A note started making the rounds in my Telegram channels last Tuesday. A macro analyst — the aggregator stripped the byline, which should tell you something about the confidence level — declared that the AI bull market is over. Two reasons: leverage liquidation and compute oversupply. My first instinct was to file it with the other macro parlor tricks. But I don't trade instincts. I query the chain.
The data shows the analyst is pointing at the right failure modes — just in the wrong market. The same two risks have migrated into crypto's AI trade. The GPU DePIN tokens. The AI-agent perps. The compute-marketplace liquidity pools. Here, the leverage is public, timestamped, and reproducible. It leaves a footprint.
Start with one number: in the past 30 days, aggregate open interest across the five largest AI-token perpetual pairs — RNDR, TAO, AKT, FET, ARKM — grew 62% while spot volume on the same pairs fell 18%. That divergence is not conviction. It is borrowed money parked in a trade the spot market can no longer absorb. Silence is just data waiting for the right query, and this particular query says the crypto AI trade is top-heavy with short-dated leverage.
Before I show the evidence chain, we need to agree on what "the crypto AI trade" means. I group it into three buckets.
First, decentralized compute networks. Render operates a marketplace for GPU rendering and, increasingly, video-generation inference. Akash runs a lease-based platform for machine-learning workloads. A handful of smaller entrants circle the same niche. Second, the agentic-AI token ecosystem — Bittensor's subnet auction model, Fetch.ai's agent frameworks — where the token is a bet that autonomous software will generate fees. Third, the infrastructure layer: zero-knowledge ML projects claiming to verify model inference, data-provenance rails, and GPU-backed RWA tokens that let retail buy a slice of a data-center lease.
The macro thesis that the analyst applied to Nvidia and CoreWeave maps structurally onto all three. Bull markets in this sector have been financed by leverage since the DeFi summer of 2020. The underlying asset is depreciating hardware with a 12- to 18-month lag between deployment and utilization. Both pillars of the bearish case — leverage fragility and compute oversupply — are real risk factors. Unlike the equity market, the crypto versions of both risk factors can be queried directly. No 10-Q waiting period; the books are live.
I have been running this kind of investigation since 2017, when I spent three weeks manually cross-referencing Ethereum mainnet logs for a token called Aether and found that 40% of reported whale movements were internal swaps. That report killed a $2 million allocation and taught me the rule I still operate by: truth is found in the hash, not the headline. In 2021, I mapped the transfer history of a wash-traded NFT collection and watched the floor price drop 60% when the circular graph went public. In 2022, I spent the bear market auditing lending protocols for undercollateralized positions — one oracle-manipulation alert alone prevented a $5 million loss for our fund. Most recently, I spent six months mapping 50,000 wallets to regulatory-compliant labels for an institutional asset manager, which taught me the difference between a data story and one that survives SEC scrutiny: reproducibility.
The framework I apply in every one of those cases is a pre-mortem. Identify the specific red flags that would appear in the ledger before a failure, then wait for them. This article is the pre-mortem for the crypto AI trade.
The leverage layer, quantified
The first evidence chain is funding rates. Perpetual futures on AI tokens have spent most of the past quarter in persistent positive funding — longs paying shorts to hold their side of the trade. At the cycle peak, the annualized cost of holding a long position on RNDR reached roughly 145%. That number identifies the marginal buyer. It is not a GPU user with conviction about Render's job queue. It is a speculator paying rent to maintain a position that spot volume is too thin to absorb.
I tracked the money behind those longs with a Dune query that isolates wallets borrowing USDC against ETH or WBTC collateral on Aave v3 and transferring funds to exchange hot wallets within 24 hours. These are leveraged longs by construction — the user did not sell their ETH; they borrowed stablecoins to buy another asset. Over the past two months, the volume of these borrow-linked inflows into the five AI-token pairs rose more than 200% versus the prior quarter. Reproducible. Public. Timestamped.
The structure has a name in traditional markets: the yen carry trade. An analyst reading AI equities sees margin debt and cross-currency funding. The crypto mirror is stablecoin borrowing on decentralized money markets, where the margin loan book is a smart contract and the liquidation engine is a public list of oracle price feeds. When the funding premium flips negative while the ETH/USDC borrow rate spikes, the unwind path is visible in advance: the highest-leverage wallets get liquidated first, the cascades hit the spot book, and the governance-token purchases stop.
This is the same pattern I analyzed during the Terra collapse, when I was asked to audit the solvency of three lending protocols. The pre-mortem markers were overcollateralized loans shedding value, oracle lags, and concentrated whale positions. Markers in the AI-token leverage book today: open interest out of proportion to spot liquidity; funding regimes implying a one-directional book; and a stablecoin borrow curve rising into falling prices. Each of those appeared in the 72 hours before every cascade I have documented since 2020.
One nuance for institutional readers. A meaningful share of AI-token perpetual volume settles on L2s, and the sequencers for those chains are, in every case I have audited, operated by a single entity. Decentralized sequencing has been a PowerPoint slide for two years. If a liquidation cascade hits during a sequencer outage, the forced-sale backlog will make the correction deeper — and the on-chain record becomes the only reliable account of who exited first.
The compute glut is a depreciation curve, not a demand collapse
The second pillar — compute oversupply — needs translation before it can be tested against the ledger. The claim is that AI compute is becoming abundant and the scarcity premium that justified astronomical valuations is evaporating. The data from actual compute networks tells a more granular story: the oversupply is concentrated in one layer of the stack, and it is a depreciation curve, not an idle-asset collapse.
The distinction that gets lost is between training compute and inference compute. Training demand is pulse-shaped: a lab accumulates GPUs, runs a massive pre-training run, then the cluster idles during iteration. Inference demand is a linear ramp: every agent call, every video-generation request, every API invocation consumes steady, incremental compute. A macro analyst looking at data-center utilization sees idle Hopper racks and concludes "glut." A Dune analyst looking at Docker pod start times sees the split.
My query on Akash's lease history shows a bimodal distribution. A small number of long-duration, high-GPU leases — training runs — spiked in the first quarter and then flattened. A larger number of short-duration, low-GPU leases — fine-tuning and inference — grew steadily month over month. Render's job queue tells the same story: the video-generation wave in February produced a burst of high-value jobs, followed by a quieter plateau. The plateau is not demand collapse. It is the pause between a training-led paradigm and an inference-led one.
The genuine glut is concentrated in the previous generation of hardware. Nvidia's transition to B200/GB200 is pushing H100/H200 units into secondary markets at declining rental rates. On-chain, this shows up as two artifacts. First, GPU-backed RWA projects are repricing their collateral down — a governance proposal to lower the valuation of H100 collateral is the canary. Second, DePIN providers are cutting staking yields because revenue per GPU is falling as new hardware comes online.
This is where my DeFi-summer experience becomes directly relevant. In 2020, I wrote SQL queries tracking impermanent loss across 500 wallets and found that 15% of yield was extracted by front-running bots. The lesson: when a return stream is arbitraged by insiders, the public participants are exit liquidity. The same mathematics applies to GPU tokens. Annualized token emissions across the top five DePIN networks outpace network fee revenue by a factor of three to five. When emissions are the primary yield and fees are a rounding error, the token is not a bet on compute — it is a bet on the treasury's willingness to keep subsidizing its own TVL. That willingness is exactly what a glut tests.
The standardized metric I use is the Emissions-to-Fees ratio. Above 2, the token price is a subsidy chart, not a revenue chart. Every GPU DePIN project I monitor sits above 3. The macro analyst calls it oversupply. I call it an unfunded liability.
The CoreWeave dynamic, with a public ledger
The analyst's note singled out leverage at infrastructure companies — CoreWeave being the archetype — that finance GPU purchases with debt and rely on future lease revenue to service it. The crypto analog is more transparent. Several projects tokenize GPU-backed debt, selling yield to retail while pledging hardware as collateral. The ledger records every payment, every liquidation threshold, and every governance vote on collateral parameters.
I applied the wallet clustering technique from my CryptoClones investigation to the top 500 holders of a mid-cap compute-backed token. About 30% of "independent" holders shared funding addresses with the project's treasury or its seed investors. That circularity inflates perceived distribution, and it inflates the perceived market demand for the underlying GPU capacity. The signature is always the same: project entities trading a project asset to manufacture the appearance of demand.
The specific danger is the collateral loop. A GPU-backed debt token's value depends on GPU resale prices. GPU resale prices depend on data-center capex. Data-center capex depends on AI revenue growth. AI revenue growth, in the current cycle, is partially financed by the same leveraged capital paying rent in perpetual funding. When funding resets and collateral values drop together, the loop unspools from the weakest node. On-chain, the weakest nodes are visible: the wallets at loan-to-value ratios above 80% in compute-backed lending pools.
Underneath it all is a governance-token problem. Holders of these compute tokens have no cash-flow rights. No dividend. No redemption. No claim on lease revenue. The only hope of appreciation is a later buyer paying more — a definition that sits uncomfortably close to a Ponzi structure. The macro analyst calls it a bubble. I call it a non-dividend security with an emissions schedule. The ledger makes the distinction precise.
What the unwind looks like in the ledger
When this breaks, it will not break quietly. The sequence is legible. First, funding rates drop from the triple-digit annualized range toward zero as longs close — that is already underway. Second, the stablecoin borrow curve flattens or inverts; leverage has no reason to refinance an unprofitable position. Third, the first liquidation event hits the compute-backed lending pools — a collateral NFT under water is the tell. Fourth, emergency governance proposals appear: cut emissions, lower collateral valuations, extend unlock schedules. That is the moment the "glut" narrative stops being a macro story and becomes a bookkeeping event.
Correlation is not causation, and the macro analyst is committing a category error. A leverage flush in AI equities does not invalidate the demand curve for inference compute. A margin call in a DeFi token does not invalidate the utility of decentralized exchanges. The stock tape and the hash ledger measure different things, and conflating them is how you sell at the bottom of a liquidity-driven correction.
The second point is the Jevons paradox. If the compute glut is real — if rental rates keep falling as new hardware floods the market — the historically consistent outcome is not a collapse in AI demand. It is an expansion of AI consumption. The API price history is the proof: cost per million tokens fell over 80% between early 2023 and late 2024, and API call volume exploded in the same window. Cheaper inference subsidizes exactly the agentic workloads that crypto compute networks are designed to serve. A macro analyst sees falling prices and says "glut." A data analyst sees falling prices and says "marginal cost is approaching the elastic threshold of demand."
The blind spot in the bearish case is the assumption that the AI trade is a single asset. The ledger disagrees. Training and inference, infrastructure and application, the leverage book and the spot book — separate tables in the same database. The precise read is not "the AI bull market is ending." It is "the leverage layer is unsustainably priced, and its correction will not discriminate between strong projects and weak ones."
The signal I am watching over the next three months is not token price. It is the funding-rate reset on AI perps. The LTV distribution in compute-backed lending pools. The lease-utilization trend on Akash. The moment a GPU-backed debt protocol sends its first emergency governance proposal — to cut emissions or revalue collateral — the word "glut" stops being a macro narrative and becomes a measurable event. Truth is found in the hash, not the headline. The hash is telling me that this unwind is not a question of if. It is a question of which block the first liquidation lands in.
