IBM Announces US$2 Billion Investment in AI Campus, Hinting at Reentry into the Semiconductor Business

The AI Hardware Center at SUNY Polytechnic Institute campus in New York will focus on computer chip research, development, prototyping, testing, and simulation. The new center is being built in partnership with New York State, which hopes the center will attract new companies in the field of AI and create several hundred new jobs. The investment marks IBM focusing more on the hardware side of its business. It also shows IBM’s increased appetite to invest in human capital and to address one of the stickiest problems currently facing the AI community.

IMB expanding its research into the area of AI hardware should not come as a shock. Venture Beats reported back in June 2018 that IBM had designed a power-efficient chip for AI processing. IBM published a paper describing their designs and how they were using analog memory in the process of training neural networks. Use of analogue memory is unique to IBM, which is now using press releases and marketing material to suggest that a technique it’s developed called Phase Change Memory can improve power efficiency of current generation AI chips by 90%. IBM says this technology is scalable for both AI training and inference and will most likely be using its investment in the SUNY Polytechnic Institute campus to test and scale these techniques with the intention of reentering the semiconductor industry. IBM sold its semiconductor business in 2014 to Global Foundries, but promising technology developments and significant investments in AI hardware research suggest a return.

IBM has also seen its hardware business tick up in recent quarters and has said that the growth in the hardware segment had been predicated on an upsurge in customers looking to implement AI on hybrid systems, where some training or inference is taking place at the edge or on local server solutions together with some elements of IBM’s cloud software. IBM refers to this as hybrid cloud. Hardware systems have also been complemented by growth in IBM’s software solutions, integration software growing by 2% to US$4.5 billion in 2018, while cognitive solutions software (the group primarily associated with Watson services) has grown its gross profitability by 1.1%. Systems hardware revenue has grown by 2% this year, toUS$8 billion, with gross profit on this same segment increasing by 3.8% to 40.7%. The margins IBM has generated from its integration software business are particularly impressive at 81.3%. IBM clearly sees particular opportunities for continued future growth in the hardware element of its business.

IBM has worked closely with NVIDIA to build GPU-based server solutions under its LinuxONE range of servers. These products have proved popular, and in June of last year were used in the Department of Energy’s latest high performance computer (HPC), which at launch took the record for the world’s largest. However, where IBM has proved most popular with enterprises is in complex installations where a hybrid implementation is required.

Given that IBM sold its chip business to GLOBALFOUNDRIES in 2014, this research effort could hint at a return to the chip business, focusing on custom AI chips this time. However, it is more likely that this research will focus on supporting the HPC computer server solutions that IBM is building around AI. IBM’s competitors in the server solutions business, Dell and HPE, have yet to make quite as bold an investment in one single research center. Dell has chosen to invest in promising AI chip company Graphcore, working with them to bring a test iteration of Graphcore’s IPU chip to life in a server solution. HPE has promised an investment of US$4 billion to improve its edge AI solutions, and it was also rumored at one time in 2017 that HPE was developing its own AI chip, although HPE hasn't validated these rumors. Chinese HPC vendor Huawei has developed a range of AI accelerator cards, called Atlas, based on its AI Ascend chip line which it launched at the end of last year. Huawei is currently the only HPC vendor that offers AI accelerators developed internally. Lenovo also has an initiative called Lenovo Intelligent Computing Orchestration (LiCO). LiCO is a software that simplifies AI model development on HPCs by allowing data scientists to manage and distribute AI model training across segmented clusters. IBM is facing serious competition in the HPC space and will be hopeful that its investment in hardware research will allow it to release market share defining innovations.

One of the important side effects of IBM’s investment in the research center is that it will help grow the size of the AI talent pool. There is currently a shortage of data scientists, meaning that the cost of researching and developing commercial AI systems is on the high side. Analysis published by jobs website Indeed shows that, in certain metropolitan areas, average data science salaries have risen to around US$150,000 per annum. Indeed also found that the number of data science job postings had grown significantly in comparison to the number of data science job searches, from 599 postings for every 509 searches per million searches in 2016 to 1014 postings for every 611 searches per million searches in 2018. While Glassdoor estimates that average salaries for AI-related jobs advertised on company career sites rose 11 per cent between October 2017 and September 2018 to US$123,069 annually. Chinese Internet giant Tencent estimated in 2017 that there were only 300,000 AI engineers globally, which could represent a skills shortage of millions of jobs.

IBM’s research center is a significant investment in AI skills, which is a step in the right direction in terms ofaddressing the AI skills gap. It also marks a full turn around in IBM’s approach to AI in terms of skills. For several years, IBM was criticized for not sharing knowledge about its AI and trying to retain as much information about the area internally as was possible. In the last year or two IBM has clearly departed from this view. It has now released a considerable amount of educational content explaining different aspects of AI and how its products relate to them. The company has also become a lot more friendly toward open source approaches and plans to become a more active contributor towards them, an important aspect of attracting AI talent as many in the research community prioritize sharing their research with the wider community.

A further indication of IBM’s increased openness and flexibility is the recent introduction of Kubernetes, an open source containerized framework that will allow companies to run Watson microservices on IBM Cloud and other public, hybrid, or multi-cloud environments. Kubernetes will give customers the flexibility to choose how they want to implement Watson’s services, whether that means using IBM’s technology or their competitors’.

IBM’s change in approach is noteworthy and should make for a good example for other traditional IT companies, even if they have been slow to get here in comparison to Google, Facebook, and Microsoft. If the skills shortage in the AI space is to be addressed, the other companies need to follow IBM’s lead in partnering with universities to extend research into AI. There also needs to be a greater effort by governments to coordinate the skills of students within the education system so they have some basic computer science knowledge for employers to build upon.