How NVIDIA, AMD, Intel and others are battling for the AI data center.

From GPUs to “AI Superchips”

A few years ago, AI workloads mostly ran on repurposed gaming GPUs and CPUs. Today, the flagship chips are purpose-built AI “superchips” designed from the ground up for massive model training and inference. NVIDIA’s Blackwell-architecture GPUs pack around 208 billion transistors and use a custom TSMC 4NP process, with two huge dies linked by a 10 TB/s interconnect. NVIDIA

At rack scale, the Grace Blackwell GB200 NVL72 system ties 72 Blackwell GPUs and 36 Grace CPUs into a single NVLink domain, effectively behaving like one exascale-sized accelerator for trillion-parameter models. NVIDIA Documentation+1

NVIDIA: Still the Benchmark to Beat

NVIDIA’s Blackwell platform is framed as the successor to Hopper, promising far higher performance per rack and aggressive energy savings, especially in its liquid-cooled GB200 NVL72 configuration. NVIDIA claims as much as 25× more performance at the same power compared with air-cooled H100 infrastructure, plus dramatic water and cooling cost reductions. NVIDIA+2NVIDIA Blog+2

Recent news that the first Blackwell wafers are now being manufactured in the United States at TSMC’s Arizona fab highlights how AI demand is reshaping not only the data center but also geopolitics and semiconductor supply chains. Reuters

AMD: Chiplets, APUs and the MI300 Gambit

AMD is betting on advanced packaging and chiplet design to claw share from NVIDIA. Its Instinct MI300 series accelerators combine CDNA3 GPU tiles with high-bandwidth memory, and the MI300A even fuses Zen 4 CPU cores and GPU compute into a single APU package for tightly coupled HPC and AI workloads. AMD+2AMD+2

While some analysts argue MI300 still lags NVIDIA’s best GPUs on AI benchmarks, large customers and OSATs are actively evaluating the platform for AI workloads, seeing gains in cost and energy efficiency when paired with AMD’s EPYC CPUs. Barron’s+1

Intel: Gaudi 3 and the Open Accelerator Pitch

Intel’s Gaudi line is its most serious shot at AI accelerators in years. The Gaudi 3 chip is marketed as delivering on average 50% better inference performance and 40% better power efficiency than NVIDIA’s H100, at a significantly lower list price. Intel Corporation

Gaudi 3 is being rolled out in major ecosystems like Dell’s AI Factory and IBM Cloud, where enterprise customers can rent clusters with an open software stack instead of being locked into CUDA. Newsroom+1

Cloud Titans and Custom Silicon

The hyperscalers aren’t sitting out this fight. Google’s Cloud TPU v5p offers more than 2× the FLOPS and 3× the high-bandwidth memory of the previous v4 generation, underpinning its “AI Hypercomputer” architecture. Google Cloud+2Google Cloud+2

Google recently followed up with its Ironwood TPU generation, reaching exaFLOPS-scale FP8 performance in pods that it claims can beat NVIDIA’s latest GB-class platforms on certain training and inference workloads. Tom’s Hardware

Closing Thoughts and Looking Forward

The AI chip market is no longer a simple “GPU vs CPU” story. It’s an ecosystem war across:

• Chip architecture and packaging (3D stacking, chiplets, HBM, advanced interconnects).

• System-level design (liquid-cooled racks, network fabrics, storage).

• Software and ecosystem lock-in (CUDA vs ROCm vs open ML stacks).

Expect the next few years to bring even sharper competition: NVIDIA defending its lead with Blackwell and beyond, AMD doubling down on chiplet innovation, Intel pushing cost-efficient accelerators, and cloud providers racing ahead with custom silicon that blurs the line between “chip vendor” and “cloud platform.” For buyers, the upside is more choice—but also more homework.

Reference Sites

1. “NVIDIA Blackwell Architecture” – NVIDIA – https://www.nvidia.com/en-us/data-center/technologies/blackwell-architecture/

2. “GB200 NVL72: The Blackwell Rack-Scale Architecture” – NVIDIA – https://www.nvidia.com/en-us/data-center/gb200-nvl72/

3. “AMD Instinct MI300 Series Accelerators” – AMD – https://www.amd.com/en/products/accelerators/instinct/mi300.html

4. “Intel Unleashes Enterprise AI with Gaudi 3” – Intel Newsroom – https://www.intc.com/news-events/press-releases/detail/1689/intel-unleashes-enterprise-ai-with-gaudi-3-ai-open-systems

5. “Introducing Cloud TPU v5p and AI Hypercomputer” – Google Cloud Blog – https://cloud.google.com/blog/products/ai-machine-learning/introducing-cloud-tpu-v5p-and-ai-hypercomputer

Author: Serge Boudreaux – AI Hardware Technologies, Montreal, Quebec
Co-Editor: Peter Jonathan Wilcheck – Miami, Florida