Log entry Nº 543 · the stack watch
Bearing SE 135° · November 10, 2026 · 4 min read
AI in MarketingSE 135°Watch log
Google adds agentic and confidential eighth-generation TPU variants
Google’s eighth-generation TPU split adds a memory-heavy 8i Max for agents and an 8tx for confidential inference, but pricing and launch dates remain undisclosed.
By Amara Osei4 min read837 words

By Amara Osei
Google has added two eighth-generation TPU variants: the 8i Max delivers 768 MB of on-chip static RAM and 17.2 TB/s of high-bandwidth memory (HBM) bandwidth for agentic work, while the 8tx targets confidential inference at half the TPU 8t’s latency.
Google created its TPU accelerator line for the company’s own AI work and, increasingly, cloud customers. Each generation has addressed the bottlenecks of its time, rather than treating compute capacity as the only constraint.
Mark Lohmeyer, Google Cloud’s vice president and general manager for AI and computing infrastructure, framed the shift this way:
“[AI] used to be mostly about how much compute you had available, but now it’s also about how quickly data can move, how efficiently you can use the memory, how do you minimize latency, and how you manage the cost and security of running all these workloads at scale.”
What changes for agentic workloads?
Google designed the TPU 8i Max for the massive context windows that agentic workloads require. It adds a second TPU to a multi-chip package, allowing groups of agents to work together on tasks such as software development and video creation.
- Memory: The two-chip design supplies 768 MB of on-chip static RAM and 17.2 TB/s of HBM bandwidth.
- Data-center density: Deployed in large numbers, these TPUs can maximize compute per square foot.
- Open-source orchestration: TPU Sync, software Google recently open-sourced, coordinates chip HBM, host DRAM, shared storage and the data-center network as a single memory hierarchy.
Patrick Moorhead, founder, CEO and chief analyst of Moor Insights & Strategy, wrote in an email: “With roughly two-thirds of AI compute now going to inference, specializing [in] inference hardware is the right call.”
Takeaway: The 8i Max favors large-context, multi-agent work, while TPU Sync coordinates the memory around it.
How does the 8tx protect sensitive work?
Google designed the TPU 8tx for secure and confidential inference, a workload it is hearing from government customers and enterprises in regulated industries. The system supports fine-tuning with private data, model distillation and reinforcement-learning training for agents without moving sensitive data between separate clusters.
Moorhead described the operational result: “A bank or government agency can fine-tune, distill and run reinforcement learning on the same confidential system that serves the model, without moving sensitive data between clusters.”
Amin Vahdat, senior vice president and chief technologist of AI and infrastructure at Google, said in a Monday blog post that teams can fine-tune, distill and train agents without migrating sensitive data. During processing, data and model weights stay encrypted and isolated at the hardware level.
Lohmeyer said the 8tx incurs a near-zero performance penalty and runs at half the latency of the TPU 8t.
Takeaway: The 8tx lets regulated teams keep sensitive training and inference inside one isolated system.
Who is using the new TPU generation?
Google built each TPU generation with input from Google DeepMind, but the company now also consults an expanding outside user base reached through Google Cloud and channel sales. More than 100 customers and partners use TPUs, according to Google; they include frontier labs, trading firms, retailers and universities.
Among them, Anthropic will expand its use of next-generation TPUs beginning in 2027. Moorhead learned in a conversation with the team overseeing that operation that the California Institute of Technology has shifted some cancer research from its own HPC system to Google Cloud TPUs.
Google has not provided availability dates or pricing for either the 8i Max or the 8tx. The earlier 8t and 8i remain due to go online by the end of this year.
Takeaway: Adoption is broadening, but buyers still lack prices and firm dates for the two newest chips.
Why does workload specialization matter?
Google launched the first batch of eighth-generation TPUs in April with two variants: the 8t for large-scale training and the 8i for inference. Lohmeyer said this generation was the first to split AI workloads into two specialties with separate hardware modifications.
Rack-level engineering remains central. The eighth generation optimizes CPUs, system DRAM, networking and shared storage for each workload, rather than improving only the chip. According to Vahdat, the 8t delivers up to 2.7 times better training performance, while the 8i delivers 80% better inference price-performance than previous TPU generations.
Lohmeyer summarized the constraint: “You can’t optimize that infrastructure just by focusing on the chip.” The new 8i Max and 8tx extend that approach to agent memory and confidential workloads.
Moorhead sees a broader hardware shift. “Agents put the CPU and host platform back at the center,” he wrote, arguing that GPUs no longer stand as the only focal point of AI computing.
Takeaway: Workload fit, not chip-only gains, now drives Google’s TPU design.
For practitioners, the buying decision is clearer by workload, but not by dollar: Google still has not published prices or launch dates for the 8i Max and 8tx.
Plot points
- Google added 2 eighth-generation variants: TPU 8i Max and TPU 8tx.
- The 8i Max provides 768 MB of on-chip SRAM and 17.2 TB/s of HBM bandwidth.
- The 8tx targets confidential inference at half the latency of the TPU 8t.
- More than 100 customers and partners use TPUs; Anthropic plans to expand its use from 2027.
- The 8t offers up to 2.7x better training performance, while the 8i delivers 80% better inference price-performance than prior generations.
via techtarget.com (Source)
More from Amara Osei
Show full bio
Market editor covering technology and product news in marketing technology at Martech Beacon.
51 articles
Neighbouring fixes
- Salesforce Unveils Agentic Marketing Cloud Next
- Alphabet Made $127.6 Billion in Ads Last Quarter — and Barely Mentioned It
- Salesforce and Google Cloud Confirm New Platform Integrations
- Google Vids adds free 15-second Omni 1.2 Flash video generation
- NiCE Ships Next-Generation Cognigy: Reusable Skills, Tracing, Bundled Models