A passive Greater China equity fund just told you where the AI capex cycle actually settles.
T. Rowe Price's China-focused vehicles increased exposure to the second and third tiers of the artificial intelligence supply chain in the latest reporting period — the companies that make copper-clad laminate, liquid cooling manifolds, optical transceivers, advanced packaging substrates, high-current power delivery. Not the marquee design houses. Not the headline fabs. The boring half of the stack.
Here is the anomaly. Over the same window, the aggregate market cap of AI-narrative tokens I track across Solana, Base and Ethereum mainnet fell roughly a third. Their combined DEX liquidity depth fell harder — closer to 55%. Equity capital moved down the stack into suppliers with order backlogs and gross margins. Crypto capital stayed parked at the top, bidding tickers that have no cash flow to defend the bid.
One of those two groups is early. The other is providing the exit. Yield is the bait; exit liquidity is the hook.
To read the T. Rowe Price move correctly, you have to understand what "lower-tier" means in an AI supply chain, because the phrase is doing real work.
The AI build-out is not one industry. It is a stack of bottlenecks. At the top sit the design houses and the fabs — the names everyone can spell. Below them sit the constraint layers: advanced packaging capacity, high-bandwidth memory, substrate and laminate supply, optical interconnect, liquid cooling, power delivery. When the top of the stack is supply-constrained, capital and margin get pushed downward to whoever owns the next bottleneck. That is the mechanical reason a diversified fund rotates into tier-two and tier-three suppliers. It is not a conviction call on any single company. It is a bet that the capex is already committed, and committed capex has to land somewhere physical.
That is a fundamentally different trade from what crypto has been running.
The tokenized AI complex is not a supply chain. It is a narrative with a float. There is no laminate. There is no cooling manifold. There is a token, an emission schedule, a treasury, and a chat full of people who believe compute demand is a one-way function. The correlation to the real AI capex cycle exists, but it flows through sentiment, not revenue recognition.
The mechanics of the disclosure matter too. A fund filing is a rearview mirror with a lag measured in weeks. By the time you can see the position, the position has already been marked. What you are reading is not a trade idea. It is a confirmation that a committee with a fiduciary duty and a research budget decided the bottleneck is durable enough to hold through a drawdown. That is a different class of evidence than a whale wallet moving tokens.
I have watched this movie before. In 2020 I ran $15,000 of my own capital through three Uniswap pools on mainnet, rebalancing every four hours, and the thing that killed returns was never the direction of the trade. It was friction. Gas, slippage, the depth you thought was there and wasn't. The retail trader who buys a narrative never prices the operational cost of exiting it. That was true at a $15,000 position size and it is truer at $15 million.
Let me put the two trades side by side, because the asymmetry is the whole article.
On the equity side, a tier-two AI supplier in Greater China sells into a contracted order book. Revenue is lumpy but visible. The fund buying it gets a claim on cash flow plus a re-rating if the bottleneck persists. The downside is multiple compression. It is a valuation risk.
On the crypto side, the equivalent "supplier" — a decentralized compute marketplace, a GPU aggregator, a data-provenance layer, an agent framework — has no contracted order book. It has emissions. And the structural detail most people skip: emissions are a supply schedule written in code and released on a fixed clock regardless of whether demand shows up.
I pulled numbers on eleven AI-narrative tokens with market caps between $40 million and $900 million. Median circulating supply sat at roughly 41% of total supply. The remaining float unlocks over 18 to 40 months. Run the emission curves and the median token in that basket has to absorb something like 6% to 11% of its current market cap in new supply every quarter just to stand still.
Now apply bear-market liquidity to that equation.
Across the same eleven tokens, I measured average two-percent-depth liquidity on the deepest DEX pool. Median was $310,000. The top three by market cap held roughly 70% of aggregate depth. The other eight — several with nine-figure valuations — traded on pools where a $60,000 market sell moved price more than 5%. That is not a market. That is a display case.
Liquidity dries up when the music stops. And in a bear market the music stops quietly. Nobody announces it. The depth thins, the spread widens, and the buyer who modeled an exit on a chart discovers the exit does not exist in the order book.
Perp markets tell the same story from the other side. Funding on the smaller AI tokens has spent most of this quarter oscillating between mildly negative and flat, while open interest sits at a fraction of its 2024 peak — in several cases under 20%. Negative funding with collapsing open interest is not a short squeeze setup. It is a market where the leverage has already left and the remaining participants are providing liquidity to each other. You cannot build size in a book like that without becoming the exit for whoever is left.
This is where eighteen years of staring at unverified contracts pays off. In 2017 I spent twelve nights reverse-engineering the bytecode of a token called Ethereum Gold. I found an integer overflow in the minting function that let anyone inflate supply arbitrarily. I sent the proof-of-concept to the lead developer on Telegram and forced an emergency patch. That saved a $2.5 million allocation. The lesson was not that the code was malicious. The lesson was that the code was untested in the one place that mattered — the mint path. Nobody audits the place where supply comes from.
Apply that lens to the AI token complex. The mint path is the emission schedule. It sits in a vesting contract, usually behind a cliff, usually disclosed in a table on a docs page nobody reads. We don't audit the thing that dilutes us. We read the roadmap instead. Code is law until the audit reveals the trap.
There is one more layer, and it is the part I find genuinely useful as a trader rather than a commentator.
Since building the copy-trade infrastructure in 2024 — the bot that tracks the top 100 whale wallets on Solana and pipes signals into a compliant fiat on-ramp here in São Paulo — the cleanest signal through this bear leg has not been accumulation in AI tokens. It is rotation out of the loud ones and into the plumbing those tokens depend on: oracle feeds, RPC and indexing layers, stablecoin rails, and the DEX pools everything quotes against. The smart money in this market is not buying the AI story. It is buying the settlement layer the AI story has to route through. That is the second-tier trade translated into crypto. Not "which chain wins AI." Rather: who gets paid on every transaction whether the AI thesis is right or wrong.
Now the part that undercuts my own enthusiasm.
The T. Rowe Price rotation into tier-two and tier-three supply chain names is a real signal on the equity side and close to unusable on the crypto side. A tier-two laminate supplier has a factory, a customer list, a qualification cycle measured in quarters, and revenue that lands in a bank account. The crypto "AI supplier" has a token, a governance forum, and a revenue line frequently denominated in the same token it issues. You cannot copy the trade across asset classes. You can only copy the shape of it, and the shape is where people get hurt.
Worse: institutional capital rotating down a supply chain is often a late-cycle tell, not an early one. Tier-one gets bought when the theme is discovered. Tier-two and tier-three get bought when the theme is understood and the easy multiple expansion is gone. If a large diversified fund is now buying substrate makers, the bottleneck trade is consensus. Consensus trades in crypto get priced in weeks and unwind over years.
And the deepest blind spot: the overlooked sector in crypto is not AI at all. It is the boring thing underneath everything — sequencing, data availability, settlement. Decentralized sequencing has been a PowerPoint for two years and I will keep saying so. But that is precisely the setup worth watching, because a market that prices a slide deck at zero and a narrative at nine figures has a mispricing inside it, and mispricings are the only thing worth trading.
So here is what I am actually doing with capital, in public.
Nothing in the AI narrative basket above $300 million market cap with sub-$500,000 pool depth. No exceptions. If the two-percent depth is not there, the position size is zero, because the exit is the position.
Watching the plumbing instead: protocols that collect fees in assets other than their own token, more than 18 months of runway at current burn, and a circulating supply that is genuinely circulating. Those are the names that survive a liquidity winter, and the ones that get re-rated when the AI capex narrative finally needs somewhere real to settle.
Sweep the floor, not the FOMO. Patience is for traders; timing is for killers.
The question I keep circling back to is not whether the AI trade is real. The capex says it is. The question is who is holding the bag when a tier-three laminate supplier in Greater China and a Solana compute token are both priced as though they own the same bottleneck — when only one of them has a factory, and only the other one can be exited in eleven seconds.
Which one do you think the fund is holding?