Z.ai GLM-5.2 matches Anthropic at 75% lower cost, Jefferies says

A Chinese AI model now matches Anthropic's corporate performance at a quarter of the cost, threatening the pricing power of Western AI developers ahead of their planned listings.

Z.ai's GLM-5.2 model delivers corporate-grade AI performance nearly matching Anthropic's Claude at 75% lower cost per token, Jefferies strategist Christopher Wood said, calling it a "DeepSeek moment" that threatens the pricing power of Western AI incumbents.

"The GLM-5.2 model from Hong Kong-listed Z.ai is almost a match for Anthropic in the corporate market at a quarter of the cost per token," Wood, author of Jefferies' widely read Greed & Fear note, said.

Anthropic's annualized run-rate revenue surged to $47 billion in May from $9 billion at the end of 2025, growth Wood expects to slow as companies push back against heavy token consumption. Chinese AI systems processed 21.37 trillion tokens on the OpenRouter aggregator platform in the week to June 21, up from 4.37 trillion in late April, compared with 5.76 trillion for leading US models. The top Chinese systems now account for nearly 80% of the platform's total token volume, up from about 43% two months earlier.

The challenge lands as Anthropic prepares for a planned stock market listing and rival OpenAI also weighs going public. Cheaper Chinese models are already gaining share, reinforcing a view that large language models will become commoditized, Wood said, while giving companies an incentive to move smaller models onto their own servers to protect data.

Despite the competitive pressure on model developers, Wood remains positive on the "picks and shovels" suppliers that have driven AI-related stock gains. He cited the Jevons paradox, under which cheaper tokens spur greater overall demand for computing power and memory chips. Memory makers are the principal beneficiaries, he said, arguing SK Hynix, Samsung Electronics and Micron Technology should now be valued on earnings rather than book value and still looked cheap on that measure. High-bandwidth memory chips used in AI training clusters have become a key bottleneck, with SK Hynix commanding roughly 50% of the market.

Wood said he was raising exposure to technology hardware across the Greed & Fear portfolios, adding SK Hynix and Kioxia to the global long-only portfolio while removing Alphabet and Alibaba. The main risk to the wider trade, he said, is a sudden realization among investors that hyperscalers and leading AI developers cannot earn an adequate return on their spending, a fear compounded by circular financing arrangements such as Nvidia funding OpenAI's chip purchases. For now, Wood said, such concerns remained theoretical, with no sign yet of the AI capital spending race slowing. Hyperscalers including Microsoft, Amazon and Google are projected to spend more than $250 billion combined on AI infrastructure this year.

The emergence of a cost-competitive Chinese alternative compresses the valuation premium that Western AI developers have commanded ahead of their IPOs. Anthropic's $47 billion revenue run-rate, while impressive, faces margin pressure as token prices fall. Memory makers including SK Hynix and Micron, by contrast, stand to benefit from rising volume demand even as per-unit pricing declines, a dynamic that could sustain their current earnings-based valuations. Oracle and Meta Platforms, as major AI infrastructure spenders, could see their capital allocation strategies tested if cheaper models reduce the urgency of proprietary AI development.

This article is for informational purposes only and does not constitute investment advice.