Your AI Supplier Is Becoming the Stack

OpenAI has published the first measured results for Jalapeño, its custom inference chip. On three public model workloads, the company reports more work per watt, lower end-to-end latency and higher performance for highly interactive use than leading commercially available systems in the comparison. Initial deployment is expected by the end of 2026. ¹²

The specifications are not the executive story. Jalapeño is designed to sit beneath OpenAI's models, serving software and products. Google and AWS are also joining custom accelerators, model platforms and serving layers. ³⁴⁵ The unit of competition is widening from a model to an operating stack.

Inference is the work a trained model performs each time it answers, classifies, plans or acts. If that work becomes faster and more energy efficient, value can appear in ordinary operating terms: shorter wait times, more attempts within a budget, and longer workflows that remain economical. OpenAI reports 1.5–1.9 times more work per watt at peak throughput, 1.7–3.6 times lower end-to-end latency and 2.1–4.1 times higher performance for highly interactive workloads in its tests. ¹

Those results are supplier-authored and cover selected open models on public workloads. They do not establish production reliability, customer pricing, regional availability or a universal advantage. The right response is a bounded business test: compare the total cost and elapsed time required to produce work that an accountable owner accepts.

A tightly integrated stack can be easier to buy and operate. The supplier can tune chips, models and software as one system, remove hand-offs and expose a single commercial surface. That simplicity has value.

It also changes the failure and bargaining surface. A pricing change, capacity shortage, regional restriction, service incident or product redesign can now affect several operating layers at once. Performance achieved through the complete stack may not transfer cleanly to another supplier. Boards should therefore price dependency as part of the return, not treat it as a later architecture concern.

A benchmark can reveal direction, but it rarely represents the complete business process. Ask what it costs to produce one accepted analysis, resolved case, approved design or completed customer interaction. Include model and infrastructure charges, retries, human review, integration, exception handling, controls and delay.

Then ask which part of that result depends on the supplier's integration. A low unit price can be rational even with material dependency. The discipline is to make the trade visible and compare it with a credible alternative before approving scale. The IEA's view of accelerating capital expenditure alongside rapidly falling energy per task is a reminder that efficiency can expand total use rather than reduce infrastructure demand. ⁶

First, keep the organisation's data context, permissions and retention rules separable from the supplier's interface. Second, own the workflow definition: the business rules, tool connections, escalation paths and exception handling. Third, retain the evaluations that decide whether work is acceptable, safe and lawful. Fourth, negotiate commercial terms that make capacity, export and transition possible in practice.

Portability does not mean changing suppliers every quarter. It means being able to test an alternate route, negotiate with credible information and respond when risk or economics changes. The useful proof is a working test, not a clause that has never been exercised.

The newest research batch shows why faster infrastructure cannot replace operating assurance. In SWE Refactor Bench, only 28 of 520 agent runs completed a complex software migration while passing behavioural tests and an independent agentic audit. Another preprint finds that sharing full solutions among multiple agents can remove useful diversity. A third demonstrates that one interaction can poison an agent's persistent memory. ⁷⁸⁹

These are version-one preprints, not production verdicts. Their shared executive meaning is practical: verify the whole workflow, design information boundaries deliberately and treat persistent memory as a security surface. Faster execution makes those controls more important, not less.

Full-stack cloud capacity is deepest in major data-centre markets. Elsewhere, network quality, power availability, language coverage and procurement capability can dominate model performance. Data-localisation and sovereignty requirements may restrict which stack layers can be used or where they run. A global contract does not guarantee equivalent regional economics or service.

The European Union now has enforced AI Act provisions and an operational enforcement framework. ¹⁰ NIST is advancing voluntary interoperability and security standards for agents. ¹¹ Australia's new assurance framework applies to government and its suppliers; it is a useful public-sector comparison, not a general private-sector rule or a global proxy. ¹² Each region needs its own availability, legal and exit test.

Choose one important workflow that already has a clear acceptance test. Record its complete cost, delay, human review and failure path on the preferred stack. Then run the smallest credible alternative: a different model, accelerator route or serving layer using the same data boundaries and evaluations.

The purpose is not to crown a permanent winner. It is to discover what is genuinely portable, what performance depends on integration and which dependency the organisation is willing to fund. Assign an operating owner, a commercial owner and a risk owner. Bring the result back as a decision about value and resilience, not as a technical comparison.