The Personalized AI Model Revolution | Generated by AI

Home 2026.05

Question: I see a very clear future where every company and individual has their own large model that is updated monthly, communicates with external APIs, exchanges information, and provides value. The cost of training large models will definitely decrease—DeepSeek v2 is about five million (USD or RMB?), and in the future it may drop to only a tenth. What about Chinese chips, AMD, and ASIC custom chips?

Answer:

Yes, your vision is quite plausible. Here’s a breakdown of the main points:

  1. Individual & Company LLMs

    • In the near future, having personal or enterprise-specific large models (LLMs) that are continuously updated and can interface with APIs is realistic.
    • These models could act as personal assistants, knowledge bases, or automation agents. The key trend is “modular LLM ecosystems” where models exchange information safely, like you described.

  2. Cost Trends

    • Training costs for large models are dropping steadily due to:

      • Algorithmic improvements (sparse models, LoRA, QLoRA, etc.)
      • More efficient hardware
      • Better software stack and optimization

    • If DeepSeek v2 cost around $5M (USD) for training today, future iterations could indeed drop to a fraction—maybe $500k—thanks to these improvements.

  3. Hardware Options

    • Domestic Chinese chips: Companies like Biren, Cambricon, and Hygon are developing AI-focused accelerators. Performance may approach NVIDIA’s high-end chips in a few years, especially for inference and fine-tuning rather than full-scale pretraining.
    • AMD GPUs: With MI300 and next-gen RDNA architectures, AMD is becoming more competitive, particularly for mixed workloads. They are less dominant than NVIDIA in AI training but still viable.

    • ASIC / custom chips:

      • Custom AI ASICs (like Google TPU, Cerebras, or Graphcore) drastically reduce per-token training cost.
      • In China, several startups and state-backed companies are pushing ASICs for inference-heavy workloads, which aligns perfectly with your vision of monthly-updated models rather than constant full-scale retraining.

  4. Economic Implications

    • Once costs drop to a fraction of current levels, it will be feasible for small companies and even individuals to maintain specialized models.
    • API-based knowledge exchange between models could create a decentralized “AI economy,” where value is exchanged efficiently between models without manual intervention.

So yes, your scenario is not only possible, it’s likely a few years away given the convergence of cheaper training, improved algorithms, and specialized hardware. The only limiting factor might be data access and regulatory constraints, especially for private or corporate models.

References:

If you want, I can make a timeline projection of when personal and corporate LLMs become cheap enough for everyone to have one, factoring in hardware and training costs.

Back Donate