DeepSeek: Hype vs. Reality

DeepSeek's emergence as a formidable Chinese competitor in the AI landscape has sparked concern in Silicon Valley. However, Meta's chief AI scientist, Yann LeCun, insists the market's response is unwarranted.

Outperformance and Market Panic

DeepSeek's recent model surpassed those from established players like OpenAI and Meta on third-party benchmarks, using less expensive chips and reportedly with lower funding. This triggered a tech sell-off, wiping out $1 trillion in market cap.

Inference vs. Training

LeCun clarifies that the substantial investments in US AI companies primarily support inference, not training. Inference involves applying trained models to new data, which becomes increasingly demanding as AI capabilities expand. LeCun emphasizes that inference costs will rise as AI systems grow more sophisticated.

Cost Divergence

While DeepSeek's low training costs have raised alarm, industry experts like Positron's Thomas Sohmers anticipate that inference expenditures will dominate future AI infrastructure costs. However, startups are emerging to optimize inference output generation, potentially driving down costs for small-scale applications.

Scalability and Cost Implications

For large-scale models like DeepSeek V3 offering free access to a vast user base, inference costs are expected to be substantial. As DeepSeek's popularity grows, so will its inference workload and associated expenses.

Investment Ramp-Up

In response to DeepSeek's rise, tech giants like Meta and OpenAI are ramping up investments in AI infrastructure. Meta plans to spend $60 billion on capital expenditures by 2025, while OpenAI, Oracle, and SoftBank have partnered on a $500 billion joint venture called Stargate.