Article title: Report: Google Adjusts TPU Supply Chain Strategy: Split Design Duties with MediaTek, Bypass Broadcom to Buy HBM Directly from Memory Makers
Body:
- Day/Night Follow System Light Dark
Report: Google Adjusts TPU Supply Chain Strategy: Split Design Duties with MediaTek, Bypass Broadcom to Buy HBM Directly from Memory Makers
IT之家 November 20 news, according to Korean media Chosun Biz, as investment in artificial intelligence infrastructure increases, large US tech companies are gradually adjusting their chip procurement strategies.
To reduce dependence on Nvidia GPUs, these companies are adopting self-developed AI chips or pursuing supplier diversification on the one hand, and choosing to purchase high-bandwidth memory (HBM) directly to cut out intermediary costs on the other.
According to industry news on the 18th, this change has become an important variable affecting the profitability of the three major memory chip makers: Samsung Electronics, SK Hynix, and Micron. As customers directly control procurement costs, negotiations over supply prices and contract terms will also intensify.
An industry source said: "Right now, due to the HBM shortage, the seller's advantage continues, but some long-term contracts have already included price caps," and "as time passes, buyers' bargaining power may strengthen."
It is reported that Google is reducing procurement expenses by separating the design of its self-developed AI chip TPU from memory procurement. Instead of handing all of design, packaging, and HBM procurement to Broadcom, it is splitting design duties with MediaTek and purchasing HBM directly.
IT之家 notes that, according to a further breakdown by semiconductor research firm Counterpoint Research, Google is not "completely abandoning" Broadcom across all TPU product lines, but is adopting a "dual-track / split-duty model" in some new-generation TPUs — Google designs the core compute die itself, MediaTek handles I/O chip design, and Google purchases HBM directly; while Broadcom still participates in the development and packaging of Google's other high-end mainline TPU models.
Related Articles
- Google confirms native "App Lock" feature cannot be pushed to older Android devices via Google Play, requires upgrade to Android 17 QPR3
- Running a 270-million-parameter model in about 14MB of memory: Google engineers put AI into a printer, about 60 seconds per token
- Changing just one piece of code, Google AI agent task success rate improves 8x after Harness optimization
- Google announces global rollout of Android developer verification program in January 2027, covering Google Play, Samsung, Honor, OPPO, vivo, Xiaomi, Transsion
- Google Gemini can now directly operate Apple Mac computers, enabling cross-app operations, creating PPTs, and filling out forms
- Google partners with Blue Origin, plans to launch AI server satellites in 2029
Sites under 软媒: IT之家 最会买 - 返利返现优惠券 iPhone之家 Win7之家 Win10之家 Win11之家
Software under 软媒: 软媒手机APP应用 魔方 最会买 要知