Dispatch 08 / Written from 2035
Scenario confidence: high · Evidence cutoff: 12 August 2026
Compute Becomes a Creative Material
By 2035, energy, latency, locality, and model scale will shape the aesthetic—not merely the cloud bill.
AI-generated conceptual scenario art, not documentary or forecast evidence.
The thesis
The next decade will make inference simultaneously cheaper per useful operation and more consequential in aggregate. Creative organizations will treat compute placement as a medium-level decision, composing across device, studio, regional edge, and cloud according to privacy, responsiveness, continuity, energy availability, and the expressive needs of the work.
Evidence / 2026
Signals visible from here.
The IEA projects that data-centre electricity consumption will more than double by 2030 in its base case and models substantial uncertainty through 2035. AI is a major driver, but electricity supply, grids, geography, and deployment timing constrain what nominally cheap computation can become.
Stanford’s AI Index documents rapidly falling inference cost and improving smaller models. MLPerf records major gains across datacentre, interactive, and edge scenarios. These trends support broader placement choices; they do not guarantee that total demand or environmental impact will fall.
Creative teams already experience the trade: a remote frontier model may offer capability, a local model may protect unreleased material, an edge model may enable live interaction, and a queued batch may use energy and budget more responsibly than instant generation.
Inference / 2035
A future history.
A production plan includes a compute score beside the shot list. Live, embodied work reserves low-latency local inference; confidential development stays within controlled environments; exploratory bulk generation follows budget and carbon-aware schedules; exceptional scenes earn frontier-scale computation.
Directors understand model scale as they understand lens, stock, or ensemble size. Deliberately constrained models create coherent vernaculars. Offline generation becomes an aesthetic and resilience choice. Expensive inference is used for moments where additional capability remains perceptible to the audience.
The operating system routes workloads against policy and real conditions, then records the decision. A campaign can explain not only what model touched an asset but where, under whose authority, at what quality threshold, and with what measured resource cost.
Interpretation
What changes for leaders and makers.
- Cloud-only and local-only doctrines both become liabilities. A portable workload and evaluation layer creates negotiating power and continuity.
- Sustainability claims must connect to measured workloads and energy sources rather than generic model labels. Efficiency per inference can coexist with rapidly increasing aggregate use.
- Creative and infrastructure planning should converge early. A spatial live work, private source-library model, and global localization pipeline have fundamentally different compute geometries.
Recommendation / Now
Moves to make before the future hardens.
- 01Classify workflows by sensitivity, latency, quality threshold, availability, volume, and energy tolerance before selecting a provider or model.
- 02Benchmark representative creative tasks across at least two model sizes and two deployment locations; retain quality judgments alongside latency, cost, and energy data.
- 03Make queues, graceful degradation, and offline continuation part of the creative brief for experiences that must survive network or provider loss.
- 04Set an explicit escalation rule for frontier-scale inference so greater expenditure requires an observable creative benefit.
Open questions
What could change this conclusion.
Model efficiency, energy supply, hardware availability, regulation, and rebound effects can move in opposing directions. Forecasts should be revisited annually rather than embedded as a permanent infrastructure assumption.
Comparable energy and emissions data at workload level remains difficult. Teams should state measurement boundaries and avoid false precision.
Primary sources
The evidence beneath the scenario.
- Energy and AIInternational Energy Agency · 2025-04
Models global data-centre electricity demand, supply, emissions, and uncertainty through 2035, including AI as a principal growth driver.
- Artificial Intelligence Index Report 2025Stanford Institute for Human-Centered Artificial Intelligence · 2025-04
Documents rapid reductions in inference cost, improving small-model capability, and the narrowing performance gap between open-weight and closed models.
- MLPerf Inference v5.0 Benchmark ResultsMLCommons · 2025-04
Reports reproducible results from 23 submitting organizations, major year-over-year generative-AI performance gains, and new interactive and edge benchmarks.