A four-sentence wire crossed my terminal on a Tuesday afternoon. No dataset attached. No chart. No follow-up research note. Just four words from a strategist whose language moves allocations: "AI bubble trading has entered a new phase."
Hong Hao — chief economist at Grow Investment Group, formerly head of research at BOCOM International — speaks a dialect that US market participants frequently misread. He is not a technology analyst. He is a macro strategist whose daily inputs are flows, positioning data, and the internal plumbing of global capital markets. He does not publish statements like this without intent.
Here's what seventeen years of extracting alpha from other people's throwaway comments has taught me: when an influential voice transmits a low-information, high-signal message, the content is almost secondary. The signal lives in the timing, the word selection, and the absence of supporting evidence. That absence tells you the speaker is seeing something he is not yet willing to detail publicly.
In 2017, I built a Python script to monitor the Ethereum mempool during the ICO crowdsale frenzy. Four hundred micro-transactions later, my team had secured a 22% net profit on $500,000 of deployed capital. The permanent lesson: markets leak information through mechanics, not through commentary. Hong Hao's "new phase" is a mechanics statement. Let me show you what it actually means.
Context: Who Is Speaking, and Why It Matters
Hong Hao's public record on AI is consistent and well-documented. In July 2025, he told interviewers that "the end of AI is electricity," pointing to power infrastructure as the binding constraint on the AI compute build-out. He flagged that the AI trade had become crowded. He argued that a bubble burst requires a trigger event — a mega-cap earnings miss, a profitability failure, or a macro-environment shift. These are not the statements of a technology skeptic. They are the observations of a market-structure analyst watching a crowded trade approach its inflection.
"New phase" is a deliberate escalation of that language. He is not saying the bubble will pop tomorrow. He is saying the trade has changed its character — and that the change is structural, not temporary.
The mid-2025 market context confirms why he chose his words carefully. Nvidia's market capitalization had crossed $4 trillion — a figure that, less than three years earlier, would have been dismissed as absurd math. Microsoft, Amazon, and Alphabet were running annualized capital expenditure north of $60 billion each on AI infrastructure. The Magnificent Seven had fractured into two camps: the AI infrastructure winners grinding through increasingly violent two-sided volatility, and the laggards shedding valuation month by month.
The options market had already voted before Hong Hao opened his mouth. Put demand on AI megacaps rose steadily through the second quarter. The CBOE SKEW index — a measurement of tail-risk hedging demand — reached levels historically associated with crowded positioning. Implied volatility on individual AI names decoupled from the index, trading at persistent premiums. Dispersion between AI winners and losers hit multi-year highs.
Translation from market mechanics to plain English: the market had stopped pricing a one-way AI bet. It was pricing a two-sided, high-variance regime. Hong Hao simply gave that regime shift a name.
Core: The Four Shifts Inside "New Phase"
Let me break down what "new phase" means in trading mechanics. Not in headlines. Four identifiable shifts.
Shift One: From Momentum to Mechanics
The 2023–2024 AI trade was a long-only momentum trade. Liquidity-driven, narrative-driven. The playbook was simple: buy the complex, hold through the noise, let beta do the work. Entry timing was secondary. Conviction was the alpha.
That phase is over. I can see it in the data.
Short interest on AI names climbed through Q2 2025 to levels not seen since the 2022 drawdown. Single-stock variance on Nvidia began trading at a persistent premium to index variance — the market charging more for idiosyncratic risk than systematic risk. The gamma profile on major AI names shifted, meaning market makers began amplifying directional moves rather than dampening them.
These are mechanical signals. None of them appear in a headline. Together, they describe a trade that has moved from "accumulate" to "manage."
I recognize this transition. In March 2020, when DeFi's over-collateralized lending stack threatened to cascade, I led a 15-person quant team deploying automated liquidation bots on Aave v1. We triggered over 500 liquidations in 48 hours, recovering 110% of exposed principal by selling distressed assets at a discount. We survived not because we predicted the bottom, but because we recognized the trade had shifted from positioning to execution.
That same shift is happening in AI equities right now. The edge is no longer in the direction of the trade. It is in speed, sizing, and volatility management. Don't trade the dip; trade the volume.
Shift Two: From Story to Receipts
The second shift is the hardest for retail participants to internalize.
From 2023 to 2024, AI companies were priced on narrative multiples. Discounted total addressable market. Projections treated as guarantees. OpenAI's private valuation at $500 billion in 2025 was priced on potential, not profit — a structure that works only in the early innings of a boom.
By mid-2025, the market's tolerance for stories had thinned visibly. The question had shifted from "how big is the opportunity?" to "when do you get paid?" Earnings season became a referendum on whether the AI build-out could generate returns that justify the capital deployed.
The specific data point I watch: capital expenditure guidance from Microsoft, Google, and Meta. Combined hyperscaler capex has crossed the $300 billion annual threshold. Any hint of a slowdown in AI spending triggers systematic selling across the entire complex. Any signal that spending accelerates — even at the expense of near-term margins — provides a bid.
This is the verification phase of a bubble. The market is no longer buying the dream. It is discounting the delivery timeline. If delivery arrives — if AI revenue growth compounds and models become commercially viable at scale — the bubble deflates gracefully, absorbed by real earnings. If it doesn't, the repricing is sudden, deep, and correlated.
Shift Three: From Uniform to Fractured
Third: dispersion.
Early bubbles lift everything in the sector. Beta is the strategy. Late bubbles get selective. Capital separates the water from the wine.
The AI complex has entered that latter stage. Upstream infrastructure — GPU manufacturers, ASIC designers, optical module suppliers, power utilities, data center REITs — continues to outperform. These are the shovel sellers of the AI gold rush, and their revenue sits on contractual orders, not aspirations.
Midstream model companies occupy a different position. API pricing wars have crushed unit economics. Every price cut from one lab forces a matching cut from the next. The margins that looked inevitable in 2023 are now a competitive battleground with no clear winner.
Downstream applications are still hunting for their killer product. Enterprise AI deployment remains measured, conservative, and ROI-driven rather than FOMO-driven.
The result is that dispersion between layers of the stack is rising. If Hong Hao is correct about a "new phase," it means stock selection within AI matters more than sector allocation. Directional beta fades; structural analysis of who actually captures value in the compute stack becomes the dominant P&L driver.
Shift Four: From Idiosyncratic to Systemic
The final shift: AI has become macro.
When a handful of companies constitute 30% or more of the S&P 500's market cap, the sector stops being a sector. It becomes the market. And when a sector is the market, it trades on the same variables that drive everything else: the Fed's policy path, inflation prints, liquidity conditions, financial stability considerations.
In the second half of 2025, AI stocks gapped down on Fed speakers. CPI prints moved Nvidia more than product announcements. That is not a failure of AI's fundamentals. It is a change in its beta.
This matters because it changes the risk calculus. A bubble that intersects with a tightening cycle is not just a correction — it becomes a systemic event. When an overleveraged complex meets declining liquidity, the unwind is faster, deeper, and more correlated across supposedly independent names.
Volatility is where the signal lives. In the macro phase, the signal is no longer company-specific. It is the global cost of capital.
Contrarian: The Bubble Is Not What You Think It Is
The reflexive read on "AI bubble enters new phase" is: sell everything. That read is likely wrong.
Hong Hao's language is precise. He said "trading," not "collapse." He said "new phase," not "end phase." A bubble in its trading phase is not a bubble that is popping. It is a bubble that is being traded. That distinction is everything.
Bubbles persist far longer than anyone's carry capacity. The Kindleberger-Minsky model identifies the final stage as discredit, but before that comes mania — characterized by retail participation and leverage ratios that, in mid-2025, still had room to expand. Hong Hao sounding the alarm may say less about the bubble's imminence and more about the fact that the easy, one-directional money has been harvested.
There is also a second-order effect that almost nobody is discussing. The "AI bubble" narrative itself is a cognitive hazard. When capital markets assign valuations far beyond demonstrated capability, expectations distort. The public begins believing AI is further along than it actually is. When reality reasserts itself — a model underdelivers, a product launch fails, adoption slows — the correction does not just hit prices. It hits beliefs. And belief corrections overshoot. Liquidity dries up faster than hope.
The smart-money reading of Hong Hao's message is not "sell." It is "change your mechanics." The AI trade still offers returns — but only to participants who can navigate widening dispersion, rising volatility, and systemic macro overlap. The buy-and-hold era is over.
Takeaway: What I Am Tracking
Here is what I am watching from my desk.
Short-term: earnings guidance from hyperscalers — specifically whether capital expenditure holds or cracks. Long-term: private market valuations. If OpenAI and Anthropic begin raising at flat or down rounds, that is the early signal the public market has not yet priced.
For the crypto side specifically: AI tokens and GPU-backed DePIN networks are downstream of the same capital flows. If the AI equity complex reprices, expect that correlation to transmit into digital assets faster than most crypto traders' models assume. The link is not narrative. It is the shared cost of capital.
Hong Hao has not told us the bubble is over. He has told us the part where you "buy and hold and forget" is over. The AI trade now belongs to people who treat markets as machines to be executed, not stories to be believed.
I have survived 2020. I have survived 2022. I have survived mempool wars. The pattern repeats: the crowd arrives late, the mechanics shift, and the unprepared get run over. Volatility is where the signal lives. Don't trade the dip. Trade the volume.