Google Unveils Two New TPU Systems, Completing Four-System Eighth Generation in One Year

Google announced two additions to its eighth-generation Tensor Processing Unit lineup on Monday: TPU 8tx for confidential computing and TPU 8i Max for massive context windows. The systems join TPU 8t and TPU 8i, unveiled at Google Cloud Next in April, marking what the company calls the first time it has introduced four TPU systems in a single year.

The announcement, authored by Amin Vahdat, Google's SVP and Chief Technologist for AI and Infrastructure, positions the expanded portfolio as purpose-built for what Google terms "the agentic era" — workloads where autonomous AI agents reason across large contexts, execute multi-step tasks, and handle sensitive data.

Two Systems, Two Gaps #

TPU 8tx uses the same silicon as the training-focused TPU 8t, paired with 5th Generation AMD EPYC CPU hosts. Its differentiation is at the system level: hardware-level confidential computing that keeps data and model weights encrypted even while in active use. Normally, encrypted data must be decrypted in memory for processing. Google says TPU 8tx closes that gap, targeting regulated enterprises and governments that want to fine-tune, distill, or run reinforcement learning on proprietary data without moving it between clusters. The system also adopts the Boardfly low-latency network built for TPU 8i, which Google claims delivers up to 2x lower latency than TPU 8t with a near-zero performance penalty from the security features.

TPU 8i Max addresses a different bottleneck: memory bandwidth. It packages two TPU 8i inference chips together, doubling compute, on-chip SRAM, and HBM bandwidth within the same rack footprint. Google says this lets data centers deploy significantly more compute per square foot and per watt, targeting workloads where many agents operate across large codebases or hours of video — scenarios where memory bandwidth stalls before compute does.

Neither system is a new chip design. Both are reconfigurations of existing silicon through changes to packaging, networking, memory, and security — a modular approach Google says lets it deliver specialized systems in months rather than the multi-year cycle of new silicon.

Software Push Targets Nvidia Lock-In #

The hardware announcement came with a slate of software releases aimed at lowering the barrier to TPU adoption. Google open-sourced TorchTPU last month, enabling PyTorch code and Hugging Face models to run natively on TPUs with backends for vLLM and SGLang. A new TPU agent suite (in preview) plugs into coding tools including Claude Code, Cursor, and Google Antigravity to automate onboarding, bottleneck diagnosis, and kernel tuning. Agent Compute Engine suspends idle agent sessions and moves working memory off-chip in milliseconds. TPU Sync, now open-sourced, tiers KV caches across chip HBM, host DRAM, shared storage, and the data-center network — Google claims over 90% reduction in KV transfer latency and time-to-first-token. Cloud Pathways and GKE dynamic slicing are now generally available, with Google claiming up to 5x faster job startup and 4.5x faster fault recovery for the latter.

Availability, Pricing, and the Analyst Forecast #

Google's blog post gives no general-availability date, pricing, or regions for either system. Bloomberg, which interviewed Vahdat, reports both systems arrive in 2027. The same Bloomberg story cites a Bloomberg Intelligence forecast that TPU sales — including related hardware and software, but not Google Cloud usage — will rise from about $12 billion in 2026 to more than $135 billion by 2028. That is an analyst forecast, not a Google figure.

Bloomberg also names Apple, Meta, Oracle, and Anthropic as TPU customers, with Anthropic holding an option to purchase up to $150 billion worth of chips between 2028 and 2031, and OpenAI planning purchases. Google claims "nearly 100 customers" across frontier labs, trading firms, retailers, and universities, but named none. Jump Trading confirmed it uses thousands of TPUs for training research models, though it still mainly uses Nvidia for training.

Co-Design and the AlphaChip Loop #

Google says it increasingly uses AI to design the next generation of AI chips. The company cites its AlphaChip reinforcement learning system — originally published by DeepMind in 2020 and now used across three TPU generations — and Gemini for delivering chip layouts in hours rather than weeks. AlphaChip has also been adopted externally by MediaTek. Google describes this as a unique co-design loop: TPU-powered AI services accelerate the development of the next TPUs.

Competitive Context #

The announcement landed the same morning Nvidia shares fell in premarket trading over its board's investigation into CEO Jensen Huang. Google's accelerated cadence — four systems in one year versus a historical two-year cycle — reflects both internal demand from Gemini and external pressure from cloud rivals. Amazon ships Trainium and Inferentia; Microsoft has Maia; Nvidia remains the dominant merchant silicon vendor with H100/H200 and is itself moving toward more specialized systems, including CPUs and inference-focused chips using Groq technology.

Nvidia already offers confidential computing on its GPUs. Google's 8tx enters that space with a different architecture: training-grade silicon paired with AMD CPUs and hardware-enforced isolation. The 8i Max targets long-context inference, an area where a startup co-founded by former TPU engineers (Tenspresso) is also working, though at least a year from design completion per Bloomberg.

What Remains Unknown #

Google has not published absolute latency, throughput, or power figures for either new system. The April announcement said TPU 8t and 8i would be generally available later in 2026; it is unclear whether that timeline holds. No pricing, regional availability, or committed customer names for the new systems were disclosed. The $135 billion sales forecast is an external analyst projection that includes hardware, software, and related sales — not a Google target.

Sources #

1. [Our newest TPUs: 8tx and 8i Max, built for the agentic era](https://blog.google/innovation-and-ai/infrastructure-and-cloud/global-network/tpu-8i-max-and-8tx) (blog.google)
2. [Google TPU 8tx and 8i Max: Four TPUs in One Year | CellCog](https://cellcog.ai/blog/google-tpu-8tx-8i-max) (cellcog.ai)

3. Google adds TPU 8tx for confidential AI and a two-chip TPU 8i ... (mlq.ai)
4. Google adds TPU 8tx and TPU 8i Max to its eighth-generation ... (mezha.net)
5. Google expands eighth-generation TPU portfolio with two new ... (datacenterdynamics.com)
6. Google Unveils TPU 8tx and 8i Max for Agentic AI Workloads (techbuzz.ai)
7. Google Escalates Nvidia Rivalry With Specialized Versions of AI Chips (finance.yahoo.com)
8. How AlphaChip transformed computer chip design — Google DeepMind (deepmind.google)