After reading this SpaceX S-1 excerpt from Simon Willison, the real signal to watch is in llm/tools: Grok 5 is training on COLOSSUS II, and COLOSSUS/COLOSSUS II can also sell computing power.
The Anthropic contract is directly written to US$1.25 billion per month, lasting until May 2029, and can be terminated 90 days in advance.
This is indeed a good business.
Training top models can also accumulate rare and ultra-large-scale AI Infra experience👍 There are two points here that are most worthy of disassembly for the engineering team: First, whether the capability boundaries are clearly written, input prerequisites, tool selection, call links, permission boundaries, resource quotas, and failure conditions all determine whether a method can be moved from demo to production; Second, verify whether the closed loop is strong enough.
The indicator caliber, evaluation set, grayscale strategy, rollback path, cost attribution, and observability will determine whether it can be stable after going online.
llm/tools This line has moved from "what does the model do" to "how does the organization link computing power, tools, processes, and responsibilities into a production system?" What will really widen the gap in the future may be the engineering capabilities that put uncertainty in a cage and deliver certainty to the production line.