A Chinese strategist just said the quiet part out loud. Hong Hao — partner and chief economist at GROW Investment Group, the former head of research at BOCOM International — delivered a verdict on the most crowded trade on Earth. AI bubble trading, he said, has entered a new stage.
Full stop. No valuation tables. No charts. No carefully hedged institutional caveats. Just a regime change announced with the casual confidence of a man who's seen this movie before.
That's precisely why it matters. Hong Hao isn't a tech blogger or an AI-adjacent podcaster hunting for engagement. He's a macro strategist who has spent two decades reading global liquidity tapes from Shanghai and Hong Kong, a man whose quarterly calls move capital at institutions managing billions. When a strategist of that caliber shifts his vocabulary — from "growth" to "bubble stages" — he's not offering a prediction. He's telling you the internal logic of the trade has changed. The narrative has rotated.
In my world, where I've spent five years capturing the emotional weather patterns of markets rather than just their chart patterns, narrative rotation is a leading indicator. Not a lagging one. Don't buy the chart. Buy the chaos.
Let me decode what "new stage" actually means. And then I'll show you why most investors are going to get it dangerously wrong.
Hong Hao's Signal and the Two-Stage Lifecycle of Every Narrative
Hong Hao's positioning matters because he sits at the intersection of two great market narratives. As the former head of research at BOCOM International, he cut his teeth in China's institutional markets, where macro calls are a blood sport and analysts live or die by their ability to read the global liquidity cycle. At GROW Investment Group, he has translated that experience into a global mandate. When a strategist with this background invokes the vocabulary of bubble phases, he's drawing on a long lineage of market-cycle thinking — from Shiller's irrational exuberance to the quiet warnings that preceded the 2022 crypto collapse.
The backdrop he's responding to is surreal by any historical measure. NVIDIA has spent months orbiting a $3 trillion market cap — a valuation larger than the GDP of most nations on Earth. OpenAI's annualized revenue has scaled from zero to billions in less than three years. Every enterprise software company on the planet has spent its 2024 marketing budget rebranding as an "AI company." And in crypto, the AI-tagged complex — decentralized compute networks, agent economies, data marketplaces — has ridden the same narrative wave, with tokens producing five-to-twenty-fold moves on sentiment alone.
Underneath that backdrop, the mechanics of the AI narrative are shifting. Every market story I have ever tracked follows the same arc: ignition, amplification, rationalization, verification. The first three phases are faith-driven — anyone can ride them if they're early enough. The fourth phase is different. Verification is data-driven. It's the moment the market stops asking "what could this become?" and starts asking "what is this right now, and does the math work?" The transition from rationalization to verification is the most dangerous moment in any speculative cycle. It's a fog of war. The old playbook — buy the story, ignore the fundamentals — stops working, but nobody knows the new playbook yet. That's what Hong Hao is flagging. The AI trade has entered its verification stage. The story is no longer the price driver. The numbers are. And the numbers are about to get brutal for a large slice of the market.
The Commercialization Verification Window Has Opened
My investment framework is built on something I call the Sentiment-to-Value Chain. I developed it in 2025 after analyzing 30+ modular blockchain projects against their narrative virality scores. The finding wasn't subtle: projects with strong, community-driven narratives outperformed technically superior peers by 300% during early adoption. But the more important discovery was the decay curve. When a narrative hits saturation — when the story has been told to everyone who will listen — the sentiment premium collapses unless real adoption metrics arrive. In crypto, we call it "sell the news." In equities, it's called profitability gating. In AI, it's happening right now.
The evidence is unmissable if you're looking through the right lens. OpenAI and Anthropic have grown annualized revenue into the billions — real commercial traction by any historical standard. But they're still burning cash at rates that would have sunk a 1998-era startup. The distance between current losses and eventual profitability is the exact distance between the current AI valuation and its fundamental support. Microsoft's Copilot, meanwhile, has hit enterprise headwinds. The first wave of "let's try AI" enthusiasm has collided with the second wave of "show me the ROI" procurement review. CIO surveys across the industry show AI spending optimism cooling from "transformational" to "incremental" — a one-word vocabulary shift that represents a fivefold change in valuation support.
And the structure of the value chain is deeply distorted. NVIDIA's data center revenue grew more than 100% year-over-year in recent fiscal periods. The computing layer — the chipmakers, the cloud providers, the data center operators — is capturing the vast majority of AI's actual profit. The model layer and application layer are largely losing money. This is the classic "selling shovels in a gold rush" structure. In every technological bubble from railways to telecommunications to the internet, the infrastructure providers profited first, were bid up most aggressively, and then became the most fragile when the music stopped. NVIDIA's current valuation is simultaneously the strongest argument for the AI thesis and the most stretched point in the entire bubble.
Based on my audit experience — and I've spent four years pulling apart token models, not just financial statements — this is where the real damage begins. Not at the top. The leaders will survive. The damage happens in the middle. That's where you find the "AI-powered" startups that are just wrapping an API call in a better sales deck. It's where you find the crypto AI protocols whose value proposition is a blog post merging two fashionable words. It's where you find the enterprise vendors that added a chatbot and renamed themselves a platform. The verification stage is a carnivore. It eats the middle.
Code Breaks: The Technical Compression Underneath
There's a technical reality underneath the bubble conversation that the mainstream bull case glosses over. The frontier's marginal breakthrough rate is slowing. The capability jump from GPT-4 to GPT-4o was a fraction of the jump from GPT-3 to GPT-4. Benchmark margins between rival flagship models have compressed into single digits. DeepMind's own published research has openly discussed diminishing returns on scaling. The industry's pivot to inference-time computation and test-time training isn't just innovation — it's an admission that brute-force scale is hitting a ceiling.
The open-source gap is closing too. Llama, Qwen, DeepSeek — the open-weight models have compressed the distance to the closed frontier to somewhere between six and twelve months. That's a fundamental reordering of the competitive landscape. The pricing power the closed-source labs have assumed as permanent is eroding every quarter. In the verification stage, eroded moats get repriced as risk. Not opportunity.
Beneath all of this sits the inference cost problem. For AI to become a mass-market, high-margin business, inference costs need to fall by roughly two to three orders of magnitude from current levels. That reduction is happening — but slower than the story demands. The gap between "the technology is amazing" and "the technology is cheap enough to generate real margins" remains the single largest unresolved question in AI economics. A bubble's new stage is exactly the moment when this gap gets scored under the harshest possible light.
From Technology Race to Capital Exhaustion
Here's where Hong Hao's vantage point adds a dimension that Western coverage tends to miss. The AI competition has shifted from a technology race to a capital-exhaustion contest. Training a frontier model costs hundreds of millions of dollars per run. Staying at the frontier requires billions in annual R&D. That's not a competition anymore — it's a filter. Only the hyperscalers, the trillion-dollar platforms, and a handful of deeply funded labs can stay at the table. Everyone else is competing in a different league with a hopeless hand.
The "new stage" is when that filter produces visible casualties. The second tier of AI firms — good demos, thin revenue, dependent on continuous fundraising — face a brutal financing environment. The same dynamic plays out in crypto's AI layer, and I've watched it before. During the 2022 LUNA collapse, while major analysts were panic-selling, I spent three weeks mapping wallet interactions across the assets that absorbed the fleeing liquidity. The conclusion: trust is social, not algorithmic. When social consensus cracks, the collateral cracks with it. The AI-crypto token complex — decentralized compute stories, AI agent economies, autonomous-trading narratives — is structurally the thinnest layer of the AI market. The story is beautiful. The revenue is hypothetical. Under a faith-driven market, that's survivable. Under verification, it's a fast path to irrelevance.
I say this with personal scars. In 2024, I co-founded NeuralLedger Labs in Austin — an experiment building decentralized identity verification, merging local AI startups with blockchain infrastructure. Five engineers, $50,000 in angel funding, a working beta in four months. It failed on scalability. But the failure taught me something permanent: the technology was never the bottleneck. The story was ahead of the infrastructure. And a market that is forced to verify will not wait for the infrastructure to catch up.
The Regulatory Layer Nobody Is Pricing
My post-ETF work gave me a lens for this that most market commentators lack. I spent the weeks after the January 2024 Bitcoin ETF approval manually parsing over 500 pages of SEC filings, looking for language shifts beneath the institutional hype. The insight that emerged — long before the liquidity trap hit — was that regulatory frameworks don't lag markets by years anymore. They lag by months. And they arrive as shock, not as guidance.
The AI regulatory calendar is now the market calendar. The EU AI Act's implementation details are being written as I write this. U.S. agencies are drafting enforcement contours through the same regulation-by-enforcement pattern that crypto has survived. And history tells us what happens next after a bubble's verification stage: the post-2000 internet crash birthed Sarbanes-Oxley. The 2008 financial crisis produced Dodd-Frank. When AI bubbles correct forcefully, AI legislation will accelerate — not for logic's sake, but for legitimacy's sake. Politicians need a villain. The AI market will volunteer one.
That regulatory overhang compounds the fundamental problem. AI companies already face a future of compressed margins, open-source competition, and profitability gating. Add compliance costs and liability frameworks, and the second tier of the AI market doesn't just face a drought. It faces an existential question.
The Contrarian Angle: This Is a Differentiation Trade, Not a Burst
Here's the counter-intuitive reading that most people will miss. A "new stage" in bubble trading is not the same as a bubble bursting. Professional strategists use the vocabulary of bubble trading to mean stage identification and stage-specific positioning. Hong Hao's framing points to a phase shift, not to an endpoint. The AI market is rotating from a tide that lifted all boats to a current that separates them.
That rotation creates a sharp, zero-sum trade. Capital consolidates into assets that can show revenue, margins, and a defensible path to sustained profitability. NVIDIA has pricing power, real earnings, and a visible order pipeline. The frontier labs have ARR and enterprise contracts. But the middle layer — the AI-enabled-everything buffet, the narrative tokens, the API-wrapping startups — is about to experience what I call a narrative drought. Not a crash. A drought. Slow capital evaporation punctuated by sharp downward revisions.
History is brutal on this point. The 2000 internet crash erased nearly 80% of the Nasdaq and wiped out the Pets.com layer. But the companies we remember — Amazon, Google, Apple — emerged stronger, because capital didn't leave the internet economy. It reallocated to businesses that had actual earnings. The same pattern runs through the railway bubbles, the telecom capacity mania, and every crypto cycle I've lived through. The infrastructure gets overbuilt during the mania. The overbuilding looks insane during the bust. But that overbuilt infrastructure becomes the foundation for a decade of value creation.
The same is true in AI and in crypto's AI layer. The narrative will contract. But the infrastructure being built right now — the chips, the data centers, the agent frameworks, the identity protocols that make autonomous commerce possible — doesn't disappear when the story loses its audience. It gets repurposed. The question isn't "is AI a bubble?" It's "which parts of this bubble become the foundation, and which parts become the footnote?"
What I'm Watching Now
Here's my scorecard as the verification stage unfolds. Track these signals with me.
NVIDIA's quarterly data center revenue growth is the lead indicator of whether the compute narrative has real demand behind it. If growth holds above 50%, the infrastructure trade survives. If it compresses, the entire stack reprices in sympathy.
Cloud capex guidance from Microsoft, Google, Amazon, and Meta is the second tell. These capital expenditure lines are the real AI capex lines. They reveal whether the platform giants are investing through the bubble or starting to blink.
The revenue and loss trajectories of the frontier labs are the fundamental anchor. The distance between revenue growth and cash burn is the distance between valuation and fundamentals. When those lines converge — through revenue acceleration or cost discipline — the landing gets softer.
Private market funding data is the quietest signal. When venture checks shrink, the second tier of AI labs starts dying quietly. That's the first casualty list of the new stage.
Implied volatility and short positioning in AI mega-caps round out the equity-side dashboard. When hedging demand builds, the crowd is positioning for a burst. Ironically, that positioning can delay the burst — and then amplify it.
And on the crypto side, I'm watching the AI-token complex with a simple filter: revenue, not tweet volume. The protocols drawing real usage and fees can survive a drought. The ones whose value proposition is the word "AI" attached to a ticker? Code breaks. Stories don't. But the stories that survive require a balance sheet to back them up.
The Final Trade Is about Receipts
Don't buy the chart. Buy the chaos.
The AI trade has entered its most violent phase. Not because the technology is failing — it isn't. But because the market's capacity for faith has run its course, and the verification era demands different conviction. The narrative that survives this stage won't be the one with the most impressive benchmark scores. It will be the one with a balance sheet that makes the story mathematically honest. Narrative remains the primary driver of value — but in the verification stage, the narrative that wins is the one that can withstand its own receipts.
I've watched this rotation three times in my career. In the WASM wars, developers didn't choose the best technology — they chose the story they wanted to belong to. In the LUNA collapse, social consensus disintegrated faster than any algorithm could price it. And in the post-ETF market, institutional narratives inverted retail expectations almost overnight. The pattern is always the same: stories carry the market further than fundamentals justify, and then the fundamentals arrive to collect.
The collection has begun. The question isn't whether you believe in AI anymore. It's whether you own the assets that can survive their own audit.