China presses its case at WAIC 2026

The center of gravity this week was Shanghai. At the World Artificial Intelligence Conference, President Xi Jinping laid out a strategy built around open, accessible models, and China used the stage to pitch itself as the architect of a new global AI order. The conference, which opened Friday, July 17, doubled as a venue for unveiling new models and for advancing shared rules of the road: drafting is underway on a set of Global Guidelines on AI, part of a broader push toward international governance frameworks backed by both governments and industry groups.

That diplomatic ambition is backed by hard usage numbers. China's National Data Administration reports that the country's daily AI 'token' consumption has exploded, a sign that open models are scaling toward the frontier and being adopted at startling speed.

A benchmark that rattled assumptions

The headline that traveled furthest came from Chinese startup Moonshot AI. Its new Kimi K3 model reportedly writes GPU code roughly 14.82x faster than PyTorch — a coding feat drawing global attention and challenging comfortable assumptions about America's lead in AI. It lands alongside a fresh wave of large language models trading blows on hard math problems, results now being scrutinized by mathematicians and AI watchers alike. The same episode is prompting a sharper question: as models saturate these tests, how much can any single benchmark still tell us?

Washington tightens its grip

Against that backdrop, U.S. policy is shifting. The Trump administration has moved away from the 'light-touch' posture it started with toward a far more interventionist stance — one that has already produced restrictions on top AI models. It's a notable reversal, and a reminder that the contest between the U.S. and China is now playing out as much through policy as through parameters.

Compute becomes the constraint

Inside the leading labs, the binding limit is increasingly raw computing power. OpenAI underscored the point by promoting Uday Ruddaraju to a newly prominent role: CTO of Compute, overseeing the infrastructure behind its models. Anthropic, meanwhile, is rationing access to Claude Fable 5, its most in-demand frontier model, restricting it to its priciest plans as compute runs short. When the most capable systems are gated by hardware rather than ideas, who gets to use them becomes a business decision.

Safety, families, and a warning about a generation

OpenAI also moved on teen safety, adding a feature that notifies linked parents when a teen is booted from ChatGPT — an incremental but telling step as the company navigates its youngest users. The human stakes were voiced bluntly by novelist Dave Eggers, who took his message directly to OpenAI staff, telling them ChatGPT is 'silencing an entire generation' by eroding young people's ability to think and write for themselves.

Building the pipeline

Finally, a quieter marker of how thoroughly AI is settling into the mainstream: Morgan State University in Baltimore will launch a Bachelor of Science in Artificial Intelligence this fall, joining a growing roster of Maryland schools formalizing AI as a degree. The frontier grabs the headlines, but the workforce that will run these systems is being trained now.

The takeaway

Today's throughline is competition on every front at once — models, benchmarks, policy, compute, and talent. China is betting openness and scale can close the gap; Washington is answering with restrictions; and the labs themselves are quietly discovering that the next bottleneck may be electricity and silicon, not algorithms.