AI Infrastructure

Google Unveils TPU 8tx and TPU 8i Max — Four TPU Systems in One Year, Challenging NVIDIA

Google TPU 8tx and TPU 8i Max chip packages with four-system portfolio diagram challenging NVIDIA GPU

Google Unveils TPU 8tx and TPU 8i Max — Four TPU Systems in One Year, Challenging NVIDIA's Dominance

Google expanded its eighth-generation TPU portfolio to four systems on November 9, 2026 — the first time the company has launched four TPU systems in a single year. The new additions: TPU 8tx (confidential inference with hardware-level security) and TPU 8i Max (dual-chip package for massive context windows and memory bandwidth). They join TPU 8t (large-scale training) and TPU 8i (low-latency inference) launched in April.

This modular system approach — same silicon, different packaging/networking/hosts — lets Google deliver specialized accelerators in months, not years, directly challenging NVIDIA's monolithic GPU strategy amid the latter's leadership crisis (Huang resignation).

The Four-System Portfolio: Purpose-Built for Every Agentic Stage

System Silicon Host CPU Network Purpose Key Spec
TPU 8t TPU 8t Arm Axion 3D torus Large-scale pre-training 12.6 FP4 PFLOPS, 216GB HBM
TPU 8i TPU 8i Arm Axion Boardfly Sampling, serving, reasoning 10.1 FP4 PFLOPS, 288GB HBM
TPU 8tx Same as 8t 5th Gen AMD EPYC Boardfly Confidential inference; small-scale training 2× lower latency vs 8t, near-zero security penalty
TPU 8i Max Two TPU 8i chips Not stated Not stated Long-context agent work; max density ~20.2 FP4 PFLOPS, 768MB SRAM, ~17.2 TB/s HBM

Modularity thesis: "A system is about how we bring together compute, networking, and storage around the chip... we made a foundational bet on system balance and modularity, architecting our TPUs as building blocks." — Amin Vahdat, Google SVP.

TPU 8tx: Hardware-Level Confidential Computing

Problem: Data encrypted at rest/in-transit but unlocked in memory during processing — vulnerable to insider threats, cloud provider access, side-channel attacks.

Solution: TPU 8tx builds security directly into the hardware layer, keeping sensitive data and model weights protected and isolated even while in active use.

Feature Detail
Confidential inference Hardware isolation prevents snooping on data/model weights in memory
AMD EPYC hosts 5th Gen EPYC replaces Arm Axion; enables Boardfly low-latency network
Boardfly network Same as TPU 8i; 2× lower latency vs TPU 8t for multi-step agents
Training + inference unified Fine-tune on private data, distill large→small, RL training — single system
Attestation Customer-verifiable confidentiality (details TBD)
Target market Regulated enterprises, governments, banks, healthcare, defense

Competitive positioning: vs. NVIDIA H100/H200 CCX (confidential compute) — Google claims TPU 8tx delivers near-zero performance penalty from security. vs. AMD SEV-SNP / Intel TDX — TPU 8tx is accelerator-native, not CPU-added.

TPU 8i Max: Breaking the Memory Wall for Agents

Problem: "Memory wall" — when dozens of AI agents work across entire codebases or hours of video, memory bandwidth becomes the bottleneck. Compute scales but HBM/SRAM doesn't.

Solution: Two TPU 8i chips on a single package — doubling compute, SRAM, and HBM bandwidth in lockstep so processors stay fed.

Metric TPU 8i TPU 8i Max (derived) Change
Compute (FP4 PFLOPS) 10.1 ~20.2 2×
HBM bandwidth 8.6 TB/s ~17.2 TB/s 2×
SRAM 384 MB 768 MB 2×
Rack footprint 1 unit Same 2× density
Watts/perf Baseline Better More compute/sq ft/watt

Use cases: Massive context windows (millions of tokens), agentic coding across repos, long-video analysis, mathematical reasoning, scientific research, drug discovery.

vs. NVIDIA Rubin R200: Rubin rated ~50 FP4 PFLOPS, 22 TB/s HBM4 — NVIDIA still leads raw speed, but Google claims 3× better AI output per dollar for agent workloads (SemiAnalysis on Ironwood vs. Blackwell).

The Strategic Context: NVIDIA Crisis + Modular Velocity

Factor Google TPU NVIDIA GPU
Systems/year (Gen 8) 4 (8t, 8i, 8tx, 8i Max) 1–2 (H100→H200→Blackwell→Rubin)
Architecture Modular systems (same silicon, diff packaging/hosts/network) Monolithic chip generations
Time-to-variant Months Years
Software TorchTPU open-sourced (PyTorch/HF native), TPU agent suite (Claude Code, Cursor, Antigravity) CUDA moat (decade+ ecosystem)
Customers Nearly 100 (Meta, Anthropic, OpenAI, Oracle, trading firms, universities) Hyperscalers + everyone else
Key customer signal Meta on TPUs; Anthropic $161B TPU lease payments (IPO filing) Locked-in
Leadership stability Stable (Kavukcuoglu, Vahdat) Crisis (Huang resignation, board investigation)

Anthropic IPO filing reveals ~$161B in TPU lease payments — making Anthropic Google's largest custom-chip customer (via Broadcom). Meta's TPU adoption is a major validation signal for external customers.

Software: Lowering the CUDA Switching Cost

Google's biggest historical barrier: CUDA ecosystem lock-in. New initiatives:

  1. TorchTPU open-sourced — PyTorch code + Hugging Face models run natively on TPUs
  2. TPU Agent Suite (preview) — Supports Claude Code, Cursor, Google Antigravity coding tools
  3. Automated tools — Deployment, profiling, optimization
  4. On-prem deployment — Customers can deploy TPUs in own data centers, not just Google Cloud

Market Projections: $12B → $135B by 2028

Source Projection
Bloomberg Intelligence TPU sales: $12B (2026) → $135B+ (2028) — 11× in 2 years
NVIDIA comparison Data center: $15B (FY23) → $115B (FY25) — 8× in 2 years
Epoch AI TPU compute shipments doubled in single quarter to ~800K chips
SemiAnalysis Ironwood TPU: ~3× better AI output/$ vs. Blackwell for agents

Caveat: Forecasts ≠ sales. NVIDIA still sells far more AI chips. But growth rate and unit economics favor Google's modular approach.

What Comes Next

  1. Availability dates — Currently "coming soon" for all four systems; ship dates TBD
  2. Full specs — HBM capacity, pod sizes, pricing, attestation details for 8tx
  3. Benchmarks — Independent MLPerf, real-world agent workloads
  4. Customer wins — Beyond Meta/Anthropic, enterprise/government adoption of 8tx
  5. NVIDIA Rubin response — H2 2026 launch; will it match modular velocity?
  6. Gen 9 timeline — If Gen 8 delivered 4 systems in 7 months, Gen 9 could accelerate further

The Bigger Picture

Google's four TPU systems in one year is a direct architectural counter to NVIDIA: instead of waiting 2–3 years for a new chip generation, Google reconfigures proven silicon into specialized systems for emergent workloads (confidential compute, agentic memory walls).

This modular velocity — combined with open software (TorchTPU), on-prem option, and NVIDIA's leadership crisis — creates the strongest competitive window for TPUs since their inception. The $135B revenue projection by 2028 assumes Google executes on availability, benchmarks, and CUDA migration.

If TPU 8tx/8i Max ship on schedule with claimed specs, Google transitions from "internal accelerator" to credible merchant silicon alternative — the first since NVIDIA's CUDA moat formed.


Related AIPress Coverage

Sources

  • Google Blog (Nov 9, 2026): "Introducing TPU 8tx and TPU 8i Max" (Amin Vahdat)
  • The Decoder (Nov 10, 2026): "Google ships four TPU variants in one year"
  • TechCrunch (Nov 10, 2026): "Google broadens its AI hardware portfolio with two new specialized chips"
  • Yahoo Finance (Nov 10, 2026): "Google's chip business is on a faster track than Nvidia's own AI boom"
  • The Hindu (Nov 10, 2026): "Google launches two AI chips as it seeks to challenge NVIDIA's dominance"
  • Tom's Hardware (Nov 12, 2026): Detailed specs table, derived 8i Max numbers
  • AA.com.tr (Nov 2026): Bloomberg report summary
  • TrendForce (Nov 10, 2026): "Google Marks First 4-TPU Year with 8tx and 8i Max"
  • Bloomberg Intelligence (Nov 2026): TPU revenue projections
  • SemiAnalysis (Nov 2026): Ironwood vs. Blackwell cost analysis
  • Epoch AI (Nov 2026): TPU shipment compute doubling

Building something with AI?

DevsIsle designs and ships AI systems, agents and integrations for teams that need it done properly.

Talk to our team →