Key Findings
1. NVIDIA’s GPU monopoly economics is the single organizing hub of this entire research effort. It links to 64 other concepts — far more than anything else — and carries the highest strength rating found anywhere in the analysis. At the same time, it is caught in a genuine contradiction: at least a dozen forces actively reinforce NVIDIA’s dominance (the hyperscaler capex supercycle, NVIDIA’s own architecture refresh cadence, circular financing arrangements, Project Stargate, packaging and memory supply bottlenecks, among others), while at least a dozen others are actively eroding it (custom silicon, DeepSeek’s efficiency breakthrough, US chip export controls, the shift toward inference-centric computing, and erosion of the CUDA software moat, among others). The research captures NVIDIA as simultaneously dominant and multiply contested — it does not tell us which side wins.
2. The gap between AI infrastructure spending and the revenue it generates is the primary place where systemic risk converges. Almost every threat pathway in the analysis eventually runs into or out of this capex-revenue gap. It’s amplified by GPU depreciation accounting problems, the hyperscaler capex prisoner’s dilemma, NVIDIA’s circular financing arrangements, the race to zero on inference token pricing, the AI infrastructure “pork cycle,” and Meta’s LLaMA-driven commoditization, among others. At the same time, it is partly masked by manipulated depreciation assumptions and partly offset by the Jevons paradox (efficiency gains that increase rather than decrease demand). The research does not resolve whether the masking or the amplifying forces are larger in magnitude.
3. The Jevons paradox shows up in five separate guises and functions as the primary force stabilizing the system. Five distinctly named concepts — the AI Jevons paradox, the inference Jevons paradox, Jevons paradox in AI compute, “DeepSeek paradox: efficiency amplifies capex,” and the 1000x collapse in AI inference costs — all encode the same underlying logic: making AI compute more efficient increases total demand for it rather than reducing it. Collectively, these are the main counterforce to GPU overbuild risk, to the revenue-to-capex gap, and to a collapse in GPU rental rates. This is the central reason the infrastructure build-out hasn’t self-corrected despite an apparent mismatch between supply and demand.
4. Physical power infrastructure is an independent constraint system, separate from financial risk. A cluster of related concepts — grid interconnection queues and their bottlenecks, AI power demand limits, data center power constraints, and power as a hard ceiling on deployment — form a distinct constraint system. These concepts limit the hyperscaler capex supercycle, GPU overbuild risk, the sovereign AI movement, and the Jevons paradox itself, through pathways that are structurally independent of the financial-risk story. As of Q1 2026, the research encodes electricity delivery — not capital or chip supply — as the binding constraint on the industry.
5. Depreciation accounting is the mechanism linking balance-sheet presentation to real-world asset risk. A cluster of closely related concepts — manipulated useful-life assumptions, the resulting accounting gap, and the risk this creates — captures a specific problem: hyperscalers depreciate GPUs over 5-7 years on their books, while GPUs’ actual economic useful life is closer to 3-4 years. This cluster feeds directly into the capex-revenue gap (by masking it), into the capacity-overshoot cascade (by triggering it), into CoreWeave’s debt exposure (by amplifying it), and into concentration risk for passive index investors (by triggering it). This isn’t merely a reporting quirk — the research treats it as a mechanism that defers risk now and amplifies it later.
Feedback Loops
Loop 1: NVIDIA and the hyperscaler prisoner’s dilemma reinforce each other.
NVIDIA’s monopoly economics strongly amplifies the hyperscaler capex prisoner’s dilemma, and that dilemma in turn strongly enables NVIDIA’s monopoly — and also strongly amplifies the capex-revenue gap. Each hyperscaler’s individually rational decision to spend more on GPUs reinforces NVIDIA’s pricing power, which raises the cost of the next purchase cycle. Nothing internal dampens this loop; the only exits are external shocks — custom silicon reaching real scale, or power constraints binding hard enough to force a stop.
Loop 2: Jevons paradox, the capex supercycle, and overbuild risk — a loop with no clear direction.
The Jevons paradox in AI compute strongly amplifies the hyperscaler capex supercycle, which strongly amplifies GPU overbuild risk — but the same Jevons paradox concept also strongly counteracts GPU overbuild risk directly, at an equally strong rating. Whether the net effect on overbuild is dampening or amplifying can’t be determined from the strengths given, since both pathways are rated identically. The loop is left unresolved.
Loop 3: Token-price deflation, Jevons, and capex reinforce each other with no ceiling.
The race to deflate LLM token prices strongly triggers the Jevons paradox, which strongly amplifies the capex supercycle, which strongly amplifies overbuild risk — and overbuilt supply pushes prices down further, feeding back into token deflation. This is a reinforcing loop with no built-in limit. The one identified counterforce is the demand multiplier from agentic AI workloads, which counteracts the token deflation race directly.
Loop 4: Circular financing, NVIDIA, and CoreWeave’s debt — reinforcing, but fragile.
NVIDIA’s circular financing arrangements fund its own monopoly economics, which strongly enables the GPU-backed debt model used by neocloud operators, which strongly amplifies the capex-revenue gap. Separately, the circular financing loop strongly amplifies CoreWeave’s debt exposure, which itself depends strongly on NVIDIA’s monopoly position holding. This loop reinforces itself as long as GPU values hold — but it has a single point of failure: a collapse in GPU rental rates would very strongly undermine the entire GPU-collateralized debt model.
Loop 5: NVIDIA’s refresh treadmill both causes and counteracts overbuild — net effect ambiguous.
NVIDIA’s architecture treadmill strongly causes the depreciation “time bomb” (obsoleting prior-generation hardware), which strongly amplifies GPU overbuild risk. But the same treadmill also moderately counteracts overbuild risk directly, by shrinking the effective supply of still-viable older hardware. The net sign of this loop can’t be determined from the numbers given.
Loop 6: Export controls backfire through DeepSeek, efficiency, and Jevons.
The US-China chip split very strongly triggered DeepSeek’s algorithmic efficiency breakthrough, which strongly triggered the Jevons paradox, which strongly amplifies the capex supercycle, which strongly funds NVIDIA’s monopoly economics — even as the export-control paradox itself strongly undermines that same NVIDIA position through a separate, direct pathway. Export controls designed to constrain Chinese AI development produced an efficiency breakthrough that, through the Jevons mechanism, increased global compute demand and ended up strengthening the very company the controls nominally targeted. The research treats this as an explicitly self-defeating dynamic.
Non-Obvious Connections
Meta’s LLaMA strongly amplifies NVIDIA’s monopoly economics.
Open-source model releases would normally be expected to commoditize inference and erode hardware pricing power. The research encodes the opposite: by accelerating the token-price deflation race (which triggers Jevons, raising total compute demand) and by enabling inference deployment at much greater scale (raising GPU unit sales), Meta’s open-source strategy structurally benefits NVIDIA’s overall revenue even while it attacks per-token margins. The pathway is indirect, but rated strong enough that the research treats it as significant.
Depreciation accounting manipulation moderately enables NVIDIA’s monopoly economics.1
The same accounting practices that mask the capex-revenue gap also sustain NVIDIA’s revenue, by preventing the demand destruction that would occur if hyperscalers recognized GPU losses at their true depreciation rate. Artificially extending assumed useful life defers write-downs and keeps procurement cycles running. It’s a second-order link connecting a balance-sheet practice to hardware market structure.
Hyperscalers’ nuclear power deals moderately undermine the sovereign AI movement.2
Long-term power purchase agreements signed by US hyperscalers reduce the grid interconnection capacity available overall, pulling power allocation away from national AI infrastructure projects elsewhere. The very mechanism that secures hyperscalers’ energy advantage simultaneously constrains sovereign AI ambitions it appears to have no obvious relationship to — the link runs entirely through power-grid capacity being a fixed, shared resource.
NVIDIA’s architecture treadmill moderately counteracts GPU overbuild risk.3
NVIDIA’s roughly 24-month refresh cadence is usually read purely as a demand-generation tactic. The research encodes a second function: by obsoleting the prior generation of chips, it removes effective supply from the market, working against oversupply. Shorter refresh cycles thus produce both higher depreciation risk for operators and a shorter overbuild window for the market as a whole.
The private-credit financing regime for AI infrastructure strongly enables the externalization of GPU depreciation risk.4
The roughly $800 billion private-credit financing layer is the structural mechanism by which GPU depreciation risk moves off hyperscaler balance sheets and onto private-credit holders instead. This one link connects a financial-market structure, an accounting practice, and a hardware risk — three domains with no obvious surface-level relationship to each other.
Passive-investor concentration risk depends on manipulated depreciation assumptions.
Index-fund investors’ concentration in AI exposure is, in effect, a function of the accounting choices that inflate hyperscaler earnings. If depreciation schedules were shortened to match GPUs’ real economic life, reported earnings would fall, index weightings would shift, and passive investors’ exposure would look different. The link between passive-investing risk and hardware accounting is real in the research but rarely discussed together.
Central Mechanisms
NVIDIA’s GPU monopoly economics is both the primary beneficiary of the infrastructure build-out and the system’s primary point of dependency. It sits upstream of the neocloud debt model, the hyperscaler compute-subsidy moat, circular AI financing, and the profit-distribution stack across AI infrastructure — while itself depending on two other concentrated, high-strength market structures: the HBM memory supply triopoly and the CoWoS packaging bottleneck. With 64 connections into the rest of the research, almost any significant change to NVIDIA’s position would ripple across most of the analysis.
The capex-revenue gap is the primary tension point — 40 connections, rated at the top of the strength scale. It doesn’t drive the build-out (that’s the hyperscaler capex supercycle’s role); it aggregates the risk the build-out creates. More than fifteen concepts amplify it, and at least eight downstream concepts threaten to worsen it further. Three separate mechanisms — manipulated useful-life assumptions, the resulting accounting gap, and circular financing — actively mask it. The gap between how much amplifies this risk and how much masks it is the single biggest structural uncertainty in the research.
The hyperscaler compute-subsidy moat has 33 connections but carries a notably low strength rating (5.6) compared to the 7-9 range typical of the other major hubs. That divergence — highly connected but comparatively weak — is itself notable. It’s eroded by Gulf sovereign AI capital, Meta’s LLaMA-driven commoditization, GPU overbuild risk, the revenue-to-capex gap, the confinement of NVIDIA’s moat to training workloads, and disruption from custom ASICs. The pattern suggests this moat is structurally important to the story but fragile in practice.
The Jevons paradox in AI compute (16 connections) is the research’s primary stabilizing mechanism. It counteracts GPU overbuild risk, the revenue-to-capex gap, and a potential collapse in GPU rental rates — while simultaneously amplifying the capex supercycle, the power-grid hard ceiling, “the great decoupling,” the hyperscaler prisoner’s dilemma, and the shift from training to inference economics. It counteracts the very downside risks it also helps create. Its net effect on the system as a whole isn’t something the research can settle on its own.
The sovereign AI movement, like the compute-subsidy moat, is highly connected (27 links) but rated weak (5.6). It’s amplified by the US-China chip split, DeepSeek’s efficiency dividend, AI-infrastructure-as-military-target dynamics, Gulf sovereign AI capital, and power constraints as a deployment ceiling — while being constrained by the CoWoS packaging bottleneck, data-center power limits, the power grid as an AI bottleneck, and hyperscalers’ first-mover nuclear power deals. The low rating may reflect that sovereignty ambitions depend on scarce physical resources that US hyperscalers already have priority access to.
Tensions & Open Questions
Tension 1: The Jevons paradox both stabilizes and destabilizes at once. It strongly counteracts GPU overbuild risk directly, while also strongly amplifying the hyperscaler capex supercycle, which itself strongly amplifies overbuild risk. The same concept fights and fuels the same downstream problem through two different equally-rated pathways, and the research doesn’t tell us which path is faster, bigger, or more durable.
Tension 2: NVIDIA’s inference position is under threat and defended at nearly the same strength. The split between inference and training markets very strongly undermines NVIDIA’s monopoly economics — while NVIDIA’s own move into inference (via its Groq-related extension) strongly extends that same moat. Both the threat and NVIDIA’s countermeasure are rated at nearly equal strength; whether NVIDIA’s response absorbs the erosion or just delays it isn’t resolved.
Tension 3: DeepSeek’s net effect on NVIDIA is structurally ambiguous. DeepSeek’s efficiency breakthrough strongly undermines NVIDIA’s monopoly directly — but it also, through a separate and even more strongly rated pathway, amplifies the Jevons paradox, which amplifies AI capex spending, which funds NVIDIA. The indirect route carries a higher strength rating than the direct one, but takes two extra steps to land. The net effect on NVIDIA can’t be resolved from the numbers alone.
Tension 4: An explicit contradiction between nuclear power commitments and demand volatility. Nuclear power purchase agreements treated as long-term AI demand commitments directly contradict the “bullwhip effect” describing dramatic AI infrastructure demand swings — one of only a handful of outright contradictions the research records. Twenty-year power commitments lock in long-term demand signals, while the bullwhip effect describes the opposite: sharp volatility. Both are recorded as real, without resolving which one dominates over the medium term.
Tension 5: Agentic AI could resolve overcapacity while accelerating labor displacement. The demand multiplier from agentic AI inference counteracts GPU overbuild risk and could absorb the bullwhip-effect volatility described above — but the same concept also moderately accelerates the shift of income share away from labor and toward capital. If agentic AI solves the overcapacity crisis, it may simultaneously speed up labor displacement. The research records both effects without weighing their relative timing or size.
Tension 6: The hyperscaler compute-subsidy moat funds its own competition. The moat moderately enables hyperscalers’ custom silicon (XPU) strategy, which strongly undermines NVIDIA’s monopoly economics — while the moat itself moderately depends on NVIDIA GPU access in the first place. It’s funded by access to NVIDIA chips and then used to build the alternatives to those chips.
Open question: why do three of the five most-connected concepts carry unusually low strength ratings? The hyperscaler compute-subsidy moat, the sovereign AI movement, and “the great decoupling” are all highly connected but rated at only 5.6 — well below NVIDIA’s monopoly economics and the capex-revenue gap, both rated at the maximum. The research encodes these as structurally central but substantively weaker concepts, without telling us whether that reflects genuine uncertainty about their staying power, contested evidence, or simply their relative novelty as phenomena.
Hypotheses
H1: GPU spot rental prices are a leading indicator for CoreWeave’s debt coverage. The GPU rental market’s capacity barometer strongly triggers stress on CoreWeave’s debt wall, and a collapse in GPU rental rates would very strongly undermine the GPU-collateralized debt model overall. Testable prediction: an H100/H200 spot rental price decline of X% below contract prices should precede covenant stress at CoreWeave by a measurable lag. The full chain — from spot prices, to debt distress, to cascading contagion — is fully specified.
H2: How fast agentic AI workloads scale will determine whether the capex-revenue gap closes or widens in 2026-2027. The demand multiplier from agentic AI inference is the primary counterforce to both GPU overbuild risk and the revenue-to-capex gap. If agentic workloads scale to absorb 30-40% of idle GPU capacity, the Jevons mechanism closes the gap; if they don’t scale — due to model reliability, enterprise adoption friction, or unfavorable cost-per-task economics — the gap widens further. The research supplies the mechanism but no timing signal.
H3: Power grid interconnection queues, not chip supply or capital, will be the binding constraint on AI infrastructure growth for the next 18-36 months. Power-grid interconnection bottlenecks constrain both the hyperscaler capex supercycle and GPU overbuild risk, rated as strong as the HBM memory triopoly and stronger than capital-related constraints. Testable: data center capacity additions (in megawatts) should correlate more strongly with grid interconnection approvals than with GPU delivery schedules or debt issuance volumes.
H4: NVIDIA’s refresh cadence trades a shorter overbuild window for deeper stranded-asset losses. The architecture treadmill counteracts overbuild risk by obsoleting prior-generation supply, but the same treadmill economics strongly causes the GPU depreciation time bomb. Shorter refresh cycles should produce shorter overbuild windows but larger stranded-asset losses per cycle. Testable: compare overbuild duration and neocloud write-down size across the Ampere, Hopper, and Blackwell chip generations.
H5: A depreciation reckoning arrives once one major hyperscaler shortens its useful-life assumptions. Manipulated useful-life assumptions strongly trigger the capacity-overshoot cascade. If a single hyperscaler — likely under SEC or auditor pressure — cuts its assumed GPU useful life from roughly 6 years to 3-4, it should force disclosure of the same gap across its peers, triggering simultaneous write-downs, with knock-on effects for passive-investor concentration risk. Testable: watch 10-K filings for changes to useful-life assumptions in property/plant/equipment footnotes.
H6: Custom silicon adoption should track the growing share of inference versus training workloads. The split between inference and training markets strongly amplifies hyperscalers’ custom-silicon (XPU) strategy, because custom ASICs carry structural cost advantages specifically for inference — fixed workload patterns, optimizable memory access. As the inference share of compute spend grows, the payoff from custom-silicon investment should rise with it. Testable: compare TPU/Trainium/Maia deployment growth against the inference-versus-training revenue splits hyperscalers disclose quarterly.
H7: The sovereign AI movement’s real constraint is grid interconnection and packaging allocation, not capital. The CoWoS packaging bottleneck and the power grid as an AI bottleneck both strongly constrain the sovereign AI movement, even as capital keeps flowing in via Gulf sovereign AI capital and sovereign wealth fund investment. Capital is arriving; physical constraints are what’s actually binding. Testable: compare announced sovereign AI data center capacity against megawatts actually commissioned — the gap should track grid interconnection queue position, not funding announcements.