Reading a conditional scenario
Planning around powerful AI requires a way to think beyond the next product announcement. The reviewed paper uses five horizons measured from 5 November 2026. It combines announced industrial milestones with its own explicit calculations. These are conditional scenarios, with no probabilities assigned. The fastest case is one demanding possibility to examine, and circumstances could produce a faster, slower or interrupted path. The status of the five scenarios · What is assumed and what is reported
Three different quantities are tracked. Terrestrial electricity is the energy used by data centres on Earth. Terrestrial processing capacity is compared using NVIDIA H100-equivalents, a way of putting different equipment on a common performance basis. Orbital capacity is tracked through rated power, deployment and energy, because the future mix of processors is unknown. Energy use and H100-equivalents describe equipment and operation. Neither measures intelligence directly. The three measurement series · The H100-equivalent starting point · Separate sensitivity models · Why orbital processing remains separate
The terrestrial electricity reference follows the IEA's April 2026 path from 485 terawatt-hours in 2025 to 950 in 2030. The paper then assumes 5 per cent annual growth through 2036. Separate low and high paths test different rates. Its processing calculation starts from an estimated 31.6 million H100-equivalents for September 2026, derived from Epoch AI research. Growth in powered capacity and improvements in processing per unit of power compound together. The electricity and processing calculations cover different equipment populations and remain separate sensitivity models. Terrestrial electricity growth assumptions · The estimated processing baseline · Processing growth through 2031 · Later processing assumptions · IEA energy outlook · Epoch AI processing-capacity research
The orbital ambition case begins with annual commissioning rates rising to one gigawatt by the end of 2027, ten by the end of 2028 and one hundred by the end of 2029. The paper selects rates of 200 in 2030 and 300 in 2031, then holds the latter through 2036. Those later selections are its assumptions, informed by the scale in dated company discussions. Rates rise gradually during each year. A rate reaching one gigawatt a year in December therefore adds half a gigawatt over that first year. The orbital commissioning-rate selections · Smooth commissioning and installed stock
The humanoid figures use three cases, always in this order. Delayed adoption assumes slower commercial uptake. Factory expansion assumes repeated expansion of productive manufacturing lines. Accelerated feedback explores a stronger interaction in which AI and robots help expand industrial production. Their starting outputs and growth rates are assumptions selected for the paper. The three figures are different possible paths, and readers can examine or change the assumptions. Read the three production hypotheses and the fleet calculation.
The first three delivery tests
2027 tests delivery. The reference terrestrial energy figure is about 635 terawatt-hours, with processing around 1.6 times the September 2026 baseline. The orbital ambition case adds about 1.2 terawatt-hours during the year. The evidence to watch is complete equipment entering useful service, including promised computer facilities, first orbital computing flights and the initial output of new chip and memory plants. The three humanoid cases start with worldwide production of 100,000, 500,000 or one million machines. The paper’s 2027 calculations · The announced delivery tests · The factory milestones · Humanoid production starting assumptions
2028 tests manufacturing. Terrestrial reference electricity reaches about 726 terawatt-hours. The orbital ambition case reaches six gigawatts of installed fleet capacity and adds about 17.5 terawatt-hours of computing energy. Chip-factory, packaging and satellite programmes have important milestones in this period. A building opening late in the year may make most of its contribution later. Humanoid production reaches 150,000, one million or 2.5 million in the three cases. The paper’s 2028 calculations · Other industrial milestones · Factory and packaging schedules · The 2028 humanoid hypotheses
2029 tests combination. Terrestrial reference electricity reaches about 830 terawatt-hours, while the orbital ambition case uses another 182. Their combined figure is about 1,013 terawatt-hours. The relevant questions concern whole services customers can use and maintain. In the manufacturing outlook, the wafer-capacity index reaches approximately 126 against 100 in 2026. The humanoid hypotheses reach annual output of 225,000, two million or 6.25 million machines. The paper’s 2029 calculations · Wafer capacity and processing · The robot production cases · The orbital cost test
What could be established by 2036?
2031 tests accumulated dependence. The terrestrial reference is about 998 terawatt-hours. Combined with the orbital ambition case, it reaches roughly 3,294 terawatt-hours. Annual humanoid production reaches about half a million, eight million or 39 million. More of an organisation's records, training and procedures may by then sit within a connected supplier system. Arrangements for moving data and maintaining local operation become harder to add once all of those dependencies are established. The paper’s 2031 calculations · Humanoid production and fleets · Accumulated dependence
2036 tests the consequences of compounding. The terrestrial reference reaches about 1,273 terawatt-hours. The combined total is about 2,313 with the smaller orbital case or 11,668 with the orbital ambition case. The orbital component carries its own generation and places no equivalent load on Earth's grid. The humanoid hypotheses give surviving fleets of about 5.2 million, 158 million or 988 million, assuming five-year service lives. Useful labour still depends on reliability, tasks, operating hours and human support. The paper’s 2036 computing-energy cases · The 2036 humanoid cases · Surviving units and useful labour
Assumptions readers can challenge
The orbital calculations assume average operation at 80 per cent of rated computing power and replacement after five years. They also assume orbital equipment is additional to the terrestrial programme. Shifting the same chips or workloads into orbit would require reducing the Earth-based path. Future quantum systems, materials discoveries and manufacturing improvements could change these trajectories when they demonstrate useful performance. Operating fraction and replacement life · Additional orbital equipment and shared supply · Quantum roadmap targets · Atomic-scale manufacturing · IBM quantum roadmap · NIST atomic-scale manufacturing programme
These assumptions make the scenarios open to correction. A reader can watch completed output, production yields, useful work, replacement rates and expansion time. The point of the horizons is to make decisions while options remain open and to revise the plan as the physical evidence arrives. The physical requirements · Indicators readers can inspect · Testing the feedback strength