CF40 Forum: China AI Supply Chain in ‘Marathon’ First Half, K-Shaped Rally Masks Bubble in Long Tail

AI’s Revenue Rocket and the Capex Gulf Between the US and China

Since OpenAI launched GPT-3.5 at the end of 2022, large language models have iterated every 6 to 12 months, ballooning from hundreds of billions of parameters to around 4 trillion today. Scaling laws are holding, and the market has shifted from a single dominant player to multipolar competition. While Anthropic’s Claude 5 now leads in coding, Chinese models—such as Zhipu’s GLM 5.2 and Kimi K3—have narrowed the gap to just 3–6 months, or roughly one generation, behind their US counterparts.

Coding has emerged as the clearest business model, with Anthropic’s monthly annualized revenue (ARR) reaching $700 billion by July 2026 and forecasts pointing to $1,200 billion by year-end. The market’s ‘bubble’ fear, participants at the CF40 forum argued, is not about whether AI can make money—it clearly can—but whether the explosive revenue growth can be sustained. In effect, concerns center on the slope of the growth curve, not the existence of profits.

On the infrastructure side, North American AI capital expenditure hit roughly $7,325 billion in 2026, up 76% year-on-year. China’s annual capex, at about RMB 700–800 billion ($100–110 billion), is only a sixth to a seventh of the US total. This gap is far wider than during the mobile internet era, when China’s cloud capex reached 60–70% of the US level. The forum’s consensus was that the domestic AI supply chain is only now moving from initial breakthroughs to mass scaling, with a software-hardware flywheel beginning to spin.

Why China’s AI Chain Looks Undervalued Despite US Dominance

Anthropic’s Revenue Rocket: Defying Bubble Fears?

The bullish case rests on a simple calculation: there are about 900 million white-collar workers globally, and if AI coding tools capture even 50% of related tasks, the addressable annual revenue could reach $1,500 billion. With Anthropic tracking toward $1,200 billion ARR, the top end of the market is still well within that theoretical ceiling. The bearish case, however, is that monthly ARR growth could decelerate from the current 20–30% to 10–15%, triggering a repricing. The company’s $1 trillion private-market valuation implies a roughly 10x price-to-sales multiple—a level that only holds if profit growth continues to outpace revenue, as was the case for AWS in 2015. The forum noted that the key metric to watch is the monthly rate of change in Anthropic’s ARR, not the absolute number.

China’s AI Capex Gap: A Macro Opportunity, Not a Weakness

In the mobile internet era, Chinese cloud providers spent roughly two-thirds of what US peers did; today they spend less than one-sixth. If scaling laws remain valid—meaning bigger models require proportionally more compute—then China’s current capex level is orders of magnitude too low to produce truly world-class models. The flipside is that the runway for growth is enormous. The forum argued that domestic capex has only just stabilized and started rising since mid-2023, and that the supply of locally produced, A100-equivalent chips is only about 5% of the US volume. The next phase is therefore a shift from “0 to 1” breakthroughs in advanced packaging and fabrication to “1 to N” mass production, which would close both the hardware and model gaps.

The K-Shaped Equity Story: Winners and the 98%

Headline Chinese AI stocks already reflect a reasonable relationship between price and fundamentals. Optical module leader Zhongji Innolight trades at a 2026 dynamic PE of 35x and 2027 of 20x, sitting at the 80th percentile of its three-year range, while GPU player Cambricon’s 63x multiple is only at the 13th percentile of its own history—meaning earnings are catching up fast. By contrast, second- and third-tier optical names have median 2026 PEs above 100x and are at the 80th-plus percentile of their five-year ranges. The forum framed this as a classic K-shaped market: the top 1–2% of firms that will ultimately win may be fairly priced, but the long tail is pricing in a linear extrapolation of a supply-chain boom that cannot last. This is where the real bubble risk resides, not in the sector leaders.

Non-Linear Shocks: Meta’s Glut and the Open-Source Deflation Risk

Recent anxieties have been amplified by several concrete developments. Meta has begun renting out surplus AI compute, altering the supply-demand balance. Korean markets saw a spike in leverage and liquidity strains. Amazon’s $25 billion bond sale attracted only a 1.6x bid-to-cover ratio, versus a typical 4–5x, raising fears that AI capital absorption is saturating global liquidity. Most strategically, the forum highlighted that China’s open-source models may export deflation to the American monetization model: if top-tier coding capabilities become freely available, the high-margin subscription edifice in the US could come under pressure. These are all non-linear risks that simple extrapolation models fail to capture.

For Investors and Industry: Position for the AI Marathon, Not the Sprint

Monitor Anthropic’s monthly ARR growth rate, not just the headline number. A deceleration from 20–30% month-on-month to 10–15% could reset valuations across the US AI ecosystem, with immediate spillovers to Chinese sentiment.

Watch for the inflection in US capex growth. Consensus expects the year-on-year rate to slow from 76% to 20–30% as the base effect kicks in. A sharper drop would signal that the infrastructure build-out is peaking, favoring downstream application plays over pure hardware.

In China, focus on the top tier and ignore the long tail. Companies like Zhongji Innolight and Cambricon have seen earnings grow faster than their share prices, keeping multiples reasonable. By contrast, many second-tier suppliers trade at 100x earnings based on a temporary pricing boom. In an industry where only 1–2% of players may survive, the safer bet is to stay with those already proving fundamental delivery.

Track domestic chip scale-up milestones. The shift from “0 to 1” to “1 to N” in domestic GPU supply will be the single largest variable for Chinese AI capex. Key indicators include the monthly output of advanced packaging lines and the yield rates at the leading fabrication node—metrics that will determine whether the 5%-of-US supply gap can begin to close.

Position for a marathon, not a sprint. The forum’s core message: this AI cycle is only halfway through its first leg. The reset that follows any growth scare is likely to reward patient capital that can hold through the K-shaped differentiation, rather than those chasing the hype in the long tail.