过去多年，Nvidia 一直是 AI 晶片的绝对龙头，市值已超过 4 兆美元；每一代新晶片都让企业在超大规模资料中心中以数百到数千颗处理器训练更大模型。然而，这种优势正受来自软体层的挑战。Wafer 以强化学习微调开源模型，让 Claude、GPT 等模型更会直接为特定矽晶圆撰写 kernel 程式码，并与 AMD、Amazon 合作优化效能；其种子融资已达 400 万美元，由 Jeff Dean 与 Wojciech Zaremba 等人士支持。这代表 AI 正在把「会为新矽芯片写程式」从稀缺人才资源，转为可扩充的技术能力。

另一个趋势是，Nvidia 的主导力不再只来自硬体原始效能，而来自可程式化生态。Andere 指出，最高等级的 AMD 晶片、Amazon Trainium 与 Google TPU 在理论浮点运算（flops）上可与 Nvidia GPU 对齐，因此瓶颈更可能是每瓦智慧（intelligence per watt）与编译最佳化效率，而非单点算力。同时，具有能够稳定高效优化各平台程式码的工程师仍稀缺且昂贵；像 Anthropic 在 Trainium 上合作时仍需自下而上重写程式码，显示软体可携性与优化效率仍是关键护城河，而这个护城河正被 AI 自动化工具快速侵蚀。

在晶片设计端，Ricursive（由 Azalia Mirhoseini 与 Anna Goldie 创立）将 AI 推向更上游：针对实体设计与设计验证两个核心环节做自动化，并加入大型语言模型，使工程师可用自然语言查询与修改设计。其目标是可「对话式」设计晶片，甚至让 AI 参与程式码与晶片的递回式共同进化；该公司已在数月内完成 3.35 亿美元融资、估值 40 亿美元，融资额约占估值的 8.375%。若可持续缩短设计周期、扩大「可调整能力」，整个产业的门槛就可能被重塑。

For years, Nvidia has remained the uncontested AI-chip leader with a market capitalization above US$4 trillion; each new generation lets firms train larger models using hundreds to thousands of processors in hyperscale data centers. That edge is now being challenged from the software layer. Wafer uses reinforcement learning on open-source models to make Claude, GPT, and similar systems write kernel code directly for specific silicon, and is working with AMD and Amazon to optimize performance. It has raised US$4 million in seed funding from figures including Jeff Dean and Wojciech Zaremba, signaling that AI is turning low-level chip software engineering from a scarce human craft into a scalable capability.

The competition is shifting from raw hardware power to programmability. Andere says top AMD chips, Amazon Trainium, and Google TPUs can match Nvidia GPUs on theoretical floating-point performance (flops), so the key battleground becomes performance per watt and optimization quality. Expertise in reliable, efficient cross-chip code optimization remains expensive and scarce, and even large firms face friction: Anthropic reportedly had to rewrite model code from scratch to run efficiently on Trainium. Nvidia’s historic software ecosystem has been a moat, but AI-assisted code-generation tools are making that moat increasingly porous.

The trend extends further upstream to chip design itself. Ricursive, founded by Azalia Mirhoseini and Anna Goldie, targets physical design and verification and integrates LLMs so engineers can describe changes or ask questions about designs in natural language. The firm claims progress toward automating more of the design flow, potentially enabling conversational chip creation and even recursive co-design where AI tunes both silicon and algorithms. In a few months, Ricursive raised US$335 million at a US$4 billion valuation (about 8.375% of value), suggesting investor belief that AI can compress design cycles and create a scaling law for faster, better chip development.