The AI Capex Mirage: Reading Nvidia's Earnings Through a Liquidity Lens

There is a specific kind of document that tells you more by what it omits than by what it contains. Earlier this cycle, a research note surfaced — attributed to DBS, relayed by Crypto Briefing — asserting that AI equities are not in a bubble. It carried no publication date, no fiscal quarter, no analyst name, and not a single quantitative figure. Nvidia's earnings were characterized as "robust." The growth outlook was called "sustainable." And that was the entire argument. I have spent nine years watching how macro narratives get laundered through layers of retelling, and I have learned to treat absence as data. The missing valuation anchor, the missing discount rate, the missing invalidation condition — these are not editorial oversights. They are the structural signature of a sell-side sentiment text, a document engineered to manage positioning rather than to inform it. The question worth asking is not whether the note is right. It is what the note's silences reveal about the liquidity regime that produced it. Liquidity is a mood, not a metric. This is the first thing I teach anyone who asks me to read a market. Before you look at a single earnings line, you look at the tide — the aggregate dollar liquidity that determines whether risk assets rise or fall together, regardless of their individual fundamentals. Right now the tide is high, and it is rising for reasons that have almost nothing to do with the fundamentals of any single company. Institutional capital has been flowing into both AI equities and digital assets through the same channels: passive ETF mandates, wealth-management allocation models, and the corporate treasury desks that treat both as "innovation exposure." In March 2024, I worked with three senior portfolio managers at a Warsaw asset management firm to model how $15 billion in institutional inflows would reshape spot market dynamics. The most important thing we learned had nothing to do with Bitcoin. It was that the traditional risk frameworks these institutions use — value-at-risk, drawdown bands, correlation matrices — completely fail to account for on-chain velocity. The models were blind to the fastest-moving part of the system. That blindness matters, because it means institutional allocators cannot see the leverage they are actually taking. They believe they hold two uncorrelated bets: an AI basket and a crypto basket. They hold one bet, expressed twice. Structure is the skeleton; liquidity is the blood. Nvidia's balance sheet is the skeleton of this cycle. The blood is the global dollar liquidity that funds the capex behind it. And right now, the skeleton has a fracture that the sell-side note refuses to name. Start with circular financing. Nvidia has invested directly in a constellation of customers — CoreWeave, Nebius, xAI, and others — who then use that capital to purchase Nvidia GPUs. Public reporting suggests one such customer has committed to compute purchases on the order of a trillion dollars, while its own revenue run-rate remains in the tens of billions. This is not demand in the ordinary sense. It is a supplier funding its own order book. The pattern should be familiar to anyone who watched 2020's DeFi summer, when I spent forty hours manually tracing $2.5 million in USDC from Compound Finance into Uniswap V2 pools and discovered that decentralized "liquidity" was quietly replicating fractional reserve banking. The mechanism differed; the hidden leverage did not. Then there is depreciation. A GPU's useful life — three years or six — determines whether cloud operators' reported profits are real or manufactured. In the second half of 2025, this became a genuine controversy, with high-profile short-sellers arguing that hyperscalers were extending depreciation schedules to inflate earnings. This is the technical core of the "are AI profits real" debate, and the DBS note does not touch it. It cannot. To discuss depreciation honestly is to admit that the capex cycle's returns are back-loaded into a future that has not arrived. Then there is power. The real constraints on AI compute have migrated away from chips. The binding bottlenecks are now CoWoS advanced packaging, HBM memory supply, and — most critically — grid interconnection and data center power. Hyperscalers have begun signing nuclear power purchase agreements precisely because electricity, not silicon, is the ceiling. A note that still narrates this cycle through the old "chip demand" frame is describing a war with last decade's map. The market's continued fixation on chip supply is a lagging indicator dressed as a leading one. And then there is the question nobody wants to ask: where does the demand end? A September 2025 MIT study, widely cited and methodologically contested, found that roughly 95% of enterprise generative-AI pilots produced no measurable return. This is the empirical battlefield on which the entire capex thesis stands or falls. The sell-side note is silent on it. Competition compounds every one of these fragilities. Nvidia's moat is not uniform; it is bimodal. On the training side, the CUDA software ecosystem, NVLink interconnect, and mature tooling make the position nearly unassailable in the near term. On the inference side, the moat is far shallower — inference is less software-dependent and more cost-sensitive, which is precisely why hyperscaler custom silicon (Google's TPU, Amazon's Trainium and Inferentia, Meta's MTIA, Microsoft's Maia) and AMD's MI series are gaining ground in inference workloads. As inference grows as a share of total compute, Nvidia's pricing power faces structural erosion. The brutal irony is that Nvidia's own margins fund the R&D of its largest customers — who are also its most dangerous future competitors. The higher the margin, the stronger the incentive to replace it. Here is where crypto enters the frame, and where most analysts miss the connection. The AI capex cycle and the crypto cycle are not parallel stories. They are the same story told in two dialects. Both are levered expressions of dollar liquidity. Both depend on the willingness of institutional capital to fund long-duration, cash-flow-negative infrastructure on the promise of future utility. And both are increasingly shaped by the same actor: the algorithm. In August 2026, I published a white paper arguing that AI-driven trading algorithms had captured roughly 60% of high-frequency liquidity in crypto derivatives markets. The reaction was split — some called it techno-pessimism, others called it foresight. But the core finding was structural: automated systems optimizing for short-horizon gains create feedback loops that amplify macro volatility and decouple asset prices from the economic indicators that supposedly anchor them. When the same algorithms run in both AI equities and crypto, the correlation between the two is not a coincidence. It is an artifact of shared plumbing. The crypto side of this equation is where the story gets genuinely interesting, and where the note's blind spot becomes an opportunity for those paying attention. The same capex dynamics that inflate AI equities have spawned an entire class of tokens — decentralized compute markets, GPU rental protocols, DePIN infrastructure networks — that trade as high-beta proxies for the AI trade. When capital floods into AI narratives, these tokens rally hardest. When the capex cycle stalls, they will fall first, because they are the most reflexive expression of a reflexive thesis. The valuation logic here is circular in the same way Nvidia's customer financing is circular: the token price rises because the narrative rises, and the narrative rises because the token price rises. Patterns repeat, but the context never does. In 2000, Cisco was still growing revenue at the top of the bubble, and its product was genuinely essential to the internet. Growth and bubble are not opposites. The definition of a bubble is price running ahead of fundamentals — not fundamentals failing to exist. The prevailing belief is that crypto has "decoupled" from traditional markets, and that AI equities and digital assets represent distinct risk buckets. This is the most dangerous assumption in the current cycle. Illusions fade when the tide of liquidity recedes, and the decoupling thesis is an illusion sustained by abundance. When dollar liquidity is ample, everything rises and correlations appear low. When it contracts, the shared plumbing becomes visible, and the "two bets" collapse into one. The honest reading is this: crypto is not a hedge against an AI unwind. It is a higher-beta expression of the same trade. If the AI capex cycle rolls over — if hyperscaler capex guidance decelerates, if circular financing cracks under a tighter funding environment, if the application-layer ROI is finally falsified — crypto will not be spared. It will be amplified. The protocols that survive will be the ones whose value capture does not depend on narrative reflexivity, and those are far fewer than the market currently prices. The future is written in the present liquidity. Watch the four hyperscalers' quarterly capex guidance, the financing events of AI-native compute firms, and Nvidia's gross margin trajectory. Those three signals will tell you when the tide turns — long before any earnings call admits it. The macro is the mirror of the micro, and right now the micro is showing you a structure funded by its own reflection.