Search the Practicum

Type two letters · ↑↓ select · Enter open · Esc close

Type at least two letters.

AI rails map: where is your backup route?

This is a live Practicum page for Volume 2, Chapter 9, "Physical AI and Who Owns the Rails." Verified: 2026-08-18. Updated: 2026-08-18.

The chapter's principle: use large platforms, but do not build a critical business as if pricing, access rules, and compute availability will never change.

Cheap intelligence runs on someone else's rails: models, clouds, APIs, data centers, energy, payment systems, marketplaces, and customer channels. This page helps you see where a critical process depends on one owner and choose a backup route before you need it.

Three layers of the map

Layer What to check Backup route
Model Which model runs the critical process: one closed API, several providers, an openly available model, or a local deployment? Keep prompts and data formats compatible, and test two or three models on the same task.
Data and prompts Can you export your prompts, logs, quality ratings, data sets, and test results? Store source material and evaluation data outside the provider's interface.
Integration Where does the business logic live: your code, the provider's no-code workflow, or a closed agent interface? Move critical rules into a layer you own, such as a repository, rule table, contract, or evidence trail.

Quick dependency test

Answer five questions for every AI process that already matters to the business.

  1. What breaks if the main AI provider triples its price?
  2. What breaks if the API is unavailable for a week?
  3. Could you move the task to another model in one day without rewriting the process?
  4. Where do the data, prompts, and quality criteria live: with you or inside the provider's interface?
  5. Which one node could you move from rented to owned this quarter?

How to read the model market

You do not need to become an ML engineer. Four categories are enough:

Category When it helps Constraint
Closed market leaders best available quality, quick start, strong multimodal and agent features price, limits, access policy, vendor lock-in
Several API providers reduces dependence on one interface requires shared quality tests and compatible prompts
Openly available models control, portability, local adaptation, sensitive data infrastructure cost, support, and quality must be measured
Local inference data stays with you; cost is predictable under stable load hardware, DevOps, model updates, and responsibility for security

Minimum artifact

Build a one-page table:

Process Main rail Backup rail What we own What we check once a quarter
Customer support Prompts, knowledge base, evaluation set Price per request, quality, limits
Text generation Templates, tone guide, archive of strong answers Portability, editor changes
Analysis and reporting Source data, review rules Errors, cost, response time
Code and automation Repository, tests, review process Security, license, fallback

If the "What we own" column is empty, you are not building a process on a platform. You are renting the entire process.

Order-of-magnitude estimates for 2026

These figures explain why a backup route matters. The cost of current token services may be an adoption phase, and prices may rise. Every figure below is an analyst forecast. Check the primary source, including Goldman Sachs work from 2025-2026:

  • Big Five hyperscaler AI infrastructure spending is forecast at about $725 billion in 2026, almost three times the roughly $256 billion spent in 2024.
  • US data-center electricity demand is forecast to rise from 31 GW in 2025 to 41 GW in 2026, then to about 66 GW. Its share of peak summer demand may rise from about 4% to about 8.5%.
  • The power grid is the bottleneck. Estimates call for about $720 billion in grid investment. Current US power shortages are estimated above 11 GW, rising to about 40 GW by 2028.

What changed in 2026: the queue moved from chips to transformers

This is the most useful correction on the page, and it changes where you should look for risk.

In 2022 and 2023 the binding constraint on building an AI data center was almost always the GPU. That eased: TSMC has repeatedly expanded its advanced packaging capacity since 2024. The queue did not disappear, it moved. The constraint now is transformers, switchgear, and grid connection.

  • Roughly half of the US data centers planned for 2026 have been delayed or cancelled, and only about a third of the 12 GW slated for this year is under active construction.
  • Transformer lead times average about 128 weeks, roughly two and a half years. Generator step-up units run to about 144 weeks.
  • The manufacturing bottleneck is narrower than it looks: it centres on grain-oriented electrical steel, and roughly 60% of global transformer production capacity sits in China.

Why a reader outside the industry should care. A chip shortage is a supply problem that money and a new fab eventually fix. A transformer shortage is a physical queue measured in years, and it sits behind every promise about cheap inference. When someone tells you compute will keep getting cheaper on schedule, this is the question to ask: scheduled against whose grid connection?

It also relocates the geopolitical question. The public argument is about export controls on chips. The quieter dependency is a specialty steel and a transformer supply chain concentrated in one country. Treat all of these as analyst estimates and check the primary source before quoting them.

The business conclusion is the same as in the chapter. Do not panic. Keep a backup route and know how much AI and platforms contribute to your cost base. Use the Platform Rent Map for a one-number diagnosis.

Update rule

Update this page after a major change in:

  • prices or limits at leading AI platforms;
  • availability of openly distributed models;
  • local inference costs;
  • data-processing and corporate-security rules;
  • the practical abilities of agent interfaces.

Sources (verified 2026-08-18)

Compute, energy, and hardware sources for ownership of the rails:

Treat capital-spending and energy figures as estimates from a named organization and year, not facts. Keep forecasts separate from measured data, and check disputed claims in the primary source.

Version: 2026-07. Next review: before the final publication of Volume 2.

AI rails map: where is your backup route?