Cheaper Tokens, Bigger Bills
Unit token prices have fallen dramatically since the GPT-3 (2020) baseline, yet total enterprise AI bills have continued to climb as usage scales. Inference workloads increasingly shape data-centre energy demand, tying their economics to power geography and goodput.
Budget against total tokens and goodput, not headline unit price. A large share of real-world cost sits outside the GPU invoice, and inference economics are inseparable from power.
~99.7%
Unit token price fall
Since the GPT-3 era (SIG-INF-01)
3×
Enterprise AI bills
Total spend tripled despite falling unit prices
GPT-3 (2020)
Unit-price baseline
Baseline for the ~99.7% token-price fall (CLM-INF-001)
40–70%
Neocloud savings
Specialised providers vs. hyperscaler defaults
$90bn → $465bn
Market 2026 → 2033
AI infrastructure market projection
~54%
Hardware share of 2026 spend
Balance in software, services & operations
The shape of inference economics
AI infrastructure market trajectory
USD bn · ~90 in 2026 to a projected ~465 by 2033
Where inference cost actually accumulates
Directional share of total workload cost (%)
Rental economics by access tier
USD/hr range, H100-class · spot to hyperscaler
The goodput shift
The market has moved from pricing raw GPU-hours to measuring goodput — the useful work performed per pound spent. With a substantial share of true cost sitting in orchestration, retrieval, data movement and observability, and utilisation chronically low, the GPU invoice is far from the whole story. The precise split is a CUDO scenario, not a measured figure.
Inference signals
Evidence-backed signals with confidence and source references. Data as of 15 July 2026.
Token prices collapse while total AI bills rise
Unit token prices fell ~99.7% since the GPT-3 era, yet total enterprise bills tripled as agentic workflows multiply the tokens consumed per task. Falling unit prices are outrun by exploding agentic consumption.
Source: SRC-009
Inference ties AI economics to power geography
Inference now runs continuously at scale, so its economics are inseparable from power price, deliverability and time-to-power. We previously cited an 80–90% share of AI-related energy; that figure is withdrawn pending a source with matching scope and definition, and the direction is retained here only as CUDO interpretation, with no public figure shown.
Source: CUDO view
Much AI cost sits outside the inference invoice
A material share of AI cost sits outside the GPU-hour line item — in orchestration, RAG, vector databases and observability — while enterprise GPU utilisation is often far below optimal. A previously cited 60–80% figure is withdrawn pending a reproducible method; goodput — useful work per pound — not headline GPU price, is the metric that matters.
Source: CUDO view
AI infrastructure market scaling to 2033
The global AI infrastructure market is projected to grow from ~$90bn in 2026 to ~$465bn by 2033, with hardware around 54% of 2026 spend. Inference-led demand underpins the trajectory.
Source: SRC-044, SRC-045
Reading the economics
Turn these signals into a placement decision
Discuss power, location, cluster architecture and TCO with a CUDO AI infrastructure specialist.