The chip war heats up! NVIDIA GPU dominance is surrounded by Google and Amazon

Nvidia GPU shipped 6 million units last year, with a single Blackwell server priced up to 3 million USD and weekly shipments of 1,000 units, dominating the AI chip market. However, this chip war is shifting as custom chips like Google TPU, AWS Tranium, and Broadcom ASIC rise. Analysts estimate Broadcom’s market share in the custom ASIC chip market could reach 70-80%.

GPU’s Golden Decade from Gaming Card to AI Core

Nvidia GPUs evolving from gaming cards to AI core chips can be traced back to 2012’s AlexNet. The research team first applied Nvidia GPU’s parallel computing power to neural network training, significantly leading in image recognition competitions, ushering in the era of deep learning. The core advantage of GPU chips comes from thousands of parallel processing cores, efficiently executing matrix multiplications and tensor operations, making them ideal for AI training and inference.

In large server racks, 72 GPU chips can be combined into a single computing unit via NVLink technology, resembling a giant GPU. Nvidia not only supplies GPUs to OpenAI, governments, and enterprises but also directly builds complete server systems. Competitor AMD relies on Instinct GPUs and open-source software ecosystems to accelerate progress, recently supported by OpenAI and Oracle. AMD GPUs primarily use open-source software, while Nvidia GPUs are optimized around CUDA.

Custom ASIC Chips Become the Key to Breaking the Deadlock for Cloud Giants

From Google, Amazon, Meta, Microsoft to OpenAI, major cloud giants are investing in R&D of custom ASIC (Application-Specific Integrated Circuit) chips. These purpose-built chips are expected to become the fastest-growing category for AI chips in the coming years. As large language models mature, inference demands are rapidly surpassing training. The costs, energy consumption, and stability of inference become pain points for cloud platforms, and this is where ASIC chips dominate.

Unlike versatile GPUs, ASICs are like “specialized ultra-precise tools,” optimized in hardware for specific AI workloads, resulting in faster speeds and lower power consumption. The downside is less flexibility and extremely high development costs, often hundreds of millions of dollars per chip, affordable only for cloud giants. Custom AI ASIC chips are extremely expensive, costing at least hundreds of millions USD, but for large cloud service providers, they offer higher efficiency and reduce dependence on Nvidia.

Google was the first major player in ASIC chips, pioneering custom AI accelerators, and in 2015 coined the term Tensor Processing Unit (TPU). TPU chips helped Google invent the Transformer architecture in 2017, which underpins AI models like ChatGPT and Claude. Today, Google has developed to the 7th generation TPU Ironwood chips, training models like Claude using millions of TPUs.

AWS, after acquiring Annapurna Labs, has invested heavily in developing its own AI chips. Tranium and Inferentia chips have become vital components of AWS’s training and inference platforms. In 2024, for example, Anthropic trained models at AWS North Indiana data center using 500,000 Tranium 2 chips, entirely without Nvidia GPUs, highlighting the rising prominence of ASICs.

Chip Diversification Trends and Cost-Effectiveness Trade-offs

Broadcom and Marvell, as chip foundry design companies, are core strategic partners of mega cloud companies. Broadcom is deeply involved in building Google TPU and Meta AI inference and training chips. Analysts project Broadcom’s market share in custom ASIC chips could reach 70% to 80%.

Edge AI chips are also extending to personal devices. NPU (Neural Processing Unit), designed for running edge AI on devices, are now integrated into Qualcomm Snapdragon, AMD, Intel, and Apple M-series SoCs, used in smartphones, laptops, smart homes, cars, and robots. Edge AI devices will bring higher privacy, lower latency, and greater control.

Comparison of the Three Main Chip Categories

GPU Chips: Highly versatile, suitable for various workloads, but high power consumption and costly, with a single rack costing around 3 million USD.

ASIC Chips: Highly specialized, faster, lower power, with development costs reaching hundreds of millions USD but offering long-term cost-effectiveness of 30-40%.

FPGA/NPU Chips: Reconfigurable, positioned between the two, suitable for edge devices and testing phases.

TSMC Controls the Chip Supply Chain

Whether it’s Nvidia’s Blackwell chips, Google TPU chips, or AWS Tranium chips, most AI chips are ultimately manufactured by TSMC. This tightly links AI computing power supply with global geopolitical considerations. The US is attempting to bring some chip manufacturing back onshore through TSMC’s Arizona plant and Intel’s 18A process. Meanwhile, Chinese companies like Huawei and Alibaba are actively developing their own ASIC chips to seek domestic alternatives under export restrictions.