The Blackwell supply chain recovered after a design defect in the processor die connecting the two Blackwell GPUs forced TSMC to stop mass production to address yield issues. This flaw, discovered late in the production process, reduced yields and delayed large-scale shipments until the first quarter of 2027. The defect specifically affected the processor die connecting the Grace CPU and Blackwell GPU. While this defect caused immediate pain, TSMC Vice President Jun He stated at the August 11 OCP APAC Summit that 5.5x reticle-size products now achieve stable yields above 98%, with some reaching 99%. This 5.5x reticle-size technology covers approximately 4,565 square millimeters. This defect caused NVIDIA’s stock to fall 23% from its recent high. I find that the supply bottleneck has moved from internal packaging defects to external component shortages. Specifically, the high demand for large, multi-layer organic substrates puts pressure on Ajinomoto and Ibiden. Ajinomoto plans to expand its Gunma plant capacity through 2030, but Ibiden’s new equipment for mass production only begins incrementally in fiscal 2027. This shortage mirrors broader industry concerns, as Intel warned that DRAM, NAND, and substrate shortages could limit revenue opportunities throughout the year.
The GB300 NVL72 provides the most effective scale for AI inference, particularly as agentic AI queries grew from 11% to 50% of all AI queries this year. Microsoft Azure’s deployment of these clusters uses 72 Blackwell Ultra GPUs and 36 Grace CPUs per rack. The GB300 delivers 50x higher throughput per megawatt and 35x lower cost per token than the Hopper generation. You likely already know that Blackwell handles trillion-parameter models, but the GB300 specifically improves performance for reasoning and multimodal tasks. For long-context scenarios involving 128,000-token inputs, the GB300 delivers 1.5x lower cost per token compared to the GB200. The system provides 130TB/s of NVLink bandwidth and 37TB of fast memory. It uses 800Gb/s per GPU cross-rack scale-out bandwidth via the next-generation NVIDIA Quantum-X800 InfiniBand. This hardware achieves 1,440 PFLOPS of FP4 performance. The performance improvement over the H100 reaches 2.45x in theoretical peak metrics.
| Parameter | GB300 NVL72 Specification |
|---|---|
| GPU Count | 72 Blackwell Ultra GPUs |
| CPU Count | 36 NVIDIA Grace CPUs |
| Memory Capacity | 37 TB fast memory |
| NVLink Bandwidth | 130 TB/s intra-rack |
| Cross-rack Bandwidth | 800 Gb/s via Quantum-X800 |
| Compute Performance | 1,440 PFLOPS FP4 |
NVIDIA plans to secure 595,000 CoWoS wafers by 2026, which covers 60% of the 1 million unit global demand forecasted for 2026. This captures 510,000 wafers specifically for CoWoS-L technology. Other players also fight for the remaining capacity. Broadcom plans to acquire 150,000 CoWoS wafers to serve custom chips, including 90,000 for Google’s TPUs. AMD targets 105,000 wafers. 80,000 of these come from TSMC and 25,000 from ASE/SPIL for the Venice CPU. Intel attempts to challenge the TSMC monopoly using its EMIB technology, which embeds bridges in organic substrates rather than RDL interposers on 300mm wafers. Intel’s EMIB can deliver 8x reticle size today and targets 12x by 2028. The CoWoS-S variant uses a single monolithic silicon interposer, but its size limit stays at 3.3x reticle size. However, the physical properties of large organic substrates cause warpage and thermal expansion issues that challenge high-yield production. Even as packaging yields hit 99%, the industry must reconcile the scarcity of HBM and the slow expansion of substrate manufacturing. Does the massive capital investment from TSMC and Intel guarantee a solution to the growing physical constraints of large-scale AI packaging?
