FirstMoveTUE 04 AUG 2026
TODAYWORLDMARKETSAIAI SKILLSSPORT
First Move  ·  AI Hardware & Infra  · 
The build-out — Tuesday morning, 04 August

AMD Unveils Instinct MI455X, Challenging Nvidia with 40 Petaflops AI Accelerator

The AI compute build-out intensifies as chipmakers push new architectures, while infrastructure developers explore novel solutions like floating data centers and office-to-edge conversions to meet surging power demands.

EDITION 2026-08-04 · EVERY CLAIM SOURCED · GROUNDED VIA SEARCH AT PUBLISH
ShareWhatsAppXLinkedIn

The story

AMD has fired a significant volley in the AI accelerator race, unveiling its Instinct MI455X GPU at the Advancing AI 2026 event. Designed for large-scale AI training and inference, the MI455X boasts impressive specifications, including 320 billion transistors, 432GB of HBM4 memory, and up to 40 petaflops of FP4 AI compute.

This new flagship, built on AMD's CDNA 5 architecture using TSMC's 2nm and 3nm processes, aims to power AMD's Helios rack-scale AI platform. Its memory capacity notably surpasses Nvidia's Rubin architecture by 50%, with 23.3 TB/s memory bandwidth.

The MI455X is slated to begin shipping by the end of Q3 2026, with volume ramping into 2027. The launch underscores the escalating competition in high-performance AI silicon and the relentless pursuit of greater compute capacity.

Silicon

Instinct MI455X

Maker: AMD

What: Flagship AI accelerator with 40 PFLOPS FP4, 432GB HBM4, 23.3 TB/s bandwidth, built on CDNA 5 (TSMC 2nm/3nm)

For Whom: Large-scale AI training and inference, powering AMD's Helios rack-scale AI platform

The build-out

ProjectWhoScaleWhere
Floating Data Center Engineering ContractSamsung Heavy Industries (SHI) and Mousterian CorporationEach unit 50 MW critical IT capacity, initial deployments planned for Texas and other US marketsTexas and other US markets
Edge AI Data Centers using spare office capacityPerimeter Compute (backed by Montauk Capital)0.5 MW to 20 MW per commercial buildingMajor U.S. cities (e.g., New York, Boston)
EU AI Gigafactories ProgramEuropean Union (with backing from AMD, NVIDIA, Qualcomm)Up to seven Gigafactories, €30 billion total public/private investmentAcross Europe
West Texas AI Data CenterPowerPlay AI (joint venture)400 MW initial power capacity by 2028West Texas

Supply & policy signals

ASML pushing 10% price hikes on lithography tools, including DUV systems

Implication: Increased manufacturing costs for chipmakers like TSMC, potentially impacting chip pricing and supply chain stability

SIA-Deloitte report states semiconductors comprise over 95% of the value in a leading AI server rack

Implication: Reinforces the concentration of profit and pricing power with chip designers in the AI build-out, driving hyperscalers towards custom silicon programs

Meta and Google striking deals with utilities to protect residential customers from AI data center power costs

Implication: Acknowledges growing concerns over AI infrastructure's strain on power grids and aims to shift infrastructure investment responsibility to tech giants

Onsemi raises Q3 revenue forecast, expecting AI data center revenue to more than double in 2026

Implication: Strong demand for AI data center power management chips indicates continued momentum in AI infrastructure spending

What we'll be watching

Reporting + analyst voices: grounded via Google Search at publish time.