Recorded talk

NVIDIA and Data Center Partners Review Highlights at COMPUTEX 2025

enplatform captions

Summary

Computex 2025 framed the modern computer as a unified, AI-driven data center — an "AI factory" — solving the problem of scaling inference and unstructured-data throughput across systems. Nvidia's approach centers on the Grace Blackwell platform and the new Grace Flagwell Ultra in full production, claiming 1.5x inference performance through dual-chip connectivity, TSMC POWSL packaging, and an NV link spine. The NV link spine provides 130 TB/s all-to-all GPU bandwidth across 72 Flagwell Ultra GPUs to present a single large virtual GPU, and Nvidia paired a purpose-built NVIDIA AI data platform plus AIQ to semantically query unstructured data. Availability signals focus on production deployment (MGX components from 60+ suppliers, RTX Pro servers, liquid cooling and high-power architectures), though transcript details on evaluation methodology and operational guardrails are limited and should be validated before integration.

Key technical points

Reframe: modern computing described as unified AI-driven data center ("AI factory") rather than standalone PCs/servers.
Core platform: Nvidia Grace Blackwell platform presented as foundational technology for this shift.
New product in production: Nvidia Grace Flagwell Ultra is now in full production and claims 1.5x inference performance versus prior generation.
Hardware engineering: Grace Flagwell Ultra incorporates dual-chip connectivity and uses TSMC's advanced POWSL packaging (as stated).
GPU fabric: NV link spine enables 130 terabytes/second of all-to-all GPU bandwidth across 72 Nvidia Flagwell Ultra GPUs to form a single virtual GPU.
Interoperability: Envink Fusion described as an open, flexible approach to Nilink that lets partners tightly integrate their own CPUs, AI accelerators with Nvidia Whiteall and the upcoming Ruben architecture for semi-custom infrastructure.
Ecosystem servers: Multiple partners (listed in talk: ASUS, as rockrecom, gigabyte, invents, egatron, QCT, super micro, we strong, and wewing) will offer RTX Pro servers aimed at enterprise AI factory workloads.
Enterprise AI platform: Nvidia announced an Enterprise AI platform intended to accelerate large IT investments and deployment at scale.
Storage + semantics: NVIDIA AI data platform is presented as a purpose-built storage system that can semantically query unstructured data with claimed unprecedented speed and accuracy.
AIQ pairing: The NVIDIA AI data platform is paired with Nvidia AIQ to address the bottleneck of extracting meaning from massive amounts of raw unstructured data.
MGX ecosystem: Keynote showcased MGX component availability from over 60 global suppliers, indicating supply-chain and component readiness.
MGX component types: MGX supports specific components and subsystems including Nvidia GB300, MBL72, liquid-cool systems, and high-power architectures, highlighting mechanical, thermal, and electrical solutions necessary for next-gen adoption.
Deployment signal: Multiple announcements emphasize production-readiness across silicon, systems, and data-center infrastructure, but the talk provides few technical details on evaluation metrics beyond the stated performance and bandwidth numbers.

Engineering relevance

Engineers should note three concrete infrastructure implications: (1) system architecture: Flagwell Ultra plus an NV link spine is intended to scale GPUs into a single large virtual GPU via 130 TB/s all-to-all bandwidth across 72 devices, which changes networking/fabric topologies and software placement assumptions; (2) packaging and connectivity: dual-chip connectivity and TSMC POWSL packaging are central to claimed 1.5x inference gains and will affect board-level and thermal designs; (3) data and storage: the NVIDIA AI data platform + AIQ targets the unstructured-data semantic-query bottleneck, so storage and data pipelines will need to integrate with that stack. Additionally, Envink Fusion/Nilink and the mention of Nvidia Whiteall/Ruben architectures indicate an intended semi-custom partner integration layer; validate compatibility and interoperability details with each partner. Finally, MGX component availability (GB300, MBL72, liquid cooling, high-power architectures) and dozens of suppliers imply production and deployment engineering workstreams (mechanical, thermal, electrical) are required to support next-gen nodes.

Japan relevance

For Japanese system integrators, OEMs, and data-center engineers, the announcements signal a near-term shift in procurement and design priorities: prepare for high-bandwidth GPU fabrics (NV link spine) and associated rack-level power and liquid-cooling needs called out by MGX components; assess compatibility and integration paths for Envink Fusion / Nilink with local CPU and accelerator offerings; plan storage and data-pipeline validation against the NVIDIA AI data platform + AIQ to meet semantic-query requirements for unstructured datasets; and engage MGX component suppliers now given the stated >60 supplier ecosystem and Grace Flagwell Ultra production readiness. Because the transcript provides limited evaluation and operational detail beyond headline metrics, Japanese teams should prioritize direct technical validation with Nvidia and announced partners before large-scale deployment.

Full transcript

Transcript

platform captions

[Music] Hello everyone, welcome to Computex 2025. I'm your host Ying Yin from Nvidia product management. It's a true pleasure to be here in Taiwan, a home for Nvidia for over three decades and the heart of so many of our invaluable partnerships. Our CEO Jensen Wong talked about many exciting technologies and let's get to it. Jensen began by redefining the modern computer not as a standalone PC or server but as a fullscale data center operating as a unified AIdriven computing system. These new data centers or AI factories transform energy into tokens of intelligence marking a fundamental shift in how computation powers innovation. At the heart of this transformation is Nvidia's Grace Blackwell platform. Now in full production, the new Nvidia Grace Flagwell Ultra delivers 1.5x the inference performance with groundbreaking engineering that includes dual chip connectivity, TSMC's advanced POWSL packaging and the NV link spine which enables 130 terabytes per second of all to all GPU bandwidth across 72 Nvidia flagwell ultra GPUs. effectively creating a massive single virtual GPU. Nvidia also introduced Envink Fusion, an open and flexible approach to Nilink that allows partners to tightly integrate their own CPUs, A6 and accelerators with Nvidia Whiteall and the upcoming Ruben architecture. This marks a leap forward for semi-custom AI infrastructure solutions tailored to diverse workloads and industries to further accelerate the enterprise shift to AI powered infrastructure. Nvidia's ecosystem partners spanning major players from ASUS as rockrecom gigabyte invents egatron QCT super micro we strong and wewing will offer RTX pro servers built to handle the demands of enterprise AI factory environments. One of the most exciting announcements was the Nvidia Enterprise AI platform accelerates trillion dollar global IT investment. NVIDIA AI data platform, a purpose-built storage system that can semantically query unstructured data with unprecedented speed and accuracy. Paired with Nvidia AIQ, this platform tackles one of AI's biggest bottlenecks, extracting meaning from massive amount of raw unstructured data. Finally, the keynote showcased the Nvidia MGX ecosystem component availability featuring data center infrastructure innovations from over 60 global suppliers. This includes MGX components support Nvidia GB300, MBL72, liquid cool systems, and high power architectures. All highlighting substantial mechanical, thermal, and electrical solutions, laying the groundwork of Nvidia's next generation product launches and widespread adoption. From silicon to systems to AI factory, we cannot make it without you. Thank you, Taiwan.