Google Releases TPU 8tx and 8i Max: Data Protection During Processing and Twice the Memory
Google has added two accelerators to its eighth-generation TPU family: one protects data during computation itself, and the other handles very long context. The TPU 8tx isolates data and model weights even while they are being processed, and inter-chip latency is up to two times lower than on the TPU 8t. The TPU 8i Max combines two 8i chips in a single package, doubling compute, SRAM, and HBM bandwidth. Hardware choice now depends on the task: sensitive data or long-context agents.
AI-processed from Habr AI; edited by Hamidun News
Google did not build one universal accelerator for every case. Instead, it expanded the eighth-generation TPU family with two more systems: the TPU 8tx and the TPU 8i Max. The first is for customers who run models on data that must not be shown to anyone, including the cloud provider itself. The second targets agents that need to keep a huge context in memory. Both systems grew out of existing chips, but each is built around a different pain point.
How to Protect Data While It Is Being Processed
The usual security model looks like this: data is encrypted on disk and in transit over the network. But for a model to work with that data, it must be decrypted in RAM. That is the moment when it is vulnerable. It is like keeping documents in a safe and transporting them in an armored car, only to read them at a checkpoint where strangers walk by.
The TPU 8tx closes that gap. It is built on the same chip as the TPU 8t, but Google changed the system configuration: it uses fifth-generation AMD EPYC server processors and adds hardware-level data protection during computation. According to the company, sensitive data and model weights remain isolated even while a calculation is running.
The target audience is clear: companies and government agencies that run their own models on confidential data. A single system can serve regular requests, fine-tune a model on internal documents, perform distillation (transferring knowledge from a large model into a smaller one), and train agents with reinforcement learning. In other words, there is no need to maintain a separate secure environment for each task.
Another detail: the TPU 8tx gets the Boardfly network that Google previously developed for the TPU 8i. According to the company, latency when working with multiple accelerators can be up to two times lower than on the TPU 8t. Google also claims the additional protection has almost no effect on performance. This is the company's own claim; the article contains no independent measurements.
Two Chips in One Package for Long Context
The TPU 8i Max has a different job. When an agent parses a large repository, a long contract, or several hours of video, it quickly runs short of memory and of the bandwidth needed to feed data to the compute units. Extra compute does not help here: the accelerator sits idle waiting for data.
Google tackled this head-on: it placed two TPU 8i chips in a single package. As a result, the system has twice the compute resources, SRAM (fast on-die memory), and HBM bandwidth compared with a standard TPU 8i. The two chips take up the same rack space that one system used to, so compute density has increased.
In practice, this means a single agent with a huge context can fit entirely on one node instead of being split across several machines. Less data exchange between accelerators usually means lower latency.
What the Lineup Looks Like Now and Why It Matters
The eighth-generation TPU family now includes four variants:
- TPU 8t for training large models;
- TPU 8i for fast inference;
- TPU 8tx for working with confidential data;
- TPU 8i Max for tasks with especially large context.
This segmentation shows where the market is heading. Training, routine inference, confidential data, and long context each place different demands on hardware, and general-purpose solutions are losing ground to specialized ones. Other accelerator makers are taking a similar path.
What this means for developers and founders: if you are building products for banks, healthcare, or the public sector, hardware-level data protection during computation can be a useful argument in a conversation with your security team. If you are building coding agents or tools for analyzing video and documents, it makes sense to look at configurations with generous memory headroom, not just operations per second.
What the Announcement Leaves Out
The material includes no prices, no availability dates, and no comparison with competitors' accelerators. It also lacks details on the isolation technology the TPU 8tx uses or on what counts as the trusted party in that scheme. The claims of latency "up to two times" lower and "almost no effect" on performance remain Google's own statements. Until independent tests and pricing appear, drawing conclusions about the benefit for a specific workload is premature, but the lineup itself already signals how to choose hardware for agents.
- Nov 22, 2026A Backend Developer With No Unity Experience Built a 3D Game With an AI Agent: Subscriptions Cost $112
- Nov 22, 2026OpenAI's Chief Scientist Proposes an AI Speed Limit: GPT-4 Perplexity at 0.91, Now Below 0.8
- Nov 22, 2026Anthropic Shares Fell 8.5% in a Day, Nearly Back to the $83 IPO Price
- Nov 22, 2026Linger Runs an AI Agent on a 16 GB Mac With No Token Fees and Promises a 5x Speedup
Key Questions Answered
How to Protect Data While It Is Being Processed?
The usual security model looks like this: data is encrypted on disk and in transit over the network. But for a model to work with that data, it must be decrypted in RAM. That is the moment when it is vulnerable. It is like keeping documents in a safe and transporting them in an armored car, only to read them at a checkpoint where strangers walk by. The TPU 8tx closes that gap.
What are the key details on Two Chips in One Package for Long Context?
The TPU 8i Max has a different job. When an agent parses a large repository, a long contract, or several hours of video, it quickly runs short of memory and of the bandwidth needed to feed data to the compute units. Extra compute does not help here: the accelerator sits idle waiting for data. Google tackled this head-on: it placed two TPU 8i chips in a single package. As a result, the system has twice the compute resources, SRAM (fast on-die memory), and HBM bandwidth compared with a standard TPU 8i.
What the Lineup Looks Like Now and Why It Matters?
The eighth-generation TPU family now includes four variants: TPU 8t for training large models; TPU 8i for fast inference; TPU 8tx for working with confidential data; TPU 8i Max for tasks with especially large context. This segmentation shows where the market is heading. Training, routine inference, confidential data, and long context each place different demands on hardware, and general-purpose solutions are losing ground to specialized ones. Other accelerator makers are taking a similar path.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.