Google Unveils Two New TPU Systems for the Agentic Era
Last week, Google revealed two new systems based on its eighth generation of Tensor Processing Units (TPUs): the TPU 8tx, designed for confidential inference, and the TPU 8i Max, with doubled memory capacity for massive context windows. For the first time, the company is introducing four TPU systems within a single year, dramatically expanding its specialized AI hardware strategy.
From internal chip to a business of thousands of clients
TPUs were originally designed more than a decade ago for Google's internal demands, such as voice recognition and language translation. Today, the situation is entirely different. According to Google, approximately 100 external customers now use TPUs on the cloud — including AI labs, trading firms, universities, and most recently, companies that have installed the chips directly in their own data centers.
The expansion to four systems in 2026 marks a shift in mindset. Instead of launching one chip per year and hoping it works for everything, Google is segmenting its portfolio: there is a system for each stage of the AI agent lifecycle. The TPU 8t (launched in April) focuses on training massive models; the TPU 8i (also from April), on real-time inference; the new 8tx, on hardware-level data security; and the 8i Max, on context windows that can ingest entire codebases or hours of video.
TPU 8tx: hardware security for sensitive data
The TPU 8tx combines the TPU 8t training chip with fifth-generation AMD EPYC processors, creating the first cloud TPU system with hardware-enforced confidential computing. The goal is to protect both the data and the model weights even while they are actively being processed — something critical for highly regulated sectors like healthcare and finance.
Additionally, the 8tx features a custom inference network that allows reasoning to run alongside training, enabling on-the-fly distillation and training of smaller models. This is relevant for companies that want to train models on proprietary data without moving information across devices — reducing both breach risk and architectural complexity.
Latency has been cut in half compared to the standard TPU 8t, and the ability to process confidential data at production scale positions the 8tx as a direct alternative to NVIDIA's GPU solutions with hardware enclaves, such as Confidential Computing.
TPU 8i Max: double the hardware in the same footprint
The TPU 8i Max takes a different approach. Instead of redesigning the chip, Google placed two TPU 8i processors in a single package. The result is double the compute power, double the on-chip SRAM, and double the HBM bandwidth — all within the same rack housing.
This may seem like an engineering trick, but the practical consequence is significant: data centers can pack considerably more compute per square foot and per watt, meaning lower infrastructure costs for ever-larger workloads. The 8i Max is aimed at AI agents that need to process massive amounts of information at once — a common example being an agent that reads an entire codebase or watches hours of video to answer a query.
The AI chip race shows no signs of slowing
The announcement of the new TPUs comes amid competitive pressure. Major companies like Microsoft, Amazon, and even labs like OpenAI and Anthropic are developing their own AI chips to reduce reliance on NVIDIA. Bloomberg reported that Google's new chips cost roughly 30% less than NVIDIA equivalents while delivering similar performance and lower power consumption for certain AI tasks.
Google no longer keeps its TPUs exclusively for itself. Google Cloud's infrastructure now offers TPUs as a service, and the company expects hardware sales to continue growing in Q4, with "meaningful revenue ramp-up in 2027," according to CFO Anat Ashkenazi during the Q3 earnings call.
Co-design: Google's differentiator
One aspect that differentiates TPUs from competitors is the co-design process between silicon architecture teams, networking engineers, compiler developers, and Google DeepMind researchers. Google already uses its own AlphaChip (a reinforcement learning system from DeepMind) to generate chip layouts, reducing weeks of development to hours. Gemini is also used to deliver layouts in days.
This feedback loop — using AI to design AI chips — is something competitors are still trying to replicate. NVIDIA does the same with its own ASICs and CUDA software, but Google has the advantage of operating at both the chip level and the model level, creating a feedback cycle that can accelerate the evolution of each generation.
What happens next
With four TPU systems in a single year, Google is signaling that the future of AI chips is not a single chip that does everything — it's a segmented portfolio for different stages of the model lifecycle. The question is whether this fragmentation will finally force an open standard, or if we'll end up with three or four proprietary ecosystems competing, each with its own software, drivers, and lock-in. Google says it wants to be open-source with PyTorch and XLA, but computing history shows that real interoperability is harder than launch-day rhetoric.
Sources: TechCrunch, Google Cloud Blog, Stock Analysis
✓ Independent sources cross-checked and verified before publishing
