Meta is making a major long-term bet on Nvidia hardware, and the most important part may not be the GPUs. The Facebook parent has signed a multiyear agreement to buy millions of Nvidia chips, including current Blackwell GPUs, upcoming Rubin GPUs, and standalone Grace and Vera CPUs.
Neither company disclosed a price. Ben Bajarin, CEO and principal analyst at tech consultancy Creative Strategies, estimated the agreement would be worth billions of dollars. According to The Register, it is likely to add tens of billions to Nvidia's bottom line.
A bigger Nvidia role inside Meta's AI buildout
The agreement lands against the backdrop of Meta's expanding AI infrastructure plans. Meta CEO Mark Zuckerberg had previously announced plans to nearly double the company's AI infrastructure spending in 2026 to as much as $135 billion.
That makes the chip mix especially important. Meta is not only buying Nvidia's most visible AI accelerators; it is also adopting Nvidia processors as standalone products at large scale. For Nvidia, that expands the relationship beyond the GPU-heavy systems that have defined much of the recent AI boom.
The deal includes several layers of Nvidia hardware:
- current Blackwell GPUs;
- upcoming Rubin GPUs;
- standalone Grace CPUs;
- standalone Vera CPUs.
The CPU portion is the strategic change. Until recently, Nvidia's Grace processors were usually sold as part of "Superchips" that combine a CPU and GPU on one module. Nvidia changed that sales strategy in January 2026 and began offering the CPUs separately. The first named customer at that time was neocloud provider CoreWeave.
Why Nvidia wants CPUs in the inference era
Nvidia's CPU push is tied to a shift in how AI systems are used. The source article describes an industry that had been dominated by GPU-heavy training of large models, while attention is now moving toward inference, the process of running trained models.
That distinction matters because inference does not always require the same hardware profile as model training. For many inference tasks, GPUs are overkill. Nvidia is therefore trying to position its CPUs for workloads that sit around and behind AI systems, not only the large-scale training jobs that made GPUs central to the market.
"We were in the 'training' era, and now we are moving more to the 'inference era,' which demands a completely different approach," Bajarin told the Financial Times.
Ian Buck, Nvidia's VP and General Manager of Hyperscale and HPC, said according to The Register that the Grace processor can "deliver 2x the performance per watt on those back end workloads" such as running databases. He also said that "Meta has already had a chance to get on Vera and run some of those workloads, and the results look very promising."
In plain terms, Nvidia is trying to make itself useful in more places inside AI infrastructure. GPUs remain central to Meta's purchase, but standalone CPUs give Nvidia a broader server footprint and put it into direct competition with Intel and AMD in the server market.
Grace, Vera, and WhatsApp AI features
The Grace CPU features 72 Arm Neoverse V2 cores and uses LPDDR5x memory, which offers advantages in bandwidth and space. Nvidia's next-generation Vera CPU brings 88 custom Arm cores, simultaneous multi-threading, and confidential computing capabilities.
According to Nvidia, Meta plans to use Vera for private processing and AI features in its WhatsApp encrypted messaging service. Vera deployment is planned for 2027.
That detail shows why the CPU part of the agreement matters. Nvidia is not only selling chips for headline AI model development. It is also pushing into the infrastructure needed to run services and features after models have already been trained.
The move also gives Meta another hardware path for AI workloads. Meta is working on its own AI chips, but according to the Financial Times, that in-house chip strategy had "suffered some technical challenges and rollout delays." Buying standalone CPUs from Nvidia gives Meta access to hardware that fits its near-term infrastructure needs while its own chip work continues.
Meta is taking a different path from other hyperscalers
Meta's decision stands out because other large cloud and platform companies have been leaning into their own processors. Amazon relies on Graviton processors, while Google uses Axion. Meta, by contrast, is buying standalone Nvidia CPUs even as it works on its own AI chips.
That does not make Meta an Nvidia-only customer. According to The Register, Meta operates a fleet of AMD Instinct GPUs and was directly involved in the design of AMD's Helios rack systems, which are due out later this year.
The broader competitive picture is also tightening for Nvidia. Google, Amazon, and Microsoft have all announced new in-house chips in recent months. OpenAI has co-developed a chip with Broadcom and struck a significant deal with AMD. Several startups such as Cerebras are offering specialized inference chips that could challenge Nvidia's dominance.
Nvidia has also moved to defend its position in inference. In December, the company acquired talent from inference chip company Groq in a licensing deal to shore up its position in the market.
The Meta agreement therefore cuts both ways. It reinforces Nvidia's role as a major supplier to one of the largest AI infrastructure buyers. At the same time, the emphasis on standalone CPUs shows that Nvidia is adapting to a market where customers are diversifying, inference is becoming more important, and rivals are pushing hard with internal and specialized chips.
The strategic meaning of the deal
The immediate takeaway is simple: Meta is buying millions of Nvidia chips under a multiyear agreement, and the purchase includes more than GPUs. The longer-term signal is that Nvidia wants its hardware to cover more of the AI infrastructure stack.
For Meta, the agreement supports a huge AI infrastructure expansion while preserving flexibility across Nvidia, AMD, and its own chip efforts. For Nvidia, it opens a larger role in server workloads and inference at a time when competition from Intel, AMD, hyperscaler chips, and inference-focused startups is becoming harder to ignore.