The new AI bottleneck isn't GPUs. It's the data production loop.

The bottleneck in AI has moved upstream of the GPUs. Human question-writers cannot draft practice problems fast enough to keep training data ahead of the model, so the data production system itself becomes the optimization target. The race is no longer who trains the largest network. It is who can run a tight loop in which a data system produces harder, verifiable tasks and the model consumes them, with both sides improving together.

QbitAI's report on the Endless Frontier trio at Shanghai Jiao Tong, DP Technology, and the Shanghai Algorithm Innovation Institute shows the shift. BigBang-V1 is a 35B model trained entirely on AI-synthesized, verifiable tasks. QbitAI reports the model at 76.5 on BrowseComp and 54.2 on SWE-Bench Pro, with capability transfer to long-horizon search and software engineering. The 35B-versus-1T framing grabs the headline, but parameter count is the least interesting number on the page.

The mechanism is a three-agent loop. A Generator proposes synthesis strategies. A Critic rejects problems that fail basic constraints or that the model already solves. A Meta-Critic re-tunes the Generator based on what the trained model can and cannot do. The substrate holds the loop together: tasks must be hard enough to teach frontier behavior and verifiable enough to score, which is also where the trick fails. Creative, non-verifiable domains, where no clean answer key exists, are where the loop degrades into reward hacking.

The wire will frame the next six months of model releases as a size war. The honest frame is a substrate war: who can build the cleanest verifier and the tightest co-evolutionary loop. Labs that treat training data as a static corpus will keep paying for parameter count they do not need.