Compute demand is increasing rapidly as larger models and latency-sensitive applications become mainstream. Industry forecasts suggest that AI infrastructure spending will exceed half a trillion dollars within the next five years. This surge is driving a sustained wave of innovation across the hardware stack, from processors to interconnects.
NVIDIA currently dominates training, with over 90% of the GPU market, but competition is intensifying. New and established players are developing specialised accelerators, including wafer-scale engines, photonic devices, and neuromorphic chips. Cloud providers are also designing in-house silicon, such as Microsoft’s Azure Maia and AWS’s Graviton series, to optimise workloads at scale.
Networking has become a parallel focus of innovation. High-performance network switch ASICs (i.e. application-specific integrated circuits) such as Broadcom’s Tomahawk 6, capable of 102.4 Tbps throughput, are reshaping how thousands of xPUs are interconnected. These advances are essential to overcoming bottlenecks in distributed training and large-scale inference.
Semiconductor manufacturing and equipment firms underpin much of this progress. ASML’s extreme-ultraviolet (EUV) lithography systems are enabling chips at the 3 nm node and below. TSMC, the largest patent holder in semiconductor manufacturing, continues to expand capacity, while design houses such as Intel, AMD, and Qualcomm invest heavily in new CPU and GPU architectures. Startups, including UK-based Graphcore, are exploring novel approaches to AI acceleration.
As firms race to advance AI hardware, intellectual property (IP) strategies are central to securing returns on investment. Patents and trade secrets are the primary tools, but IP strategies include several other elements.
Chip Design
At the architectural level, chip design for AI must reconcile competing demands: massively parallel compute for training large-scale models, low-latency inference at the edge, and efficient memory bandwidth to prevent bottlenecks. The design of these chips involves trade-offs between precision (e.g., FP32 vs FP16 vs INT8), throughput, and power consumption. Protecting chip design matters because innovation at this level requires immense investment, yet the resulting technology can be quickly copied once disclosed or reverse engineered.
Patents can cover novel circuit architectures, manufacturing processes and system designs, giving exclusive protection for 20 years. For example, a unique AI accelerator design or a new method of stacking chips could be patentable. However, functional aspects of integrated circuits can be difficult to define, and the patentability threshold for these techniques can be higher.
Trade secrets require ongoing internal controls and contractual safeguards to remain enforceable. While some manufacturers continue to protect processes as secrets, others increasingly patent them to prevent reverse engineering.
Defensive measures also play a role. In addition to using IP as leverage in negotiations or as deterrent against lawsuits, companies may also choose to publish details of non-core innovations. These publications create ‘prior art’, which can prevent competitors from obtaining patents for similar ideas.
Freedom-to-operate reviews before launch are commonplace, and some companies participate in patent pools to manage litigation risks. Internally, access restrictions and data security protocols remain critical to safeguard proprietary know-how.
An effective IP strategy for chip designers is often multi-layered: protecting core innovation using a combination of patents and trade secrets, and using defensive publications and licensing to secure freedom to operate. Additionally, robust IP portfolios are increasingly seen by investors as indicators of long-term value.
Edge AI
More and more use cases are being developed where AI models are run on edge devices, i.e. devices on the edge of a network, with or without an internet connection. These devices range from mobile phones, electric vehicles, and other devices with significant compute and memory resources, to highly constrained devices such as sensors and other internet of things (IoT) devices.
Advances in edge AI come from developments in mainly two areas. First, chip designers are developing processor units that cram more compute capacity into a smaller space, or custom hardware adapted to run highly specialised AI models. Second, developers are optimising AI models for deployment, often by reducing model size and refining weight parameters to focus on those that matter most.
The expansion of AI into new industries through deployment on edge devices should be fruitful ground for companies seeking patent protection. In general, aspects of computer programs and mathematical methods without a technical effect can fall into categories of technology excluded from patentability. However, when it comes to edge AI, new methods, systems and products are often developed with a specific, real-world application in mind, which can make it easier to demonstrate a technical effect for purposes of patentability.
Even edge AI related innovation that may not be designed for a specific use case, but for lots of use cases, can often still be patentable in at least some aspects. This is the case, for example, if a process is designed to leverage custom hardware, bringing out advantages in terms of computational efficiency, reduced latency or greater privacy. Advantages such as these can confer technical character to various aspects of AI innovation, making it more likely to be eligible for patent protection.
Supply Chains
The risks associated with AI hardware supply chains are increasing rapidly. Whether it be export control regulations or geopolitical tensions, many major and small players are seeking to diversify their supply chains to mitigate risk. At the same time, governments are seeking to capitalise on this uncertainty by offering up large subsidies to set up manufacturing operations in their respective countries, in an effort to grab a piece of previously well-established supply chains. For AI hardware innovators, this landscape brings about important considerations associated with patent filing strategies and patent application drafting.
Patents are territorial rights, meaning that they only provide protection for an invention in a given territory. For example, a patent obtained from the Taiwanese patent office will not normally be enforceable against a manufacturer operating in any other country in the world. When devising an IP strategy, it’s important for companies to consider how their supply chains are changing and use their budget for protection in the most relevant countries, where enforcement is more likely and/or consequential.
Early on in the patent filing process, it may be worth filing in a wider variety of countries to keep options open. While this filing strategy can increase costs at the outset, it enables broader coverage, allowing more time for decision makers to identify primary jurisdictions of interest. For instance, once the most valuable jurisdictions have been identified, patent applications (or even granted patents) in countries that are no longer of interest can then be allowed to lapse to cut costs (including post-grant renewal fees). In this way, the increased cost upfront can be worth it to help mitigate risk and deal with uncertainty.
Whilst major patent laws are harmonised across many countries, there are some important differences. For example, the European patent office (EPO) and the Japanese patent office (JPO) typically prefer a clear description of advantages associated with different features of an invention included in the patent application text, whereas the United States patent and trademark office (USPTO) considers this less of an issue. When it comes to preparing patent applications for filing, ensuring that they are drafted to meet the requirements of all their intended filing countries is another important consideration.
What the UK’s AI strategy means for AI hardware innovation
The UK government has released several publications emphasising their commitment to supporting the development and expansion of AI across the economy. Recently, these include the AI Opportunities Action Plan, the spending review, the UK’s Modern Industrial Strategy, and the Digital and Technologies Sector Plan. In each publication, support and funding is committed to AI.
Increasing the UK’s compute power is consistently mentioned. The government aims to increase the UK’s AI compute capacity to support AI research by 20-fold by 2030. This will be supported by committing £1 billion in extra funding to increase compute power and a further £750 million to build a new national supercomputer at the University of Edinburgh.
The AI Opportunities Action Plan calls out providing government-funded programme directors with significant autonomy to back R&D programmes directed at supporting AI. The plan explicitly references that such directors could be supported in a similar manner to the US government’s Defense Advanced Research Projects Agency (DARPA) programme.
DARPA is credited with providing initial seed funding for some of the world’s most transformative technologies, making substantial contributions towards the global positioning system (GPS) and advances in semiconductors. DARPA’s approach involves investing in high-risk, high-reward research, focusing on projects that serve US national security but may also have broad civilian applications. The Advanced Research and Invention Agency (ARIA) is one such UK body with the ability to autonomously fund projects, including breakthrough research.
Rather than pursuing mass-market chip fabrication, the UK strategy builds on established strengths in design and intellectual property. The National Semiconductor Strategy highlights firms such as ARM and Imagination as global suppliers of IP cores and supports startups through incubators and targeted funding.
Intellectual property is identified as a key enabler of innovation. Policymakers stress the need for UK champions in critical layers of the AI stack, supported by strong IP frameworks. This means patents will remain central to securing funding and defending competitive positions.
Overall, the UK’s AI strategy is likely to stimulate demand for advanced chips while shaping the environment in which hardware innovation occurs. Public investment in compute will provide access to both mainstream GPUs and emerging architectures. For innovators, the opportunity lies in combining R&D with robust IP strategies, engaging with national programs, and aligning with international frameworks. In a globally competitive field, these elements will be decisive for UK-based ventures seeking to thrive in the AI hardware battleground.