nVidia gambles on AI-rich future with Arm

From the headline purchase price down there is so much about the announcement that nVidia will buy Arm from the Softbank Vision Fund that looks good but which is clearly there to paper over issues with the future of all three players. 

The release trumpets a deal value of $40bn in total. At first blush, it seems Softbank, which has run into trouble with major investments such as WeWork in recent months as well as trades on Nasdaq derivatives that are more like those of a hedge fund than a technology company, has made a profit on the $32bn it paid for Arm several years ago. However, that $40bn includes an earnout of $5bn that is not guaranteed and close to half the remainder is in nVidia stock rather than cash, and subject to the vagaries of the Nasdaq which went into reverse in the past few days after surging ahead thanks to monetary easing by the US Federal Reserve and general optimism over all things cloud computing.

The second major point is that nVidia wants to build an “AI supercomputer” and base the R&D for that at Arm’s home in Cambridge, which should please those who are concerned about the company's future in the UK. However, this brings into focus the third point in the announcement, which was that Arm would retain its “customer neutrality” that has indeed helped the UK-based operation become the leading provider of processor core designs to the world’s chipmakers. It was a position that provided the rationale for Softbank to buy it in the first place. But things change. Even in silicon design, where major projects do take years to complete, things can change fast. And even armed with the best of intentions, subtle changes in direction mean the customers will decide to leave you behind. 

The question of whether a silicon vendor would attempt to buy Arm dogged the company for years. However, such a deal looked to have the attraction of mutually assured destruction. A chipmaker might look longingly at the customer base and think: “We could do so much if we controlled that”. But by the time they tried, the customers would start running for the hills. Though there were concerns about the purchase, Softbank was one of the few companies that would not immediately disrupt the equilibrium. 

As a purchaser, nVidia is arguably one of the safer options for any existing Arm customer. Don’t think so? Imagine Intel taking its place. But chipmakers are notoriously secretive and distrustful. They have shared a lot of information with Arm over the years. Some provide technology and ideas to Arm that are then incorporated into the wider core line-up. Is that something that Arm can count on in the coming years as chipmakers begin to worry about not what they are telling Arm but what they might be telling nVidia?

The prevailing winds in the market play a role in determining how important these concerns are in reality: it’s going to be a tension between a number of factors. 

It’s a common belief in the sector that had Motorola offered its once popular 68K family for IP licensing in the early 1990s, Arm would have found it far more difficult to gain as much traction as it did. But during that period Motorola much preferred to sell its own chips, seeing the margins as far better. By the time Motorola offered a licensable core – the MCore – almost no-one took it seriously because Arm had landed important deals with the likes of Nokia, and though they might have considered a 68K-based core to go with their existing software they had no interest in being beholden to a large chipmaker for access to future processors.

Asked on a conference call about the role of RISC-V at nVidia, which the company uses as a management processor in its GPUs, founder and CEO Jen-Hsun Huang said he regards Arm and RISC-V as very different things even though both wind up as processor cores. His argument was that the richer software ecosystem around Arm, which has been established over two decades, makes it a better option for applications-level programming. RISC-V can take care of low-level grunt work that most users of GPUs and similar processors simply never see: everything is buried under APIs and environments like CUDA and OpenCL.

Again, though, if you wind back the clock, Arm faced similar issues. In the 1990s, tools vendors didn’t want to know about it because all their seats were focused on the 68K and Intel’s x86. It was “that funny little processor Nokia uses”. RISC-V is in no position to compete with Arm directly today but it is gradually gaining support and has been helped along by Donald Trump’s trade war against China. In forcing a decoupling of the US and China technology sectors, offerings that are not tied directly to Western companies and, even better, are supplied in open-source form in many cases look far more attractive now than they did a couple of years ago.

If Arm’s chipmaker customers do not like the direction nVidia takes with its proposed purchase, they have even more choices now than they did when Tudor Brown, former CTO of the IP vendor, described the company’s business model as a “poison pill” for most acquirers. 

Against that we have to also look at the reality of software developers being asked to switch out one instruction set architecture (ISA) for another. By and large, the customers of the chipmakers do not care who owns Arm. They care about getting something familiar at a good price and with a reasonable roadmap. None of that looks in jeopardy right now, though nVidia’s overarching strategy may ultimately cause concern. The developers in Arm’s traditional core market of embedded processors are probably not going to jump ship anytime soon because porting to a new ISA is such a painful process using traditional tools. And, unless nVidia does something daft like suddenly push up royalty rates, there will not be a strong financial incentive to move.

In higher-performance computing, the trend is very much towards managed-code environments and software frameworks that will target pretty much anything. Here, Arm is far more vulnerable, which would not be so bad if it were not the focus market now. Cloud computing technologies will move into edge computers and use much the same software tools. As long as they run the right frameworks, the people using them will not care what processor is inside. 

Then you have to consider nVidia’s overriding heavily AI-focused strategy. In building an Arm-powered supercomputer in Cambridge as – at the very least – a testbed, the company aims to steal a march on its competition. But you have to consider that the world’s current supercomputing champion – the Riken-Fujitsu Fugaku – already uses Arm processors and a bunch of specialised accelerators. As an Arm licensee, there was not much stopping nVidia from trying to build its own without spending up to $40bn on an acquisition, other than it will get to influence what goes into the Arm ISA itself. This is where the competitors to nVidia will start to look more closely. Is nVidia getting a jump on them? And if they believe that, why not go with their own option in the shape of something based on RISC-V or even their very own ISA? Hardware architecture in high-performance computing is going through a revolution, and it seems unlikely that anyone with the cash to do their own investment will see nVidia and Arm as being part of their future unless they make significant improvements in performance on their own.

At the same time, traditional Arm customers may begin to worry that their concerns are not addressed by focus on AI. This is probably not a major problem in the short to medium term for Arm because of the pain of transition. By the time that concern becomes evident, Arm would already have suffered a migration away in other sectors. The likely result given all these factors is that RISC-V moves in on territory Arm and nVidia covet most strongly while the rump business in embedded processors remains strong, though possibly with growing levels of discontent. 

At the surface level, the deal has the kind of attractive sheen that will convince Wall Street. But the details lurking underneath make it less enticing overall.