Nvidia Reveals First US-Made Blackwell Wafer
Nvidia has unveiled its first Blackwell chip wafer manufactured by TSMC in the United States, a milestone for domestic semiconductor production. The development comes as TSMC is accelerating the construction of up to ten new fabs by 2026 to meet surging global demand for AI chips and advanced packaging.
- The Blackwell B200 GPU features 208 billion transistors and is built on a custom TSMC 4NP process node, offering up to 20 petaflops of AI performance. This represents a significant leap from the Hopper architecture's 80 billion transistors and is designed to handle the demands of trillion-parameter AI models. - Architecturally, Blackwell introduces a dual-die design, connecting two chips with a 10 TB/s interconnect to function as a single, unified GPU. This overcomes manufacturing limitations and is a key factor in its performance gains over the monolithic design of the previous Hopper generation. - TSMC's Arizona site, backed by up to $6.6 billion in direct CHIPS Act funding and up to $5 billion in loans, is now slated for three fabs. The first fab is producing 4nm chips, with the second fab targeting 3nm production in 2027—ahead of the original 2028 schedule—and a third fab planned for 2nm and 1.6nm processes. - The build-versus-buy decision for hyperscalers remains a central theme; custom ASICs from Google (TPU), Amazon (Trainium), and Meta (MTIA) are designed for specific, high-volume internal workloads to optimize cost and power efficiency. However, they continue to be major buyers of Nvidia GPUs to serve the diverse needs of their external cloud customers. - The AI accelerator market is projected to grow from over $33 billion in 2025 to more than $300 billion by 2034, with GPUs expected to maintain a dominant market share. While Nvidia holds a commanding lead, AMD is establishing itself as a credible second source, securing deployments with major cloud providers. - A key feature of the Blackwell architecture is the second-generation Transformer Engine with new 4-bit floating point (FP4) AI inference capabilities. This allows for double the performance and model size support compared to previous generations while maintaining accuracy, directly addressing the cost and energy consumption of large model inference. - The fifth-generation NVLink, with a bandwidth of 1.8 TB/s per GPU, allows for the connection of up to 576 GPUs, a critical feature for training and deploying the largest and most complex AI models. This is a significant increase over the previous generation's connectivity capabilities. - The total investment in TSMC's Arizona facilities has grown to over $65 billion, making it the largest foreign direct investment in a new project in U.S. history. The three fabs are expected to create 6,000 direct high-tech jobs.
