AI Chip Companies Continue to Raise More Capital

ai-chip-companies-continue-to-raise-more-capital

High mask costs made it hard for chipset startups to raise capital during the early 2000s. Today, mask costs can range up to $25 million and, by some estimates, the design costs for a 12-nanometer (nm) node chip can run as high as $174 million. This made it hard for investors to justify the return on investment (ROI), as most of them demanded a 10X return. Only a few markets offered high volumes for chipsets to generate that kind of revenue.

The emergence of AI has opened up many new possibilities for AI chipsets and Tractica estimates that the market will reach $66 billion by 2025. The widespread applicability of AI means that many applications will need different chipsets, leading to different requirements. This has led many semiconductor companies and startups to jump into the market with their own solutions.

The startups started appearing in 2016 and, as of 2018, many cloud companies, top semiconductor companies, startups, and field programmable gate array (FPGA) companies have announced their intention to make AI chips. Many of the seed rounds raised by startups were of the order of few million dollars and that did not receive a whole lot of attention. As of 2018, some companies have developed products and are able to command high valuation from chip companies. Cambricon, a Chinese-based application-specific integrated circuit (ASIC) company, became the first official unicorn when it raised “hundreds of millions” dollars recently for a total valuation of $2.5 billion. Cambricon is not the only company generating headlines. Many others have also recently raised large amounts of capital, including:

  • Wave Computing, a California company recently announced that it had raised $86 million in a Series E round of funding.
  • Habana, which recently stormed onto the ASIC scene by releasing its own chip for data center inferencing, announced that it has raised $75 million.
  • SambaNova, a startup out of the University of California (UC) Berkeley raised $56 million.
  • Groq, a startup from the designers of the tensor processing unit (TPU), announced that it raised $52 million in September 2018.
  • Thinci, an El Dorado, California-based company raised $65 million to build its products and solution in September 2018.
  • Graphcore, yet another high-flying startup based in Cambridge, UK, made an announcement that it has raised $200 million, giving it a valuation of $1.7 billion.

The list goes on and on. While some of these have released samples and products, some startups have been able to raise capital only on the basis of PowerPoint and their teams—something unheard of since the late 1990s when semiconductor companies were the darlings of the venture capital (VC) world.

Considering the problems with AI hardware, the investment is justified. A chip company not only has to manufacture a chip, but also provide customers with reference boards, intellectual property (IP), development tools, and software to their customers. The software and IP carry more value than hardware for original equipment manufacturers (OEMs), as very few people at this time have the skills to design AI applications from scratch. This means that semiconductor companies have to go the extra mile to produce software and a framework.

Today, NVIDIA is the de facto leader in the AI world, followed by Intel. Both companies are generating over a billion dollars in revenue from AI chipset products. The value add of AI startups is in their hardware that offers power and performance benefits for a given application. Currently, a generic solution, such as a central processing unit (CPU) or a graphics processing unit (GPU) is being used by most companies worldwide, but the need for specialized chipsets is widely recognized. Startups are not quite shipping their products yet, but when they do start delivering the performance level they are promising, it will be an interesting battle.

Comments are closed.