The AI Race Is Becoming A Silicon Race
Iri Trashanski, Chief Strategy Officer at Ceva, is shaping the future of the Smart Edge with extensive experience across tech sectors.
gettyGoogle, Amazon, Meta, Microsoft, Apple, OpenAI and other technology leaders are investing in custom silicon as AI becomes increasingly central to their products and platforms. The reason is straightforward: Competitive differentiation in AI is moving deeper into the technology stack. When AI becomes strategic, the architecture underneath it becomes strategic too.
The takeaway for other companies is broader than custom chip development. Companies need to decide where proprietary technology creates meaningful competitive advantage, where proven intellectual property can accelerate development and where their engineering resources can create the most value.
Semiconductor IP can play an important role in that strategy. Instead of spending years developing every underlying capability, companies can start with proven building blocks and focus their investment on the architecture, algorithms, software and experiences that differentiate the finished product.
For much of the software era, companies could differentiate at the application layer while relying on relatively standardized computing platforms underneath. But as AI moves into PCs, vehicles, robots, industrial equipment, wearables and billions of connected devices, product performance increasingly depends on decisions made at the silicon level.
Consider what makes an AI experience effective. It needs to respond quickly. It may need to operate without a reliable cloud connection. It needs to protect sensitive data. In battery-powered products, it needs to accomplish all of this within tight power constraints.
A robot can’t wait for a round trip to the cloud before reacting to its environment. An always-on wearable can’t continuously send sensor data to a data center without affecting battery life. A voice interface that depends entirely on network availability will struggle to deliver an immediate experience.
These requirements change the engineering challenge. At the edge, simply adding more compute is rarely the answer. The challenge is delivering the appropriate compute within the application’s power, memory, latency and cost constraints.
As a result, silicon architecture increasingly influences the capabilities companies can deliver at the product level.
The investments being made by hyperscalers might suggest that companies competing in AI need to develop their own silicon. But for most companies, it’s more useful to determine which technologies they need to own to create differentiated value.
Modern chips are enormously complex. A single system may incorporate CPUs, AI accelerators, DSPs, wireless connectivity, sensing capabilities and specialized software. Developing and optimizing all these technologies internally requires significant engineering resources, specialized expertise and time.
Proven semiconductor IP gives companies another path. This is how the silicon race broadens beyond the hyperscalers. Companies don’t need Google’s or Apple’s engineering scale to make silicon a strategic part of their differentiation.
Companies can license established building blocks such as AI accelerators, connectivity technologies and DSPs, integrate them into their own architecture, and concentrate their engineers on the technologies that make their products unique.
This changes the economics and speed of product development. Engineering teams don’t have to recreate mature technologies before they can begin creating differentiated value. They can build on an established foundation, potentially shortening development cycles, reducing execution risk and reaching the market faster.
That matters particularly in AI, where models, architectures and use cases evolve rapidly. Semiconductor design talent is finite, and the number of technologies required to build intelligent products continues to expand.
Competitive advantage increasingly comes from knowing which technologies are strategic enough to develop internally and which can serve as building blocks for innovation higher in the stack.
This becomes even more important as AI moves from data centers into the physical world. Robots, vehicles, industrial machines and wearables must sense their surroundings, interpret information, communicate and act in real time. That can require cameras, microphones and other sensors to work alongside wireless connectivity and local AI inference within the constraints of the device.
The engineering challenge therefore extends beyond optimizing AI compute alone. Processing, sensing, connectivity and software increasingly need to work together as part of the overall system architecture.
There is also enormous diversity among these systems. A data-center accelerator and an always-on sensor may both run AI, but their requirements for performance, power, memory and cost are fundamentally different. There’s no single architecture that optimally addresses every application.
As AI moves into more specialized physical systems, the need for architectural customization will grow, not diminish. Companies can select proven building blocks appropriate for a particular product and then optimize the system around their own requirements.
Speed becomes increasingly important as well. AI is evolving quickly, and product cycles continue moving forward while the underlying technology changes. Companies that can combine proven technologies with their own intellectual property, optimize the resulting architecture for their application and move efficiently from concept to production can respond faster as the market evolves.
Every major computing transition changes where companies create differentiated value. The PC elevated operating systems. The internet elevated platforms. Smartphones rewarded close integration across hardware, software and services. AI is pushing that integration deeper into the underlying computing architecture.
Models, software and silicon are becoming increasingly interdependent. As intelligence moves into the physical world, sensing and connectivity become part of that architecture as well.
This further reinforces the need to make deliberate choices about what to build, what to license and where to focus engineering resources to create the most value. Proven semiconductor IP can provide a faster starting point, while proprietary investment moves toward the technologies and experiences that ultimately differentiate the product.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
