Nvidia Corp. (NVDA, Financials) may face new competition as DeepSeek, a Chinese artificial intelligence model, challenges the need for high-end AI chips, Fitch Ratings said.
Should DeepSeek's assertions of competitive performance with less expensive chips come true, the fast rate of AI infrastructure investment might slow down, therefore affecting revenue growth for artificial intelligence chipmakers.
DeepSeek claims that utilizing less sophisticated infrastructure, their AI model runs well, which raises questions about the sustainability of the around $50 billion in quarterly expenditures made by hyperscale cloud providers. Fitch expects a slowing down in artificial intelligence expenditure as businesses try to earn a profit, maybe resulting in a longer inventory correction.
Fitch said that a change toward cost-effective AI development will help the larger technological industry. Forecasting mid-single-digit sales growth, the company has a "neutral" 2025 view for the semiconductor sector. Although artificial intelligence infrastructure expenditure is still high, a near-term personal computer refresh cycle is also predicted to bolster demand.
The emergence of open-source AI substitutes might hasten acceptance and stimulate semiconductor industry innovation. Should the latter reach technical parity with top producers, this tendency may benefit firms manufacturing less complex semiconductors, including Advanced Micro Devices (AMD, Financials) and Intel (INTC, Financials). But if DeepSeek's efficiency promises hold true, Nvidia, which rules AI model training, may find itself in trouble.
Citing their important positions in specialized artificial intelligence silicon and networking for hyperscale cloud providers, Fitch recently improved ratings for Broadcom (AVGO, Financials) and Marvell (MRVL, Financials). With 40% from software and 25% from Apple Inc. (AAPL, Financials) products, Broadcom's revenue diversity helps to lessen its vulnerability to changes in AI infrastructure expenditure.
Though DeepSeek's influence raises questions, Fitch said that its efficacy in artificial intelligence training is yet unknown. The model seems best suited for inference tasks instead of training, in which case conventional artificial intelligence techniques might still have value. Furthermore limiting DeepSeek's acceptance in Western countries might be geopolitical and security issues.
Factoring in a slow down in AI infrastructure investments, Fitch's semiconductor view takes a more cautious profits trajectory than management guidance and market consensus.
