AI Agents Northeastern and Stanford Give Agents a Git History: Shepherd Lifts Pair-Coding Pass Rates From 28.8 to 54.7 Percent Coding Assistants China Telecom Trains a 285B-Parameter Agentic Model on 8,192 Ascend 910C Chips With No Nvidia Silicon Anywhere in the Stack News Google Cuts 2027 Broadcom-Designed TPU 8i Orders by About a Million Units and Pushes Volume Toward MediaTek Business China's Data Regulator Will Set National Rules for Metering and Billing AI Tokens After Daily Use Tripled to 500 Trillion Business Shanghai's Yuanshi Technology Raises Nearly RMB 300 Million for AI Virtual Cells, Two Weeks After the Last Chinese Record Fell Security GAO Finds No Widely Accepted Test for Generative AI Cyber Skills and Hands Congress 16 Options, Including Human Sign-Off on Consequential Actions News Microsoft Turns Three USB4-Cabled PCs Into One AI Machine: Qwen3-235B Runs at Up to 22 Tokens a Second Across 384GB of Ordinary Memory Business Dell, HPE, Lenovo and Supermicro Are Demanding Up to 45% More for Signed AI Server Deals, and Trading Firms Are Paying AI Agents Northeastern and Stanford Give Agents a Git History: Shepherd Lifts Pair-Coding Pass Rates From 28.8 to 54.7 Percent Coding Assistants China Telecom Trains a 285B-Parameter Agentic Model on 8,192 Ascend 910C Chips With No Nvidia Silicon Anywhere in the Stack News Google Cuts 2027 Broadcom-Designed TPU 8i Orders by About a Million Units and Pushes Volume Toward MediaTek Business China's Data Regulator Will Set National Rules for Metering and Billing AI Tokens After Daily Use Tripled to 500 Trillion Business Shanghai's Yuanshi Technology Raises Nearly RMB 300 Million for AI Virtual Cells, Two Weeks After the Last Chinese Record Fell Security GAO Finds No Widely Accepted Test for Generative AI Cyber Skills and Hands Congress 16 Options, Including Human Sign-Off on Consequential Actions News Microsoft Turns Three USB4-Cabled PCs Into One AI Machine: Qwen3-235B Runs at Up to 22 Tokens a Second Across 384GB of Ordinary Memory Business Dell, HPE, Lenovo and Supermicro Are Demanding Up to 45% More for Signed AI Server Deals, and Trading Firms Are Paying

Google Cuts 2027 Broadcom-Designed TPU 8i Orders by About a Million Units and Pushes Volume Toward MediaTek

Google Cuts 2027 Broadcom-Designed TPU 8i Orders by About a Million Units and Pushes Volume Toward MediaTek

TF International analyst Ming-Chi Kuo said Google cut about 1 million Broadcom-designed TPU 8i chips from its 2027 plans, roughly 20 percent of that series, concentrated in the pricier 8i Max. Google is expected to take more MediaTek-designed TPU 8t volume instead, as power and packaging constraints push it to rebalance costlier parts of its custom silicon stack.

Google has cut roughly one million Broadcom-designed TPU 8i chips from its 2027 plans, about 20 percent of the projected volume for that series, according to a supply-chain survey published by TF International Securities analyst Ming-Chi Kuo on the evening of Nov 20. The reduction falls mainly on the higher-priced 8i Max variant, known internally by the codename Whalefish.

At the same time, Kuo said Google expects to take more of the MediaTek-designed TPU 8t series in 2027, though the size of that increase has not been fixed. MediaTek has supplied a share of Google's custom accelerator work since the 8t generation, and the shift would move volume toward the cheaper end of the lineup rather than replacing Broadcom outright.

The stated reason is arithmetic rather than a rupture in the relationship. Google has said repeatedly over recent months that frontier model development gets first call on available compute. With data-centre power constrained and advanced packaging capacity sold out well into next year, trimming the most expensive per-unit part of the roadmap is the fastest way to protect research capacity without cutting total deployed compute.

Korean outlet Chosun Biz reported a parallel structural change: Google is separating TPU design from memory procurement so that it no longer hands design, packaging and HBM sourcing to Broadcom as a single bundle. Under that arrangement Google would split part of the design work with MediaTek and buy high-bandwidth memory directly, a move aimed at pulling cost out of each accelerator.

Broadcom fell 8.9 percent on the day the survey circulated, according to Chinese market reports, a reminder of how concentrated its custom silicon business has become. Broadcom's AI revenue is now dominated by a small number of hyperscaler programmes, and a 20 percent trim on one of the largest of them reads to investors as a signal about 2027 rather than a one-quarter wobble.

Both figures should be treated as supply-chain estimates rather than confirmed orders. Kuo's surveys have a strong record on Apple's roadmap and a thinner one on hyperscaler silicon, and neither Google nor Broadcom has disclosed 2027 unit plans. The 8t volume increase in particular is described as undetermined.

The broader pattern is that custom accelerators are becoming a portfolio problem rather than a single-bet problem. Google now runs several TPU generations in parallel with different design partners, mixing high-end parts for training with cheaper parts for inference and serving, which makes each supplier's share volatile even when total capex keeps rising.

What this means: the story the market will take from this is not "Google dumps Broadcom" but "Google is optimising a stack it can no longer grow at will". When power and packaging, not demand, set the ceiling, the winning supplier is the one that can deliver acceptable performance per watt per dollar, and that competition is now openly running through MediaTek.

Comments (0)

Log in to join the discussion

Log In

No comments yet