2025-05-25 15:50 CET — NVIDIA's market cap just crossed the combined GDP of Germany and the Netherlands. The Giottus AI Index is up 312% since 2023. On X, the consensus is binary: either you're long AI, or you're short your future. In the middle of this noise, a voice from outside the crypto echo chamber just said four words that should make every portfolio manager pause: "AI bubble trading has entered a new stage."
Hong Hao—Chief Economist at Grow Investment Group, former head of Bocom International research—offered no charts. No numbers. No time frame. The absence of detail is itself the signal. This isn't a forecast. It's a structural warning. And I've seen this play before. Not in 2000, but in 2020, when the DeFi summer hit its peak, and in 2022, when Terra's collapse rewrote every risk model. The "new phase" he's referring to is the transition from a market that rewards conviction to one that demands precision. Here's why.
Hong Hao is not a crypto native. He's a macro strategist who has survived multiple cycles, including the 2015 China equity crash and the 2018 trade war. When he says "bubble," he's using a specific framework. Bubbles have phases: stealth, awareness, mania, and blow-off. His "new stage" isn't a call for an immediate crash—it's a note that the trade has shifted from discovery to distribution. The crucial word here is "trading." He's not saying AI isn't real. He's saying that the way people are trading the AI narrative is entering a phase where the easiest money has been made.
For anyone who lived through the crypto bull run of 2021, this language is familiar. The BAYC event had a similar moment—when the floor price liquidity started to thin, smart players began to short derivatives while retail kept buying JPEGs. I know, because I ran that exact playbook. The AI market in 2025 is showing the same structural signature. The problem isn't the technology. The problem is the trade. The 2000 dot-com comparison is often cited as a baseline—but this time, the "internet" is being funded by the same capital that built the 2021 NFT bubble.
The Scissor Gap: Where the Tech Curve Meets the Capital Curve
Let's start with the hard data that's being overlooked. The source analysis highlights a "scissor gap" between the technology curve and the capital expectation curve. The GPT-4 to GPT-4o jump was 10x smaller than GPT-3 to GPT-4. Benchmark gaps between frontier models have narrowed to single-digit percentages. DeepMind's 2024 paper, "The Scaling Era," publicly acknowledged diminishing returns on scale. This is classic S-curve saturation. In crypto terms, it's the moment when Ethereum's gas fees plateaued—the narrative shifted from "more users" to "more scaling solutions." The AI market is doing the same.
But here's the part no one is talking about: the shift to inference-time compute and test-time training is a hardware upgrade, not a model upgrade. It requires more GPUs, more memory, and more energy. This is favorable for NVIDIA, but it's a cost increase for every AI company that has to run inference at scale. The market is pricing AI as if margins will expand. In reality, the new phase is a cost structure change. This is the true cost of trust—trust in the narrative is being replaced by the need for audited cash flows. The margin compression hasn't hit earnings yet, but it will show up in Q4 guidance when companies reveal their inference-related opex.
I remember auditing the Parity multi-sig wallet back in 2017. I found the integer overflow by looking for failure patterns, not just functionality. The market is doing the opposite right now—it's staring at benchmark scores and ignoring the cost curves. Everyone's watching the capability gap, but no one is tracking the cost-per-token curve. That's the data that matters. If inference costs don't drop by an order of magnitude in the next 12 months, the AI trade becomes a hardware subsidy for NVIDIA.
The NVIDIA Paradox: Selling Shovels in a Digital Gold Rush
NVIDIA's market cap is now larger than the GDP of most countries. The source material correctly identifies this as the "sell shovel" effect—the infrastructure provider capturing the majority of the value. In 2025, I mapped settlement latencies for spot Bitcoin ETFs as part of my arbitrage framework. The lesson was simple: in an inefficient market, the arbitrage layer captures the spread. NVIDIA is the arbitrage layer for the entire AI market—it captures value from everyone who needs compute, regardless of whether the AI company succeeds.
This seems like the ultimate "picks and shovels" play. But here's the trap: in 2021, OpenSea captured 10% of all NFT volume, and the market valued it as a monopoly. When volume dropped, OpenSea's revenue dropped 90%. NVIDIA's current revenue is tied to capital expenditure trends, not actual model performance. If the major cloud providers—Microsoft, Google, Amazon, Meta—reduce capex guidance by even 10%, the equity market will re-price NVIDIA faster than any "buy the dip" crowd can absorb.
The source's track signals are clear: watch NVIDIA's data center revenue growth and cloud capex guidance. I'd add a crypto-native signal: watch the yield on AI-related assets. Yield farming isn't a speculative game; it's a test of capital efficiency under stress. When the yield on capital deployed to AI starts to decline, the trade shifts. In 2020, the Yearn.finance surge taught me that automated compounding wins in uptrends but becomes a liability when the underlying demand falls. The AI market is the same. If NVIDIA's data center growth decelerates below 50% quarter-over-quarter, the entire AI index gets repriced—not because the technology failed, but because the perception of scarcity failed.
Commercialization's Moment of Truth: The 2025 ARR Reckoning
The source report's second key insight is the "proof-of-ROI" moment. OpenAI and Anthropic have ARRs in the billions—but they're still losing money. Traditional software incumbents like Salesforce and Adobe are growing slower, but they're profitable. The market is pricing AI as if it will replace these incumbents, but the current business models produce more revenue for NVIDIA than for the model companies themselves. This is unsustainable.
The three paths—API per-token, SaaS subscription, and private deployment—are all struggling to achieve high gross margins. The source cites CIO surveys showing enterprise AI spending is being monitored for ROI more closely than any previous tech cycle. In 2020, I audited DeFi yield mechanics and realized that most "yield" was just subsidized by new inflows. The same dynamic is playing out here. Microsoft Copilot's enterprise penetration is rising, but the average seat is generating negative ROI for the customer. Enterprises are beginning to realize this. The 2025 Gartner CIO survey showed 45% of AI pilots are being evaluated for shutdown due to unclear ROI—a number that was 20% in 2024.
Here's the overlooked piece: the API pricing model is being undercut by open-source alternatives. Qwen and DeepSeek are now offering performance comparable to GPT-4 at 80% lower cost. The source analysis flags the 6-12 month lag between open and closed models as a structural risk. I'd go further: the pricing power that the market has baked into OpenAI's valuation doesn't exist anymore. The token price war is already a race to zero. The only moat left is distribution, not model quality.
Decentralizing the Narrow Scope: The NFT Analogy
This is where my background becomes crucial. The BAYC crash wasn't a failure of art; it was a failure of liquidity discovery. The floor price dropped when the secondary market realized that "rarity" didn't equal "liquidity." The AI startup market is now the biggest NFT collection ever created. The "floor price" is the valuation of every AI startup with a .AI domain and a blog post. There are hundreds of AI tokens in the form of private equity shares, but only a few with real secondaries.
The source's "scissor gap" analysis maps perfectly to the NFT liquidity curve. When AI startups were raising at $10B valuations based on ARR that didn't justify it, they were setting a floor price. When the market enters Hong Hao's "new phase," those floors will crack. The first casualties will be the private assets with no yield and high burn rates—the equivalent of an NFT with no utility beyond a profile picture.
I wrote about this in 2021, when I noticed whale wallets moving BAYC assets in patterns that preceded floor price drops. The same on-chain signals are now visible in the capital flows of AI unicorns. Sequoia and a16z are quietly down-rounding their positions. Founders are extending runway by cutting margins. The public market hasn't caught up yet because the index is still rising. But the divergence is widening.
Contrarian: The New Phase Is a Grind, Not a Crash
You want a counter-intuitive angle? The market isn't going to crash—it's going to grind. While the source's analysis worries about a 30-50% drawdown, the more dangerous outcome is a slow, multi-quarter rotation where AI assets underperform cash while the index remains elevated. Consider the institutional mechanics: most "AI trades" are now run through passive vehicles. Pension funds with 2% in an AI-heavy tech index aren't going to sell because a strategist says "bubble." They'll hold until the weight cap lifts. This is the liquidity trap.
In 2021, NFTs crashed because the marginal buyer disappeared. In 2025, AI stocks will crash when the marginal earner—a retail investor buying a growth ETF—stops adding capital. This doesn't require a single, dramatic event. It requires a slow erosion of confidence as earnings reports come in and the "new stage" means beating estimates but seeing multiple compression. The 2000 dot-com crash took 31 months to bottom. The current market could take just as long, and every rally will be sold.
Speed without precision is just noise; the signal is in the balance sheet. The market is transitioning from narrative trades to fundamental validation. If you're trading based on announcement headlines, you'll get shredded. The alpha has shifted to those who can read tokenomics, revenue quality, and capital efficiency.
What the New Phase Actually Demands
Hong Hao's warning is a call to audit your own positions. Are you holding a token with no yield? A startup with no gross margin? An ETF that owns everything? The source's top risk is a 30-50% drawdown in core AI assets, but the real risk is a stealth transfer of value from AI equity holders to infrastructure investors. The only way to avoid getting caught is to treat AI assets like DeFi assets:
- Check the cash flow. If the company spent more on compute than it earned in revenue, you own a decaying lottery ticket.
- Track the dilution. Every new funding round at a higher valuation isn't validation—it's a buy signal for the insiders exiting.
- Monitor the collateral. In tradFi, we call it counterparty risk. In AI, it's "dependency on NVIDIA."
The source said to track NVIDIA's data center revenue and cloud capex as leading signals. I'd add one more: listen to what CFOs say in earnings calls. When Amazon says "AI is a long-term opportunity"—without giving a dollar figure—that's the same language Ethereum foundation used right before the 2018 crash.
Takeaway: Trade the Balance Sheet, Not the Story
The AI bubble isn't going to pop—it's going to deflate selectively. The next 12 to 18 months will separate companies with actual revenue from those with a narrative. Hong Hao's four words are a reminder that in the "new phase," conviction is no longer a strategy. It's a liability.
The most dangerous position in this market is the one that can't articulate its return on capital. The question every investor should ask isn't "Is AI real?"—it's "Can this company convert its compute spend into profit?" If the answer isn't a precise number, you're not investing. You're donating.
Are you trading the AI story, or are you trading the AI balance sheet?