One extended theme today: the latest China-driven AI selloff. NLW zooms out to catalog every recurring way investors panic about AI — cheap models, circular financing, revenue-vs-CapEx, token caps, scaling walls — and argues the market's compulsive bubble-hunting is exactly what keeps a real bubble from forming.
AI market freakouts follow a pattern — and the pattern is why the bubble hasn't formed.
Since ChatGPT, investors have cycled through the same anxieties: cheap Chinese models, circular financing, revenue not justifying spend, CapEx running too hot, token caps, scaling walls. NLW's field guide walks each one, notes they often coincide with predictable summer seasonality, and lands on a counterintuitive conclusion — the market's relentless determination to call everything a bubble is itself the pressure-release valve keeping a runaway 1999-style bubble from ever taking hold.
Treasury Secretary Scott Bessent proposed that Chinese AI companies could face sanctions and entity-list designations as a response to 'covert industrial-scale distillation attacks that cross the line into IP theft.' It's the first time an executive-branch member suggested the government might actually act on distillation.
OSTP Director Michael Kratsios alleged Moonshot built 'a sophisticated internal platform to conduct large-scale distillation against US models' for its K3 model, switching between access methods to avoid detection, and accessed GB300 servers in Thailand. Anthropic confirmed it's working with the administration on the matter.
— Signal, on the proposed China AI restrictions. Signal's viral summary argued the crackdown forces US companies to buy pricier American AI while the rest of the world uses cheaper models of equal intelligence — and distillation doesn't stop just because Washington publishes a list.
Per Wired, parts of the White House want stricter controls while Commerce sees them as unworkable. Commerce Secretary Howard Lutnick reportedly favors combating Chinese AI with competition — pitching direct incentives for US labs to open-source their models rather than regulating anyone.
Per Bloomberg, AI investment represents 25% of US GDP growth — the largest single-sector contribution in history. JPMorgan credits AI with 75% of S&P 500 returns, 80% of earnings growth, and 90% of capital-spending growth since ChatGPT's release.
With roughly 50% of the S&P 500 now AI or AI-exposed, NLW argues everyone needs to understand the narrative — even a 401(k) contributor or passive indexer carries significant AI exposure whether they realize it or not.
This round echoes the early-2025 DeepSeek panic, when markets believed a Chinese lab built frontier AI for a few million dollars. Investors later learned the difference between training and inference compute — and that R1 wasn't actually ahead, just the first free reasoning model.
K3 is being served at around a third of Fable's price or half of Opus — a meaningful saving, but far from the near-free clone many analysts imagine. The 'Chinese AI is cheap, US AI is eye-watering' framing is oversimplified.
The worry that Nvidia's investments flow straight back as chip revenue misreads history: unlike Cisco's 1990s vendor financing, OpenAI pays cash for GPUs. If OpenAI stumbles, Nvidia holds devalued stock, not a debt default and a balance-sheet hole.
Google posted 24% overall growth and 82% cloud growth — but the stock fell 1.2% because CapEx guidance hit $200 billion. 'The two hundred was the do-not-cross line,' said Zacks' Brian Mulberry, likely the psychological shock of a round number more than the fundamentals.
Hyperscalers have raised CapEx guidance every quarter for three years, with combined spend on track to exceed a trillion dollars in 2027. The flip side of 'revenue not growing fast enough' is 'CapEx growing too fast' — the bigger the spend, the steeper the revenue hill.
Late-2025 bubble talk faded when analysts stopped multiplying knowledge-worker seats by $20-30 and saw agentic use cases arrive — employees spending hundreds or thousands per month. It made clear AI isn't just another category of SaaS.
NLW dismisses the notorious report as 'absolute garbage social science MIT should be embarrassed to be associated with' — but notes it didn't matter, because the headline landed on every Wall Street desk anyway.
Uber caps users at $1,500/month and Tesla allows $200/week with the ability to request more. Since most knowledge workers use nowhere near those budgets, even universal adoption of such policies leaves enormous room for growth.
Fall-2024 panic held that pre-training had plateaued and both labs scuttled flagship runs. Then o1 revealed reasoning as a new performance vector, and pre-training clearly didn't hit a wall — few would argue Fable V isn't a completely different beast from Opus III.
Summer momentum breakdowns are well documented, especially when semis drive the rally. Morgan Stanley notes momentum stocks are down 40% this month — the worst on record, beating early-2021's 29% — while Goldman says hedge funds sold tech in record numbers.
NLW argues that a generation raised on The Big Short calling everything a bubble is a feature, not a bug: every time the market gets over its skis, there's a pressure release. We're not seeing the frenetic 1999-2000 boom-and-bust IPO frenzy.
Half of this year's announced data centers were canceled or delayed — which bears read as weak demand, but NLW says the simpler explanation is that permitting and construction are genuinely hard and builders have badly mishandled community concerns. These road bumps buy the economy time to adjust.
NLW argues this China round resolves once people realize inference capacity matters more than price: Moonshot was completely tapped out of compute on K3's opening weekend, and it's questionable whether all Chinese labs combined can serve even a fraction of US companies' users.
— Nick Carter, investor. Investor Nick Carter argued that if token economics collapse, American enterprise will be fine — benefiting from hyper-deflation in the cost of cognition. The hyperscalers, neoclouds and internet companies survive; only OpenAI and Anthropic's current forms are at risk, and they can adapt.
He expects every frontier token to be bought at a premium for at least five years even as cheaper workloads come online, because premium state-of-the-art supply stays ahead of demand. A flourishing of routers, fine-tuning and vertical model plays distributes AI's revenue gains and relieves pressure on the big two labs to carry the whole market.