INSIGHTS · 28 OCTOBER 2026
AI agents are turning the chip race into a whole-system competition
AIstratica Editorial · Published 28 October 2026 · Evidence checked 28 October 2026
More accelerator capacity is only one part of making AI useful. Agents also retrieve information, execute software and wait for tools. Recent infrastructure announcements make that distinction important for both technology buyers and the economics of the industry.
What is new: infrastructure for the work between model calls
On 27 October 2026, NVIDIA and Oracle announced multi-gigawatt expansion plans for OCI Superclusters and general availability of Vera Rubin NVL72 and Vera CPUs on OCI. Their announcement distinguishes GPU-heavy model work from CPU-heavy agent tasks and describes independent CPU/GPU scaling. Planned expansion is not the same as completed capacity. Primary source: NVIDIA, 27 October.
Separately, NVIDIA’s 26 October release says Meta plans broader NVIDIA infrastructure deployment and joint optimisation of Muse models and agents. It also announces a US$3 billion NVIDIA investment in Meta’s equity offering. Potential security collaborations are described as under consideration, not delivered features. Primary source: NVIDIA, 26 October.
The U.S.–China dimension
A 19 October Institute for Progress report argues that compute access and model distillation affect the U.S.–China capability gap. It advocates tighter controls and models possible effects. This is a policy argument with assumptions, not an enacted rule or a guaranteed forecast. Primary research: IFP, Tim Fist, Georgia Adamson and Konstantin Pilz.
AIstratica analysis: the competition extends beyond where a chip is manufactured. Access to cloud capacity, networking, software compatibility, power and deployment expertise can affect which organisations turn model capability into dependable services. Domestic substitution may reduce one supply dependency while creating software-conversion costs. Open weights can widen deployment options, but cannot eliminate the need for compute.
Economic implications: capacity is an input, not a payoff
AIstratica analysis: infrastructure spending can benefit equipment and service suppliers before end users see productivity gains. For an adopter, the relevant denominator is a completed, accepted task. An apparently inexpensive model can be costly if it retries repeatedly or needs extensive human correction; expensive infrastructure can be productive when it is highly utilised on valuable work. Announced gigawatts alone resolve neither question.
Track three signals: actual availability rather than roadmaps; utilisation rather than installed capacity; and customer savings after review, integration and support costs. Financing links between suppliers and customers also warrant attention when interpreting demand. They are not, by themselves, evidence of improper activity or a particular investment outcome.
A buying checklist for the next six months
- Benchmark a complete workflow with its tools and review step.
- Ask which capacity is available in your region now and which is planned.
- Keep an exportable copy of prompts, evaluations and business data to reduce switching costs.
- Price a second deployment option and a manual fallback for critical tasks.
- Treat policy proposals separately from enforceable access restrictions.
Read today’s sourced briefing and our dated chip-policy explainer. This article offers technology and economic analysis, not investment advice.