Follow The AI Leader

In the 1980s, a common expression was “nobody ever got fired for buying IBM.” It was considered the safe option, long after new technologies had emerged. While it may not have been the most advanced option available, it remained the safe bet. It had an established ecosystem, and it was a known quantity.

But who or what is the safe bet when it comes to AI? Who has the necessary data? Who has the right people with the best knowledge and the ability to create good solutions? Who has the money? For each of these questions, you might point to a different group of companies, but the problem is that none of them likes to share more information than they have to. That means that everyone is playing with less than a full deck.

EDA companies have long been the supplier of tools to the industry because they can create good economies of scale. But this was not always the case, and it may not be happening today. In the early days of the semiconductor industry, every company produced its own tools. There were no standards, and tools often were seen as a differentiator. As standards emerged and each company realized its tools were similar to those of other companies, it was no longer cost-effective for them to maintain their own. In many cases, they gave the technology to EDA companies in exchange for free maintenance for several years.

This may be happening again. We have seen systems and software companies become hardware developers again, both at the semiconductor and systems levels. Google has been developing its Tensor processors for several generations, along with some proprietary tools. It has trained AI models to understand Verilog and aspects of its designs, and has reached the point where AI can design systems better than manual efforts. It has developed test generation tools, some of which have been made available for RISC-V.

Nvidia has developed several tools, including ChipNeMo. According to Nvidia’s website, “Instead of directly deploying off-the-shelf commercial or open-source LLMs, we instead adopt the following domain adaptation techniques: custom tokenizers, domain-adaptive continued pretraining, supervised fine-tuning (SFT) with domain-specific instructions, and domain-adapted retrieval models.” The list of researchers attached to this single piece of work is significantly larger than any engineering team within an EDA company that I am aware of. This is just one piece of work. Others include simulators, static timing analysis engines, place-and-route, and more.

Where does that leave everyone else? These tools are not for sale to the general public. And even if they were, it isn’t clear if they are equally applicable to designs outside of the area in which they are specifically trained. It is much easier to build a tool that only has to do well on a small range of designs. EDA companies do not have that luxury.

While EDA companies have the most experience in the development and maintenance of tools that serve the mass market needs, they do not have access to all the information necessary to do this, or even to know the types of problems that internal tools are being developed for. Some have limited visibility in areas such as high-speed interfaces, but they also are likely to have developed tools to enable them to build these interfaces faster and better than their competition. Making this public may not be in their best interests.

This is why we are seeing more tentative steps towards melding EDA and AI. EDA companies can, and are, developing agentic solutions around their own tools. They are providing APIs and MCPs to allow their customers to develop their own solutions that may or may not integrate their own internally developed tools, but they may not fully understand why certain demands are made of them.

Researchers and startups have the freedom to explore new ideas, new possibilities, often without having to consider legacy, but that can make their solutions very difficult to integrate into existing flows. They also do not have the financial ability to train models or to support large numbers of customers, making adoption slow – something that nobody has the luxury of these days. This is why we are seeing tens of millions of dollars in venture capital pouring into EDA startups, in the hopes that this will enable them to grow quickly. This is not the first time we have seen large investments coming in for technologies that are expected to be disruptive. In the past, these did not pan out.

Today’s landscape is like the Wild West, and everyone is trying to work out who they can trust, who they can partner with, and how to satisfy the needs of the industry. But while we can probably predict the long-term status quo for the industry (although maybe not the companies that will be at the top), the short term is a lot more difficult to predict.