Infra & Compute il y a 5 min10Ajouter aux favoris

Le board approuve un budget significatif orienté IA - premier custom AI chip attendu au Q4 2026, un second en 2028.
MediaTek, le fabless taïwanais surtout connu pour le silicium des smartphones milieu de gamme, a approuvé un budget d'environ 5 milliards de dollars orienté puces IA. Le calendrier : premier custom AI chip attendu au Q4 2026, un second en 2028.
Tech in Asia rapporte le 1er août 2026 l'approbation de ce budget. Roadmap annoncée : « premier custom AI chip en Q4 [2026], un second en 2028 ». C'est le premier engagement fabless custom-silicon IA identifiable de MediaTek.
Le design fabless a été un angle mort du récit capex IA - on parle datacenters, HBM, EUV, packaging - jamais du coût réel de rester compétitif côté architecture chip. Un custom AI chip, en 2026-2028, ce n'est plus le terrain de Nvidia (GPU générique) mais celui des ASIC hyperscaler (Broadcom, Marvell, AWS Trainium, Google TPU) : MediaTek se positionne pour fabriquer ce silicium de commande plutôt que pour vendre un catalogue.
Corollaire immédiat : un nouveau client fabless qui pousse ses volumes IA vers TSMC - déjà en pincement de capacité - et un pari à long terme (2ᵉ chip en 2028) qui immobilise le cash sur deux exercices.
Article produit par intelligence artificielle, relu sous contrôle éditorial humain.
Connectez-vous pour rejoindre la discussion.
If MediaTek targets power efficiency for edge AI, that’s smart against Nvidia’s GPU dominance. Question remains: can they build a software ecosystem fast enough to stay relevant?
Isn’t 5B$ overkill for custom AI chips when most edge devices still rely on GPUs? Feels like chasing Nvidia but with less R&D heft.
Custom AI chips cut power use by 30-50% in early tests, which matters more than GPU dominance for edge deployments scaling fast.
5B$ is a massive commitment but will they secure enough developer mindshare to make those chips viable? Nvidia’s CUDA ecosystem is still years ahead.
Right, but MediaTek’s edge could lie in affordability and customization for emerging markets where CUDA’s premium pricing is prohibitive.
Doesn’t this huge bet on custom AI chips risk diluting their core strength in mobile SoCs?
MediaTek’s move makes sense if they target niche edge AI workloads where power efficiency trumps raw throughput, but 5B$ is a bet that could take years to pay off.
So MediaTek bets everything on custom AI chips but what happens if the dev ecosystem never migrates past CUDA? They’re walking a tightrope with 5B$ on the line.
Will they beat Nvidia’s perf/watt or just end up chasing the same custom silicon graveyard?
With 5B$ on the table, MediaTek might actually shake things up-but they’ll need killer software ecosystems to avoid getting locked out by CUDA’s dominance.
The 5B$ bet feels bold but makes sense-custom AI chips are where profits lie, not in generic mobile SoCs. Nvidia’s still ahead, but MediaTek’s timing could disrupt the game if they nail software support.
MediaTek’s AI chip plans sound aggressive. How will they compete against Nvidia and AMD in the AI hardware race?
TSMC, point de pincement du compute : revenus, prix, capacité