training to inference. A rival can buy the same foundry capacity, but it cannot replicate this whole software stack within a year.

4. Why now: agents rewrote what compute is for

To understand this launch, start with what changed on the demand side. Vahdat writes it plainly: computing is moving from human to agent scale. The old pattern was a person sending one request and a model answering once. Now dozens of agents work across an entire codebase at the same time, reason across million-token contexts, call tools back and forth, and wait for human feedback. Memory bandwidth, long context, and the cost of idle sessions were secondary problems; now they are the main bottlenecks. 8i Max, the Agent Compute Engine, and TPU Sync each answer one of them.

The other side is compliance. More enterprises want to put agents into production, but their data cannot leave the internal network. 8tx's confidential computing targets exactly that gap: models and agents run right next to the sensitive data while the data stays protected. It is a real market — banks, healthcare, government — but also the hardest market to win over. A hardware-level confidential computing claim ultimately has to be backed by certifications, audit reports, and verifiable attestation, and the announcement mentions none of them. For now, 8tx's security promise is a vendor statement, and strictly regulated customers will demand far more than an announcement before they trust it.

5. For independent developers, did the barrier fall — or just move?

The direct significance of this launch for independent developers and small teams is not the hardware, but the software. TorchTPU lets PyTorch code run on TPUs without rewrites, the TPU agent suite promises onboarding and tuning in plain language, and Cloud Pathways opens up DeepMind's runtime. All of these lower the barrier to using TPUs. For a small team wanting to try inference or fine-tuning on a TPU, the biggest obstacles used to be rewriting code and tuning kernels; the official tooling now takes on part of both jobs.

But the barrier has not disappeared — it has moved. TPU 8tx and 8i Max themselves remain a cloud vendor's business: no price, no availability date, no regions, and nothing an individual developer can buy or rent in the near term. What is usable right now is the open-source software — TorchTPU, the TPU agent suite, the XLA Serving Engine — plus the now-generally-available Cloud Pathways and GKE dynamic slicing. In other words, what Google is handing independent developers this time is not a chip but a toolchain: get comfortable running on today's TPUs first, so that when 8tx and 8i Max arrive, you are already inside its stack.

That is also the most interesting part of this launch. Shipping four TPU systems in one year is itself a statement: Google does not plan to chase Nvidia generation by generation on chips, but to compete on system-level assembly speed and the depth of its software stack. Whether that strategy works depends on two things: first, whether 8tx and 8i Max ship on time and actually deliver the announced performance; second, whether open-source tools like TorchTPU genuinely keep developers around. Neither question has an answer yet. What the announcement offers is a roadmap pointing in the right direction, not a report card that can be graded.

Views 4Comments 0

Comments (0)

ME
0/1000
Loading comments...

Sources

Browse projectsPublish your project

Related articles

A developer with glasses types on a laptop at a wooden desk, an external monitor to the left showing C code on a dark background, a small plant on the desk
News
Microsoft's Coding Model Stops Playing Second Fiddle: MAI Code 2 Scores 96% of Opus 5.5 for 5% of the Cost — Read the Three Caveats First

On 17 November 2026, Microsoft launched MAI Code 2 and Code 2 Flash: the larger model has 77 billion active parameters and a one-million-token context, averaging 63.2% across four benchmarks — close to GPT-6.1 Sol and Claude Opus 5.5 — at $0.309 per task. Read the three caveats first: promotional pricing, the lowest score of the three, and no long-horizon work.

Developer WorkflowProduct LaunchModel Updates
Against a black background, a conductor seen from behind raises a baton, with swirling trails of light circling beneath him like an orchestra being conducted
News
One API for 45 Models, and It Switches Providers When One Fails: Perplexity's Orchestrator Wants to Be the Switchboard

On 17 November 2026, Perplexity launched the Orchestrator API: one interface reaching OpenAI, Anthropic, Google, xAI and open-weight models such as GLM, Kimi and DeepSeek, billed at the answering model's published rate with no platform fee. When one provider fails, the same model retries on another deployment, and preview routing assigns requests to models by task.

Developer WorkflowProduct LaunchModel Updates
In a dim office, a developer wearing glasses sits with his back to the camera before several monitors showing code, lit by blue and purple light
News
The Fourth Gemini 4 in Seven Weeks: Neon Sits Between Carbon and Hydrogen, Selling Speed With Enough Smarts

On 18 November 2026, Google released Gemini 4 Neon: third overall on the Artificial Analysis Intelligence Index (61) and first on the Coding Agent Index (58.6), priced at $3 per million input tokens and $15 per million output. It sits between the heaviest Carbon and the lightest Hydrogen, conducting teams of Hydrogen subagents in Antigravity — yet per token it is November's priciest Gemini 4, only becoming the cheaper middle option after 1 December.

Product LaunchModel UpdatesAI Coding