The Cloud Era Abides: Why Nvidia Has Not Subverted the Status Quo
In closing yesterday’s post, I wrote the following:
One analyst yesterday said that Microsoft’s mildly disappointing cloud results, combined with its ever-expanding capital expenditures, represented a “transfer of wealth from Microsoft shareholders to Nvidia shareholders.”
Tomorrow I will challenge such a facile view, which is, at best, glancingly accurate and only for a limited time, qualifiers that seriously diminish its utility as valid benchmark for buy-and-hold investors.
Okay, it’s tomorrow today. Well, you know what I mean. (And how solipsistic is it for me to quote myself? If I start referring to myself in the third person, take it as a cry for help.) It’s now time for me to explain myself. Why do I think the view summarized above is only partly correct and facile?
First, I want to restate something I’ve said before (not again!): Nvidia is a great company, highly proficient at engineering outstanding products that have steadily boosted its revenue and profitability. The Nvidia executive team is stellar – with exemplary leadership extending beyond Jensen Huang, its charismatic but relatable leather-jacketed CEO.
Nvidia's strength in depth, by the way, applies to any exceptionally successful tech company. Business journalists are drawn to the charms of a visionary founder/CEO, a narrative facilitated in no small measure by publicists who intuitively understand that forceful personal stories are easier for writers and readers to grasp than stories that reflect a more complex reality.
Anybody who’s worked at an outstanding technology company will tell you that organizational success never begins and ends with one person, no matter how brilliant. Behind the marquee name are several other executives, supported by an impressive breadth and depth of talent in nearly every functional department. Those people, the supporting players who mainly work behind the scenes, rarely get the credit they deserve, but they are indispensable to market leadership and commercial success.
So, Nvidia executives and employees – all of you – take a bow. You’ve worked the oracle (not that Oracle), and you deserve a rich reward. When the time comes, you can take some of your well-earned bounty and retire in style. Kudos to you and all you’ve accomplished.
Reader, you knew the dreaded “but” was coming, right? Well, yes, it’s about to make its impertinent appearance, but in no way will it detract from what I’ve just written. You see, Nvidia’s very success has set the stage for its downfall, which is certain to happen. The only questions are when and how far the company will fall.
The Cloud Era: Circa 2007 to 2035?
To explain, I have to summarize some relatively recent history in what I define loosely as the cloud era of the IT industry. The exact date when the cloud era began isn’t all that relevant to our analysis here, but let’s say, for the sake of argument, that it started circa 2007. We’re still in the cloud era, which will likely continue for at least another decade. These sweeping generational eras in the IT industry typically run for 25 to 30 years.
What’s that you ask? AI? Not, this is not the AI era. We are in a period of intensive genAI activity, but as Nvidia’s sales of GPUs to cloud giants and the proliferation of genAI workloads on clouds indicate, we remain firmly ensconced in the cloud era. The majority of actual genAI projects – as opposed to genAI brainfarts that never get beyond the conceptual stage – are running in the major IaaS clouds of Microsoft Azure, AWS, and Google Cloud.
GenAI has not displaced cloud. If anything, it has bolstered cloud’s hegemony.
What will come after the cloud era? Maybe, at some indeterminate point, we will see the rise of artificial general intelligence (AGI), but that hazy prospect remains in the realm of science fiction, more than a decade (perhaps much more) from now. At that point, if it should come to pass, AGI will bring intelligently autonomous devices to offices, industrial settings, and homes across the globe. But it has many hurdles to clear before it gets real, and the challenges it confronts appear formidable.
Let’s get back to the present, the cloud era. One defining attribute of the cloud era is how the cloud giants shun reliance on any single supplier of technologies deemed essential the execution of their businesses. Why is that? Unlike enterprises, everything a major cloud provider does in its on-premises computing environments (datacenters and edges) directly contributes to the top and bottom lines of cloud business. Datacenters, and the IT infrastructure that fills them, are the cloud’s integral, indispensable engines.
Given that reality, cloud service providers have an intimate understanding of their precise technology requirements at any given time; they invariably know exactly how the infrastructure they use should be designed, manufactured, deployed, and operated. They don’t need vendors to advise them on what to do, and they don’t want to rely on any single vendor as a key supplier of an essential technology. If, as a cloud giant, you have such a dependence, you lose control of your own business metrics.
Dependence on a sole-source, third-party vendor results in multiple suboptimal consequences for cloud providers. First, the cloud provider must pay a premium for the sole-sourced products or components, manifesting in unsustainably high capital expenditures. Two, the cloud provider loses a measure of control over the timing and volume of product deliveries and technology deployments. Third, the cloud provider loses at least some capacity to meaningfully differentiate its services, since the sole-source supplier also delivers the same product to other cloud providers and enterprises. Fourth, the cloud provider must trust the integrity, quality, and security of the third-party vendor’s technology, regardless of whether it is sold as software or as software embedded in hardware.
The Datacenter is the Business
For those reasons, the cloud giants, more than a decade ago, took ownership of the key specifications, technologies, and supply chains relating to the IT infrastructure that was designed, developed, and implemented in their datacenters and other on-premises environments. These initiatives covered servers, storage, networking (primarily datacenter switches and the software that runs on them), as well as orchestration and automation software.
That is why, in networking, Cisco’s market share in datacenter networking declined gradually on an annualized basis as we moved through from 2012 through into the early 2020s. Cloud providers specified their network-hardware requirements and bought bare-metal switches from white-box ODMs, shunning name-brand OEMs such as Cisco. For the reasons mentioned earlier, cloud giants did not want to pay a premium for Cisco network gear and they did not want to rely unduly on Cisco as a supplier. There was no animus against Cisco, and nothing personal in the decision to move away from Cisco. It was purely a business decision, based on a thorough cost-benefit analysis. The same, of course, applied to servers and storage, and even to the CPUs and other processors that powered servers.
Ask yourself: What is so different about Nvidia’s position that makes its GPUs irreplaceable for cloud providers? The whole palaver about the alleged value of Nvidia’s software is nothing more than propaganda. All the vendor hardware I discussed above came with software. In fact, in the earliest years of the cloud era, the cloud providers, like the enterprises that preceded them, used the proprietary vendor software that came with vendor hardware. At a certain point, however, as the cloud giants scaled their datacenters and regions, it made no logical business sense to continue doing so – so they stopped doing it.
Similarly, we are at a point now – in fact, we’ve passed it – where it no longer makes sense for cloud giants to continue buying GPUs and AI accelerators from Nvidia. Again, it is a business decision, and the business pressures that drove it have become painfully acute, as evidenced by the intensified beatings that cloud behemoths are taking from financial analysts and institutional investors.
You need look no further than this past week’s quarterly results from Google and Microsoft – and from Amazon today – to see why the cloud giants will gradually, though not all at once, walk away from Nvidia.
What is Wall Street’s most vehement complaint, getting louder by the day? The good burghers of Wall Street are hopping mad about the rising Capex costs associated with genAI buildouts, especially since the latter has yet to generate anywhere near sufficient revenue to recoup those costs. Even if the genAI wave comes ashore, with enterprises merrily ramping up AI spending in the cloud, the hyperscalers will want to rationally scale their genAI business in a way that delivers robust and predictable profit margins. Can they do that if Nvidia is demanding its fat, juicy premium on pricy GPUs? It’s not likely, is it?
At this point, you might counter and say, “Well, Nvidia could drop its GPU prices, reducing the margins associated with their sale.” Yes, theoretically that could happen, but Nvidia has its own eagle-eyed investors and it has to answer questions from Wall Street analysts, too.
Besides, the logic of the cloud era dictates, as I have explained, that hyperscale cloud providers have and will take responsibility for the design, delivery, and deployment of technologies that are integral to their business success. To its credit, Nvidia understands the relentless business logic that governs the cloud era. The company knows its cloud patronage is here for a good time, not a long time. Give Nvidia credit for the foresight and execution that has resulted in a stratospheric market valuation but understand also that the company has maximized its advantage to the breaking point in cloudland.
All good things come to an end, and, at some point in the foreseeable future, Nvidia will stumble, after the cloud giants, one after another, begin specifying and deploying their AI chips, manufactured by commercial foundries (a business Intel, among others, is desperate to break into) at a fraction of the cost of Nvidia’s chips. Costs and control are the determinants here.
The cloud giants have been developing their own AI chips for a while now. Google, Microsoft, Amazon, and – in the sphere of consumer-oriented cloud – Meta (Facebook) all are advancing AI-chip development initiatives. Some, such as Google and Amazon, are further along than others, but they’re all moving in the same direction. Similarly, the cloud giants have begun buying GPUs from Nvidia’s competitors, principally AMD, which is still a long way behind Nvidia in market share and is already having trouble keeping up with demand. Ideally, the hyperscale clouds would like to take matters into their own hands rather than fostering dependence on a competitive foil to Nvidia, but they welcome any and all alternatives that defray costs and allow them to reassert control.
Gears Shifted, Wheels in Motion
These efforts take time to come to fruition, which means we’re at stage in tech-world realpolitik where the cloud giants will continue to tap Nvidia as a GPU supplier while they assiduously work toward a diminution of the relationship. This is why you will rarely, if ever, see or hear any criticism or disparagement of Nvidia from the cloud giants. (Tesla’s Elon Musk, occasionally a loose cannon on earnings calls and on X, railed against Nvidia’s GPU tyranny recently while touting his company’s own plans for AI accelerators, but that outburst was the exception to the rule of diplomacy and reticence.)
All the cloud behemoths have relationships with Nvidia that are not particularly warm, but they are necessary and, for the interim, mutually convenient. Nvidia understands this ambiguous dynamic, and it will continue to supply the hyperscale clouds while also attempting to cultivate a tier-2/3 AI-cloud market intended to take up some of the demand slack the cloud giants will leave in their wake.
Do not mistake silence for inactivity, however. As the cloud behemoths persistently emphasize a long-term need to invest heavily in AI infrastructure while pushing out the dates when they expect to see returns on those investments through enterprise demand for genAI services, they are advancing multiple initiatives, including AI chips, that will reduce or limit surging Capex outlays. Much of the work is not extensively publicized, but the need to report tangible progress grows more pressing as AI costs balloon while prospective revenue realization from commercial AI services recedes further into the future horizon. There’s not much the cloud giants can do about the precise sequencing of commercial uptake of AI services, which will or won’t happen in the fullness of time, but there are near-term measures they can take to meaningfully check rampant costs.
One of those measures, which demonstrates how seriously the cloud giants want to liberate themselves from dependence on Nvidia, is the UALink Promoter Group, which features Intel, Google, Microsoft, and Meta, as well as IT infrastructure stalwarts including Cisco, AMD, Broadcom, and HPE. The focus of the UALink Promoter Group is to direct development of components that connect AI accelerator chips in datacenters. For different reasons, neither Nvidia nor Amazon is involved in the initiative; Nvidia because it knows it’s a target of the group, and Amazon likely because it believes it has its own house in order.
To summarize, my key message here is simple: We’re still in the cloud era, strongly characterized by cloud hyperscalers’ dominance over the ecosystem and supply chain of IT infrastructure. It might appear to some that Nvidia somehow has subverted the cloud hierarchy, and that a purveyor of “picks and shovels” has jumped to the forefront of a new genAI revolution, akin to the picks-and-shovels predominance of vendors such as Cisco in the Internet era.
Such an interpretation, however, is illusory. We’re still in the cloud era, and genAI is merely a potentially lucrative extension to the cloud platform. It’s true that genAI requires new picks and shovels (GPUs and AI accelerators), but AI workloads, for reasons I’ve addressed here before, are aggregating in cloud datacenters, not in on-premises enterprise datacenters – or anywhere else for that matter.
Anticipating the rise of genAI, and the utility of its own GPUs as effective AI processors, Nvidia deserves full credit for its prescience and for catching the cloud giants napping. In the final analysis, however, Nvidia remains an OEM supplier of a key technology that cloud hyperscalers now identify as essential to the cost containment, growth, and profitability of their datacenter-centric businesses. Like the OEM vendors of servers, storage, and network infrastructure before them, Nvidia is susceptible to inevitable displacement by GPUs and AI accelerators designed in-house by the cloud titans themselves.
Nvidia has done exceptionally well, exploiting a lucrative opening and generating untold wealth for its executives, shareholders, and employees. It has not, however, subverted the natural order of the cloud era. Despite Nvidia’s attempts to avoid the fate of fate of Cisco in the 21st century, there’s probably nothing the company can do to escape a similar fate. There’s an inevitability to how this will play out, and the wheels are already in motion.
In the meantime, Nvidia might still be a stock worth holding from quarter, even with its bloated valuation. Trading algorithms see the trend as their friend, and the trend for Nvidia, buoyed by an exuberant narrative, remains favorable at this particular moment. At some not-too-distant point, though, the ground will shift, as one hyperscale after another incrementally and then definitively reduces its reliance on Nvidia as a GPU/AI accelerator supplier.