Technologie · AMD vs Nvidia · ArchivesHelios d'AMD réunit 72 GPU dans un seul rack, et quatre hyperscalers disent ouiRédigé 7 septembre 2026 à 21:22 UTCRegarderTranscription0:00AMD just fired a direct shot at Nvidia's0:02stranglehold on the AI chip market. The0:04company formally launched Helios, its0:06first rack-scale AI system, and it0:08already has some of the biggest names in0:10tech signed up as customers. Microsoft,0:12Meta, OpenAI, and Oracle are all on0:14board. This is AMD's biggest bet yet to0:17break Nvidia's near-total dominance of0:18AI hardware. Helios is a rack-scale0:21system, meaning it's not just a chip.0:23It's an entire integrated hardware stack0:25designed to power the largest AI0:27workloads. Each rack packs 72 GPUs and0:3018 CPUs, is liquid-cooled, and costs an0:33estimated 5 to 5 and 1/2 million0:35dollars. Power draw per rack, between0:37225,000 and 245,000 watts. AMD has put0:428,000 employees on this project, and0:44they expect tens of billions in revenue0:46from data center AI hardware starting0:48next year. What makes Helios a genuine0:50threat to Nvidia is twofold. First, the0:53customer list reads like a who's who of0:55AI. Microsoft, Meta, OpenAI, Oracle.0:58These are the companies building the0:59largest AI clusters in the world.1:01Second, Helios is built on an open1:03standard rack design, directly1:05contrasting with Nvidia's proprietary1:07Vera Rubin architecture. That open1:09standard gives hyperscalers more1:11flexibility and more leverage in1:12negotiations. The specs are aggressive.1:15Each rack handles enormous compute1:17density, and the open standard approach1:19means cloud giants aren't locked into a1:20proprietary ecosystem the way they are1:22with Nvidia's architecture.1:24That alone is a powerful selling point1:26for companies looking to diversify their1:28AI hardware supply chain. For the1:30hyperscalers, Microsoft, Meta, Oracle,1:32Helios represents something they've been1:34wanting for years, a credible second1:36source for AI compute. It gives them1:38negotiating leverage with Nvidia and1:40reduces their dependency on a single1:42supplier. For Nvidia, this is the first1:45time it faces a genuine rack-scale1:46competitor.1:47Vera Rubin is still the market leader,1:49but AMD now has a product, customers,1:51and a massive manufacturing pipeline.1:53Microsoft and Meta are expected to be1:55among the earliest deployers. OpenAI,1:58which runs some of the most1:59compute-intensive models in existence,2:01has also committed, and Oracle, which2:03has been aggressively building out its2:05cloud infrastructure, rounds out the2:07anchor customer base. AMD CEO Lisa Su2:10has described Helios as the culmination2:12of years of investment in data center2:14AI. The company's data center head,2:16Forrest Norrod, has emphasized that this2:17is not just another chip launch, it is a2:19platform play designed to compete with2:21Nvidia at every layer of the stack. The2:23question now is whether AMD can ramp2:25production fast enough to meet demand2:27and start eating into Nvidia's revenue2:29share.Lire l’articleChronologie de l’histoireHeyDexLe format court ↗
