Les sources de cet événement
5-
Entreprise
Anthropic
The safety-focused lab behind Claude, running on Trainium, TPUs, NVIDIA and AMD GPUs at once.
Founded in 2021 by former OpenAI researchers led by Dario and Daniela Amodei, Anthropic builds the Claude model family alongside a safety research agenda (constitutional AI, interpretability), and Claude Code made it a leader in enterprise coding. Company-stated run-rate revenue rose from about $9B at the end of 2025 to more than $47B in May 2026, when a…
Ouvrir la fiche → -
Entreprise
Google (Alphabet)
The only company vertically integrated from TPU silicon to frontier models.
Alphabet owns the whole stack: TPUs designed in-house, the Jupiter network and global data centers, Google Cloud, DeepMind's Gemini models, and distribution through Search, Android and Workspace. In April 2026 it split its eighth TPU generation into a training chip (TPU 8t) and an inference chip (TPU 8i), and Anthropic agreed to multiple gigawatts of TPU…
Ouvrir la fiche → -
Entreprise
Broadcom
The quiet giant of custom AI silicon and the switch chips inside every fabric.
Broadcom co-designs the hyperscalers' in-house accelerators, Google's TPU line most famously, with Meta and others as disclosed custom-ASIC customers and an OpenAI partnership announced in 2025, while its Tomahawk and Jericho switch silicon underpins AI Ethernet fabrics everywhere. Its AI semiconductor revenue guidance kept re-rating upward through 2025…
Ouvrir la fiche → -
Entreprise
Amazon Web Services
The largest cloud, building custom silicon and gigawatt campuses for Anthropic and OpenAI.
AWS remains the biggest cloud business and answered the AI wave with vertical integration: Trainium chips from Annapurna Labs (Trainium3 generally available since December 2025), the Bedrock model platform, up to 1,920 MW of Susquehanna nuclear power under its June 2025 Talen agreement, and Project Rainier, the Indiana campus training Anthropic's models. In…
Ouvrir la fiche → -
Puce
AWS Trainium3
AWS's first 3 nm AI chip: 2.52 FP8 PFLOPS and 144 GB of HBM3E, in 144-chip UltraServers.
Trainium3 became generally available in Trn3 UltraServers at re:Invent on December 2, 2025. Each chip carries 144 GB of HBM3E at 4.9 TB/s and 2.52 PFLOPS of FP8, and up to 144 chips share one NeuronSwitch-v1 fabric. AWS cites up to 4.4x the compute and 4x the energy efficiency of Trn2 UltraServers. Amazon Bedrock serves production traffic on it, and…
Ouvrir la fiche →


