GLM-5.3 shows how agents break CUDA lock-in for model inference

The agent that built its own infrastructure

According to Teortaxes, a researcher close to DeepSeek labs, GLM-5.3 was used to build and optimize the inference stack that serves the distilled GLM-5.3-Flash. The system went from first successful run to production-ready in under two weeks, with end-to-end throughput tripling against the initial baseline. The source calls this the first expression of what he terms TCD — the ability of agents to write and port kernels automatically to any available silicon.

Three weeks to pilot, two weeks to production

The striking figure isn't model size but iteration velocity. A process that used to take engineering teams months of hand-written CUDA has collapsed to a window of days. The agent identified bottlenecks, wrote architecture-specific variants, and benchmarked them without human intervention beyond the goal definition. "The agents are too good now," Teortaxes says, and every piece of silicon will be pushed to its limit.

The real wall is software, not hardware

The core claim: Nvidia's moat was never CUDA the language, but the software-stack lock-in wrapped around it. As long as efficient kernels required rare human expertise, "just buy Nvidia" was the rational default. Once agents can translate and port kernels automatically to any architecture — AMD, Intel, Chinese chips, dedicated accelerators — that default collapses fast.

What this means for the chip market

The scenario described here isn't incremental improvement but structural change. Switching cost between hardware platforms drops to zero when the grunt work of kernel optimization moves to the model. Chip makers that suffered from a lack of mature software support suddenly get a fast path to production; customers locked into Nvidia gain real negotiating leverage. The test is whether this capability graduates from internal demo to a product available to every developer — and then the market changes from the ground up.