Google TPU 8i: The Complete Guide

As of this check, Google’s own product page still lists TPU 8i as “Coming Soon” 2, even as Broadcom, Google’s manufacturing partner, told investors on its Q3 fiscal 2026 earnings call that it had already begun production shipments (see Made, below).

Google’s own stated design goal: more memory bandwidth to serve the most latency-sensitive inference workloads, optimized for post-training, reinforcement learning, and low-latency serving of large mixture-of-experts models 12.

Google’s technical deep-dive describes Boardfly precisely: chips connect in fully-connected boards that are then aggregated into groups using a high-radix design, connecting up to 1,152 chips per group 3. For a comparable 1,024-chip pod, Google shows the worst-case hop count for a data packet crossing the system falling from 16 hops (a standard 3D-torus layout) to just 7 hops — a 56% reduction, which Google says delivers up to a 50% latency improvement for communication-heavy workloads 3. The 1,152-chip figure describes Boardfly’s maximum per-group interconnect capacity; the 16-to-7-hop reduction is demonstrated specifically for a 1,024-chip pod, and this page reports both numbers separately rather than merging them into one claim. Google has not published a process node or foundry for TPU 8i.

Figures per Google’s technical deep-dive 3. Google’s launch blog separately states 19.2 Tb/s of interconnect bandwidth — a different measurement from the 8,601 GB/s HBM-bandwidth figure above, and the two should not be conflated 1. Google’s product page states roughly 80% better performance-per-dollar versus the previous generation 2.

Broadcom’s April 2026 SEC filing describes its long-term agreement to develop and supply custom TPUs for Google, running through up to 2031, without naming TPU 8i specifically or disclosing a process node 4. On Broadcom’s Q3 fiscal 2026 earnings call (2 September 2026), CEO Hock Tan is reported as saying Broadcom “began production shipments of the next-generation TPU version 8i for Google,” describing it as designed with more memory and bandwidth than Ironwood and, like Ironwood, optimized for inference 5. This quote is sourced from a published transcript of Broadcom’s public earnings call, not from a Broadcom-issued press release or SEC exhibit, so this page attributes it accordingly rather than presenting it as a formal written disclosure.

The only customer-adjacent detail available concerns Anthropic, and it comes from Broadcom’s earnings call rather than a Google or Anthropic statement: Broadcom’s CEO is reported describing a path from Anthropic’s 2026 Ironwood deployment toward several gigawatts of TPU 8i capacity in 2027 5. Neither Google nor Anthropic has named TPU 8i directly in their own official material as of this check.

Google Cloud’s pricing page, checked September 2026, lists no TPU 8i entry. This page reports that plainly rather than estimating a rate, even though Broadcom has publicly discussed shipping the chip in volume.

As with TPU 8t, no MLPerf or third-party benchmark submission exists for TPU 8i as of this check, consistent with its “Coming Soon” status on Google’s own product page.

TPU 8i succeeds Ironwood as the inference-focused half of Google’s eighth generation. As of this check, Google has announced nothing beyond the eighth generation.

Broadcom’s own CEO told investors on a 2 September 2026 earnings call that production shipments of TPU 8i had already begun 5 — while, as of this site’s September 2026 check, Google Cloud’s own customer-facing product page still lists the same chip as “Coming Soon,” with no price, region, or GA date published. The gap between what a supply chain discloses to investors and what a customer-facing product page says is ready is rarely this visible, and here it is visible only because Broadcom, not Google, is the one who is required to talk to public shareholders about it.