The Economics of AI Inference: Inflation Dynamics, Welfare Costs, and Optimal Monetary Policy under the Inference-Cost Phillips Curve

TEXT ANALYSIS: arXiv econ.GN — "The Economics of AI Inference"

1. THE DISSECTION

What the text is really doing: Conning itself into believing the existing macroeconomic framework is structurally adequate for an epochal rupture.

This paper performs a precise, technical, and entirely wrong operation: it takes the New Keynesian DSGE toolkit — designed for an economy where inflation dynamics trace through nominal rigidities and marginal cost channels anchored to labor — and graft-inserts an "AI inference component λ-bar" into the firm-level marginal cost function. The result is a closed-form Inference-Cost Phillips Curve (ICPC) with structural slope κ*_inf = λ-bar × κ. The econometrics is meticulous. The GMM identification is defensible within its own framework. The R² = 0.998 on rolling windows is genuinely impressive statistical work.

None of which addresses the fundamental problem: you cannot patch a framework that assumes mass employment to describe an economy where mass employment is the variable being eliminated.

The paper's entire architecture requires that firms set prices using marginal costs that include AI inference inputs, that these costs pass through to output prices, that the central bank can influence this via nominal interest rate anchors, and that consumer welfare can be decomposed via a Hicks-Kaldor measure. Every step assumes a functional labor market where human wages anchor aggregate demand. The ICPC is a refinement of a Phillips Curve that already has one foot in the grave under standard post-2020 inflation dynamics — and the paper doesn't even notice it's modeling the death throes of the very curve it wants to calibrate.

2. THE CORE FALLACY

The paper assumes AI inference costs are a factor-input shock analogous to oil price shocks or supply chain disruptions — exogenous disturbances that monetary policy can, in principle, stabilize around.

This is catastrophically wrong. AI inference costs are not a supply shock. They are the mechanism by which the supply side of the economy severs its dependence on human labor. When the authors write "λ-bar × κ" and treat λ-bar as a structurally recoverable parameter amenable to policy optimization, they are modeling the wrong variable. The relevant question is not how monetary policy should respond to inference-cost-driven inflation. The relevant question is what happens to the Phillips Curve itself when the labor market ceases to be the transmission mechanism — when AI systems produce differentiated goods at marginal cost approaching zero, and human wage labor becomes economically optional for the producers who matter.

The paper's "generalized Taylor principle" and "optimal monetary policy response coefficient ψ*_inf" are solutions to a system that has already been superseded. You cannot optimize monetary policy for an economy where the wage-consumption circuit is structurally intact using the same framework that assumes it is.

The paper is an elegant autopsy dressed as a policy prescription.

3. HIDDEN ASSUMPTIONS

• Mass employment persists as the primary channel of aggregate demand. The entire welfare decomposition — Hicks-Kaldor, consumer welfare under inference-cost shocks — assumes human labor income remains the dominant demand anchor. It does not model a scenario where AI-generated income displaces labor income and no transfer mechanism exists to close the gap.

• AI inference costs are exogenous and vary with monetary conditions. The paper treats λ-bar as a structural parameter to be identified from data. But the trajectory of AI inference costs is overwhelmingly determined by semiconductor physics, data center scaling, and algorithmic efficiency — variables that are functionally uncorrelated with the Fed funds rate. The monetary transmission mechanism cannot reach where the inflation is coming from.

• Monetary policy retains meaningful leverage over real economic outcomes in the relevant domain. The Taylor principle, the second-order welfare loss formula, the optimal ψ*_inf — all assume a world where central bank credibility anchors inflation expectations sufficiently to matter. For AI-native production, the relevant price dynamics occur in asset markets (compute, data, model ownership) that are already outside the transmission mechanism of conventional monetary policy.

• G7 cross-country homogeneity (p = 0.78) validates the framework's universality. This is the most statistically sophisticated sleight of hand in the paper. Cross-country uniformity in the measured slope of an inflation pass-through curve tells you that the measurement instrument is consistent — not that the underlying economic structure is stable or that the curve represents what you think it does.

• Welfare measurement via consumer welfare decomposition remains meaningful when productive participation is itself the variable under threat. The entire welfare function assumes a representative agent whose consumption utility is degraded by inference-cost-driven price increases. It does not model a representative agent who is being structurally excluded from productive economic participation.

4. SOCIAL FUNCTION

Classification: Transition Management Theater / Elite Self-Exoneration via Technical Sophistication

This is a paper that performs institutional legitimacy. It says, in effect: "The macroeconomic establishment has the tools to understand AI's economic impact, and those tools are the ones we already have." The authors have produced 80 KB of technically rigorous work that allows central banks, economics departments, and policy forums to believe they are engaging with the AI transition on adequate intellectual terms. They are not. They are describing the heat signature of a building fire using thermometers designed for mild climate variation.

The social function is to absolve the mainstream macroeconomic apparatus of the obligation to develop genuinely new frameworks by demonstrating that existing frameworks, sufficiently augmented, can accommodate the new phenomenon. This is intellectual maintenance of an order that is already structurally compromised. The precision of the GMM estimation and the closed-form welfare loss formula create an impression of control and comprehensiveness that is, structurally, a fantasy.

Secondary function: Career capital accumulation for the authors within the existing academic incentive structure. "The Inference-Cost Phillips Curve" is a title that will generate citations, conference invitations, and policy advisory engagement. It is optimized for the academic reward function, not for illumination of the actual structural transition underway.

5. THE VERDICT

The paper is a sophisticated description of a transient phase dressed as a permanent framework.

The empirical findings — κ-hat_inf = 0.087, near-unit-elasticity pass-through, G7 homogeneity — are likely genuine measurements of the current, transitional inflation dynamic. For the period 2022–2026, as AI inference costs begin to enter supply chains and firms adjust prices with nominal rigidities still operative, the ICPC probably does describe something real. The paper deserves credit for identifying and quantifying this channel.

But this describes the mechanical lag phase, not the structural endpoint. The DT lens is unambiguous: AI inference costs will continue to fall until the marginal cost of AI-native production approaches zero. At that point, the Phillips Curve — even the augmented ICPC — ceases to describe the relevant inflation dynamics. The relevant price level will be determined by compute ownership, energy access, and the distribution of AI capital, not by the interaction of nominal rigidities and marginal cost pass-throughs that the paper's framework assumes.

The authors are measuring the temperature of a patient in the early stages of a disease that will kill the patient. Their thermometers are excellent. Their diagnosis of the disease is wrong. And their prescription — adjust ψ*_inf upward under commitment — is treating the fever while the underlying condition destroys the immune system.

The paper's greatest achievement is demonstrating that mainstream macroeconomics can generate technically impressive nonsense at scale. The GMM is clean. The standard errors are HAC-corrected. The model is in closed form. And none of it touches the structural reality that post-WWII capitalism's consumption-anchoring mechanism is being severed at the source, not stressed at the margin.

Structural Rating: Autopsy Artifact — Relevant for lag-phase analysis, structurally blind to terminal transition.