Google Cuts Broadcom TPU 8i Orders by 20% and Shifts to MediaTek: A Strategic Reorganization Signal Prioritizing Training Over Inference

Google has cut its 2027 Broadcom-designed TPU 8i shipment plan by about 1 million units, roughly 20%, while increasing orders for MediaTek-designed TPU 8t.

On November 20, 2026, TF International Securities analyst Ming-Chi Kuo disclosed that Google had cut its 2027 shipment volume for the Broadcom-designed TPU 8i series by about 1 million units, a reduction of roughly 20% of the series’ full-year plan, lowering the original forecast of 5 million units to 4 million. At the same time, Google plans to increase orders for the MediaTek-designed TPU 8t series. Broadcom’s stock plunged more than 12% that day, wiping out about $191 billion in market value.

Inference Makes Way, Training Takes Priority

The TPU 8i series is designed by Broadcom and mainly targets inference workloads; the MediaTek-designed TPU 8t series is better suited to training workloads. In his analysis report, Kuo cited Google’s stance in recent months: “Frontier model research and development has absolute priority for compute.” Under the dual constraints of data center power and chip production capacity, Google is compressing inference-side orders and reserving resources for the training side.

Over the past two years, the AI industry narrative has centered on inference cost reduction and commercial deployment. Google’s move indicates that in internal compute allocation priorities, the pace of frontier model R&D ranks above the scale replication of existing capabilities.

Power Constraints Have Moved from Theoretical Risk to Practical Binding

Kuo attributed the adjustment to supply chain tightness and data center power limits. The cuts are most concentrated in the 8i Max (codename Whalefish), the highest-priced and highest-power variant in the series. With both power and rack space constrained, high-power premium inference chips have become the sacrifice. Because the 8i Max unit price is higher than other models in the series, the actual cut’s impact on Broadcom’s revenue is greater than the shipment number suggests.

Whether the incremental volume of MediaTek’s 8t series can fully offset the reduction in Broadcom’s 8i remains uncertain, but total compute volume is broadly unchanged, while the internal mix tilts toward training.

MediaTek’s Long-Term Positioning Is Taking Shape

Kuo estimates that MediaTek will become the sole design supplier for Google’s v9-generation TPUs (codename Humufish, Triggerfish) and v10 generation (codename Quillfish). Its share of Google’s total TPU supply will rise from about 45% to 50% in 2027 to 50% to 55% in 2028, and further to 55% to 60% in 2029. The original Icefish design has advanced to v11, targeting mass production in 2030, with the design service order awarded to MediaTek.

Kuo forecasts TPU shipments: 8 million to 9 million units in 2027, 14 million to 15 million units in 2028, and 16 million to 18 million units in 2029. Starting with the v9 generation, Google TPUs will adopt Intel’s EMIB-T advanced packaging technology, and MediaTek’s role will expand from chip design service provider to key technology integrator.

Is Broadcom Facing a Cyclical Adjustment or a Structural Shift?

On November 20, the Taiwan Weighted Index fell more than 6% overall, while MediaTek closed limit-down, down 9.95%. Broadcom enjoys relatively high gross margins on the 8i design, and its strategic focus is shifting toward non-TPU XPUs and AI networking and communication chips; this revenue could surpass TPU revenue as early as 2027.

Starting with the v9 generation, MediaTek takes over sole design rights for Google TPUs. Losing design leadership for v9 and subsequent products is the starting point that affects the long-term landscape. Kuo defines the v8 series as the core profit engine for MediaTek and Broadcom over the next few years, but this window will narrow significantly around 2029.

The Diversification Logic of the AI Chip Supply Chain Has Entered Execution

Google’s adjustment is an execution case of the trend toward reducing single-source dependence in AI infrastructure supply chains. This logic previously manifested as cloud providers ramping up in-house chips to reduce reliance on Nvidia; this time it extends within the in-house system. OpenAI and Anthropic are important Google TPU customers; the former buys directly, while the latter combines purchasing and leasing. Anthropic listed on Nasdaq on November 19 under the ticker ANT. New order demand from the two companies extends Kuo’s order visibility for the TPU system to 2029, exceeding visibility for Nvidia GPU orders.

Independent Judgment

The core signal of this adjustment is that the priority of training compute is being institutionalized in AI chip procurement. As power and production capacity become hard constraints, chip procurement shifts from “buying more” to “buying the right ones”—under Google’s strategic framework, training compute that serves frontier model building. MediaTek’s exclusive supply of v9- and v10-generation TPUs, with its supply share rising year by year, is a structural shift with a specific timetable. The real issue Broadcom faces is the design leadership it loses starting with the v9 generation. The AI infrastructure supply chain is moving from single-oligarch dominance toward a layered, function-oriented diversified landscape, and power constraints will continue to play the role of allocation mechanism.