George Noble
Investment Writer, Former Fund Manager at Fidelity, Former Assistant to Peter Lynch
Zhipu, listed in Hong Kong as Knowledge Atlas Technology, has one of the most absurd valuations imaginable. It did about $105 million in revenue last year and lost more than four times that, yet the market valued it at roughly $128 billion this week, above 1,000x sales at the peak, on a float where barely 4% of the stock trades. Its open GLM-5.2 model may undercut Big Tech's AI moat, but valuation cannot be ignored forever. x.comÂ
Zhipu (Knowledge Atlas) is already valued at nutty levels. At HK$2,094, buying it on $1B of ARR is like paying $6T for Anthropic on $50B of ARR. Its GLM models are genuinely impressive, but that valuation comes before factoring in the headaches of owning Chinese stocks. x.comÂ
Zhipu AI's stock is unattractive because the Chinese AI model space is simply too competitive to be an appealing investment. Its company vision still deserves respect, and its current valuation is not a bubble. x.comÂ
Zhipu AI is a Chinese AI arbitrage, shipping models near frontier Western performance at a fraction of the valuation. Its GLM-5.1 scored 45.3 on a coding benchmark versus 47.9 for Claude Opus 4.6, a 2.6-point gap far smaller than the valuation gap. Zhipu is up 357% since its January IPO, double MiniMax's return, yet both trade at about $40B. x.comÂ
Meta looks undervalued against Anthropic and OpenAI given its revenue, profit margin and state-of-the-art AI model. The two labs reportedly have $150 billion in combined ARR and a $3.5 trillion combined valuation, while Meta has $228 billion in trailing revenue, a 30% profit margin and a $1.9 trillion valuation. x.comÂ
Alphabet is the most complete vehicle for betting on AI, and it is undervalued. x.comÂ
Alibaba offers opportunity at $123 as a front-runner in China's AI buildout, particularly if its AI spending shows real returns and it proves it can compete with the US. Over $56 billion of AI and cloud investment over three years has made current free cash flow negative, like US hyperscalers. Cloud growth is expected near 45%, AI products are already 30% of external cloud revenue, and it holds $38 billion of net cash. Cloud growth, capex and guidance will drive the earnings reaction. x.comÂ
NVIDIA looks close to the top of an AI bubble, with a two-year megaphone pattern signaling exhaustion of growth. Such expanding, volatile patterns historically don't end painlessly. Its roughly 22x valuation on a $4.8 trillion market cap and $216 billion profit is approaching Cisco's roughly 30x in 2000. Nvidia just raised $25 billion in debt, drawing $85 billion of orders. x.com x.comÂ
Hyperscalers' incremental return on capital is cratering, so hard spending decisions are coming. In the June quarter, Amazon, Microsoft and Alphabet grew LTM EBIT 20% but needed 38% more deployed capital, before token prices cratered and with much demand from non-self-sustaining OpenAI. Compute and models may become a utility-like commodity. Insiders are selling, with Jeff Bezos filing to sell $4B of Amazon stock, and Google's $25B bond deal widened hyperscaler credit spreads. x.com x.comÂ
Bittensor's TAO is cheap at a $2.9 billion valuation because decentralized AI will be at the forefront while many AI stocks trade at crazy valuations. Per Polymarket, Microsoft CEO Satya Nadella warns that an AI future dominated by a few models is "not stable." x.comÂ
Alibaba's Qwen models look able to beat costlier US AI: two weeks of testing showed no difference from US AI agents, at 80% less per token. US companies spending hundreds of billions and driving up compute costs risk losing the battle to cheaper Chinese models. A selloff to the low $100s would be a chance to buy much more. x.comÂ
