Nvidia’s $40bn proposed acquisition of ARM represents a greater threat to the UK-based company’s unique model, and to the ecosystem it has enabled, than its previous takeover, by Japan’s Softbank, which bought ARM in 2016 for $31bn.
This summer, UK politicians were echoing their counterparts in the USA, Japan and elsewhere and calling for development of more homegrown 5G technology and intellectual property to reduce reliance on Huawei. In the UK, this just evoked a sad sense of lost opportunity, and memories of how all the technology that has been developed at home, only to be lost to acquisition or lack of support.
One of the greatest losses to ‘UK plc’, and indeed ‘UK 5G’, was ARM, when it was acquired by Japan’s Softbank in 2016. In reality, the days are long gone when the UK could aim to be a 5G leader, and in a platform that needs to be fully global to succeed, the country that owns the fundamental technology players does matter, as long as the companies remain cooperative, share their intellectual property where appropriate, and are supported in pushing innovations forward. Softbank retained ARM’s licensing model and, until the company hit its recent financial crisis, provided a welcome financial cushion, even if some of the development priorities seemed to be skewed more towards Softbank chairman Masayoshi Son’s grand vision, and less towards ARM’s core customer requirements.
But when the crisis hit – sparked by misjudged investments in companies such as WeWork, and exacerbated by over-extension and by the pandemic – Softbank had one crown jewel left to monetise, and that was ARM. It is to sell ARM to US chip vendor Nvidia, which – if allowed by regulators, which is a big ‘if’ – would mark the slow death of a licensing model that has enabled common underpinnings and massive scale in the mobile device business.
ARM made a rare success of licensing intellectual property (IP) for processor cores, including for CPUs, graphics processors and microcontrollers. It rode the wave of miniaturization computing capabilities within mobile formats and undisputedly, its greatest success has been the smartphone, providing CPU and GPU cores for increasingly powerful handsets, and enabling an ecosystem of massive scale. Over 180bn devices have been shipped based on ARM designs, and Nvidia says the deal will expand its developer base almost sevenfold, adding ARM’s 13m to its own 2m.
The ARM model in smartphones saw large numbers of licensees able to develop devices without the time and cost involved in creating a processor from scratch, and with commonality of testing, software and skills. They could innovate in many areas from system-on-chip (SoC) integration to form factor, while a few companies – like Qualcomm and Apple – acquired architectural licences in order to develop their own core designs within the ARM framework.
Nvidia pulled out of the smartphone SoC market in 2016 after buying modem maker Icera in 2011, but struggling to make any inroads into Qualcomm’s dominance. So ARM’s most famous business is presumably not the attraction, but instead, its Neoverse infrastructure CPU architecture. Nvidia has the opportunity to eat into Intel’s server incumbency with its very different architectural take on high performance cloud computing, to support artificial intelligence and other demanding use cases. Its GPU-focused platform is even sparking interest in the difficult virtualized RAN area, notably with a joint development with Ericsson of GPU-based infrastructure to support near-real time base station processes in the cloud.
But none of this needs an ARM acquisition. Nvidia licenses ARM IP, and it could invest in a full architectural licence if it wants more freedom in how it differentiates its CPUs.
This is about full control of an ecosystem that can be as broad as Intel’s in server and cloud chips. In GPUs, Nvidia has been licensing, or open sourcing, some elements of its own technology, and has built an ecosystem around its CUDA parallel processing platform and programming model. Buying ARM may be a way to accelerate the extension of its technologies into the huge ARM ecosystem, effectively creating a single GPU-based platform from devices to supercomputers and converging the surrounding tools and APIs.
The ARM purchase would mean Nvidia had technology in CPUs, GPUs and high speed networking design (courtesy of its Mellanox acquisition), plus a licensing business. CEO Jensen Huang told a press conference to announce the bid: “We will now have in one company three franchises and three incredible platforms. The amount of computer science horsepower that is inside this company would be quite extraordinary, and it allows us to create solutions and platforms that can then be made available to this vast network of partners and developers.”
Huang claimed that it would be feasible to combine Nvidia’s research efforts with those of ARM, while also preserving ARM’s open licensing and customer neutrality, and extending its licensing processes to Nvidia IP too.
“The customer should realize how excited I am about the business model,” Huang said.
In fact, we’re so excited about the business model we’re going to take Nvidia technology and put it through ARM’s vast channel networks.”
Huang was particularly focused on ARM’s Neoverse architecture, which underpins cores for server processors, and which fits into Nvidia’s key market in high performance computing (HPC) and cloud infrastructure. It has proved challenging to get ARM processors into servers at scale, and Qualcomm and Marvell are among the companies which have defocused on general purpose ARM-based server chips in favor of more tailored products. But Amazon does use ARM cores in its Graviton chips, and that route into a hyperscaler would be worth a good chunk of the price of ARM from Nvidia’s point of view. And converging Neoverse with its own GPU-based architecture would give it a richer and more powerful platform with which to compete with Intel in the latter’s stronghold, as well as with other major infrastructure processor vendors like Marvell.
Of course, it is early days for Nvidia to explain, or even know, all the details of how it would run ARM in future, but it will soon have to start drawing up detailed plans to share with customers and regulators. Huang is keeping his options open for now, but even the understandably vague answers he gave to journalists at the press conference hinted at a major change in business model – one in which Nvidia would build CPUs for companies which licensed the IP.
“Some people would like to license the cores and build a CPU themselves,” he said. “Some people may decide to license the cores and ask us to build the CPUs or modify ours and we could build it ourselves. All of those options exist.”
The vision is ambitious – bringing Nvidia and Mellanox IP into the ARM licensing framework in order to provide tailored licensing deals, or reference designs, or even whole chips, which use different combinations of the three platforms, according to a customer’s requirement.
“There are going to be so many different types of data centers in the future. There are gigantic cloud ones; there are smaller ones for enterprise; there are high power density ones for supercomputing; there are ones that are going to fit in locations all over the world in distributed data centers,” Huang said. “It’s not possible for one company to build every single version of them, but we’ll have the entire network of partners around ARM that can take the architectures we come up with.”
Success would see Nvidia dominating the core technologies underpinning future servers. It might enable additional competitors to itself, but it would also control the strategic direction for a whole ecosystem, and so put itself in a stronger position to seize that initiative from Intel, and to appeal to the critical customers, the hyperscalers.
But while it can certainly be argued that such an approach could drive more competition into the high end server market, it still leaves the bulk of ARM’s commercial business looking like a poor relation. It is hard to believe that the cores targeted at handset processors and embedded applications will get the same love from Nvidia as Neoverse, so licensees will fear that, even if Nvidia keeps the business going, they will not see the same levels of innovation and investment that their cores have enjoyed in the past.
And if it does envisage a microcontroller-to-server platform and revenues, it needs significant buy-in from the ARM community, and there are some obvious heavyweight members which will have significant conflicts of interest in supporting an IP framework that is fully controlled by a rival. Qualcomm, Marvell and others do not just compete with Nvidia for individual contracts, they put forward an entirely different architectural philosophy built around their key areas of expertise. So even if Nvidia preserves ARM’s licensing business and runs it as a fully separate business, no Chinese walls will be high enough to stop the gradual seeping of the Nvidia view of the word into the ARM roadmap.
Apple and Qualcomm are just two ARM licensees which have their own GPU roadmaps and would not welcome a convergence of ARM Mali with Nvidia’s cores. Other ARM licensees, like Marvell, compete head-on with Nvidia in server and infrastructure chips. And still more ARM customers will question whether ARM will be allowed to prioritize development of cores in markets that do not interest Nvidia, which could result in weaker offerings for licensees in those markets. Ironically, the smartphone sector could be particularly vulnerable to any diverting of resources towards Nvidia’s key areas of focus. It is unlikely that Nvidia will try to re-enter that space, and the growth in the market is declining, so the secret of ARM’s huge success could find itself on the sidelines.
All this will send some ARM licensees looking for alternatives once current contracts, and technology generations, come near their end, and ARM will certainly lose some of the close partner cooperations which have helped it to enhance its cores in so many directions. For a chip company to work with ARM on a development that accelerates progress for the whole industry is a very different matter from handing over key expertise to a single competitor.
All this may become moot if regulators block the deal. It requires approval from antitrust authorities in the USA, China, the UK and the European Union. It seems almost impossible that China will approve the deal, unless it can extract significant concessions over the USA’s current bar on US companies selling components to Huawei and other Chinese firms on the country’s ‘entity list’. Chinese approval delays had such an impact on the value and timescales of Qualcomm’s proposal to buy NXP, in 2018, that the deal fell through.
Amid escalating intellectual property, cybersecurity and trade wars, China voiced concerns that some key patents and technologies owned by NXP would fall into the hands of a US company and possibly be denied to Chinese customers. Two years later, the situation is far graver, and the ARM technology far more fundamental to Chinese chip, device and server makers. The status of ARM China, a separate legal entity in which some intellectual property is ringfenced for Chinese use, is one of the many complex and as-yet unexplained aspects of the Nvidia deal, but even if excluded from the acquisition, it would quickly lose its value if it could not access new IP coming from the main company.
It is unlikely that the UK regulator will put up barriers, though some parties believe it should seek safeguards of UK jobs, given that ARM still has its largest bases in the country. Nvidia claims that ARM will remain in its Cambridge headquarters, and that the US firm will invest in a new AI R&D lab there – but past experience teaches UK citizens not to put too much faith in the promises extracted by its competition authority (when US-based Kraft Food acquired the UK’s Cadbury’s, it took only months for promises of job protection to be broken). And in this case, ARM is no longer even a UK-owned company.
Beyond geopolitics and jobs, the regulators in all jurisdictions, including the USA, will be lobbied hard by ARM licensees which are also Nvidia competitors (see above), and some of these have powerful voices. It remains to be seen whether the US Department of Justice will be more swayed by the prospect of claiming such a foundational hi-tech company as ARM for the America-first 5G dream, or by the objections from Qualcomm and others that this deal would damage their own competitiveness.
The proposed deal:
Nvidia proposes paying $40bn for ARM. Softbank will retain ARM’s IoT services group, and a stake in the combined firm of less than 10%. The price is seen as very generous for a company whose revenues were $1.9bn last year, according to Softbank’s annual report. Nvidia’s revenues were about $11bn.
Nvidia will pay over $21.5bn in shares and $12bn in cash, with a further payment of “up to” $5bn in either cash or stock if as-yet undisclosed financial performance targets are hit.
ARM employees will be allocated shares worth a total of $1.5bn as an incentive to engineers to stay on board.
Nvidia said it would “expand ARM’s R&D presence in Cambridge, UK, by establishing a world-class AI research and education center, and building an ARM/Nvidia-powered AI supercomputer for groundbreaking research.” It also promised to “continue ARM’s open licensing model and customer neutrality and expand ARM’s IP licensing portfolio with Nvidia technology.”
Softbank sells other telecoms assets too:
ARM is not the only hi-tech asset which Softbank is divesting as it seeks to right its lurching financial ship. It has already, of course, sold most of its stake in Sprint to T-Mobile USA, and now it has announced the sale of a majority stake in mobile phone distributor Brightstar, which generated over $9bn in revenues last year.
Control of the distributor will be sold to a private equity firm of the same name, Brightstar Capital Partners (BCP), which will buy 75% of the business for an undisclosed fee, leaving Softbank with the rest. That leaves the Japanese company with no major telecoms businesses outside its home market MNO – and it has even sold 22% of the Softbank operator through a secondary market offering to raise funds.
Brightstar’s performance has been poor in recent years, although it did reduce its operating loss in its last fiscal year, to ¥5.3bn ($50.8m) from about ¥44.7bn ($428.5m) two years earlier. But sales fell to ¥955bn ($9.2bn) last year, from ¥1.08 trillion ($10.4bn) the year before.