In GPU We Antitrust

With all of this chatter about China looking into possible violations of antitrust law by Nvidia, and regulators in both the United States and the European Union also having done the same, let’s play the “What if?” game for a just a moment.

These “What ifs” are based on the idea that a court, most likely in the United States because America, in theory, has more jurisdiction over companies based within its borders than do those outside, might get a complaint about Nvidia, possible from the Justice Department, possibly from a regulator or lawyer in one of the 50 states. We are not saying this is probable, but just doing a little war gaming to see what this might look like when there is a trade war and when one company has so much sway over the stock market and the revenue and profit pools in the IT sector these days, which will continue to raise the eyebrows of antitrust regulators.

• What if a court decides that the CUDA software stack is so strategic to the future of computing that it has to be open sourced or licensable at a low cost and must be allowed to be ported to other compute engines to stimulate demand and lower software barriers to promote lower cost computing and competition in the AI server sector?
• What if a court decides that the NVLink protocol has to be open and licensable at what it deems is a fair price?
• What if a court decides that Nvidia has to really and fully unbundle its systems software from its GPU hardware and price out all of its software and charge a reasonable price for it that reflects its actual value?
• What if Nvidia is prohibited from selling complete servers or finished clusters?
• What if a court decides that the Nvidia datacenter business has to be broken free from the rest of Nvidia, which is focused on graphics for game consoles, PCs, and workstations?

We threw that last one in because it is the kind of thing that antitrust lawsuits often pursue, and often because they are easy to do even if they don’t really go after some of the anticompetitive behavior regulators are against. This is, for instance, a parallel to Microsoft having to break its Internet Explorer browser free of Windows two decades ago and why the US Department of Justice and a bunch of states are now proposing that Google break off the Chrome browser and the Android operating system from the rest of the company to try to break its hegemony in online search and advertising.

Courts do weird things, regulators do weird things, and it is important to keep this in mind as you grow your business to near-monopoly status. The situation could get weirder still if regulators in the United States, Europe, and China all agree to attack Nvidia at the same time even if for different reasons and in different ways.

Which is why the decision by the Chinese State Administration for Market Regulation (SAMR) to launch a probe into Nvidia over monopoly concerns with regard to its $6.9 billion acquisition of Mellanox Technology, which was completed in April 2020 after passing muster with all of the global antitrust regulators, is perplexing. We were not expecting any issues on that front, but apparently there were some conditions to the SAMR’s approval of that Mellanox deal that are no longer being met. Those conditions were not explained in any of the statements we have seen. We reached out to Nvidia for comment and the company did not anwer any questions directly about this issue.

We are curious about all of the details in the ongoing probes into Nvidia, and we are also wondering what an incoming Trump administration thinks about Nvidia and its clear hegemony in AI servers – and what it might do about it. Every presidential administration has its own ideas about competition and monopoly power, and can leverage the power of various agencies to compel corporations to either do something or to stop doing something.

Nvidia’s GPU hardware and software is at the heart of a new datacenter computing architecture, and one that is absolutely as profound as the advent of the System/360 mainframe more than six decades by IBM. Given IBM’s woes with the US Department of Justice, the comparison is perhaps more apt than Nvidia co-founder and chief executive officer Jensen Huang might like.

During the Great Depression and in the wake of World War II, IBM became the dominant supplier or punch card and tabulating equipment that comprised datacenter computing at the time. In 1952, the US Justice Department’s Antitrust Division sued Big Blue, and after four years of fighting in the courts, IBM signed a consent decree in 1956 governing its behavior. This settlement was amended in 1969 to cover more modern mainframe and minicomputer systems, which comprised a much larger and much more profitable market for IBM, and to address some concerns in a second antitrust lawsuit that was filed against IBM by the Justice Department and that dragged on until 1982.

One could argue that IBM’s monopoly power in tabulating equipment allowed it to gain a monopoly in mainframe computing, but we are not interested in that right now. Nvidia might do something in the future, like start a datacenter construction business, that might make such an analogy interesting. But that is not today’s discussion.

With the evolution of client/server computing and the widening of the systems market, IBM was able to wiggle out of many of the provisions of that consent decree in 1994, and the rest of them were removed in 2001. The timing was important for the consent decree being lifted. This was when IBM was having its “near death experience” and when a new chief executive officer, Louis Gertner, kept the company together rather than split it up and pivoted to selling services directly rather than at arm’s length as it was compelled to do by that consent decree four decades earlier.

This consent decree and its update was invasive and was absolutely meant to control IBM’s behavior in the market; it shaped Big Blue’s thinking about what it could and could not do for decades.

Among other things, this Consent Decree required that IBM:

• Provide both sale and rental prices for its equipment, with a reasonable relationship between the two.
• Open up the specifications of peripheral controllers so third party equipment could plug into IBM systems.
• Support systems that had third party memory or third party peripherals – storage, printers, other controllers, terminals, and so forth.
• Allow for customers to run IBM operating systems on clone hardware at a reasonable price.
• Allow for third party lessors to buy IBM systems and lease them to end users.
• Keep its services business at arm’s length to allow for the emergence of other service providers on IBM mainframes and minicomputers.

Our point in bringing this up is two fold. First, regulators and the courts can be clever, or get it wrong, and still compel behavior that IBM then or Nvidia now does not like. And second, every argument about monopoly comes down to relevant market size and every argument about antitrust usually comes down to tying of one product to another one.

If you look at CPU shipments over time from the Dot-Com Boom to the GenAI boom, for many of those years Intel’s X86 server chips had well over 90 percent shipment share, with varying degrees of direct competition from AMD and indirect competition from Sun Microsystems, Hewlett Packard, IBM, SGI, and others. Intel was able to get close to 50 percent operating income out of its datacenter business.

IBM’s mainframe platform business has well north of 85 percent operating margins on its software even to this day, and usually somewhere around 30 percent to 40 percent operating margins on hardware. (The hardware margins were a lot higher two decades ago, and so was the revenue stream for mainframe hardware.) Clearly, having software to tie to hardware is profitable, and Intel did not have that even though Microsoft and Intel operated as a duopoly of sorts, taking on other systems in the datacenter and capturing those mainframe and minicomputer profits gradually over the 1990s and 2000s as customers unplugged this gear or added a lot more client/server capacity wrapped around their IBM systems.

As we have said many times before, 75 percent of Nvidia’s employees work on software, but software represents maybe 1 percent of its revenue stream. It is not unreasonable to want Nvidia pricing to reflect what customers are really paying for and for Nvidia’s revenue streams to be allocated to more accurately reflect its value and profits. But Nvidia doesn’t like that model, saying that it gives the software away with its hardware excepting the $4,500 per GPU per year charge it recently instituted for its AI Enterprise stack.

It could turn out that US antitrust regulators reckon that Nvidia is tying its software to its hardware, and hiding the value and profits it extracts from that software. This might be why there is an AI Enterprise license at all, in fact, to start the process before regulators get to thinking about it all.

IBM had more than 95 percent share of tabulating equipment and then almost the same share of mainframe processing and was for all intents and purposes a regulated monopoly. Intel had more than 95 percent share of datacenter CPU shipments for a lot of years and was not sued and not regulated (for that anyway). Nvidia has had probably north of 85 percent market share for datacenter GPU computing since 2008, and it is probably close to 95 percent of revenues and shipments. That GPU shared is going to soon drive half of server revenue worldwide and very nearly all the profits from servers. If the Trump administration wants to compel competition in AI serving, there are ways it can do so and there are historical precedents for it.

The question is, do regulators think it looks more like IBM decades ago or more like Intel a decade ago? Is the Justice Department or the Federal Trade Commission more or less aggressive under Trump than it has been under Reagan, Bush, Clinton, Bush, Obama, Trump, and Biden?

We are a lot less worried about what Chinese or European regulators might do than we are about what US regulators could do. Europe and China can impose fines and compel some changes in behavior, but Nvidia doesn’t need China. And Europeans need GPUs more than they need Nvidia to do anything more but supply more of them.

For the past two decades, there has not been aggressive enforcement of antitrust laws and perhaps no one wants to mess with the Nvidia that is laying the golden eggs right now excepting some curbing on sales of technology to China and Russia.

But, you never know what Trump might think, or Elon Musk, his chief technology officer and chief operating officer, for that matter, too.

What we do know is this: If governments want to regulate Nvidia in the datacenter, they had better shake a leg. With all of the hyperscalers and big cloud builders working on their own AI accelerators, and the AI startups hanging in there looking like real competition, this might be peak Nvidia happening here in fiscal 2025 and fiscal 2026. We think competition will increase for Nvidia from here on out, in no small part thanks to AMD and not having anything to do with Intel. (Gaudi2 and Gaudi3 accelerators are a fraction of a percent of Nvidia’s GPU sales and shipments.)

And maybe that is the lesson. Let the market regulate itself except where it absolutely cannot in a timely fashion. We think open source CUDA and provisions for allowing the emulation of CUDA directly on non-Nvidia hardware would be very interesting indeed. . . .

As we have said a bunch, we don’t hate monopolies, and sometimes due to the scale of the problem, we think monopolies are natural, inevitable, over a short term and sometimes over a long term. But we do believe that monopolies have to be regulated to prevent abuses of power. Thus far, we think Nvidia has created a market – with a lot of help, to be sure – and is benefitting from first mover status and a hell of a lot of hard work and a little luck. There is no law against this and there should not be. But if Nvidia over-reaches, or decides to tie InfiniBand or Spectrum-X sales to GPUs sales, or its software to its hardware, well, that’s different.

One last thing: After we went to press, we got a statement from Nvidia after asking for some explanation about the claims SAMR was making. This is what the company said: “Nvidia wins on merit, as reflected in our benchmark results and value to customers, and customers can choose whatever solution is best for them. We work hard to provide the best products we can in every region and honor our commitments everywhere we do business.  We are happy to answer any questions regulators may have about our business.”