🧠 Food for thought

Implications, context, and why it matters.

Why lower inference prices change the economics for AI agents

• AI agents such as OpenClaw can burn through huge token volumes during multi-step work, which can leave users with surprisingly high bills 1.
• DeepSeek prices V4 Flash at $0.28 per million output tokens, while OpenAI charges $30 per million output tokens for GPT-5.5 2.
• The lower price comes from efficiency gains in DeepSeek V4 and from Huawei Ascend chips, which DeepSeek said it used to train V4 3.
• Cheap access plus open weights under an open-source license could make advanced AI feel more like shared infrastructure, opening room for applications that need tight cost control 2.

A new AI hardware ecosystem emerges alongside Nvidia

• DeepSeek said it trained V4 on Huawei Ascend processors. Huawei said Ascend would offer “full support” for DeepSeek’s models, which suggests another path for heavy AI workloads beyond Nvidia’s CUDA software platform 3.
• U.S. export controls aimed to slow China’s AI progress, yet they may have pushed Chinese firms to build leaner models. DeepSeek’s work on Huawei chips marks one move away from U.S. suppliers such as Nvidia and AMD 3.
• The market is starting to split between pricey integrated providers such as OpenAI and lower-cost open infrastructure groups such as DeepSeek 2.
• That fading middle layer may lead software developers to route work by task, saving expensive models for planning while cheaper ones handle volume 2.