The headline reads: Howard Marks is not convinced the AI boom is a bubble. That sentence is a compression function. It takes a conditional, price-dependent judgment and outputs a bullish endorsement. Between input and output, the conditional clause โ "at what price" โ has been garbage-collected.
I have spent enough time reverse-engineering interfaces to distrust them. The interface is a lie; the backend is the truth. When a financial outlet, and then a crypto outlet, relays a value investor's position, you are not reading the position. You are reading the serialization layer the outlet chose to expose. This particular serialization dropped the only field that made the claim operational. Marks' framework is not "is the technology real." It is "what did you pay for it." Those are different opcodes. The headline executed one and printed the other.
To understand why that matters, you need his actual evaluation function. Across his memos, the bubble criteria are structural, not emotional: psychological overheating, price fully or over-fully discounting the optimistic scenario, and exhaustion of the marginal buyer. Notice what is absent from that list โ magnitude of the technology, novelty of the narrative, importance to civilization. None are inputs. A technology can be genuinely transformative and still be a terrible purchase at the wrong price. Railways transformed continents; a large fraction of railway investors near the capex peak lost money. Radio, aviation, the 1999 internet โ same pattern. The industry wins. The late capital loses.
So when the headline flattens "not convinced it's a bubble" into something that reads like "the authority says it's fine," it performs exactly the narrative processing that Marks' own framework would flag as a warning sign. That is the first systemic fragility here, and it is not in the AI models. It is in the media layer that translates them.
Now the second layer: why is a crypto outlet covering an AI-bubble debate at all? Because the audiences have merged into a single risk-appetite trade. AI equities and crypto assets now move on the same liquidity impulse. If AI is a bubble, the implicit question for a crypto reader is whether their own assets are the same trade with worse collateral. That framing pressure shapes topic selection and phrasing. The publication's incentive is not neutral, even when the source is. I have watched this industry invent a problem and then sell the cure โ "liquidity fragmentation" was the cleanest recent example. The AI-bubble debate has the same tell: a narrative with a distribution strategy attached.
Here is where I stop reading headlines and start reading the balance sheet, because the real answer lives in a line item almost nobody quotes: the capex-to-revenue ratio, and the depreciation schedule underneath it.
As of mid-2025 โ and I flag these figures need verification, because the debate is extremely time-sensitive โ the four major North American cloud providers were guiding to roughly $300โ350 billion in annual capital expenditure. The combined annualized revenue of the leading model companies was in the low tens of billions. Even generously counting AI-attributable cloud revenue, the capex-to-revenue ratio sits at a level rarely seen outside infrastructure manias. This is the hardest data in the entire debate, and the Marks headline article contained not a single number from it.
If capex runs an order of magnitude ahead of the revenue it is meant to serve, the valuation is a bet on revenue acceleration. Bets are allowed. But the bet is now being financed through structures that make it invisible โ and this is the part I care about as someone who has audited key-generation processes and knows how risk migrates to wherever the light is dimmest.
First, the depreciation assumption. Cloud providers depreciate AI servers over roughly five to six years. If the economic useful life of a GPU under current iteration cadence is closer to two to three years, reported profit is systematically overstated โ not through fraud, but through an assumption divergence. The profit impact can be tens of percentage points. When I audited the Dutch pension fund's HSM integration, the side-channel risk was not in the crypto primitive; it was in the key-generation assumption nobody had questioned. Same pattern. The vulnerability is never in the headline primitive. It is in the assumed parameter.
Second, the financing structure. Unlike 2000, a meaningful share of this capex is funded off-balance-sheet: data-center SPVs, joint ventures, private credit, and vendor financing where chipmakers invest in customers who then buy more chips. That is circular financing with extra steps, and it means the first fracture will likely appear in credit spreads, not equity multiples. Watch the ABS and private-credit spreads tied to data centers. Watch GPU lease spot prices. Equity is the lagging indicator; credit is the assembly. This is the same lesson the cross-chain world keeps relearning at a cost of more than $2.5 billion in cumulative bridge exploits: value concentrates at the transfer chokepoint, and the chokepoint is where the transparency is thinnest.
Third, the concentration. The top technology names are now roughly a third of S&P 500 market cap โ far beyond the single-sector concentration of 2000. That converts an "AI repricing" from a sector event into an index event, transmitted through passive flows and pension exposure into household wealth and consumption. This is the difference between a correction and a macro event.
And here is the crypto-native observation the report kept circling: the profit pool is concentrated at the compute-supply layer, while the application layer gets its margins eaten by inference cost. That is the same structural asymmetry I have documented in DeFi composability โ value accrues to whoever controls the scarce resource, and the application is the residual claimant that absorbs the volatility. When supply loosens, the profit pool reallocates, and that repricing hits harder than any demand-side wobble.
The contrarian read is not "AI is a bubble" or "AI is not a bubble." It is that the entire debate is conducted in the wrong layer โ equities โ when the actual failure mode is credit and accounting. Everyone argues forward P/E and EV/Sales. Nobody argues the depreciation schedule, and the depreciation schedule is where profit quality lives. Read the assembly, not just the documentation: the documentation says "record revenue and capex." The assembly says "depreciate the GPU over five years while iterating the workload every eighteen months." Those two statements cannot both hold in a steady state.
There is a second blind spot, and it is reflexive. If the "AI is a bubble" consensus hardens, capital withdraws early, and that withdrawal genuinely slows the technology's deployment โ the narrative becomes self-fulfilling. That is a worse outcome than the bubble itself, and it is structurally unavoidable once the debate goes mainstream. Which brings me to the signal almost nobody prices: the debate reaching crypto media, moving from growth investors to value investors, is itself a cycle-phase data point. Historically, that diffusion marks the middle-to-late stage, not the beginning.
So what do you monitor? Not the headline. The leading indicators: GPU lease spot prices on H100/H200 โ a sustained decline is the earliest demand-softening signal; data-center ABS and private-credit spreads โ the first place credit fractures; and any downward revision to GPU depreciation schedules โ the cleanest tell that profit quality is being repriced.
The vulnerability forecast is not that AI is fake. It is that the accounting parameters and the financing structures are running ahead of the verified revenue curve, and the fracture, when it comes, will surface in the layer with the least disclosure. Trace the logic gates back to the genesis block: the question was never whether the technology is real. It was what the ledger actually says you paid for it โ and whether the depreciation line is telling the truth. Most people are still reading the interface.