Low-cost AI layout, will GPU manufacturers hand over the baton?

1. Algorithm revolution: DeepSeek subverts the computing power arms race with a small amount of effort

• DeepSeek-V3: The total training cost is only$5.576 million, the performance is benchmarked against OpenAI o1, and the cost is that of the traditional solution1/10
• DeepSeek-R1: Incremental training cost$294,000, the cost of single-token inference has reached the industry benchmark.1/50

• MFU (model FLOP utilization): H800 cluster reaches23%, far exceeding the industry average of 15%
• Large-scale cluster efficiency: 2048 blocks of H800 achieve **98.7%** continuous utilization, 20 percentage points higher than the traditional solution
• Technical path: MoE sparse architecture, GRPO reinforcement learning, FP8 mixed precision, full-link software and hardware collaborative optimization

• Morgan Stanley: WillGB200Shipment forecasts for 2025 range from30,000 piecesdowngraded to20,000 pieces
• Stock price reaction: NVIDIA plummeted in a single day after the news was announced23.8%, the market value evaporated by nearly$900 billion

2. Reconstruction of the chip landscape: The rise of ASIC, GPU is no longer the only solution

1. Cloud vendors self-developed ASIC: cost reduction + control dual drive

• AWS（Trainium 2）: Designed by Marvell, with computing power between A100–H100, the core growth point of ASIC revenue in 2024
• Meta/Microsoft: Large-scale deployment of custom ASICs with lower inference power consumption than GPUs30–50%
• core logic: Dedicated ASIC inReasoning, recommendation, searchWait for the scene,Performance/watt ratio, cost, controllabilityOverall better than general-purpose GPUs

2. ASIC duopoly: Broadcom and Marvell make strong gains

• Market share:55–60%, customized AI chip market **70%** share
• ASIC revenue in 2024:>$8 billion, Google TPU contribution$7 billion
• Customers: Google (TPU V6/V7), Meta, OpenAI (10GW level project)

• Market share:13–15%, custom XPU/optical interconnect leader
• Clients: Amazon (Trainium), Google, Microsoft (Maya)
• Technology: Co-packaged optics (CPO), silicon photonics, AI server scale expanded from rack level to multi-rackMarvell Technology, Inc.

3. Competitive situation: GPU vs ASIC, from complementation to substitution

• training end：NVIDIAH100/H200/GB200Still dominant, butcost sensitive trainingOffloaded by efficient algorithm + mid- to low-end GPU
• Reasoning end：ASIC/NPU rapid penetration,Cloud vendors, Internet, terminal sidePrioritize customized solutions
• Trend: AI computing power moves from GPU dominance toGPU+ASIC+FPGA+NPUHeterogeneous coexistence

3. NVIDIA’s response: protect the ecology, launch new products, and build alliances

1. Technical defense: Blackwell/Rubin architecture + cost-effective solution

• GB200（Blackwell）: The computing power of a single chip is improved compared to H10030 times, energy efficiency25 times
• Rubin CPX: Low-cost GPU for inference, directly benchmarked against ASIC
• Strategy: High-end defense training, mid-to-low-end reasoning, software lock ecology

2. Ecological disruption: Recruit Marvell and build a semi-custom alliance

• March 2026: NVIDIA$2 billionStrategic Investment Marvell
• NVLink Fusion: Marvell’s customized XPU can be seamlessly connected to the NVIDIA ecosystemNVIDIA NVIDIA
• essence: Shift from pure GPU monopolyNVIDIA Dominance + Heterogeneous Compatibilitycomputing power grid

4. Outlook for 2025: Low-cost AI becomes mainstream, and the landscape is diversified.

1. Three major trends

• Algorithm first：MoE, sparse computing, efficient trainingbecome standard equipment,1/10 costAchieve equivalent performance
• Chip differentiation：
  • Training: NVIDIAGB200/GB300Dominate, but share from90%→70%
  • Reasoning:ASIC/NPURapid growth, accounting for **>30%** in 2025
• cost revolution: AI moves from giant games toInclusiveness, small and medium-sized enterprises and domestic models will fully benefit from

2. Winners and Challenges

• Biggest winner：Broadcom, Marvell(ASIC customization),Efficient algorithm teams such as DeepSeek
• Challenger: Nvidia (slowing growth, pressure on profit margins, ecological diversion)
• Good for the industry: AI deployment costDecrease 70–90%, terminal-side AI and industry large models are accelerating the explosion

Conclusion