Podcast Episode
Google Explores Direct TSMC Orders to Cut AI Chip Costs
May 4, 2026
0:00
2:30
Google is exploring whether it can reduce costs on its next-generation TPU by ordering wafers directly from TSMC, bypassing intermediary MediaTek. The move is part of a broader push to scale AI infrastructure affordably and compete more aggressively with Nvidia.
Google Probes Direct TSMC Relationship for Next-Gen AI Chip
Google is quietly investigating whether it can cut costs on its next-generation artificial intelligence chip by placing wafer orders directly with Taiwan Semiconductor Manufacturing Company, sidestepping the intermediary role currently played by MediaTek. The inquiry, reported by analyst Ming-Chi Kuo on 3 May, signals a shift in Google's approach to cost control as the company scales AI infrastructure to challenge Nvidia's dominance.A Cost-Driven Inquiry
Kuo reported that Google recently asked TSMC how much it could save by placing wafer orders for the main compute die of its upcoming TPU, codenamed Humufish (also designated TPU v8e), directly rather than routing them through MediaTek. MediaTek has been a key partner in Google's TPU ecosystem since joining the eighth-generation programme in late 2024, designing the inference-optimised TPU 8i (Zebrafish) alongside Broadcom, which handles the training variant (Sunfish). A direct arrangement with TSMC for Humufish, targeted for the second half of 2027, would mark a notable departure.Intel's Packaging Gamble
The Humufish chip is also at the centre of a separate contest over advanced packaging. Intel Foundry's EMIB-T technology has reached a validation yield of roughly 90%, but Kuo cautioned that the jump from 90% to the 98% mass-production threshold is considerably harder than the climb to 90%. Intel has prepared its Rio Rancho, New Mexico, facility for EMIB-T mass production and is in discussions with at least two major customers, including Google and Amazon. TSMC's CoWoS packaging also targets a 98% yield baseline.Broader AI Chip Realignment
The moves come amid a broader reshuffle of Google's chip strategy. At Google Cloud Next in late April, the company made its seventh-generation TPU, Ironwood, generally available and previewed a split eighth-generation lineup, with both Sunfish and Zebrafish targeting TSMC's 2-nanometre process. Google also announced it would sell TPUs to select customers for use in their own data centres, a direct challenge to Nvidia. Bank of America analyst Vivek Arya has estimated that Google's dual-sourcing arrangement between Broadcom and MediaTek could reduce per-chip costs by up to 30%, with a direct TSMC relationship pushing that further.Published May 4, 2026 at 4:21pm