200× cheaper tokens,
billion-dollar clusters.

01 / Inference deflation vs. frontier capex

Why marginal token prices fell 10×–200× in three years while pre-training cluster costs exploded.

■ Sparse MoE, Distilled & Open-Weight Tiers (≤$4.50 / 1M tokens)■ Legacy / Dense Frontier Reference Tiers ($27–$36 / 1M tokens)● White marker = MMLU-Pro / GPQA Reasoning Accuracy (%)

Modern AI unit economics obey a sharp two-part tariff. On the marginal side, algorithmic innovations — Sparse Mixture-of-Experts (routing 37B active parameters out of 671B), Multi-head Latent Attention (compressing KV-cache memory traffic by 75%–88%), and FP8/INT4 quantization — combined with custom inference silicon (TPU v6e, Blackwell B200) have driven the price of GPT-4-level intelligence down by ~68%–85% per year. On the fixed-cost side, pre-training a frontier base model requires 100,000+ interconnected accelerators and 100–500 MW of baseload data-center power ($130M to $1B+ per run).

02 / The 6-layer AI cost stack & Jevons Paradox

Memory bandwidth walls, test-time reasoning token multipliers, and enterprise cascade routing.

Why hasn't a 99% drop in per-token pricing caused hyperscaler AI revenue to collapse? Because of Jevons Paradox: test-time reasoning models (Gemini 4 Argon High Thinking, OpenAI o3, DeepSeek-R1) generate 5,000 to 32,000 internal chain-of-thought tokens per query to solve multi-step engineering, coding, and mathematical proofs. For production workloads, pairing a fast distilled router (85% of traffic on Gemini 3.8 Flash at $0.18/1M) with a deep reasoning escalation tier (15% on Gemini 4 Argon High Thinking at $3.50/1M) achieves an 80.6% reduction in total monthly spend ($0.68/1M blended) without sacrificing frontier accuracy.

"""Reproducible Pandas analysis for ST-08: AI Compute, Token Inference & Training Cost Economics."""
import numpy as np
import pandas as pd

# 1. Load Frontier vs. Open-Weight AI Model Cost & Capability Benchmark Table
df = pd.read_csv("studies-ai-cost-models.csv")
baseline_cost = df.loc[df["id"] == "GPT4-2023", "blendedUsdM"].iloc[0]
df["deflation_multiple_vs_2023"] = (baseline_cost / df["blendedUsdM"]).round(1)
df["accuracy_points_per_dollar"] = (df["benchmarkPct"] / df["blendedUsdM"]).round(2)
df["queries_per_kwh"] = (1000.0 / df["energyWhQuery"]).round(0)

print("=== ST-08: AI Inference Deflation & Training Capex Matrix (Pandas) ===")
print(df[["id", "blendedUsdM", "deflation_multiple_vs_2023", "trainCapexUsdM", "energyWhQuery", "benchmarkPct", "accuracy_points_per_dollar"]])

# 2. Enterprise Cascade Routing Simulation (100M monthly tokens: 85% Flash + 15% Argon High)
flash_cost = df.loc[df["id"] == "FLASH-38", "blendedUsdM"].iloc[0]
argon_cost = df.loc[df["id"] == "ARGON-HIGH", "blendedUsdM"].iloc[0]
all_argon_monthly = 100.0 * argon_cost
cascade_monthly = 100.0 * (0.85 * flash_cost + 0.15 * argon_cost)
savings_pct = (1.0 - cascade_monthly / all_argon_monthly) * 100.0
print(f"\n100M Token/Month Enterprise Workload: All-Frontier = ${all_argon_monthly:,.2f} vs. Cascade = ${cascade_monthly:,.2f} ({savings_pct:.1f}% savings)")