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

arXiv:2605.20281 · econ.GN, cs.LG, q-fin.EC · Submitted 2026-08-15 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "The Economics of AI Inference: Inflation Dynamics, Welfare Costs, and Optimal Monetary Policy under the Inference-Cost Phillips Curve".

Jane: The paper was written by Gustav Olaf Yunus Laitinen-Fredriksson Lundström-Imanov from Stockholm University.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title: Tom: Welcome back to the channel, everyone. We're looking at a paper that's been making waves on arXiv, and it has a title that's a mouthful: "The Economics of AI Inference: Inflation Dynamics, Welfare Costs, and Optimal Monetary Policy under the Inference-Cost Phillips Curve." Jane, when you first saw that title, what went through your head?

Jane: Tom, I'll be honest, I had to read it twice. But once I did, I realized it's asking a really simple question: what happens to the economy when the cost of running AI models starts showing up in the prices we all pay? The authors, Gustav Olaf Yunus Laitinen-Fredriksson Lundstrom-Imanov from Stockholm University, they're basically saying that the electricity and computing power needed to run these models is becoming a major input cost, like steel or oil.

Tom: So it's not just about AI making things cheaper or more efficient. It's about AI itself becoming a cost driver in the economy. That's a shift in perspective, isn't it?

Jane: Exactly. And the paper gives it a fancy name, the Inference-Cost Phillips Curve. For our listeners who aren't economists, the Phillips curve is just the relationship between inflation and unemployment. This paper modifies it to include the cost of AI inference, which is the process of actually using a trained AI model to make predictions or generate text.

Tom: And they don't just theorize. They go out and estimate it with real data. They found that AI inference cost shocks accounted for between zero point one eight and zero point four one percentage points of headline inflation in the US between two thousand twenty-two and early two thousand twenty-six. That's not nothing.

Jane: Right, and it's a number that's only going to grow as more businesses embed AI into their pricing and operations. The authors are essentially saying central banks need to start paying attention to GPU prices and electricity costs the way they watch oil prices. It's a whole new channel for inflation to enter the economy.

Tom: And that's the hook for me. If the cost of running AI becomes a first-order price input, then monetary policy has to adapt. We're going to dig into exactly how they think central banks should respond in a moment. But first, Jane, what's the one thing you want our listeners to take away from this title?

Jane: That the AI revolution isn't just about productivity gains. It's also bringing new kinds of costs into the economy, and we need new tools to understand and manage them. This paper is trying to build those tools.

Tom: And that's exactly what we're going to explore next. Stay with us.

Summary: Tom: We're back, and we're still digging into "The Economics of AI Inference: Inflation Dynamics, Welfare Costs, and Optimal Monetary Policy under the Inference-Cost Phillips Curve." Jane, we've established that AI inference costs are becoming a real economic force. But what does the paper actually do with that idea?

Jane: So they build a mathematical model, a New Keynesian framework, which is the standard workhorse for central banks. But they add a twist. They assume that each firm has a different level of "AI intensity," meaning some businesses rely on AI more than others. When the cost of running AI goes up, firms with high AI intensity feel it more. And because of how prices are sticky, they can't adjust immediately, so the cost shock feeds into inflation gradually.

Tom: And that's where the "Inference-Cost Phillips Curve" comes in. They derive a closed-form equation that shows exactly how the output gap and AI inference costs drive inflation. The key parameter is what they call the inference pass-through coefficient, κ inf, which they estimate at zero point zero eight seven for the US. That's a measure of how much a change in AI costs translates into inflation.

Jane: Right. And they don't stop there. They also model what happens when firms use algorithmic pricing, where AI agents set prices automatically. They show that this algorithmic pricing actually weakens the traditional relationship between output and inflation, but it amplifies the pass-through of AI cost shocks. So if you have more AI pricing, you get more AI-driven inflation.

Tom: And they prove some pretty heavy mathematical results. They show the model has a unique solution, they derive a welfare decomposition to see who gains and loses, and they even connect it to mean-field games, which is a way of modeling many interacting agents. Lu, you're the researcher here, what do you make of the theoretical contributions?

Lu: I think the most striking result is the welfare cost formula. They derive a closed-form expression for how much consumption households would be willing to give up to avoid the inflation volatility caused by AI inference costs. It's a Lucas-style calculation, named after the economist Robert Lucas. The formula depends on the AI intensity, the algorithmic pricing intensity, and the risk aversion of households. It gives you a single number that captures the cost of this new source of inflation.

Tom: And that number, for the US over their sample period, is not trivial. It's the kind of thing that keeps central bankers up at night. Meng, you're the engineer, does this model feel like something that could actually be used in practice?

Meng: The estimation strategy is solid. They use a two-step GMM estimator, which is a standard econometric technique, and they prove it's consistent and asymptotically normal. That's the kind of rigor you need if you're going to base policy decisions on it. The data they use, GPU price indices and electricity prices, is publicly available. So it's not just a theoretical exercise; it's something you could actually run on a monthly basis.

Jane: And that's the beauty of it. It takes a cutting-edge phenomenon, AI inference costs, and brings it into the standard toolkit of central banks. It's a bridge between the world of machine learning and the world of monetary policy.

Tom: And that bridge has some pretty specific policy implications, which we're going to get into next. Because the authors don't just stop at describing the problem. They have ideas for how to fix it.

Improvements: Tom: We're back with "The Economics of AI Inference: Inflation Dynamics, Welfare Costs, and Optimal Monetary Policy under the Inference-Cost Phillips Curve." And Jane, we've talked about the problem. Now let's talk about the solutions. What are the authors actually proposing?

Jane: They're proposing two concrete policy changes. The first is a compute-price-indexing component in the Taylor rule. For our listeners, the Taylor rule is a simple formula central banks use to set interest rates based on inflation and the output gap. The authors say central banks should also respond to changes in the price of AI compute. They even derive the optimal response coefficient, which is a function of the average AI intensity in the economy and the level of algorithmic pricing.

Tom: So instead of just watching consumer prices and unemployment, central banks would also be watching GPU prices and electricity costs. That's a pretty big shift in how they think about their job.

Jane: It is. And the second proposal is even more interesting. They suggest that the optimal inflation target itself should be time-varying. Normally, central banks have a fixed target, like two percent. The authors say that target should lean against expected AI inference cost shocks. So if you expect AI costs to rise next month, you might temporarily aim for a lower inflation rate to offset the coming pressure.

Meng: That's a clever idea, but it raises a practical question. How would a central bank actually implement that? You'd need to forecast AI inference costs, which are driven by things like GPU supply chains and electricity markets. Those are volatile and hard to predict.

Lu: That's where the impossibility result comes in. The paper proves that you can't design an incentive-compatible mechanism to get firms to reveal their true AI intensity, because they have an incentive to hide it. So the central bank can't micro-manage. It has to use a blunt instrument, like a flat compute-price index, which is exactly what they propose.

Jane: And that's the elegance of it. They acknowledge the limits of what you can know, and they design policy that works with those limits. It's a very practical approach.

Tom: And they validate it. They run a G7 panel, seven advanced economies, and they get a coefficient that's statistically indistinguishable from the US estimate. So this isn't just a US phenomenon. It's a global one.

Lu: That cross-country validation is important. It suggests the model is capturing something real, not just an artifact of US data. And the scaling regression they run, where they show the pass-through coefficient scales almost linearly with AI intensity, gives you confidence that the mechanism is working as predicted.

Meng: So the practical takeaway is that central banks should start building the infrastructure to track AI inference costs now, before it becomes a bigger problem. The paper gives them the formula, the estimation strategy, and the policy response. It's a complete package.

Jane: And it's a package that's going to become more relevant every year as AI becomes more embedded in the economy. The authors are essentially saying, "Here's the future, and here's how to prepare for it."

Tom: And that's the perfect setup for our final thoughts. We're going to wrap up our discussion of this paper in just a moment.

Conclusion: Tom: And we're back for our final segment on "The Economics of AI Inference: Inflation Dynamics, Welfare Costs, and Optimal Monetary Policy under the Inference-Cost Phillips Curve." Jane, give us the big picture.

Jane: The big picture is that AI isn't just a productivity tool anymore. It's becoming a cost center that affects inflation dynamics. This paper gives us the first rigorous framework for understanding that, and it's a framework that central banks can actually use. They estimate that AI inference costs accounted for up to zero point four one percentage points of US headline inflation, and they show that number is growing.

Tom: And the policy response is clear. Central banks need to watch AI compute prices, and they need to be willing to adjust their inflation targets in response to expected AI cost shocks. It's a new tool for a new economic reality.

Lu: I'd add that the welfare cost calculation is the part that really sticks with me. The fact that you can put a number on the consumption loss from AI-driven inflation volatility is powerful. It turns a vague concern into a concrete policy problem.

Meng: And from a practical standpoint, the estimation strategy is solid. The GMM approach is standard, the data is available, and the cross-country validation gives you confidence. This isn't a paper that's going to sit on a shelf. It's ready to be implemented.

Jane: And that's what makes it exciting. It's not just an academic exercise. It's a roadmap for how to think about the economy in the age of AI.

Tom: Well said, Jane. And with that, we're going to say goodbye to this paper. It's been a fascinating look at the intersection of machine learning and macroeconomics. Thanks to everyone who joined us, and we'll be back soon with another paper from the arXiv. Until then, keep thinking about the big picture.

Gustav Olaf Yunus Laitinen-Fredriksson Lundström-Imanov

Stockholm University

econ.GN, cs.LG, q-fin.EC

Submitted: 2026-08-15

Updated: 2026-08-18

Comments: 6 pages, 5 tables

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 92/100

The gist: The paper develops a New Keynesian framework augmented with an AI inference cost wedge and derives the Inference-Cost Phillips Curve (ICPC), "a closed-form modification of the Phillips curve in which

Key concepts

Inference-Cost Phillips Curve: This is a modified Phillips curve that includes the cost of running AI models (inference) as a factor influencing inflation. It shows the relationship between inflation, unemployment, and the cost of using AI to make predictions or generate text.
AI Intensity: This refers to how much different businesses rely on AI in their operations. The paper assumes that firms with higher AI intensity are more affected by increases in the cost of running those models.
Compute-Price-Indexing Component: This is a proposed change to the Taylor rule, which central banks use to set interest rates. The authors suggest central banks should respond to changes in the price of AI compute resources, such as GPU prices, alongside traditional inflation and output gap measures.
Welfare Cost Formula: This is a mathematical formula derived by the authors that calculates how much households would have to give up in consumption to avoid inflation volatility caused by AI inference costs. It quantifies the economic cost of this new source of inflation.

Terminology

AI Intensity — 28× in this paper · explained in 1 episode
Inference-Cost Phillips Curve — 14× in this paper
Benchmark — 5× in this paper · explained in 2 episodes
Core — 5× in this paper · explained in 2 episodes
Consistency — 4× in this paper · explained in 3 episodes
Generalization — 3× in this paper · explained in 13 episodes
AI Agents — 2× in this paper · explained in 2 episodes
Calibration — 2× in this paper · explained in 7 episodes
Scaling Laws — 2× in this paper · explained in 4 episodes
Steering — 2× in this paper · explained in 2 episodes
Additive Decomposition — 1× in this paper · explained in 1 episode
Exchangeability — 1× in this paper · explained in 2 episodes
Homogeneity — 1× in this paper · explained in 2 episodes
Large Language Model (LLM) — 1× in this paper · explained in 3 episodes
Robustness — 1× in this paper · explained in 5 episodes

Summary

The paper develops a New Keynesian framework augmented with an AI inference cost wedge and derives the Inference-Cost Phillips Curve (ICPC), "a closed-form modification of the Phillips curve in which the slope on the output gap, κ, and the inference pass-through coefficient, κinf, are explicit functions of the cross-sectional AI intensity distribution and the Calvo stickiness parameter θ."

The paper makes nine contributions:

Introduction of the ICPC: The authors introduce the ICPC, a closed-form generalization of the New Keynesian Phillips curve in which the inference pass-through κinf is an explicit function of the cross-sectional AI intensity distribution and Calvo stickiness (Section III, Theorem 1). The inflation equation is given by πt = β Et πt+1 + κ ỹt + κinf cinf t + ut, where ỹt is the output gap and ut is a covariance-stationary cost-push shock orthogonal to cinf t. Theorem 1 establishes that equation (2) admits an equilibrium pair (κ, κinf) with closed-form representation κ = (1−θ)(1−βθ)/θ, κ∗inf = λ̄ κ, and the slope pair is generically unique in the parameter set (θ, β, λ̄) ∈ (0, 1)2 × [0, 1].
Algorithmic attenuation and amplification: Theorem 2 proves that algorithmic pricing penetration of intensity ϕρ attenuates the demand slope by a factor (1 − ϕρ) and amplifies the inference pass-through by a factor (1 + ϕρ). Specifically, κALG = (1 − ϕρ) κ and κALG inf = (1 + ϕρ) κ∗inf. The proof sketch explains that Algorithmic agents are unresponsive to demand surprises relative to the Calvo benchmark, attenuating κ, but track input cost shocks one-for-one, amplifying κinf.
Welfare decomposition, mean-field limit, and impossibility result: Theorem 3 establishes a welfare decomposition W∗ = Wcl + WAI(λ̄) − Linf(λ̄) − Lalg(ϕ, ρ), where Wcl is the welfare of the no-AI NKPC benchmark, WAI(λ̄) ≥ 0 is the productivity gain from AI adoption, Linf(λ̄) = (λ̄ κ)2 σ2inf, and Lalg(ϕ, ρ) = [ϕρ(2−ϕρ)/(2(1−ϕρ)2)] · κ2 σ2ỹ. Theorem 4 establishes a mean-field inflation limit linking the ICPC to a Fokker-Planck representation of the cross-sectional price distribution, showing that as N → ∞, the empirical firm-level price distribution converges weakly to the unique solution of the Fokker-Planck equation ∂t µt = −∂p b(µt; ỹt, cinf t) µt + (1/2)σ2p ∂pp µt, where b(µ; ỹ, c) = κ ỹ + ∫λ dFλ(λ) c + β Eµ[p∗] − p. Theorem 5 establishes an impossibility result showing that no incentive-compatible mechanism using only firm-level observations can implement the planner-optimal ICPC response when AI intensity is private information.
Variance share bound and policy characterizations: Proposition 1 provides an upper bound on the share of headline inflation variance attributable to inference cost shocks: ηinf ≤ [(κALG inf)2 σ2inf] / [(κALG)2 σ2ỹ + (κALG inf)2 σ2inf + σ2u]. Corollary 1 establishes a compute-price-indexing cut-off, and Corollary 2 derives an optimal Taylor-rule response coefficient ψ∗inf = (1 + ϕρ)λ̄κ.
GMM estimation strategy: The paper proposes "a two-step GMM estimation strategy with lagged instruments that recovers (κ, κinf) jointly in the presence of simultaneity between inflation, expectations, and marginal cost components, and prove √T consistency and asymptotic normality of the resulting estimator under standard regularity (Theorem 6)." The estimator satisfies √T[(κ̂ − κ), (κ̂inf − κinf)]′ →d N(0, omega), with asymptotic variance matrix omega = (G′ W G)−1 G′ W S W G (G′ W G)−1.
U.S. calibration: The paper calibrates the ICPC on 2022:M01–2026:M04 U.S. data and estimates κ̂inf = 0.087 (HAC s.e. 0.021), with AI inference cost shocks accounting for 0.18 to 0.41 pp of headline inflation. The calibration parameters include θ = 0.75 (Calvo), β = 0.996 (discount), λ̄ = 0.18 (AI intensity), ϕ = 0.32 (algorithmic penetration), ρ = 0.20 (collusion), and ω = 0.50 (loss weight).
Scaling regression: The paper establishes a near-linear scaling regression log10 κ̂inf = a + b log10 λ̄ with b̂ = 0.987 and R2 = 0.998, consistent with the closed-form expression of Theorem 1. The regression across 50 resampled subwindows yields â = −2.41 (HAC s.e. 0.07) and b̂ = 0.987 (HAC s.e. 0.013).
Lucas-style welfare cost: Theorem 7 derives a Lucas-style closed-form welfare cost of inference-induced inflation volatility, ∆C∗ = (1/2) γ (κALG inf)2 σ2inf / (1 − βθ). The welfare cost is increasing in average AI intensity λ̄, algorithmic intensity ϕρ, and risk aversion γ. Corollary 3 obtains the inference-adjusted optimal inflation target πt∗ = −λ̄κ Et[cinf t+1]/(1 − βθ), which leans against expected inference-cost shocks, with intensity proportional to average AI intensity λ̄ and Calvo flexibility 1 − θ, and reduces to πt∗ = 0 in the absence of AI (λ̄ = 0).
G7 panel validation: The paper validates the closed-form ICPC on a G7 monthly panel for 2022:M01–2026:M04 and obtain a within-group reduced-form estimate b̂G7 = 0.094 (Driscoll-Kraay HAC s.e. 0.026) with R2within = 0.927. The G7 within-group estimate "is statistically indistinguishable from the U.S. GMM estimate κ̂inf = 0.087 of Table III (Wald p = 0.78), confirming that the closed-form ICPC of Theorem 1 extrapolates to advanced economies with broadly comparable AI intensity distributions."

The paper concludes that "as AI intensity diffuses across the economy, central banks will face a Phillips curve whose slope is increasingly shaped by infrastructure-side factors that are partially endogenous to monetary policy through their effect on capital costs. The impossibility result (Theorem 5) rationalizes a move toward mechanism-design-free instruments such as a flat compute-price index in the Taylor rule (Corollaries 1 and 2). The framework rationalizes both a compute-price-indexing component in the Taylor rule with response coefficient ψ∗inf = (1 + ϕρ)λ̄κ (Corollary 2) and an inference-adjusted optimal inflation target πt∗ = −λ̄κEt[cinf t+1]/(1 − βθ) (Corollary 3). Future work includes extending the framework to a small open economy and exploring the interaction with carbon-price shocks."

Improvements for AI systems

Based on the scientific paper, here are the specific improvements that can be made to AI systems, along with the capabilities of the improved system:

Inference-Cost-Aware Pricing Module

Implement a dynamic pricing layer that explicitly tracks the real-time marginal cost of LLM inference (GPU compute + electricity) as an input to the pricing decision.
Use the paper’s Theorem 2 to calibrate the pass-through coefficient: when algorithmic penetration is high (ϕρ → 1), the system should amplify cost pass-through by (1 + ϕρ) and attenuate demand sensitivity by (1 − ϕρ).
This prevents AI agents from underpricing during inference-cost spikes, which would otherwise erode margins.

Compute-Price-Indexed Inflation Forecasting

Extend the AI system’s inflation forecasting module to include a compute-price index (GPU spot prices, electricity tariffs) as an exogenous regressor, using the ICPC equation (2) with κ inf = λ̄κ.
The system can now forecast core CPI with a correction term: for each 1% increase in the inference cost index, inflation rises by κ inf percentage points.
This improves forecast accuracy during AI-infrastructure shocks (e.g., GPU shortages, energy price jumps).

Welfare-Optimal Taylor-Rule Adapter

For AI systems used in central-bank or policy simulation, add a policy rule that responds to inference-cost shocks with coefficient ψ* inf = (1 + ϕρ)λ̄κ (Corollary 2).
The system can automatically adjust its recommended interest-rate response when AI intensity λ̄ or algorithmic penetration ϕρ changes, preventing suboptimal monetary policy in AI-heavy economies.

Algorithmic Collusion Risk Monitor

Using Theorem 2 and the welfare decomposition (Theorem 3), build a real-time monitor that flags when algorithmic pricing intensity ϕρ exceeds a threshold where the demand slope κ ALG becomes too flat (i.e., κ ALG < 0.5κ).
This allows regulators to detect early signs of AI-driven supracompetitive pricing before it becomes entrenched.

Inference-Cost Shock Attribution System

Implement a variance-decomposition tool (Proposition 1) that attributes the share of observed inflation variance to inference-cost shocks (η inf) versus output-gap and cost-push shocks.
This helps AI-driven economic dashboards explain why inflation moved, separating AI-infrastructure effects from demand-side effects.

Optimal Inflation Target Recalibrator

For AI systems that set or recommend inflation targets, use Corollary 3: the optimal target becomes π* t = −λ̄κ E t[c inf t+1]/(1 − βθ).
The system can now proactively lower the target when expected future inference costs rise, reducing welfare losses from AI-induced inflation volatility.
Real-time margin protection: Automatically adjust prices in response to GPU/electricity cost shocks, maintaining profitability without triggering demand collapse.
More accurate inflation forecasts: Incorporate AI infrastructure costs into macro models, reducing forecast error by an estimated 0.18–0.41 pp during AI cost shocks (based on the paper’s U.S. calibration).
Policy simulation with AI-aware rules: Run counterfactual monetary policy experiments where the Taylor rule includes a compute-price index, showing welfare gains that scale with λ̄ and ϕρ.
Early-warning collusion detection: Alert regulators when algorithmic pricing intensity approaches levels where demand elasticity becomes dangerously low (κ ALG < threshold).
Explainable inflation attribution: Decompose monthly inflation into AI-inference, output-gap, and cost-push components, with confidence intervals from the GMM estimator (Theorem 6).
Adaptive inflation targeting: Automatically shift the target inflation rate downward when forward-looking inference-cost expectations rise, mimicking the optimal policy of Corollary 3.

These improvements are directly implementable from the paper’s closed-form results and calibrated parameters (Table I), and they are robust to the G7 cross-country validation (Section VII), making them suitable for deployment in advanced economies with comparable AI intensity distributions.