CONFIRMED: Meta’s engineers described the system in a blog post on 12 October 2026. The benefits are Meta’s own COMPANY CLAIM. Meta published no figures for them.
Meta says it has built a system that uses spare computing power in its AI fleet to pick better ads. In an engineering blog post, its team writes: “At Meta, we’ve built a system that enhances ads recommendations by leveraging idle inference capacity in our fleet” [1].
Inference is the work of running a trained AI model to get answers, here to score which ads suit a person. It is separate from training the model.
The problem: short of capacity, yet some sits idle
Meta calls this a “capacity paradox”. Inference fleets “can be globally constrained while meaningful pockets of capacity remain locally idle” [1]. It gives four causes:
- the fleet is built for peak demand, but traffic changes through the day;
- new hardware and new models arrive on different schedules;
- the mix of accelerators and CPUs in a region doesn’t always match what workloads need;
- some workloads can’t run on every type of accelerator.
Meta says it gathers these pockets “into a shared opportunistic capacity pool”. This capacity is “preemptible (can be reclaimed at any time), highly volatile in its availability, and carries no long-term guarantees” [1].
Two ways to use the spare capacity
Rank More Ads (RMA) “adjusts the number of candidates evaluated by the ranking stages” [1]. Meta’s system has to pick a few ads “from millions of candidates in less than a second”. When spare capacity exists, RMA lets more candidates through, which “increases the chance of finding ads that are relevant to the user”. RMA needs spare CPUs as well as accelerators [1].
Serve Better Models (SBM) “adjusts the share of requests routed to a more powerful model” [1]. It keeps the number of candidate ads the same, so it “can use accelerator capacity without substantially increasing upstream CPU work”. A router uses “request-value signals to prioritize the requests expected to benefit most” [1].
Because the two levers use different mixes of hardware, Meta says that “together, they can use a broader range of opportunistic capacity than either lever alone” [1].
Keeping normal ads serving safe
Meta’s core rule is that losing spare capacity “may reduce incremental ranking work, but it must not overload the capacity that serves the baseline workload” [1].
- SBM is isolated physically. The normal and more powerful models run on separate replicas. When spare capacity shrinks, more requests fall back to the normal model, “whose capacity is already provisioned for them” [1].
- RMA is isolated logically. The system sets “a global budget for RMA” and limits “the total incremental work to the amount that these replicas can support” [1].
RMA’s volume is set by a feedback controller that watches tail latency, meaning the slowest responses. When latency falls below the target, it ranks more ads; when latency rises, it ranks fewer [1].
When capacity vanishes suddenly
Meta says these central controls “operate on the order of minutes” [1], which can be too slow when capacity is taken back without warning. So servers and clients also drop lower-value optional work on their own. This is called load shedding. Meta says: “Both forms of load shedding react within a few milliseconds” [1].
What Meta says it gains (COMPANY CLAIM)
Meta says the system “drives significant advertiser value and capacity efficiency”. It says the system helps “people see more relevant ads and improves advertiser outcomes without requiring additional capacity orders” [1]. The post gives no percentages, revenue figures or latency numbers, and TSN has not independently checked these claims.
What’s next
Meta says it is working on “precomputation elasticity”, which does work before a request arrives. It is also widening “model portability and cross-region inference capabilities”, and building “an optimization-based system to automatically adjust the allocation of AI inference capacity” [1].
Sources
- Engineering at Meta (Jinfu Leng, Rohith Menon, Nicolas Stier and others), “Turning Opportunistic AI Inference Capacity Into Better Ads Recommendations”, 12 October 2026. https://engineering.fb.com/2026/10/12/ml-applications/turning-opportunistic-ai-inference-capacity-into-better-ads-recommendations/
