The UK organisation for the business of Open Technology

AI needs its netheads, and they need to learn from the bellheads.

The pace of progress in AI is extraordinary, investment is flowing at a scale few of us have seen, start-ups and scale-ups are beating all records of fundraising, and valuations that previously took tens of years to achieve are now achieved in just a couple of rounds. 

Let’s be clear, AI is an incredible technology and I think it will impact every corner of the economy and each and everyone of us. But the foundations of this technology are being built at the model level by a very small number of companies, on systems the rest of us cannot inspect, in a market that is consolidating faster than policy can respond.

I have seen this story before. I have spent my career in telecoms, and I believe the way our industry evolved is a warning AI cannot afford to ignore.

How we created modern Telecoms under two cultures

In the 1990s, telecoms engineering split into two camps. The bellheads, from the telephone world, believed intelligence belonged in the core of the network. They wanted central control, guaranteed quality and formal standards set by industry bodies. The netheads, from the internet world, believed in a simple network with intelligence at the edges, where anyone could build without asking permission. Their philosophy was summed up in David Clark’s famous line about “rough consensus and running code”, capturing the technical and political values of Internet engineers during a crucial phase in the Internet’s growth.

On the fixed internet, the netheads largely won, and the result was decades of innovation that nobody could have planned. But in mobile networks, the bellheads kept the radio. Even as networks moved to IP, the radio access network stayed vertically integrated, with hardware and software sold as sealed bundles and proprietary interfaces between components. Standards were open on paper, but implementations were not.

What regulators got right, and what they missed

It would be unfair to say telecoms were never regulated for openness. In the UK, forcing access to the local loop and separating Openreach from BT opened the network to competitors. What followed was a new generation of alternative networks (also known as “altnets”) which brought Broadband competition. This shows that where regulation was established – competition followed. 

But we applied that thinking only to operators and services. The equipment layer underneath was left to the market to self-regulate. Nobody required open interfaces between network components, interoperability between vendors or diversity in the supply chain. So the market did what markets do. Nortel, Lucent, Alcatel, Siemens and Motorola merged or disappeared, and the world was left relying on a handful of vendors: Ericsson, Nokia, Huawei and ZTE, with Samsung at the edges.

Closed architectures are the most effective moat a company can build, and any rational incumbent will defend one. That is precisely why we cannot expect markets to open themselves. Setting rules that align private interest with public interest is the job of policy.

What happened after

Concentration had real costs in the entire telco industry: the adoption of innovation slowed, equipment stayed expensive, and connectivity never reached the places where the business case didn’t suit incumbents. Finally, supply chain concentration arrived.

When the UK decided some vendor equipment posed an unacceptable risk, there was no easy substitute. Operators were ordered to remove it at great cost and over years, because the market offered so few alternatives. The Telecoms Security Act and the diversification strategy, with its support for Open RAN, were the right responses. But they came after dominance was established, making it almost impossible for smaller players to compete or even get to the level large equipment providers are.

In 2010, as Huawei’s equipment spread through UK networks, the company opened a centre in Banbury under an agreement with the UK government. Huawei owned and funded it, and government security experts oversaw its work, inspecting Huawei’s code on the premise that if they could see inside the equipment, they could manage the risk. It didn’t work out. In 2019, after nearly a decade of inspection, the centre’s oversight board reported serious and systematic flaws in Huawei’s engineering and could offer only limited assurance that the risks could be managed. Some of the problems could only be fixed by replacing hardware. 

Why does this matter? Because inspection was supposed to be the safeguard, and it proved to be a weak one. Every new software update needed new checks, the supplier controlled what was fixed and when, and the UK had no easy way to switch to alternatives. When trust finally broke down, the only remedy was to rip the equipment out and pay for it.

The lesson is that inspecting a closed system you depend on for something critical, on the supplier’s terms, does not build the trust needed to carry out such a critical element of our society: communications.

AI seems to be in a similar position

The bellhead model concentrates intelligence at the core: a small number of very large models, trained and run by a handful of companies, reached through interfaces they control, on terms they set and can change. Users get access to the intelligence, but not to the system itself.

The nethead model pushes intelligence to the edge, just as the internet did. Open models can be downloaded, inspected, adapted and run by anyone, on their own hardware, with their own data. A hospital can fine-tune a model for its clinical needs without sending patient records elsewhere. A startup can build on a model without asking permission or fearing that access will be priced out or withdrawn. Researchers can study how a model actually behaves rather than relying on what its maker chooses to disclose. And instead of one model doing everything, many smaller, specialised models and tools can be combined, much as the internet combined countless networks and applications that no central planner designed.

This is the end-to-end principle applied to intelligence. The value is created at the edges, by the people who use it, and the core does not decide what is allowed.

The trouble is that the market is drifting towards the bellhead model. Building frontier AI requires specialised chips, vast computing power and enormous capital, and each of those is concentrated in very few hands. Even open models depend on the same small pool of chip makers and cloud providers to be trained and, often, to be run. It is the internet’s lesson repeating itself: openness at one layer does little if the layers underneath consolidate.

Who carries our intelligence?

In telecoms, the question was who carries our communications. In AI, the question is who carries our intelligence.

Models are becoming the layer through which we search for information, draft decisions, write software and deliver public services. Whoever builds and runs that layer shapes how it behaves, what it can see, what it costs and whether it remains available.

The Huawei debate was framed around trust in a single manufacturer to carry our communications. The real mistake was building critical infrastructure on systems we could neither see inside nor replace, and our national security was compromised by having a large dependency on a small number of equipment manufacturers.

With AI, we are making the same mistake. The companies building frontier models decide for themselves whether independent evaluators see a model before release, how much access they get, for how long, and what is published afterwards. Some have shared models with government institutes and independent testers, and that is welcome. But it is voluntary, selective and on their terms, while training data, methods and much of the testing remain behind closed doors. 

Openness is not the complete answer to AI safety. There are real concerns about what the most capable models could do if misused, and a group of serious people argue that the most powerful model weights should not be released freely. I take that debate seriously. But the lesson from telecoms is that openness was never about giving everything away. It was about open interfaces, independent scrutiny and the ability to replace any component. Open RAN does not make every radio free. It makes every radio replaceable.

For AI, that means independent evaluation with real access, of the kind the UK AI Security Institute has begun. It means transparency about how models are built, tested and governed. It means open interfaces and portability, so that no organisation, and no country, becomes dependent on a single model it cannot leave. And it means treating AI used in critical infrastructure as critical infrastructure, with security obligations to match.

The ask

In Telecoms we have now spent years building open cloud-native network software in an industry designed around closed boxes. This is possible, but it is far harder than it should have been, because openness was never required when the market was still taking shape.

AI’s market is taking shape now. Governments should push for open interfaces, interoperability and supply diversity into procurement and public funding today, not after dominance sets in. Industry should contribute upstream and design for portability. And the UK, which missed its chance to be a telecoms builder rather than a buyer, has a second chance to lead.