Washington's Multi-Front Push Against NVIDIA And Why Every De-NVIDIA Path Still Runs Through Taiwan

From 080深度對談 · 17:03

Show by 島編

From antitrust scrutiny to hyperscaler alliances, the United States is pulling every lever to weaken NVIDIA's unipolar dominance. Yet at the level of physical manufacturing, every alternative to NVIDIA still cannot bypass Taiwan. In this episode of 080 In-Depth Analysis: The Post-AI Era, we unpack the geopolitical and supply-chain reality behind the next phase of AI compute.Episode HighlightsThree Fronts Challenging Jensen Huang and NVIDIADOJ Antitrust & Export-Control Scrutiny: The U.S. Department of Justice has escalated its antitrust investigation into whether NVIDIA uses tying and ecosystem lock-in to restrict customer switching, scrutinized its $17 billion licensing deal with Groq as a potential disguised acquisition, and opened chip-export compliance inquiries involving partners such as Megaspeed and Supermicro (Legal500, GuruFocus, Crypto.news).Hyperscalers' Full-Scale Pivot to Custom ASICs: Google (TPU), AWS (Trainium 3), Microsoft (Maia 200), and Meta (MTIA) are partnering with design houses like Broadcom to shift internal inference and specialized workloads onto custom silicon, reducing their exposure to NVIDIA's high margins (Presenc AI, Spheron, AI World Today).OpenAI & Rival Coalitions Breaking CUDA Lock-In: OpenAI and Broadcom are planning 10 GW of custom AI accelerators, while AMD continues advancing high-end GPUs and open software toolchains to erode CUDA's single-vendor lock-in (Globe Market Research, Enki AI).Why Challenging NVIDIA Still Requires Taiwan's CooperationTotal Dependence on Leading-Edge Nodes: Whether it is hyperscaler custom ASICs or AMD challenger GPUs, nearly all roads lead back to TSMC. In Q3 2026, advanced nodes (7nm and below) accounted for 79% of TSMC's wafer revenue (including 8% from 2nm and 33% from 3nm), with North American customers contributing 75% of net revenue (Safeguard).CoWoS Advanced Packaging as the Shared Bottleneck: The primary capacity constraint has shifted from wafer fabrication to advanced packaging, where TSMC's 5.5x-reticle CoWoS lines sustain 98%–99% yields—meaning every U.S. chipmaker aiming to displace NVIDIA still has to queue for TSMC's CoWoS allocation (Enki AI, Silicon Analysts).Reshoring U.S. Supply Chains Still Depends on Taiwan's Execution: TSMC has expanded its U.S. investment commitment to $165 billion (including advanced fabs and two packaging facilities, with Arizona Fab 2 targeting 3nm volume production in Q3 2027). Even GlobalFoundries' first U.S. CoWoS silicon interposer line in New York relies on a $2 billion ecosystem agreement with TSMC, alongside at least $250 billion in Taiwanese tech investment commitments under the 2026 U.S.–Taiwan Trade Agreement (Safeguard, Silicon Analysts, Artventure).An Irreplaceable Full-Stack AI Hardware Cluster: From wafer fabrication and packaging to AI server racks, power delivery, and liquid cooling, Taiwan's end-to-end manufacturing cluster remains indispensable as U.S.–Taiwan cooperation evolves from the "Silicon Shield" to the "AI Shield" (FSM, Stimson Center).加入會員,支持節目: https://hakka1893.firstory.cc/join留言告訴我你對這一集的想法: https://open.firstory.fm/user/cmeqfytxf00im01ynerq283nx/comments🏛️ 080深度對談|穿梭島嶼時空,尋找歷史與現實的座標從文化資產的斷壁殘垣,到都市變遷的異聞傳說;從建築背後的權力情感,到地緣政治的驚濤浪潮。我們不只談歷史,更談歷史如何形塑今天的我們。🎬 YouTube 深度影音:https://www.youtube.com/@taiwanstory558 🎙️ Podcast 隨時收聽:https://open.firstory.me/user/cl1w8w5sh01w301w8a74d2q7u Powered by Firstory Hosting

More episodes of 080深度對談

More episodes

PodSearch is a free, independent podcast search engine. Read about PodSearch, the privacy policy or contact us.