What if the next major families of SEPs were to emerge from computing infrastructures?

Recent announcements from South Korea deserve particular attention.

The strategic partnership between SK Hynix and NVIDIA centred on future HBM memory for AI Factories, Samsung’s announcements regarding HBM4, advanced packaging and its collaborations with major players in AI, as well as the massive investments made by South Korea to strengthen its leadership in AI infrastructure, all point to a more profound shift. The competition is not solely about models, data or GPUs. It is increasingly shifting towards computing infrastructure, which will determine the scaling of AI. Indeed, the rise of artificial intelligence – for both training and inference – makes computing infrastructure a key determinant of competitiveness and performance.
This infrastructure depends in particular on the ability to:
– store data as close as possible to the accelerators,
– move and share massive volumes of data with minimal latency,
– ensure the reliability of data exchanges, in particular through error-correction mechanisms,
– efficiently orchestrate computing, memory and communication resources within increasingly heterogeneous architectures.

These challenges bring several complementary technology families to the fore:
– Memory: DDR, HBM and stacked memory,
– Interconnects: CXL, UCIe, UALink,
– Communication technologies: Ethernet, InfiniBand, photonics (silicon photonics, co-packaged optics, and free-space optical communications),
– Reliability: ECC, FEC, LDPC and other error-correction mechanisms,
– Integration: advanced packaging, chiplets and heterogeneous integration,
– Orchestration: functions enabling the efficient coordination of computing, memory and communication resources.

This trend extends beyond the field of AI alone and more broadly concerns computing architectures designed for AI, HPC and, in the future, hybrid platforms combining classical and quantum computing. These platforms will face exactly the same challenges: memory bandwidth, inter-accelerator communications, energy efficiency, reliability of data exchange and the orchestration of distributed resources.

It is equally interesting to note that these technologies are accompanied by the rapid development of their standardisation ecosystem, driven both by industrial consortia (JEDEC, OIF, UCIe Consortium, UALink Consortium, PCI-SIG, CXL Consortium) and by international organisations such as the IEEE, ISO/IEC JTC 1 and ISO/IEC JTC 3, whose work on quantum technologies and hybrid HPC/quantum architectures illustrates this trend. In other words, computing infrastructures are currently undergoing a standardisation process comparable to that experienced by telecommunications or audio and video codecs in previous decades.

This development naturally raises questions regarding intellectual property and, more broadly, technological governance. Standards-essential patents (SEPs) are a prime example of this, but they now coexist with open-source ecosystems, which also play a major role in the development, adoption and dissemination of technologies. Indeed, for over thirty years, the main SEP ecosystems have been built around telecommunications and/or connectivity standards (2G to 6G, Wi-Fi) or audio and video compression standards (MPEG-2, AVC, HEVC, VVC). At the same time, these fields have seen the emergence of increasingly influential open-source implementations and platforms, illustrating the growing complementarity between standards, standards-essential patents and open source. Future codecs are themselves incorporating increasingly AI-assisted approaches – and may even become AI-native in the future – whilst work on 6G is placing ever greater emphasis on AI-native architecture and more open networks.

These considerations led Cobelty to carry out an independent study focusing on the market for computing infrastructure for AI and HPC. It analyses the historical development of these ecosystems, major technological trends, market concentration, investment dynamics, key standardisation initiatives and patent transactions. Cobelty would be delighted to discuss the subject with anyone interested.

Several months ago, LexisNexis approached Cobelty to present the key findings of its study on standards-essential patents (SEPs) relating to DRAM, DDR and HBM technologies – which was then in the final stages of completion – and to gather Cobelty’s expert views on this work. This study highlights the now strategic role of memory technologies in standardisation and the monetisation of intellectual property, and suggests that DRAM technologies could be one of the first technological segments in which a new licensing ecosystem associated with AI infrastructure is emerging.

As artificial intelligence becomes a key component of digital infrastructure, the focus of technological competition is expected to continue shifting towards computing infrastructure. Recent work on DDR and HBM technologies, and the call for patents launched by the Via Licensing Alliance with a view to establishing a licensing programme dedicated to DRAM technologies, are converging indicators of this trend. It now seems appropriate to investigate whether this trend is set to extend to other building blocks of computing infrastructure — interconnects, photonics, error correction, advanced packaging or even resource orchestration functions. If this were to be the case, computing infrastructures could gradually form a third major ecosystem for international standardisation, which could in turn give rise to new FRAND licensing programmes.

Beyond this initial consideration, another question strikes us as equally interesting: computing infrastructures are already characterised by the coexistence of several models of innovation governance — standards and standard-essential patents (SEPs), commercially licensed IP cores, open standards (of which RISC-V is a prime example) and open-source ecosystems. Understanding how these different models interact, complement one another and contribute to value creation will, in our view, be another major challenge in the years to come. The study recently published by LexisNexis provides, in this regard, an initial tangible indicator of this trend. Whilst memory technologies currently appear to be the first concrete manifestation of this trend, it remains to be seen to what extent it will extend to computing infrastructures as a whole, as well as to the accompanying economic and governance models.

We will return to these issues in a future article.