With Enlightened Self-Interest, Nvidia Reshapes The Tech World In Its Own Image

UPDATED: Watching Nvidia grow over the past two decades, from a gaming and visualization GPU supplier to the behemoth of the modern AI datacenter, has been nothing short of amazing. Few people alive today have ever seen anything like it, and for those who were alive to see it back then, Nvidia, it is now safe to say,  has now surpassed the mighty influence that IBM once wielded in the datacenter in the 1960s and 1970s during the first wave of corporate computing.

Nvidia’s stratospheric growth was the result of hard work, sweat, investment, foresight, and good luck. And not just once, but several times. Nvidia became a supplier of massively parallel compute engines that could also do visualization and then computation in HPC centers; then again as GPUs accelerated various kinds of databases and data analytics; then again as statistical machine learning usurped HPC; and finally as the platform where a lot of GenAI models were born and live. (But not all – Google built transformer models ahead of everyone else, and did so on its own TPU engines a lot of the time.)

Being the newshounds and analytical engines that we are, we were somewhat chagrinned that the time we were visiting our aging mother was the time that Nvidia decided to invest some cash in Intel and then followed this up with a whopping $100 billion promise to help OpenAI meet its infrastructure needs. But we are back in the office now, and are ready to run these announcements through the wetware inference engine we keep between our ears and behind our eyes.

Let’s start with the Nvidia partnership with Intel. Nvidia has shelled out $5 billion to buy that much of Intel’s stock as part of a broader partnership to create Intel Xeon CPUs that have NVLink ports so they can plug into the memory interconnect in Nvidia HGX server nodes and NVL72 rackscale systems much as its own “Grace” CG100 Arm server CPUs do today. The partnership will also see the two companies work on hybrid Intel Core and Nvidia RTX system-on-chip designs that will also use NVLink interconnects to glue the CPUs and GPUs together.

It would be easy to jump to conclusions and think that is the end of the Arc X^e integrated and discrete GPU accelerators for PCs and the future “Jaguar Shores” AI accelerator that Intel has been working on in the wake of its ill-fated “Ponte Vechhio” Max GPU accelerator and the side-lined “Falcon Shores” accelerator that was to combine the best parts of its GPU and Gaudi accelerators to take on Nvidia and AMD GPU compute engines in the datacenter.

In the wake of the partnership deal with Nvidia, an Intel spokesperson has clarified that it will continue to develop integrated and discrete graphics chips, but leaves it open for interpretation as well as keeping Intel’s options open:

“While we’re not sharing specific roadmaps at this time, everything we discussed aligns with and complements Intel’s existing strategy. This collaboration with Nvidia enables us to deliver additional custom solutions that accelerate AI workloads and broaden our reach across high-performance computing segments in client and data center. We remain committed to our GPU roadmap. We’ll be collaborating with Nvidia to serve specific market segments, but we’re also continuing to execute on our own path.”

It really comes down to return on investment. In the press conference with Intel chief executive officer Lip-Bu Tan, Nvidia chief executive officer Jensen Huang said that that three architecture teams at Nvidia and Intel “have been discussing and architecting solutions now for probably coming up to a year.” Which means that Intel and Nvidia had been talking about this many months before the Intel board compelled former Intel chief Pat Gelsinger to resign in December 2024 and Intel datacenter general manager Justin Hotard left for Nokia in February 2025. This deal has been in the works for a while, and Tan got it finalized and inked.

With all of the woes it has in trying to build a foundry business that can compete against Taiwan Semiconductor Manufacturing Co – and frankly, to stay ahead of an increasingly aggressive Semiconductor Manufacturing International Corp backed by the Chinese government – Intel could really have used a lift with Nvidia adding itself to the Intel Foundry customer list. Such a commitment might be a bit premature, given where Intel is at with its 18A and 14A RibbonFET processes, but it could happen with a nudge from the Trump Administration if it believes Intel Foundry needs customers to ramp those and other future processes correctly.

Given that this deal allows Nvidia to sell its RTX GPU chiplets to Intel so the latter can chase the high end of the integrated GPU market for PCs and better compete against AMD and it also allows Intel to be the preferred X86 server chip supplier for Nvidia rackscale systems, this would seem like a wash. The fact that Intel is selling $5 billion of its stock to Nvidia as part of the deal perhaps tells you Intel is giving more than Nvidia is in the exchange. Perhaps the stock purchase is just a vote of confidence that accrues over time. Still, AI servers might make up half of the system sales worldwide now, the server CPUs in the host machines in AI clusters only make up a relatively small part of the overall system budget – it’s about 10 percent or so, compared to 75 percent for the GPU accelerators.

Up until now if you wanted NVLink interconnects between the CPUs and the GPUs in, your only choice was an ancient IBM Power9 chip using an old NVLink generation or a Grace CPU from Nvidia. In the future, thanks to NVLink Fusion chiplets, you will be able to choose an Intel Xeon CPU if you want that running on your hosts for whatever reason. (Access to larger and faster blocks of DDR main memory to use as cache for the GPU compute engines is the likely reason, since Grace has only a half terabyte of LPDDR5 memory.)

The addition of NVLink Fusion chiplets to Xeon CPU packages will also mean that those hyperscalers and cloud builders that are making their own XPUs for AI workloads and that are also licensing NVLink Fusion ports for them will be able to buy an Intel Xeon variant with NVLink ports to act as a host. These companies can decide to make their own host CPUs or not from that point forward. Huang was very clear that Nvidia would be continuing to develop its own Arm processors for servers, robots, and autonomous vehicles.

We think that it is highly unlikely that AMD will ever add NVLink ports to its CPUs and GPUs, and in this sense, the Nvidia deal gives Intel an advantage. So does the Foveros 3D stacking technique that was first deployed in the datacenter with the Ponte Vecchio GPUs and which has been used in subsequent Intel CPUs. With Foveros, Intel knows how to mix and match chiplets made by itself and TSMC, and Nvidia uses TSMC to make the vast majority of its GPUs. (Samsung has made a few.)

The Nvidia deal is also important in that it gives Intel an easy way to just stop talking about AI accelerators and mothball Jaguar Shores to focus on other issues if Tan and the Intel board decide that is the best course of action. And while we are pretty sure that none of this was talked about as part of the negotiations between Intel and Nvidia – that would smell too much like collusion instead of co-opetition to some – Huang and Tan are highly intelligent people and can connect dots.

The fun bit is that $5 billion is literally pocket change to Nvidia, and if Nvidia bought the Intel shares just before it made the announcement, then it is already up $872 million on its Intel investment as we go to press. If Intel stabilizes and sells CPUs and also sells Intel NVLink Fusion chiplets and RTX GPU chiplets, then it is not hard to believe that Nvidia will get all of its bait back in a year or two while also giving a certain amount of heartburn to AMD.

That brings us to the Nvidia and OpenAI partnership, which is not pocket change at $100 billion.

Huang is not just handing OpenAI co-founders Sam Altman, its chief executive officer, and Greg Brockman, its president and chairman, a giant green bag of money or a giant check like Publisher’s Clearing House (which went bankrupt this year) used to do.

What Nvidia is doing is giving OpenAI a promise to make available up to $100 billion available to it (presumably in exchange for a stake in OpenAI, but that was not specified) to cover “at least” 10 gigawatts of AI computing capacity that the company will make available to OpenAI as its datacenters are built and have the power and cooling in place so they can receive the AI clusters based on Nvidia technology. The money will be allocated incrementally as the datacenter capacity is made available, rather than all at once. It is not clear how OpenAI will pay for the building of its datacenter facilities or the power for them, but the rumor is that Oracle will be building and installing the gear for OpenAI, as evidenced by the nearly $500 billion in revenue backlog announced two weeks ago by the software giant and now cloud player. Oracle does not build datacenters, but builds gear to put into datacenters and then manages it. It has several of its own datacenters, of course, but there are many more that it operates on behalf of customers.

Safra Catz, formerly chief executive officer at Oracle but as of this week now executive vice chairman, explained this on a call with Wall Street going over Oracle’s Q1 F2026 numbers:

“We do not own the property. We do not own the buildings. What we do own and what we engineer is the equipment. And that is equipment that is optimized for the Oracle Cloud. It has extremely special networking capabilities. It has technical capabilities from Larry and his team that allows us to run these workloads much, much faster. And as a result, it’s much cheaper than our competitors.”

“Now, because of that, what we do is we put in that equipment only when it’s time and usually very quickly, assuming that our customer accepts it, we’re already generating revenue right away. The faster they accept the system and that it meets their needs, the faster they start using it, the sooner we have revenue. This is, in some ways – I don’t want to call it asset-light from the finance world, but it’s asset pretty light. And that is really an advantage for us. I know some of our competitors, they like to own buildings. That’s not really our specialty. Our specialty is the unique technology, the unique networking, the storage, just the whole way we put these systems together.”

So when people say Oracle is building datacenters for OpenAI, this is not precisely true. Oracle is building datacenter infrastructure for OpenAI and is collaborating on datacenter design – it is not at all clear who owns the datacenters, aside from Crusoe and CoreWeave for some of the OpenAI deals we know about.

Anyway, back in 2016, OpenAI was the first company to get an Nvidia DGX server, and it was autographed by Huang (and is presumably either still running or shellacked in a closet somewhere to sell on eBay), and it looks like OpenAI will be first in line to get next year’s NVL144 systems from Big Green, based on the 88-core “Vera” Arm server CPU and the “Rubin” GPU complex, which will put two GPU chiplets in a single die like the “Blackwell” B100, B200, and B300 did.

The Nvidia announcement, which you can see here, says the first gigawatt of Vera-Rubin capacity will be generating its first tokens in the second half of 2026, ten years later. It is hard to say how many racks of machines 10 gigawatts represents because each generation of Nvidia rackscale systems is getting faster, but it is also getting hotter than we are used to during prior computer eras. We’re running out of lower precision to jump the gap. Assuming 140 kilowatts per rack for GB300 systems for doing training or inference, then for 10 gigawatts you are talking about maybe 70,000 racks, plus or minus 4,000 racks, to bring somewhere between 9.5 million and 10.7 million Rubin GPU chiplets to bear. Obviously, the first tranche of GPUs will be one-tenth of that. So 950,000 to 1.07 million Rubin GPUlets in somewhere between 6,600 to 7,400 racks. The current Grace-Blackwell GB300 NVL72 (it really should be NVL144, as Huang pointed out earlier this year) has 1.1 exaflops of compute at FP4 precision, and the Vera-Rubin is expected to deliver about 3.6 exaflops per rack at FP4. So somewhere around 7,000 racks will deliver around 25,200 exaflops in 1 gigawatt. Ten times that is enormous.

By the way, the five boroughs of New York City burn an average of around 10 gigawatts during the summer, with peaks going above that on very hot days. The peak power draw for NYC was 13.3 gigawatts on July 19, 2013 and energy conservation efforts have helped keep it lower since then.

Now, let’s think out loud about the Stargate project from OpenAI and friends for a second.

Stargate is supposed to be a $500 billion effort. If the datacenter and facilities is a little more than half of that, then the $300 billion that Oracle just added to its cloud revenue backlog a few weeks ago, and that is rumored to be part of a massive OpenAI deal, should cover the cost of AI systems (presumably based on Nvidia CPUs, GPUs, and networking) and running the datacenters. That would leave $200 billion to build the datacenters and their power and cooling systems and actually pay for the electricity they will consume. The only thing is, most of what Oracle is going to be paid for is to build special AI clusters for OpenAI – as far as we know, Oracle is not building datacenters. Call it $200 billion for systems and $100 billion for operations. With only a little markup on the Nvidia iron, there is still close to $100 billion in other iron that is part of Stargate. Perhaps this remaining money is for homegrown XPUs for AI inference, like the rumored “Titan” chip that Broadcom is said to be helping OpenAI create. Perhaps the MGX private equity firm based in the United Arab Emirates is footing a lot of the bill for most of this? And if so, the question becomes: In exchange for what? Well, having a Stargate UAE sovereign cloud running OpenAI models, which Group 42 Holding, another UAE sovereign wealth fund, is helping to pay for.

That is mostly speculation, we realize. But in the absence of data, that is what we are left to do. The point is, Nvidia is in the middle of all of this and it is using its accumulated wealth to create future accumulated wealth. This round-tripping of cash out to customers who then turn around and use it to buy capacity is not just something that Microsoft Azure, Amazon Web Services, and Google Cloud can do – and have done to prop up model builders OpenAI and Anthropic. The question is what else might Nvidia be getting aside from future sales and keeping the AI story going and growing by giving so much money to OpenAI?

The answer might be, surprisingly, nothing else. This might be the best investment Nvidia can make as it tries to rake in maybe $1.6 trillion in the next five fiscal years and brings maybe $750 billion to the middle line as operating income from its datacenter business. That’s what our model suggests is possible.

There is only one way to find out.

UPDATE: Subsequent to this story running, OpenAI announced five new datacenters for the Stargate Project in conjunction with Oracle and SoftBank. With the deals announced, OpenAI has $400 billion in investment lined up, covering 7 gigawatts of total capacity over the next five years, well on its way to 10 gigawatts of capacity and $500 million in spending. The Oracle deal is for more than $300 billion to cover 4.5 gigawatts of datacenter capacity, the company confirmed, across three sites – Shackelford County, Texas; Doña Ana County, New Mexico; and an unnamed site in the Midwest. There are two smaller Stargate sites that will together expand to 1.5 gigawatts, located in Lordstown, Ohio and Milam County, Texas. Oracle designed and is fulling the flagship Stargate I site in Abilene, Texas, which weighs in at 600 megawatts and which got its first GB200 NVL72 racks in June.