To integrate AI into WhatsApp, which has billions of users, Meta has spared no expense. In addition to purchasing millions of NVIDIA GPUs, it will also use "confidential computing" technology to solve the privacy challenges between AI assistants and end-to-end encryption.
Meta earlierAnnounceA long-term partnership agreement with NVIDIA plans to purchase "millions" of NVIDIA's latest Blackwell and Rubin display architecture GPUs. The most intriguing aspect of this deal isn't just the scale of the hardware purchase, but also how Meta intends to utilize this computing power—particularly on its messaging app WhatsApp.
According to reports, Meta will officially deploy NVIDIA's [technology/technology] in WhatsApp."Confidential Computing" technologyThis will give WhatsApp powerful AI capabilities (such as AI customer service, assistants, or generative responses) while ensuring that the confidentiality and integrity of user data are not compromised.
WhatsApp's AI privacy solution: Encryption even during computation
WhatsApp has always touted its end-to-end encryption (E2EE) as a selling point, which makes importing AI features to the server quite tricky—because AI usually needs to "see" your messages in order to process them.
NVIDIA's confidential computing technology provides solutions to protect data security "during computation," not just during transmission or storage. This means that when WhatsApp's AI processes messages, the data runs in an isolated hardware environment (Trusted Execution Environment), and even Meta or cloud providers cannot access its contents.
In its official blog, NVIDIA stated that this technology not only protects user privacy but also protects the intellectual property rights of software developers (such as Meta or third-party AI agent vendors), preventing the leakage of their model parameters or logic.
The first to use a "discrete" Grace CPU for inference
In addition to GPUs, Meta has also made bold attempts in the architecture of its infrastructure.
Meta will become the world's first technology company to deploy NVIDIA Grace CPUs in a "standalone way".
Typically, the Grace CPU is tied to Hopper and Blackwell GPUs (such as GH200 and GB200), but Meta chose to use it independently, specifically for handling inference and agentic workloads.
This also highlights Meta's strategy of offloading computing power for different types of AI tasks: GPUs are responsible for training and heavy computation, while CPUs handle logical inference and agent execution. Furthermore, Meta will also utilize NVIDIA's Spectrum-X Ethernet switches to meet the high-speed transmission needs of massive AI clusters.
AI spending is projected to reach $1350 billion by 2026.
This collaboration comes against the backdrop of Meta's ambitious capital expenditure plans. Earlier this year, Meta announced that it expects to invest up to $1350 billion in AI development by 2026. Analysts estimate that "tens of billions of dollars" of this funding will go into NVIDIA's coffers.
To support this computing power, Meta plans to build up to 30 new data centers by 2028, 26 of which will be located in the continental United States. This is a long-term infrastructure commitment with a total investment of up to $6000 billion.
Analysis of viewpoints
Meta's move is actually addressing the biggest pain point in "AI implementation": trust.
For products like WhatsApp, which are centered on private communication, adding features similar to ChatGPT would inevitably break the promise that "only you and the recipient can see the message."
Through NVIDIA's confidential computing, Meta is attempting to find a balance between the powerful capabilities of cloud AI and the privacy and security of local devices. If this model succeeds, it will set a benchmark for other privacy-focused services, such as medical and financial AI.
On the other hand, Meta's purchase of a large number of independent Grace CPUs suggests that future AI applications will no longer be just about "generating text or images," but will involve more applications of "AI Agents"—which need to handle complex logical judgments, process control, and long-term memory. Running these tasks on specially designed CPUs may be more cost-effective than using GPUs in terms of cost and energy efficiency.
