FuriosaAI défie Nvidia avec une nouvelle puce IA révolutionnaire - Le pari risqué de 2026

Le géant du GPU vient de se faire un nouvel ennemi. FuriosaAI, une startup sud-coréenne, dévoile un processeur neuromorphique qui promet de contourner les limitations des architectures traditionnelles.

Une architecture radicalement différente

Plutôt que de simplement imiter les puces de Nvidia, l'approche de FuriosaAI repose sur un design neuromorphique. Cela signifie que le matériel imite plus étroitement la structure du cerveau humain, promettant des gains d'efficacité exponentiels pour des charges de travail spécifiques comme l'inférence en temps réel. Les premiers benchmarks, bien que non vérifiés de manière indépendante, laissent entendre des réductions de latence allant jusqu'à 70% sur certains modèles.

Le marché réagit... avec scepticisme

L'annonce a envoyé des ondes de choc à travers l'industrie des semi-conducteurs, mais les investisseurs restent sur leurs gardes. Un analyste a noté, avec le cynisme habituel de la finance, que "chaque nouveau prétendant au trône promet la lune, mais finit généralement par se brûler les ailes sur le mur de jardin de CUDA". La dépendance de l'écosystème IA aux bibliothèques logicielles de Nvidia reste l'obstacle majeur à toute disruption matérielle.

Une opportunité pour la crypto et le DeFi ?

Pour les praticiens de la cryptographie, cette course aux armements du matériel IA ouvre des perspectives fascinantes. Une plus grande concurrence et des gains d'efficacité pourraient drastiquement réduire le coût des opérations de calcul intensif, potentiellement bénéfique pour les réseaux décentralisés et les projets DeFi qui reposent sur l'oracle et l'analyse de données complexes. Si FuriosaAI tient ses promesses, elle pourrait bien fissurer l'hégémonie actuelle et redistribuer les cartes de la puissance de calcul – un actif de plus en plus précieux à l'ère de l'IA.

Furiosa chip demonstrates efficiency for large AI models

Furiosa’s current chip, designated RNGD, short for “renegade,” is aimed at the inference phase of artificial intelligence, which involves running trained AI models. Nvidia’s graphics processing units (GPUs) currently dominate the broader AI computing market, particularly in training large models, but startups such as Furiosa are targeting inference as a potential area of competition.

The company’s latest funding round values Furiosa at close to $700 million. Meta Platforms approached Furiosa about a potential acquisition last year, but no deal was reached. OpenAI demonstrated the use of a Furiosa chip at a recent event in Seoul. while LG’s AI research division is testing the chip and said it showed “excellent real-world performance.” Furiosa said it is currently in discussions with potential clients.

According to Paik, Furiosa’s chips deliver performance comparable to Nvidia’s sophisticated GPUs while drawing less power. This would reduce the overall expense of putting AI systems into operation. The technology sector shouldn’t depend so heavily on a single chip manufacturer for AI computing, Paik argues.

“A market dominated by a single player—that’s not a healthy ecosystem, is it?” Paik said.

Paik began his career at Advanced Micro Devices (AMD), where he worked on GPU design, before returning to South Korea in 2013 to join Samsung. There, he led a small team developing new memory-chip products.

Hanjoon Kim, now Furiosa’s chief technology officer, previously worked with Paik at Samsung and later joined him in founding the company. Kim said Paik emphasized rapid decision-making and long-term goals during Furiosa’s early development.

South Korea positions itself for AI and inference growth

With strong software capabilities and semiconductor knowledge from homegrown firms like Samsung and SK Hynix, South Korea is putting a lot of emphasis on AI development. The government has made the development of AI a top priority, hoping to become a technological leader alongside China and the United States. The government of South Korea arranged for Nvidia to finalize a significant GPU supply deal, while OpenAI recently opened an office in Seoul.

During Furiosa’s beginning years, Paik frequently cited “Blitzscaling: The Lightning-Fast Path to Building Massively Valuable Companies,” a book popular in Silicon Valley. He used it to stress the importance of quick decisions and bold moves to dominate markets as an early entrant.

At Stanford’s respected Hot Chips conference in 2024, Paik introduced Furiosa’s RNGD chip during a keynote presentation, calling it an answer to what he termed “sustainable AI computing.” Paik shared information demonstrating the chip could operate Meta’s Llama large language model with power efficiency exceeding Nvidia’s top-tier chips by more than double.

“It was a moment where we felt we could really move forward with our chip with confidence,” Paik said.

Looking back now, Paik views his Achilles injury as a pivotal moment. “I think it could have been a blessing in disguise,” he said.

Get seen where it counts. Advertise in Cryptopolitan Research and reach crypto’s sharpest investors and builders.