IBM introduced System/360 in 1964 as one compatible family of computers spanning a wide range of performance. Customers could preserve software investments as they moved within the family. IBM could spread one architecture across engineering, manufacturing, support and sales.
The programme cost about $5 billion over four years by IBM’s historical account. It helped turn an already dominant supplier into the organising centre of enterprise computing.
AI now has a mainframe-shaped core.
The word describes the economic structure, not the machine. Training frontier models requires large pools of accelerators, energy, networking, capital and specialist labour. Cloud providers and model developers reinforce each other through investment, capacity agreements and distribution. Customers receive extraordinary capability without owning the foundations underneath it.
The analogy has a limit. AI software can be copied. Inference costs can fall. Open-weight models can run outside the largest labs. Applications remain more contestable than the upstream core.
Both sides matter. Concentration at the centre can coexist with competition at the edge.
Capital sets the frontier
Stanford’s 2026 AI Index counted $581.7 billion in global corporate AI investment during 2025. That category includes mergers, minority stakes and public offerings, so it should not be confused with infrastructure spending. It still shows the scale of capital moving around the sector.
Public-company filings show large technical-infrastructure capital programmes, but they do not isolate a clean AI total. The structural point is narrower: participation at the upper end requires balance sheets able to fund data centres, chips and energy before demand is fully known.
Reconstructed model-training estimates are not audited development budgets, so they are not used here as a second numerical case. The filings already establish the structural point: only a smaller number of organisations can fund repeated frontier experiments, absorb failed runs and reserve capacity years ahead.
This is the first mainframe property: scale becomes a product advantage before the product reaches the customer.
The bottlenecks reinforce one another
The OECD’s 2026 study of AI markets identified persistent concentration risks across chips and cloud. It cited estimates of three providers holding 74 percent of the global cloud market in 2023 and Nvidia holding 90 percent of that year’s GPU market. Those figures describe specific markets and a specific year, not the whole AI economy. They show where leverage can form.
Compute also has physical dependencies. Advanced accelerators depend on specialised manufacturing, memory, packaging, networking and energy. A model developer may choose among clouds while remaining exposed to the same upstream supply chain.
Cloud and model partnerships add another loop. Capital can arrive together with compute commitments, distribution and technical integration. The model helps the cloud sell capacity. The cloud gives the model a route to enterprise customers. Leaving one may mean rebuilding parts of the other.
Data creates narrower advantages. No company controls “the data.” Public datasets, licensed archives, synthetic data and customer-specific records all behave differently. Some proprietary datasets and large-scale interaction histories can still improve products in ways a new entrant cannot reproduce quickly.
Talent follows the infrastructure. Industry can offer access to compute, large technical teams and compensation that academic institutions often cannot match. The useful point is structural. Capital attracts scarce expertise. Expertise improves systems and infrastructure. Better systems attract more capital and users.
The edge is different
The mainframe analogy becomes misleading when it swallows the whole market.
Inference prices have fallen. Smaller models can handle useful tasks. Open-weight releases let organisations deploy, adapt and inspect systems outside a proprietary endpoint. Applications can combine several models and route work by cost, latency or risk.
These options do not remove dependence. A downloadable model still needs hardware, serving software, security and operational skill. Porting the model does not automatically port prompts, evaluations, tool integrations or operating history.
But the alternatives are real. They make the application edge more contestable than the classic mainframe terminal. A small team can build on capabilities that recently required a research lab. An organisation can reserve frontier models for hard tasks and use cheaper systems elsewhere.
My structural assessment as of 27 September 2026 is that the market is splitting. The core is concentrating around expensive general capability while the edge fragments around domain knowledge, workflow integration and trust.
In this assessment, firms at the core can often move outward while edge firms face higher barriers moving upstream. A sustained fall in frontier capital requirements, broader access to advanced compute, or repeated edge-to-core entrants would overturn it. That asymmetry matters more than the number of products in a directory.
Sovereignty is an infrastructure response
Governments have noticed the dependency.
The United States created a programme to export full-stack packages covering chips, servers, cloud, data pipelines, models, security and applications. The European Union describes technology sovereignty as greater control over critical technologies, data and infrastructure while reducing strategic dependence. Its AI Factories use shared public compute to widen access for researchers and companies.
These policies differ in philosophy, but they respond to the same structure. AI capability rests on industrial assets that can become geopolitical leverage.
Sovereignty does not require every component to be domestic. That standard would make most current programmes fail on contact with the supply chain. A more useful test asks whether a country or organisation has credible alternatives, negotiating power and the ability to keep essential systems operating when a supplier changes terms.
The same structural test applies inside an enterprise.
What dependency carries into the application
Access to a powerful model can become cheaper while dependency grows elsewhere.
An application may face low token prices and still have high switching costs because its evaluations, safety cases, identity flows and operational data are tied to one platform. A provider can offer export tools without preserving the semantics of the original system. Two APIs can accept similar prompts and produce operationally different behaviour.
This changes the architecture question. Model quality matters. So do portability, evidence ownership, fallback operation and the boundary between a provider’s control plane and the application’s own authority.
Concentration also brings benefits. Large providers can fund security teams, resilient infrastructure and research that smaller firms cannot reproduce. Centralised capacity can make advanced tools available to ordinary users and small companies at prices they could never achieve alone.
The tradeoff is real. Economies of scale lower access costs. They can also shift bargaining power toward the firms that own the scale.
Which version gets built
The new mainframe is not a single machine and it is not finished.
Its upstream shape is visible in capital expenditure, chip supply, cloud concentration and industry control of frontier development. Its edge remains open enough for model routing, open weights, public compute and domain-specific applications to matter.
The strategic question sits between those layers. Can standards, portable evidence and credible alternatives prevent cheap access from becoming permanent dependence?
The price of intelligence may keep falling. The distribution of control can still move in the opposite direction.