AI has been in the spotlight for a while. The original quest to develop programs and machines that can obtain, process and create information intelligently has produced both powerful applications and a flurry of hype. The 'AI' buzzword is now extensively used by startups to attract early-stage investment, regardless of the nature of the business model and whether Deep Learning (DL) algorithms – the essence of AI – are actually part of their core product. In short, there is plenty of excitement around the concept, with no lull likely to happen in the near future. 

Obviously, a brand-new generation of hardware is needed to move back and forth the ever-expanding troves of data and run sophisticated iterative learning algorithms. The microelectronics sector is the key source of innovation in this field.

AI accelerators – specialized chips meant to hasten the process of managing information – are doomed to become the main differentiating tool in intertwined ecosystems of the upcoming omnipresent ‘Industry 4.0.’ Memory, storage, logic and networking are the main directions where semiconductor companies are fine-tuning their devices to enable carefully constructed AI code to run faster, deeper and more precise analyses.

Anything tectonic?

There are two major shifts expected to occur in this realm. First, leading global corporations intend to create their own, completely independent AI-based products, which will lead to in-house chip development projects and trigger the tech giants to employ more aggressive market strategies. We expect the number of acquisitions of small, narrowly specialized upstarts by leading Internet companies to grow sharply in the following five years.

This trend poses grim challenges for semiconductor conglomerates – entities from which an overwhelming majority of micro-components had been outsourced before the recent oversaturation in the global consumer electronics market.

The second projected change is structural – and it is, in many respects, the factor that may force the abovementioned behemoths to diversify. As specific tasks in AI benefit from hardware designed to handle their needs as narrowly as possible, the role of the Central Processing Unit (CPU)  – the ‘one-man army’ of the semiconductor world – will inevitably lose much of its clout in the industry. In terms of getting up steam, Application-Specific Integrated Circuits (ASIC) have already begun a quest to conquer both edge architecture and data centers, pushing CPUs out of the scene.

ASIC-makers' ambitions are also destructive for another two kinds of components: Graphics-Processing Units (GPU) and Field-Programmable Gate Arrays (FPGA). While the former will almost certainly lose their ground in AI training in data centers, the latter might have some room in the edge due to their ability to be easily reprogrammed and flexible in terms of possible applications.

The market

There have already been several milestones in the sector that can help to assess the magnitude of the nascent market. Intel, the undisputed champion in the CPU domain, reportedly sold USD 1 billion of AI chips in 2017. However its nemesis, GPU segment veteran Nvidia, palms off a comparable amount of ICs, virtually dominating in the image recognition supporting hardware field and some other areas. Their products power Facebook, Google and a bunch of other digital pioneers.

Taiwan-based TSMC, currently the world’s biggest contract chipmaker with a market cap of over USD 250 billion, has also launched a number of projects aimed at building world-class accelerators. Together with Los Altos, California-based startup Cerebras, it has developed the largest processor ever built – the "supercomputer-on-a-chip", Wafer Scale Engine, with a record-breaking 1.2 trillion transistors.

These and other huge industry players such as Korean megacorporation Samsung, AMD (another GPU leviathan), Xilinx and Altera (both are well known for their FPGAs) formed a seven-billion-dollar artificial intelligence chip market in 2018. 

Growing at a 59.5% CAGR over the past three years, the AI sector saw increasing (from 8.2% in 2016 to 10% in 2018) value contribution from semiconductor providers. Spurred by the capital being pumped into hardware projects around the globe, this proportion is expected to balloon in the next few years, exceeding 18% by 2023.

In the ultra-short run, however, the growth won’t be that impressive, as the market is likely to face an overcapacity crisis caused by the extremely high costs reported in 2018 by the leading powerhouses: Samsung, Intel, SK Hynix, TSMC and Micron spent over USD 71 billion in semiconductor capex that year, up 16% from 2016’s USD 61.4 billion.

Startups?

While the deep-pocketed international companies can easily hop into this emerging field, lashing out heaps of capital on various research and development projects, young private companies have a hard time competing against the leading clique.

Financial stringency – one of the key distinguishing features of ‘from-zero-to-one’ entrepreneurship – seems to be an insurmountable barrier for those who try to make a difference by designing and manufacturing microelectronic hardware for AI outside of ecosystems that are constructed and maintained by the big industry names.

This, nevertheless, doesn’t imply complete independence. Even the high and the mighty create strategic alliances. Not to mention growth-stage rookies that always need extensive coaching and backing or, at least, bean-counting creditors to get their businesses off the ground.

In other sectors, a burgeoning global startup scene has proven itself a cradle for industrial innovation multiple times. Retail, education, entertainment and other consumer-oriented domains, in which business models and market positioning matter way more than the core technology, have each seen herds of unicorns – startups valued at over USD 1 billion – coming from both advanced and emerging worlds.

Offspring of the semiconductor industry, where technology is the heart, not a limb, need tremendous piles of investment to keep pace with an ever-changing international environment. Moreover, unlike the zephyrian software business, where money can follow a sufficiently beguiling pitch before a product takes shape, bright ideas are far from enough to raise cash for making chips.

What potential backers need here is confidence in product quality. From humble beginnings, such as angel, seed and early-stage investment rounds, to the phases closest to the IPO, the ‘show me what you are capable of’ principle is the first and the most popular for PE/VC money managers. The traditional ‘we first look at the team and market orientation’ concept is put on the back burner. In other words, investors in hardware are frugal by nature.

“Real men have fabs” – the oft-cited quote by AMD founder Jerry Sanders is now, apparently, more than just a motto for those acting at the intersection of AI and microelectronics. This, indeed, doesn‘t mean that all the startups devoted to producing accelerators are leaning towards one-stop instant mass production. On the contrary, only some of them provide complete end-to-end hardware solutions carried out without outsourcing mask services, assembly, functional testing, packaging or other capital-intensive phases of the semiconductor value chain.

Basically, it means that a cool founding team full of Ivy League graduates will not receive extra consideration when it comes to real investment decisions in the AI chip industry. And this is exactly why not so many rounds of funding have taken place in the sector so far.

In the previous three years, AI semiconductor startups have obtained nearly USD 1.7 billion in 20 of the largest financing deals (Series A-D). It is important to note that USD 1.1 billion, or almost 65% of this amount, has been secured by three companies: Beijing-based ‘AI supermarket’ Horizon Robotics, DL processor developer Cambricon Technologies and the Intelligence Processing Unit (IPU) pioneer, Bristol-headquartered company Graphcore. 

China-based upstarts are seizing fresh opportunities that are coming from the country’s state-level strategies designed to facilitate the next ‘great leap forward.’ Meanwhile, the United States is clearly lacking big-scale microelectronics innovators other than Intel, Qualcomm and the rest of internationally acclaimed kingpins. However, ‘does this fact make any difference?’ is a tricky question to answer.

The AI semiconductor segment is not only harsh for penniless newcomers but also highly polarized and, following the never-fading Moore’s law, we see market entry barriers lift up over time. Nonetheless, microfabrication is part of the value chain where real innovation happens. As Sir Francis Drake would say, sic parvis magna – "From small things comes greatness." It seems like more and more of the technologically and financially ambitious are taking this to heart nowadays.