> ## Documentation Index
> Fetch the complete documentation index at: https://docs.bordier.com/llms.txt
> Use this file to discover all available pages before exploring further.

# The next AI bottleneck

> An abstract of Bordier's July 2026 monthly insight: how the constraint on AI moved from chips, to memory, to the system that connects and powers them.

Published by **Bordier & Cie (Singapore) Ltd** in its Monthly Market Update series, July 2026.

## The argument

The past decade of artificial intelligence has been defined by shifting bottlenecks. First it was computing power. As supply improved, the constraint moved to memory. The piece argues the industry is now in a third transition, and that **the limit no longer sits inside the chip at all. It sits at the system and infrastructure level.**

The framing is that AI is moving from a chip-centric scaling problem to an **AI factory optimisation problem**. An AI factory means the whole industrial system needed to produce AI outputs: processors, memory, networking, cooling, power conversion and the data-centre infrastructure around them. The economics are measured as cost per token at rack and cluster level rather than as peak performance of any single component.

Three constraints converge.

**Data movement.** Frontier training runs thousands of accelerators in parallel, so interconnect bandwidth and latency govern how well the hardware is used. Copper grows heavier and harder to cool as clusters scale, which is why optical connectivity matters.

**Power delivery.** The constraint is not how many megawatts are available but how electricity is converted, distributed and managed inside the facility. As rack densities rise, traditional power architectures become inefficient and higher-voltage direct-current designs gain ground.

**Power flexibility.** Training is episodic. Inference is continuous and variable. The problem shifts from securing steady-state generation to absorbing volatile load without overbuilding fixed capacity. Storage becomes a form of power inventory, valuable for deferring infrastructure rather than for its cost per unit.

AI capital expenditure therefore broadens well beyond semiconductors, into optics and interconnect, power conversion, thermal management, storage and grid-facing infrastructure. The open question is how quickly that spending converts into revenue.

## Related

* [How thirsty is the AI industry](/insights/ai-water), the resource dependency underneath this build-out.
* [Uranium, the strategic bottleneck](/insights/uranium), where the electricity is expected to come from.
* [What Bordier publishes](/insights/overview), the research series this piece belongs to.
* [Bordier in Singapore](/facts/singapore), the entity that publishes this series.

<Note>
  General information only. Not investment advice, not a solicitation and not an offer. Eligibility, services and terms differ by jurisdiction and by client. Speak to your banker about your own circumstances.
</Note>

*Source: Bordier & Cie (Singapore) Monthly Market Update, July 2026.*


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.