← Futures
Speculation · 2035 · AMP

2035: The model lives next to the site

Escherbridge

Speculation. Not a roadmap, not a delivery date.

This essay is tied to AMP, the operational runtime in the PrimusNeo stack. The first half is current and sourced. The second half is speculation about 2035. It is not a plan, a ship date, or a promise that AMP, or anyone else, will be in that position by then.

What is true now

Data centres already use a measurable share of the world's electricity, and that share is growing faster than the grids that serve them. In its Energy and AI work, the International Energy Agency reported that data centres accounted for about 415 terawatt-hours in 2024, around 1.5 percent of global electricity consumption. The same assessment expects that figure to more than double to around 945 TWh by 2030, "slightly more than Japan's total electricity consumption today." In the IEA's Base Case, consumption reaches almost 1,200 TWh by 2035.

The constraint is not only generation. It is connection. The IEA reported that around 20 percent of planned data centre projects already face possible delays in commissioning because of long waits for grid connections, and that if grid and supply risks are not addressed, 20 percent of expected new data centre capacity from 2025 to 2030 could be delayed. It currently takes four to eight years, the agency said, to build new transmission infrastructure in advanced economies. A model that has to live in a distant hall is, more and more often, a model that has to wait in a queue.

A separate and older argument is about ownership rather than watts. In 2019, Martin Kleppmann, Adam Wiggins, Peter van Hardenberg, and Mark McGranaghan set out a case for "local-first software": software that keeps collaboration without making a remote server the home of the work. Their opening is still the clearest version of the problem. Cloud apps make it easy to work from several devices, but "by centralizing data storage on servers, cloud apps also take away ownership and agency from users. If a service shuts down, the software stops functioning, and data created with that software is lost." Among their ideals are that the network is optional, that work should survive in the long term, and that the people who made the work "retain ultimate ownership and control."

AMP, as it exists today, is a spatial and media runtime, not a completed answer to either of those pressures. The repository named in PrimusNeo materials, plan-systems/amp.SDK, is a public Go SDK for a 3D, geospatial, and media platform, with Unity and Unreal hooks through amp.lib. Its README states that the planet-scale runtime, amp.planet, is "not public." The named repository is a fork of art-media-platform/amp.SDK. As of 9 November 2026, the fork's last commit was 29 July 2025; the upstream project was still tagging releases in November 2026. The public repositories do not show AMP running at a particular site, and marketing copy is not a substitute. What the public code supports is narrower: a toolkit for putting a spatial model on a machine, and a planet-scale piece that is explicitly not in the open.

What follows is speculation

Suppose, in 2035, those two facts have not been repealed. Electricity for large, remote compute is still expensive or late. People who build and maintain places still do not want the working model of those places to vanish if a vendor changes its terms. Then the reasonable place for a living model is next to the work it describes.

Not a metaphor. A machine on the site, or in a building the site's owners can walk into, holding the current geometry, the sensor feeds that still fit on local storage, and the edits made that morning. Collaboration can still happen. Copies can still travel. But the authoritative model does not have to ride a wide-area link for every turn of the camera, and it does not have to die with a subscription. Kleppmann and colleagues were writing about documents and design files. The same prior applies to a site: if the network is optional, the model can be useful on a bad day, which is most days in the field.

AMP is the layer this guess attaches to because a site-owned model is, first, a runtime problem. Someone has to render the place, attach media and spatial data to it, and let more than one person look at the same object without shipping the whole scene to a distant GPU. A local-first instinct does not make that easy. It only says where the data should sit while you try. The closed amp.planet runtime is a reminder that some of the interesting machinery is not inspectable today. A 2035 in which "the model lives next to the site" would require the operable parts of that runtime to be in the hands of the people who own the site.

This is also where the energy numbers stop being a backdrop. If a fifth of planned data centre capacity can be delayed by interconnection, then putting every digital twin in a hyperscale region is not only a question of preference. It is a question of whether the power and the wires will be there when the model is needed. Edge compute will have its own appetite, and a rack beside a job site is not a free lunch. The speculation is only that some of the work — the model you steer a day with, not the archive you query once a year — will have been pulled closer, because waiting four to eight years for a transmission line is not a project plan.

I do not know that AMP will be the runtime in that picture. I do not know that PrimusNeo's connected workflow, which the public materials still describe as in build, will be finished, or that a particular town will be running a site-owned model. Those would be promises. This is a directional bet: if the IEA's queues remain real, and if the local-first diagnosis remains right, then the architecture that works is the one that can keep a place's working model on a machine the place can touch. Everything else is a render farm with a better brochure.

Sources

  1. IEA, Energy and AI, Executive summary (and 10 April 2025 news release)
  2. IEA, “AI is set to drive surging electricity demand from data centres…,” 10 April 2025
  3. Martin Kleppmann, Adam Wiggins, Peter van Hardenberg, and Mark McGranaghan, “Local-first software: you own your data, in spite of the cloud,” Ink & Switch / Onward! 2019
  4. plan-systems/amp.SDK README
  5. art-media-platform/amp.SDK (upstream of the named fork)