Here’s why so many tech companies believe Nvidia CEO Jensen Huang’s proposed acquisition of Arm is bad for business...
Late Friday last week, news broke that three big tech companies reportedly came out against Nvidia’s proposed $40 billion buyout of ARM, one of the biggest and most consequential tech deals ever, with broad implications from the cloud to the edge.
Note the word “reportedly.”
It’s been five months since Nvidia CEO Jensen Huang gushed about the potential for pairing Nvidia’s AI capabilities with Arm’s pervasive compute platform. And though the ecosystem broadly maligns this deal, not a single Big Tech exec has spoken out publicly against it.
At first blush, that may seem surprising. Because the doomsday scenarios industry insiders privately spin are as damaging to competition, pricing and innovation as they are far-reaching, with the potential to make vital electronic cogs more costly and less secure.
But more than anything, the execs I’ve spoken with over the past few months — the conversations have all been off the record or on background — say the outcome they fear most is that they might set themselves up for retribution by Nvidia if they object to the acquisition, and the deal is consummated anyway.
Regardless of whether there’s merit to the concerns, they do underscore the undue influence and power Nvidia would garner with this acquisition.
What’s in it for Nvidia
Which is why the FTC should demand a better explanation from Huang for why he feels he needs to spend $40 billion to buy a company with technology he could just license like everyone else. In fact, he already does license it like everyone else.
So how could Nvidia justify the exorbitant price tag, the most ever paid for a semiconductor company? Granted, much of it is in stock. But Nvidia will still have to fork over at least $12 billion in cash. And it’s hard to fathom how the company would otherwise spend that much on Arm licensing and royalties.
The return on investment for Nvidia, then, would have to come by redirecting Arm development and licensing policy to maximize its own profits. When you consider that Nvidia could either collect hundreds of dollars each for its own datacenter products, or maybe $8 apiece in royalties, the path to profitability for Arm’s hopeful new owner is clear: prioritize the company’s own datacenter designs.
If Nvidia were to redirect resources to the datacenter, that could, in turn, throttle the pace of innovation in smartphone, automotive, IoT — everywhere else. With security enhancements inevitably slowing as well, platforms would be more vulnerable to hacks.
Were that to happen, how would the ecosystem recover? Suppliers can’t just hop en masse to another architecture. And even if they did, they’d have to leave behind more than a decade of tuning Arm performance and security for those applications. Not to mention a mature developer ecosystem and an installed base of Arm-optimized applications.
Meantime, as innovation stalls in other segments, the new Arm owner would be motivated to move even more workloads to the datacenter, siphoning evermore value and profitability from the edge.
Threats and opportunities in the datacenter
The meteoric growth of distributed computing, along with high-speed networking technology to keep virtual machines and microservices humming, have created a rare opening for Arm in the datacenter. So the acquisition would present Nvidia with an outsized opportunity to reshape the landscape to its advantage — timely, as conditions are also making it more difficult for Nvidia to swat away competitors.
Nvidia has been addressing the emerging datacenter opportunity — but so have others. Nvidia bought Mellanox last spring for $7 billion, and then leveraged the newfound smartNIC capabilities to build its new Bluefield-2 data processing units, or DPUs. Bluefield-2 is ARM-based, and system designers will be able to program them using Nvidia’s upcoming DOCA SDK.
DOCA could help lock adopters into the DPU platform, in much the same way that the company’s CUDA SDK did in the early days of GPUs in HPC. But Bluefield-2-on-DOCA doesn’t have the same draw — or the same lock-in potential — as Nvidia GPUs with CUDA. For one thing, there are plenty of alternatives, many of which are built on the same ARM architecture.
And Nvidia’s AI hegemony is increasingly coming under attack, from emerging GPU competitors as well as suppliers of FPGAs and ASICs all taking aim at the lucrative market. The popularity of higher-level open-source tools like PyTorch, TensorFlow and Caffe has helped open the market by loosening CUDA’s grip on development.
To complicate matters for Nvidia, Amazon AWS, Microsoft Azure and Google’s GCP — the biggest potential customers for datacenter silicon — have all been developing their own ASICs for AI as well as Arm-based DPUs.
At Arm’s length?
With Arm on the inside, Nvidia would have myriad levers at its disposal to help give its own silicon unfair advantage. At a time, frankly, when Nvidia could really use it.
To tip the scales, for example, Nvidia might offer early access to the latest features of Arm’s next-generation 64-bit server architecture exclusively through DOCA. Hooks into Nvidia GPUs for AI training and inference could help insulate that business from competition as well.
Taken together, those moves might drain cloud vendors’ motivation for continuing to invest in their own silicon designs. And with most Arm development focused on the datacenter, some smartphone, automotive and IoT players might drop out in search of greener pastures. Which would lead to fewer choices, higher prices and eroding security in those markets.
So it’s no wonder Nvidia’s Huang is willing to fork over so much for Arm.
And it’s no wonder so many industry execs are crying foul. Softly, out of earshot. Where we can’t hear them. And neither can Huang.
— USA TODAY columnist Mike Feibus is president and principal analyst of FeibusTech.