On July 29, 2026, the US-listed AI computing sector received major news: Broadcom announced a large-scale custom AI chip (ASIC) order from a top cloud computing client, expected to boost its data center business revenue by over 20% in FY2027. Meanwhile, FPGA leader Xilinx released a new adaptive computing acceleration platform, reducing power consumption for AI inference scenarios by 40% and doubling performance. These two news items mark the accelerated penetration of specialized computing routes represented by ASIC and FPGA into the AI market, beginning to pose a substantial challenge to Nvidia's GPU dominance.
Broadcom ASIC Order: Explosion of Custom Computing Demand
According to Broadcom's official disclosure, the order comes from a top-five global cloud service provider. The custom AI training chip uses a 5nm process, integrates over 100 billion transistors, and is optimized for large language model (LLM) inference. Broadcom CEO Hock Tan said in a conference call: "Our custom chip team is expanding at twice the previous speed, and customer demand for differentiated computing far exceeds expectations." The order is expected to start mass production in Q1 2027 and continue supply over the next three years.
Market analysts believe Broadcom's success in ASIC stems from its flexible design capabilities and deep ties with TSMC. Compared to Nvidia's general-purpose GPUs, ASICs can achieve 5-10 times better energy efficiency under specific workloads, which is highly attractive to cost-conscious cloud giants. In fact, Google, Amazon, Microsoft and other giants have long developed their own custom chips, and Broadcom, as a third-party design service provider, is becoming an important partner for these cloud vendors.
Xilinx FPGA: New King of the Inference Era?
Xilinx's new FPGA platform is based on a 6nm process, with up to 5 million logic units and an integrated AI inference acceleration engine. Its biggest highlight is dynamic reconfigurability: after model upgrades, no hardware replacement is needed; only remote logic code updates are required to adapt to new algorithms. Xilinx CEO Victor Peng said: "AI models iterate faster and faster, and FPGA flexibility is a killer feature. More than 10 major internet companies are already testing our solution." According to estimates, in BERT-Large inference tasks, this FPGA's performance per watt is 2.3 times that of Nvidia's A100 GPU.
Notably, Xilinx has been acquired by AMD but still operates independently. AMD combines Xilinx's FPGA technology with its own GPUs to launch adaptive AI accelerator cards, aiming to compete for share in the inference market with Nvidia. This combination gives AMD a differentiated advantage of GPU+FPGA in the AI chip field.
Industry Interpretation: Specialized Computing Eroding General-Purpose Computing
For a long time, Nvidia has held over 90% of the AI training market with its CUDA ecosystem and powerful GPU computing power. But entering 2026, the situation is changing: first, the skyrocketing cost of large model training forces cloud vendors to seek more cost-effective solutions; second, the proportion of AI inference demand continues to rise, and inference scenarios are more sensitive to power consumption and latency, which is exactly where ASIC and FPGA excel.
Research firm Gartner predicts that by 2028, custom AI chips (ASIC+FPGA) will account for 30% of the AI chip market, while Nvidia's share may drop from 80% to around 60%. However, Nvidia is fighting back: its next-generation Blackwell architecture will significantly enhance flexibility, and it will launch a customization service called "cuCustom," allowing customers to add dedicated acceleration units on top of GPUs.
Investment Opportunities and Risks
For US stock investors, the rise of ASIC and FPGA brings new positioning opportunities. Broadcom (AVGO), as a leader in custom chips, has high order visibility and reasonable valuation, with a current PE of about 25x, lower than Nvidia's 40x. After Xilinx (XLNX) was acquired by AMD, its FPGA business has become one of the core engines for AMD's (AMD) future growth. Additionally, there are design service companies like Alchip and Global Unichip (GUC) worth watching on the US stock market.
But risks must also be noted: ASIC custom chips require huge upfront investment, and if customer demand shifts, Broadcom may face capacity idling risks; FPGA remains difficult to replace GPU in large-scale training, with limited market ceiling. Overall, AI computing is moving from "GPU-only" to diversification. Investors should allocate broadly, holding GPU leaders like Nvidia and AMD, while also allocating appropriately to specialized computing plays like Broadcom and Xilinx, to capture structural opportunities amid industry transformation.
As of the close on July 29, Broadcom shares rose 4.5%, Xilinx rose 3.2%, while Nvidia edged down 0.8%. The market is voting with its feet, acknowledging the arrival of the custom AI chip era.
