Baidu, Alibaba, and Tencent (BAT) are the top three investors in AI startups among large Chinese companies. Most of such funding goes to those focused on application layers or algorithms. While the tech giant trio has only made strategic investments so far, the field has also seen a stream of government and VC funding in recent times. Baidu and Alibaba have also moved into chip design themselves.
These new chip firms generally fit into three main categories—pure fabless semiconductor startups (which outsource manufacturing), algorithm startups that are moving into chip design, and internet firms that are exploring design too.
In total, I can think of at least 20 Chinese companies that fit into these categories, all of which are designing AI chips. I am sure there are a lot more of them out there in the vast industrial landscape. Traditional fabless design companies in China are also increasingly designing their own AI chips, making the market ever-more competitive.
It isn’t just China though where we see non-traditional chip companies like Baidu and Alibaba moving into design. In the US, countless listed giants, including Tesla, Google, Amazon, Microsoft, Facebook, and Apple, are designing their own AI chips, specific to their required applications. How can AI chip startups from both China and the US alike expect to live long when they face competition from behemoths that boast such technical and economic strength?
This vertical integration is not just a threat to startups but also to traditional chip designers that previously considered internet companies their customers. This dynamic will shape the industry in both regions for years to come.
These startups have moved quickly from burning VC cash to coming up with architectural innovations, and now they are tasked with actually finding customers in a very competitive market. Industry voices say it is becoming increasingly difficult for them to differentiate themselves and any start-up late to the party will struggle to bring anything unique, new, or different to the table.
Those that really do innovate and pursue emerging technologies like analog computing, in-memory computing, and neuromorphic computing warrant attention but have to deal with a much longer path to get a commercial product to market.
‘Chinese’ chips
From a hardware design perspective, these Chinese AI chips aren’t necessarily all that Chinese. As most of the companies work to tight deadlines, they tend to license an IP rather than develop in-house. The majority with which I have spoken will gain authorization to use networks that link design parts together, known as on-chip interconnect. They also rely on development and debug tools from companies like Lauterbach. That’s not to mention that the higher-end chips mostly use TSMC for fabrication. Of course, a large proportion of designs also use Arm cores. Some I have spoken with have developed their own GPU or custom neural network core, but the majority have neither the time nor luxury to do so. The bottom line is VCs want to see returns fast!
Returns come much faster in other industries. With the typical design process for a semiconductor taking 18 months and requiring tens of millions of dollars to finance just the first few designs, there is a real need to seek out customers from the get-go.
Some more experienced chip-focused VCs in the country understand this and are more patient, but we can expect to see some of these startups die off over the next couple of years as the pressure from VCs builds. The sector will consolidate for other reasons as well though. Industry mainstays are expected to snap up those new firms that look the most promising. The trend has been seen already when America’s Xilinx, the world’s leading designer, and supplier of programmable logic devices, acquired Beijing chip unicorn DeePhi last year. Other players will resort to acquisitions as a means to enter the market, as was seen when Alibaba took over C-Sky.
At this point, I think it’s safe to say the necessity to have dedicated AI hardware is here to stay. Most chips have some kind of ‘AI’ functionality, and this will only increase with time. Sure, some design teams will have to decide how much effort they dedicate to AI, but given that almost every major player is now involved, it is difficult to see such functions not becoming critical features of most future designs.
For China’s AI chip startups, it is a balancing act of getting to market first, being innovative, and having patient investors. The most successful will be acquired or secure more money, and the lame will die. The flipside for them is that AI and semiconductors are two areas of focus for the Chinese government in its push for technological independence. Those able to combine these verticals will garner a lot of attention. Despite this, the market is still approaching saturation, and only those that strike the right balance will survive.