The Golden Age of AI Acceleration

Looking back, the last two years were golden for AI acceleration startups. It seems every other week saw the announcement of a successful funding round.

In the USA, Cerebras systems raised  a total of $112M in three rounds, the latest early 2017, for a valuation of $820M. Last March, Austin-based Mythic, which promotes an original approach with hybrid digital/analog calculation inside flash arrays, announced it raised $40M in a series B led by Softbank, for a total of $55.2M. In March as well, SambaNova systems, a startup created by a former Oracle/SUN microsystem executive, raised $56M from Google Ventures. In September, Groq, funded in Palo Alto by former Googlers involved in the famed TPU processor raised $52M, for a total of $62.3M. In September as well, ThinCi, a startup focused on edge computing and IoT raised $65M from a round led by Japanese automotive giant Denso. And lately, Wave computing, one of the first to have working hardware that it has been selling in a server system, had its latest series E in November with an additional $86M for a total of $203M – which brings the company well into unicorn territory.

In China in July 2017 the central government unveiled a national plan to build a 1 trillion-yuan ($152.5 billion) AI core industry by 2030, with homegrown AI processors an important part of the ambitious goal.  Several startups are benefiting from the government push, starting with Horizon robotics, which raised $100M in October 2017 in a round that includes Intel Capital. Cambricon announced in August 2017 that it had raised $100 million in series A funding from Alibaba. It was followed one year later by another infusion of $100M in its series B that valued the company $2.5B.

In Europe, the most prominent AI startup, UK-based Graphcore that is now partnering with DELL, announced in November 2017 funding by Sequoia Capital for $50M, for a total of $110M and a value of around $1B. In France, Prophesee, formerly known as Chronocam, that designed a bio-inspired architecture for vision raised $19M in February this year for a total of $37.3M. Kalray, another French company developing a manycore processor array, got in May this year a new inflow of money from Renault-Nissan and the French equivalent of DARPA, the DGA, for a total of $37.9M. 

AI for what?

It strikes that so many companies attract so much money so fast and we may wonder what their business strategy is and on what foundation the valuation is built. 

There are only two markets that are attractive for AI startups, because they have the potential for revenue in the foreseeable future. The first one is hardware acceleration for cloud servers, used by internet companies to improve their services. This is the application that popularized AI, and has the most immediate promises, as a revenue stream already exists. It is also the most contested space with incumbents like Intel, Nvidia, Xilinx, IBM, and internet giants developing their own solutions as Google with its TPU and Amazon with its Inferentia chip. 

The second market is AI for vision at the edge, driven by the prospects of the autonomous car, as well as lesser volume applications like cognitive security cameras and drones. This market is still tenuous and depends on many other technical and business challenges being solved before it can blossom. While there is no dominant player in the AI side of things, all the established IC companies in the automotive and camera markets have existing vision solutions and roadmaps, either home grown or from the leading IP vendors.

A third market, mobile phones, is not even considered by AI startups, so strong is the stranglehold of Apple, Samsung, Qualcomm and Huawei on it.

Where is the Exit?

Seeing the competitive forces in presence, it is very unlikely that any of the AI startups can create a sustainable business as a semiconductor company. The strategy for them must be to be acquired, and this may not be a bad move. Movidius, Nervana and Mobileye, three early startups in hardware AI have all been acquired by Intel. It is true that Mobileye had revenue and profit before the acquisition, but they were very early in the market, when semiconductor companies were still testing the water. Now everyone is franticly developing solutions for edge and cloud, and carving a profitable niche is less and less feasible.

As a matter of fact, many startups have partners or investors that are industry giants, and you would guess that these godfathers are ready to buy if the technology meets its promises.

The table below summarizes the funding, valuation, market and backer for each of these startups:

Risky business

As I am, you may be excited by the potential of AI technologies. We all marvel the creativity and baldness of the innovators that dare to invent new hardware architectures to make AI more relevant. But don’t get fooled by your own enthusiasm: this is a risky business. Even with plenty of cash, the odds to make it are low, so complex is the problem, and so ruthless is the business environment.

San Diego based Knuedge may well be the first victim. After raising more than $100M it ran out of cash and was forced to close in May this year. Did they overburn? Did they lack a good godfather? Were they too early? Probably a little of everything. 

We are approaching the end of the funding cycle now, and I am curious to see how the remaining startups will negotiate the year ahead to avoid facing a similar fate. Let's wish them the best of luck, if nothing else, they are making semiconductors cool again!