The data centre industry has lived by a simple maxim for decades: “copper where you can, optics when you must.” Copper has historically been the default choice for transmitting information because it is cheap, simple, low latency and easy to install. Optics has been deployed where copper physically could not do the job — typically over longer distances or where bandwidth requirements exceeded electrical interconnect capabilities.

We believe that maxim is changing. Over the next five years, optics is likely to move from being a solution for the difficult parts of the data centre to becoming an increasingly fundamental part of its architecture. The driver is not simply that AI is creating more data centres. It is that the architecture of those data centres is changing, increasing the amount of data that must move between processors and making electrical connectivity progressively more constrained by power, bandwidth, distance and physical density.

For the Goodhart Global Future Leaders Fund (“the fund”), this is important. We invest with a five-year horizon and so can look beyond the current hyperscaler spending patterns and Nvidia quarterly earnings, but ask the more pertinent question: what will the data centre look like in five years, and what technologies will be required to make that architecture work? Answering this question is much more important than what the total dollar amount spent on capex is and has profound implications for many companies supplying into data centres.

Why copper won in the past

Copper is exceptionally good at transmitting over short distances. An electrical signal can travel directly from one chip to another without converting from electricity into light and back again. There are no lasers, photonic devices or optical receivers, and copper cabling is inexpensive and mechanically straightforward.

That matters because optics introduces complexity. An optical link requires electrical-to-optical conversion at one end and optical-to-electrical conversion at the other, together with lasers, photonic integrated circuits, receivers and often digital signal processing. For a short, relatively low-bandwidth connection inside a server, the additional cost and complexity are difficult to justify. Copper also benefits from extremely low latency and a mature ecosystem.

Consequently, even today, copper remains attractive for short-reach, in-rack connections.  The problem is that the economics change rapidly as bandwidth rises.

Electrical signals suffer increasing loss and distortion as data rates increase, which means either the reach of copper must fall or increasingly sophisticated and power-hungry electronics are required to compensate.

This creates a fundamental trade-off: the faster the computer becomes, the harder it is to move its data electrically.

Why we believe optics must increasingly win

AI is accelerating this problem because the workload is becoming massively parallel. A modern AI system is no longer simply a collection of independent servers. Thousands and ultimately millions of GPUs may need to behave as one computational system, all connected with networking, so they can speak to each other.

The distinction between scale-up and scale-out is therefore important.

Scale-up means connecting processors within the same tightly coupled computing system: GPU-to-GPU, GPU-to-memory or GPU-to-switch, often within a rack or pod. The objective is to make many processors behave like one very large computer.

Scale-out means connecting multiple servers, racks and pods together so that the computational system can become much larger. The distances become greater and the number of connections increases dramatically.

A third concept, increasingly discussed by the industry, is scale-across: connecting AI infrastructure across buildings, campuses or geographically separated facilities.

Historically, optics was primarily a scale-out technology. Copper worked inside the rack; optics was used when the signal had to travel between racks or buildings, but now AI is now pushing optics inwards. In modern AI data centres, the density of compute changes the topology: switches move further away from GPUs and the number of high-bandwidth connections explodes. Optical networking therefore becomes necessary not just between data centres, but between servers, racks and switches.

Why now?

First, bandwidth is rising exponentially. Data rates have progressed from 400G to 800G and now 1.6T, with 200G-per-lane architectures becoming mainstream in leading AI networking platforms.

Second, power is becoming a first-order constraint. AI data centres are increasingly limited not simply by the availability of GPUs but by electricity, cooling and physical density. Every electrical link consumes energy, and the power required to compensate for increasingly lossy electrical paths becomes meaningful at system scale. Co-packaged optics moves the optical conversion closer to the switch ASIC, shortening the electrical path and reducing power.

Third, AI architectures are becoming larger and more tightly interconnected. NVIDIA is now moving its Spectrum-X Ethernet Photonics systems into production, explicitly targeting AI factories scaling towards millions of GPUs[1].

The implication is that optics is no longer simply solving a distance problem. It is solving a power, density and bandwidth problem.

Co-packaged optics: the direction of travel

The likely endpoint is not simply more optical transceivers plugged into the front of switches. Instead, optics increasingly moves closer to the silicon.

Today, an electrical signal can travel from a switch ASIC across a relatively long electrical path to an optical module, where it is converted into light. In a co-packaged optics (CPO) architecture, the photonic engine sits alongside the ASIC in the same package or very close to it. The shorter electrical path reduces losses and power consumption while increasing bandwidth density.

Why is this interesting?

Much of the technology sector is currently being driven idiosyncratically by the latest investor perceptions of hyperscaler capex, worries around circular financing, and where the next bottleneck in the supply chain is going to emerge.

Whilst we think it is important to understand the whole AI value chain and where economic profits will emerge longer term, we believe the greatest near-term opportunities within the hardware space will be found in companies are developing new technologies that are going to be adopted, often looking past most people’s investment horizons. Although some uncertainties on timing of adoption remain, many of the chip companies lay out their technology roadmaps multiple years in advance before mass adoption that makes forecasting easier.

Silicon photonics and CPO is just one example of a technological change we have identified early. Adoption is currently low and so any company serving the space only has limited revenues, but as we move along the S-curve sales can grow rapidly, and earnings multiples could rapidly compress. Given how fast volume growth can be without capacity in place, the market can be tight short-term, potentially supporting elevated profitability, and hence it’s important to understand how much value a company may capture longer term.

Soitec has been a good example of this and has been the biggest contributor to the fund’s performance year-to-date. The company produces engineered wafers used in photonic integrated circuits and has already upgraded expectations for its Photonics SOI growth from 30-40% this fiscal year to 150-200%, taking revenues to over $250m[2]. Longer term, we think this could go much higher given that we are still in the very early innings of adoption, and with a dominant market share protected by unique IP we believe Soitec could be well placed to capture significant value as the market develops.

There are, however, meaningful risks to this thesis: adoption of silicon photonics and CPO could be slower than expected, while technological substitution, semiconductor cyclicality, customer concentration and increasing competition could affect both the pace of growth and the value ultimately captured by supplied such as Soitec.

More broadly, we aim to find companies where we see asymmetric returns under similar technological adoption scenarios that are being missed by a market which remains overly focused on short-term earnings. Given the rapid pace of adoption and compression of technology cycles, identifying these trends early, before they have fully manifested in company P&Ls, is key.

Soitec is discussed here as an illustrative example of the type of technological adoption we seek to identify and should not be regarded as a recommendation to buy or sell the company’s securities.

[1] https://www.youtube.com/watch?v=CNM6Mmqrfeg

[2] Soitec company guidance as at 02 September 2026