Who Owns the Intelligence? Leonard Lee’s Six-Layer Cake
Everyone in telecoms wants to sell intelligence. Leonard Lee asked what that word actually means, who owns each layer of it, and where operators fit when most of the stack is built elsewhere. His answer: trust, sovereignty, open AI computing and privacy.

Leonard Lee’s CASA26 talk on the architecture of intelligence and the opportunities it leaves for telecoms
The theme of CASA26 was intelligent engagement: making communications more intelligent, and moving from communications to conversations. Leonard Lee, Executive Analyst and Founder of neXt Curve, started by asking what everyone in the room meant by the word. Intelligence, in practice, means AI. AI, these days, means generative AI. And most people still experience generative AI as a chatbot, even though the technology has moved in a few years from chatbots built on large language models to the agentic AI that 2026 was supposed to be the year of, with physical AI already next in line.
The definitions have had to change as fast as the technology. And if this industry is going to talk about the architecture of intelligence, it needs a way to think about who owns which part of it.
Six layers, starting with power
NVIDIA’s CEO likes to describe AI as a five-layer cake. Leonard offered six. Power is the foundation: without electricity, there is no intelligence. Above that sits compute and memory, which is far more than chips. Then AI infrastructure, where networking belongs. Then the models. Then the token, the unit that all of it produces. And at the top, the applications that use it.
The infrastructure decides how much intelligence you can generate, and where it lives: on premises or in the cloud, sovereign or not, exposed to a kill switch or not. And Leonard’s reminder about the token was a useful corrective. A token is a predicted fragment of the next word. We treat it as a unit of intelligence, but it does not always produce intelligence. Sometimes it produces what he called artificial stupidity.
Is Europe the odd one out?
We asked Leonard to look at ownership across the United States, Asia and Europe. Have the US and China taken all the oxygen? He compared the regions on four things: power generation, leading-edge chip manufacturing capacity, AI data centre capacity and models. Asia has a large advantage in power and a dominant position in leading-edge manufacturing, largely thanks to TSMC, while chip design is dominated by one American company. Europe’s power base is closer to the others than many assume. Its leading-edge manufacturing is thin. And memory, increasingly constrained, is another factor to watch.
Unlearn what you learned about AI
This is the turn in the talk. The architecture is shifting so fast that much of what we believed in the last few years is already wrong. In 2023 the assumption was that AI would run best on one technology roadmap, on GPUs. By 2026 that has changed: inference-optimised accelerators, new cooling and new data centre power designs are reshaping even the market leader’s own roadmap.
Leonard listed what will drive the next changes. Post-training, because enterprise AI has to be specific to a domain, an application and a business. Open-weight models, which are democratising capability and disrupting AI services. Sparse models, which change system design down to the kernel. Model-specific optimisation driving purpose-built silicon. And agentic AI, which drives demand for CPUs and heterogeneous computing. The architecture will look very different, and many things that seem true today will be wrong.
Four opportunities for telecoms
If most of the cake is baked elsewhere, where do operators and their ecosystem fit? Leonard saw four places.
- Trusted intelligence. Trusted communications and identity, observability and routing, and the verification and certification of AI for customers in the regions operators serve.
- Sovereign intelligence. Inference and post-training are different from model training. Making them sovereign, secure and trusted, with sovereign technology and sourcing, is a conversation European operators are already having.
- Open AI computing. Open-source AI exists in abundance. Making it telco-grade, enterprise-grade and safe for consumers has not happened yet, and it is how regions can build their own secure stacks.
- Privacy and confidentiality. Few people realise how unsafe AI is, or how hard it is to protect privacy and confidentiality in these architectures. Understanding those gaps opens room for foundational trust services across industries and for consumers.
Every one of the four is about trust, which is where Rob Kurver opened CASA26: intelligence is getting cheap, trust is not. The operators do not need to own the cake. They need to be the reason people can safely eat it. Our thanks to Leonard for a talk that made the room unlearn a few things.
More from CASA26
CPaaSAA members have every CASA26 session in full in the member community.
Not a member yet? Find your fit in one minute and see which membership suits your company.
Following the series? New sessions appear here every week until Mobile World Congress. Subscribe on YouTube to catch each one.
CASA26 ran 21–23 September 2026 at the Tobacco Theater in Amsterdam, hosted by CPaaSAA. This is part of our Monday series on Transformation.

