A deep-dive analysis into the market dynamics that allowed Nvidia to take the AI crown and surpass Intel in annual revenue. CRN also looks at what the x86 processor giant could do to fight back in a deeply competitive environment.

Several months after Pat Gelsinger became Intels CEO in 2021, he told me that his biggest concern in the data center wasnt Arm, the British chip designer that is enabling a new wave of competition against the semiconductor giants Xeon server CPUs.

Instead, the Intel veteran saw a bigger threat in Nvidia and its uncontested hold over the AI computing space and said his company would give its all to challenge the GPU designer.

[Related: The ChatGPT-Fueled AI Gold Rush: How Solution Providers Are Cashing In]

Well, theyre going to get contested going forward, because were bringing leadership products into that segment, Gelsinger told me for a CRN magazine cover story.

More than three years later, Nvidias latest earnings demonstrated just how right it was for Gelsinger to feel concerned about the AI chip giants dominance and how much work it will take for Intel to challenge a company that has been at the center of the generative AI hype machine.

When Nvidias fourth-quarter earnings arrived last week, they showed that the company surpassed Intel in total annual revenue for its recently completed fiscal year, mainly thanks to high demand for its data center GPUs driven by generative AI.

The GPU designer finished its 2024 fiscal year with $60.9 billion in revenue, up 126 percent or more than double from the previous year, the company revealed in its fourth-quarter earnings report on Wednesday. This fiscal year ran from Jan. 30, 2023, to Jan. 28, 2024.

Meanwhile, Intel finished its 2023 fiscal year with $54.2 billion in sales, down 14 percent from the previous year. This fiscal year ran concurrent to the calendar year, from January to December.

While Nvidias fiscal year finished roughly one month after Intels, this is the closest well get to understanding how two industry titans compared in a year when demand for AI solutions propped up the data center and cloud markets in a shaky economy.

Nvidia pulled off this feat because the company had spent years building a comprehensive and integrated stack of chips, systems, software and services for accelerated computingwith a major emphasis on data centers, cloud computing and edge computingthen found itself last year at the center of a massive demand cycle due to hype around generative AI.

This demand cycle was mainly kicked off by the late 2022 arrival of OpenAIs ChatGPT, a chatbot powered by a large language model that can understand complex prompts and respond with an array of detailed answers, all offered with the caveat that it could potentially impart inaccurate, biased or made-up answers.

Despite any shortcomings, the tech industry found more promise than concern with the capabilities of ChatGPT and other generative AI applications that had emerged in 2022, like the DALL-E 2 and Stable Diffusion text-to-image models. Many of these models and applications had been trained and developed using Nvidia GPUs because the chips are far faster at computing such large amounts of data than CPUs ever could.

The enormous potential of these generative AI applications kicked off a massive wave of new investments in AI capabilities by companies of all sizes, from venture-backed startups to cloud service providers and consumer tech companies, like Amazon Web Services and Meta.

By that point, Nvidia had started shipping the H100, a powerful data center GPU that came with a new feature called the Transformer Engine. This was designed to speed up the training of so-called transformer models by as many as six times compared to the previous-generation A100, which itself had been a game-changer in 2020 for accelerating AI training and inference.

Among the transformer models that benefitted from the H100s Transformer Engine was GPT-3.5, short for Generative Pre-trained Transformer 3.5. This is OpenAIs large language model that exclusively powered ChatGPT before the introduction of the more capable GPT-4.

But this was only one piece of the puzzle that allowed Nvidia to flourish in the past year. While the company worked on introducing increasingly powerful GPUs, it was also developing internal capabilities and making acquisitions to provide a full stack of hardware and software for accelerated computing workloads such as AI and high-performance computing.

At the heart of Nvidias advantage is the CUDA parallel computing platform and programming model. Introduced in 2007, CUDA enabled the companys GPUs, which had been traditionally designed for computer games and 3-D applications, to run HPC workloads faster than CPUs by breaking them down into smaller tasks and processing those tasks simultaneously. Since then, CUDA has dominated the landscape of software that benefits accelerated computing.

Over the last several years, Nvidias stack has grown to include CPUs, SmartNICs and data processing units, high-speed networking components, pre-integrated servers and server clusters as well as a variety of software and services, which includes everything from software development kits and open-source libraries to orchestration platforms and pretrained models.

While Nvidia had spent years cultivating relationships with server vendors and cloud service providers, this activity reached new heights last year, resulting in expanded partnerships with the likes of AWS, Microsoft Azure, Google Cloud, Dell Technologies, Hewlett Packard Enterprise and Lenovo. The company also started cutting more deals in the enterprise software space with major players like VMware and ServiceNow.

All this work allowed Nvidia to grow its data center business by 217 percent to $47.5 billion in its 2024 fiscal year, which represented 78 percent of total revenue.

This was mainly supported by a 244 percent increase in data center compute sales, with high GPU demand driven mainly by the development of generative AI and large language models. Data center networking, on the other hand, grew 133 percent for the year.

Cloud service providers and consumer internet companies contributed a substantial portion of Nvidias data center revenue, with the former group representing roughly half and then more than a half in the third and fourth quarters, respectively. Nvidia also cited strong demand driven by businesses outside of the former two groups, though not as consistently.

In its earnings call last week, Nvidia CEO Jensen Huang said this represents the industrys continuing transition from general-purpose computing, where CPUs were the primary engines, to accelerated computing, where GPUs and other kinds of powerful chips are needed to provide the right combination of performance and efficiency for demanding applications.

There's just no reason to update with more CPUs when you can't fundamentally and dramatically enhance its throughput like you used to. And so you have to accelerate everything. This is what Nvidia has been pioneering for some time, he said.

Intel, by contrast, generated $15.5 billion in data center revenue for its 2023 fiscal year, which was a 20 percent decline from the previous year and made up only 28.5 percent of total sales.

This was not only three times smaller than what Nvidia earned for total data center revenue in the 12-month period ending in late January, it was also smaller than what the semiconductor giants AI chip rival made in the fourth quarter alone: $18.4 billion.

The issue for Intel is that while the company has launched data center GPUs and AI processors over the last couple years, its far behind when it comes to the level of adoption by developers, OEMs, cloud service providers, partners and customers that has allowed Nvidia to flourish.

As a result, the semiconductor giant has had to rely on its traditional data center products, mainly Xeon server CPUs, to generate a majority of revenue for this business unit.

This created multiple problems for the company.

While AI servers, including ones made by Nvidia and its OEM partners, rely on CPUs for the host processors, the average selling prices for such components are far lower than Nvidias most powerful GPUs. And these kinds of servers often contain four or eight GPUs and only two CPUs, another way GPUs enable far greater revenue growth than CPUs.

In Intels latest earnings call, Vivek Arya, a senior analyst at Bank of America, noted how these issues were digging into the companys data center CPU revenue, saying that its GPU competitors seem to be capturing nearly all of the incremental [capital expenditures] and, in some cases, even more for cloud service providers.

One dynamic at play was that some cloud service providers used their budgets last year to replace expensive Nvidia GPUs in existing systems rather than buying entirely new systems, which dragged down Intel CPU sales, Patrick Moorhead, president and principal analyst at Moor Insights & Strategy, recently told CRN.

Then there was the issue of long lead times for Nvidias GPUs, which were caused by demand far exceeding supply. Because this prevented OEMs from shipping more GPU-accelerated servers, Intel sold fewer CPUs as a result, according to Moorhead.

Intels CPU business also took a hit due to competition from AMD, which grew x86 server CPU share by 5.4 points against the company in the fourth quarter of 2023 compared to the same period a year ago, according to Mercury Research.

The semiconductor giant has also had to contend with competition from companies developing Arm-based CPUs, such as Ampere Computing and Amazon Web Services.

All of these issues, along with a lull in the broader market, dragged down revenue and earnings potential for Intels data center business.

Describing the market dynamics in 2023, Intel said in its annual 10-K filing with the U.S. Securities and Exchange Commission that server volume decreased 37 percent from the previous year due to lower demand in a softening CPU data center market.

The company said average selling prices did increase by 20 percent, mainly due to a lower mix of revenue from hyperscale customers and a higher mix of high core count processors, but that wasnt enough to offset the plummet in sales volume.

While Intel and other rivals started down the path of building products to compete against Nvidias years ago, the AI chip giants success last year showed them how lucrative it can be to build a business with super powerful and expensive processors at the center.

Intel hopes to make a substantial business out of accelerator chips between the Gaudi deep learning processors, which came from its 2019 acquisition of Habana Labs, and the data center GPUs it has developed internally. (After the release of Gaudi 3 later this year, Intel plans to converge its Max GPU and Gaudi road maps, starting with Falcon Shores in 2025.)

But the semiconductor giant has only reported a sales pipeline that grew in the double digits to more than $2 billion in last years fourth quarter. This pipeline includes Gaudi 2 and Gaudi 3 chips as well as Intels Max and Flex data center GPUs, but it doesnt amount to a forecast for how much money the company expects to make this year, an Intel spokesperson told CRN.

Even if Intel made $2 billion or even $4 billion from accelerator chips in 2024, it would amount to a small fraction of what Nvidia made last year and perhaps an even smaller one if the AI chip rival manages to grow again in the new fiscal year. Nvidia has forecasted that revenue in the first quarter could grow roughly 8.6 percent sequentially to $24 billion, and Huang said the conditions are excellent for continued growth for the rest of this year and beyond.

Then theres the fact that AMD recently launched its most capable data center GPU yet, the Instinct MI300X. The company said in its most recent earnings call that strong customer pull and expanded engagements prompted the company to upgrade its forecast for data center GPU revenue this year to more than $3.5 billion.

There are other companies developing AI chips too, including AWS, Microsoft Azure and Google Cloud as well as several startups, such as Cerebras Systems, Tenstorrent, Groq and D-Matrix. Even OpenAI is reportedly considering designing its own AI chips.

Intel will also have to contend with Nvidias decision last year to move to a one-year release cadence for new data center GPUs. This started with the successor to the H100 announced last fallthe H200and will continue with the B100 this year.

Nvidia is making its own data center CPUs, too, as part of the companys expanding full-stack computing strategy, which is creating another challenge for Intels CPU business when it comes to AI and HPC workloads. This started last year with the standalone Grace Superchip and a hybrid CPU-GPU package called the Grace Hopper Superchip.

For Intels part, the semiconductor giant expects meaningful revenue acceleration for its nascent AI chip business this year. What could help the company are the growing number of price-performance advantages found by third parties like AWS and Databricks as well as its vow to offer an open alternative to the proprietary nature of Nvidias platform.

The chipmaker also expects its upcoming Gaudi 3 chip to deliver performance leadership with four times the processing power and double the networking bandwidth over its predecessor.

But the company is taking a broader view of the AI computing market and hopes to come out on top with its AI everywhere strategy. This includes a push to grow data center CPU revenue by convincing developers and businesses to take advantage of the latest features in its Xeon server CPUs to run AI inference workloads, which the company believes is more economical and pragmatic for a broader constituency of organizations.

Intel is making a big bet on the emerging category of AI PCs, too, with its recently launched Core Ultra processors, which, for the first time in an Intel processor, comes with a neural processing unit (NPU) in addition to a CPU and GPU to power a broad array of AI workloads. But the company faces tough competition in this arena, whether its AMD and Qualcomm in the Windows PC segment or Apple for Mac computers and its in-house chip designs.

Even Nvidia is reportedly thinking about developing CPUs for PCs. But Intel does have one trump card that could allow it to generate significant amounts of revenue alongside its traditional chip design business by seizing on the collective growth of its industry.

Hours before Nvidias earnings last Wednesday, Intel launched its revitalized contract chip manufacturing business with the goal of drumming up enough business from chip designers, including its own product groups, to become the worlds second largest foundry by 2030.

Called Intel Foundry, its lofty 2030 goal means the business hopes to generate more revenue than South Koreas Samsung in only six years. This would put it only behind the worlds largest foundry, Taiwans TSMC, which generated just shy of $70 billion last year with many thanks to large manufacturing orders from the likes of Nvidia, Apple and Nvidia.

All of this relies on Intel to execute at high levels across its chip design and manufacturing businesses over the next several years. But if it succeeds, these efforts could one day make the semiconductor giant an AI superpower like Nvidia is today.

At Intel Foundrys launch last week, Gelsinger made that clear.

We're engaging in 100 percent of the AI [total addressable market], clearly through our products on the edge, in the PC and clients and then the data centers. But through our foundry, I want to manufacture every AI chip in the industry, he said.

More:

Analysis: How Nvidia Surpassed Intel In Annual Revenue And Won The AI Crown - CRN