Free Model Weights, Yet No Free Intelligence: The AI Compute Debate

The discussion began with a tweet quoting Elon Musk, who distinguished between benchmark performance and practical usefulness, arguing that Anthropic focuses on delivering maximal practical intelligence—something invisible on leaderboards but evident in revenue figures.

Carol Lin of the GLM team added that in the AGI era, benchmarks measure capability while revenue reflects adoption, yet neither captures the ultimate goal: how many people’s potential can be unlocked and how many lives improved.

Jina AI founder Han Xiao then shifted the conversation to a concrete angle. He praised the open‑source camp, noting that among freely available weights, ZhiPu (Zai) and DeepSeek stand out for their research quality, but warned that "weights do not equal useful intelligence."

His reasoning: useful intelligence requires scalable inference compute—essentially near‑infinite reasoning power—which the United States currently dominates over China. He illustrated this with his own experience that morning: after downloading GLM5.2, he could not run it effectively because his local NVMe SSD was too slow for cold‑loading, and although his M3 Ultra GPU had sufficient memory, its pre‑fill speed, generation speed, context length, and parallel slots fell far short of data‑center GPU clusters. These limitations cripple the model’s test‑time intelligence and hinder his daily Agent tasks.

The conclusion he drew was straightforward: possessing the weights lets one run GLM5.2 for a few small HTML demos, but this is not comparable to the practical intelligence obtained by calling Anthropic’s API. In other words, if China had unlimited inference compute, GLM5.2 could deliver intelligence as useful as Anthropic’s today.

Critics responded that the value of open‑source weights lies not in personal deployment but in enabling other service providers to host and compete, ultimately delivering the model to end‑users at the lowest possible price.

Han Xiao acknowledged this point and offered an engineering solution: since the U.S. has abundant GPUs while China lacks them, American companies could host Chinese models, thereby unleashing GLM5.2’s full inference power.

However, he highlighted a fatal flaw: once the model runs this way, all revenue flows to U.S. inference providers, leaving ZhiPu with nothing beyond social media likes. Over the long term, such a model is unsustainable.

The debate remains unresolved, but it brings to light a fact often obscured by leaderboard scores: open‑source lowers the barrier to accessing model weights, yet the true determinant of a model’s usefulness is the expensive row of GPUs behind the scenes.

Infrastructure, infrastructure, and more infrastructure—Microsoft CEO Satya Nadella’s recent "Token Capital" theory underscores this reality.