Meta explains “compute power” as it pushes custom chips and more AI infrastructure

Meta used its newsroom to unpack a concept that underlies almost every AI interaction: compute power. In an example built around asking Meta AI for nearby vegan restaurant options, the company describes a response that arrives in seconds, but depends on layers of processing that start with converting voice to text and end with running a large language model across data-center servers.

At the center of Meta’s explanation is a basic idea, compute power, which the company describes as how much work a chip can do and how fast it can do it. Meta says the most common way to quantify that speed is FLOPS, or floating-point operations per second, which measures how many calculations a chip can perform in a second. It also ties scale to gigawatts, the power level required to keep large numbers of chips running simultaneously.

Meta’s workflow example runs through multiple steps that are typically invisible to users. When a person asks a question using voice in the Meta AI app, Meta says the audio is captured, converted from sound waves into text, routed to computers or servers inside data centers, and then processed by a large language model (LLM), the software system designed to generate language and other outputs. The result is then delivered back to the user.

The company also argues that even everyday product features require compute beyond the obvious. Searching on Instagram, for example, involves understanding language, processing the query, scanning indexes, generating results, and delivering them before the user’s attention moves on. Meta’s message is that this “stack” of computations depends on processing chips located in data centers, not on the phone or laptop doing the talking.

Meta frames its infrastructure strategy as a response to how heterogeneous compute needs can be. Different chips, it says, are designed for different calculation types and workloads. Rather than relying on a single approach, Meta says it is building a global network of AI-optimized data centers with flexibility, intended to support both AI workloads and the other compute-heavy workloads tied to Meta’s consumer apps.

To support that approach, Meta highlights custom hardware as an essential component. It points to MTIA custom silicon and says it is developing and deploying four new generations of MTIA chips over the next two years, aimed at workloads including ranking, recommendations, and generative AI. The company also references an expanded partnership with Broadcom announced in April to co-develop multiple MTIA generations.

Meta adds that its CPU strategy includes an Arm collaboration. It says it previously announced a partnership with Arm to co-develop the Arm AGI CPU, described as the first data center processor specifically designed to handle the massive amount of data movement demanded by AI workloads. (Data movement refers to the traffic of data between memory and processing units, a bottleneck that can limit AI performance even when compute is available.)

In addition to its own chip development, Meta says it will source silicon from multiple partners, naming AWS, AMD, and NVIDIA as contributors to its broader compute portfolio. Meta links this mix to a goal of matching the right chips to the right tasks, which it says helps it move faster in building and delivering new AI experiences. The company ties these infrastructure layers to its AI model work as well, pointing to Muse Spark, which Meta says is its most advanced model to date and a “natively multimodal” large language model that processes voice, text, and images together.

Meta does not provide new financial commitments, chip shipment schedules, or capacity figures in the explainer. It also does not quantify how much of its AI compute will come from MTIA versus partner hardware, nor does it disclose the power or performance targets behind the FLOPS and gigawatt framing. The piece is therefore best read as a conceptual and strategic explanation rather than as a detailed update on build-out timelines or budgets.

Still, the message is clear enough for industry watchers: compute is not only a raw input to AI, it is also a supply-chain and systems-design problem. As Meta pushes new MTIA generations and broadens partnerships across custom silicon and off-the-shelf chips, the next indicates to watch are how quickly those hardware plans translate into model and product throughput, and whether Meta’s “diversified approach” leads to measurable improvements in performance per watt or overall system efficiency in its AI features.