Where Live-Stream Recommenders Should Spend Carbon

Live video shopping, where a host sells products on camera while viewers chat and buy inside the stream, is increasingly driven by recommenders that re-read each stream every 30 seconds with a large multimodal model. That design helps new and little-known streamers get discovered. It also assumes hyperscale hardware. A small regional platform or a merchant cooperative cannot afford that hardware, and nobody has measured its lifecycle carbon: the emissions from manufacturing the servers plus the electricity to run them. This proposal asks a narrow, answerable question. For a resource-constrained live-commerce operator, which part of that expensive content understanding actually earns its cost, and how should the operator split decisions between buying hardware and running it moment to moment?

Our claim has two parts. First, per-segment content embeddings pay off almost entirely on "cold" rooms: new streamers, first sessions, and moments when a stream changes topic. For established streamers, ordinary ID-based collaborative filtering already carries the signal. Second, carbon accounting should work on two timescales, following the rule from carbon-aware scheduling. Manufacturing (embodied) carbon decides which hardware and which model tier to provision for the demand envelope, including viral surges. Only marginal electricity carbon should decide, at runtime, which rooms get their content embedding refreshed. Together these should let a small operator keep most of the long-tail discovery benefit on older, lower-embodied-carbon accelerators, at a fraction of the compute.

Core Claim and Positioning

The closest systems work motivates the gap without closing it. LLM-Alignment Live-Streaming Recommendation (LARM) aligns 30-second multimodal segment embeddings from a 7B student model with collaborative-filtering embeddings. On Kuaishou's proprietary trace it lifts Hit Rate@1000 from 13.86% to 17.03% and boosts long-tail exposure. Its learned gate weights ID features at roughly 0.65 and LLM content at roughly 0.35. That split is our first hint that content is a minority signal, and possibly concentrated in specific rooms. LARM refreshes every stream uniformly and reports no compute or carbon cost.

HarmonRank solves a different problem: aligning multi-objective ensembles with AUC. We treat it as the ranking stage that stays fixed while we vary the content tier. KuaiLive supplies the regime. In it, 70.7% of streamers received fewer than five interactions over 21 days, while 27.1% of users repeatedly return to the same streamers. That combination is exactly the one in which content should matter for a sparse tail and matter little for a habitual head.

The carbon literature supplies both the method and a constraint. The Sunk Carbon Fallacy shows that folding amortized manufacturing carbon into real-time scheduling can raise total emissions, especially at 30–60% utilization. It prescribes marginal operational carbon for runtime decisions and embodied carbon for procurement. We adopt that split as a design rule and test it in our setting. Unveiling Environmental Impacts of Large Language Model Serving: A Functional Unit View (FUEL) measures serving carbon per unit of useful output. It finds that for modest request rates an older L40 beats a new H100 on lifecycle emissions, and that W8A8 quantization cuts emissions 5–15%. Small live-commerce operators sit squarely in that low-throughput regime.

SCARIF provides whole-server manufacturing carbon, including mainboards, memory, power supplies and chassis, rather than chip-only estimates. Unveiling the Uncertainty in Embodied and Operational Carbon of Large AI Models through a Probabilistic Carbon Accounting Model (PCAM) replaces point estimates with distributions over yield, die size and regional grid. It reports that deterministic tools err by up to 108.51% on operational carbon. We report every carbon number as a PCAM-style distribution for that reason.

EcoThink supplies the routing idea. It sends about 65% of queries to a cheap path and saves 40.4% energy on average at 90.7% accuracy. Its routing is per query on QA benchmarks. Ours is per live room over time, and the signal we route on is expected ranking change, not query difficulty.

Towards Carbon Footprint-Aware Recommender Systems for Greener Item Recommendation reranks items by the lifecycle footprint of the products themselves. It treats compute as free. We do the reverse and leave product greenness out of scope. The two are complementary, not competing.

The adoption and strategy literature plays a narrower role. The SWOT-plus-technology-acceptance study in Strategy of use Artificial Intelligence for optimization and stability cake business seasonal “Case study on small-medium scale model BR company” prescribes an "aggressive" cloud-AI posture for a micro-business with 4.3 million TikTok followers and holiday surges up to 150 times normal volume. It bases this on subjective factor scores (internal 0.78, external 1.27) and never deploys anything. Digitalization, financing constraints and firm performance and Overcoming Barriers to Intelligent Technology Adoption in Interior Design show that financing limits and recurring costs block small firms from such gains.

We borrow SWOT's one useful structural idea: separate internal quantities the firm controls from external ones it does not. We replace its subjective weights with measured quantities. Owned hardware and model tier are internal and set at procurement time. Traffic, surges and grid intensity are external and handled at runtime. We then check whether the qualitative posture agrees with the measured optimum. SWOT is an evaluation lens here, not a mechanism.

Research Program

We would build a small reference serving stack for live-room recommendation with three content tiers. The first is ID-only collaborative filtering. The second is a lightweight content encoder: a small, quantized multimodal or text model over whatever per-room content signals are available. The third is a LARM-style large-encoder tier. All three sit behind the same HarmonRank-style multi-objective ranker.

The central component is a refresh router. For each live room at each 30-second tick, it decides whether to recompute the content embedding or reuse the cached one. Room $r$ at tick $t$ is refreshed when

$$\hat{\Delta}_{r,t} \cdot v \;>\; \lambda \, e_{\text{tier}} \, I(t),$$

where $\hat{\Delta}_{r,t}$ is a cheap predictor of how much the refresh would change that room's ranking position for its likely audience, and $v$ converts ranking change into expected positive interactions. The right-hand side is the marginal operational carbon of one refresh: $e_{\text{tier}}$ is energy per refresh on the provisioned hardware, $I(t)$ is grid carbon intensity at time $t$, and $\lambda$ is the operator's carbon price. The predictor is trained from room age, interaction count, and chat-rate or topic-drift signals. Embodied carbon deliberately does not appear in this rule, following the sunk carbon argument.

Embodied carbon enters only at provisioning. For each candidate pair of hardware and tier $h$, we compute lifecycle carbon per functional unit:

$$C_h = \frac{E^{\text{emb}}_h \cdot \tau/L_h + \sum_t e_h(t)\, I(t)}{U_h},$$

where $E^{\text{emb}}_h$ is SCARIF-style whole-server manufacturing carbon, $L_h$ is service lifetime, $\tau$ is the evaluation window, and $U_h$ is the functional unit. We define $U_h$ as held-out positive interactions (clicks, comments, gifts) recovered in the top-K, following FUEL's functional-unit logic. Hardware must be sized for surge traffic, so $\tau$ includes synthetic 150× surge windows grounded in the micro-business case. Monetary cost per functional unit is computed alongside carbon, so that the adoption question and the carbon question stay separate and visible.

The analysis then asks three things. How is content-embedding value distributed across room "temperature"? What share of the tier-three gain does the router retain? Does the provisioning optimum shift toward older accelerators once refresh volume falls?

Baselines

The naive reference point is ID-only collaborative filtering under the same ranker, with no content embeddings at all. It is the cheapest possible system and the honest "do nothing" for a small operator. If it matches the others on the functional unit once carbon and cost are counted, the answer to our question is simply "don't bother." We would report that outcome.

The strongest existing method is LARM-style uniform refresh: every room, every 30 seconds, with a large encoder and gated fusion. Its reported result is Hit Rate@1000 rising from 13.86% to 17.03% on Kuaishou's proprietary trace in LLM-Alignment Live-Streaming Recommendation. We found no reported result on Lacuna for any content tier on KuaiLive, so we must reproduce it ourselves. We will run it on the same KuaiLive splits, the same ranker, and the same hyperparameter-search budget as our router. The large encoder will be the largest open model our compute allows, since the 100B teacher is unavailable.

For the ranking stage, HarmonRank's reported gain is +27.2 percentage points in summed AUC over multi-objective binary cross-entropy on TenRec-QKVideo (HarmonRank). We will reproduce that gain on TenRec before building on the ranker, to make sure our ranking stage is not a weak link.

The key ablation is fixed-rate refresh: every room refreshed every $k$ ticks, with $k$ tuned. This isolates whether per-room routing adds anything beyond simply refreshing less often. A second ablation is a cold-only rule that refreshes rooms below an interaction-count threshold and never others. This tests whether a hand-written rule captures everything the learned predictor does.

For the carbon side, the baseline is a SCI-style router that adds amortized embodied carbon to the refresh cost, reproducing the setup Bashir et al. warn against.

Beating the strongest method means one of two things, on the lifecycle-carbon-per-functional-unit distribution. Either the router achieves a lower median at matched long-tail Hit Rate@K, or it achieves higher long-tail Hit Rate@K at matched carbon, with PCAM intervals that do not overlap. The router must also beat fixed-rate and cold-only refresh on the same axes. Otherwise we adopt the simpler rule as the recommendation.

Why not just…?

Why not just skip content embeddings and run ID-only collaborative filtering? This is a real threat for small operators, because LARM's gate already weights ID signals at about 0.65. The counterweight is KuaiLive's tail: 70.7% of streamers have fewer than five interactions, and ID embeddings for them are nearly empty. Our stratified evaluation reports the ID-only gap separately for head and tail rooms. If the tail gap is negligible, the proposal's conclusion becomes "ID-only suffices," and that is still a useful, measured answer.

Why not just refresh every room less often? This may work nearly as well, and it is the alternative we take most seriously. That is why fixed-rate and cold-only refresh are first-class baselines, not afterthoughts. The router is only justified if value is concentrated in rooms that a fixed rate cannot distinguish, such as established streamers who suddenly switch products. The topic-drift feature is designed to catch exactly that case. If the frontiers for fixed-rate or cold-only refresh match the learned router's, we recommend the simpler rule.

Why not just include amortized hardware carbon in every runtime decision, as the Software Carbon Intensity metric does? The Sunk Carbon Fallacy shows this raises total emissions on a real datacenter trace at intermediate utilization. We do not assume its result transfers to bursty live traffic. We run the SCI-style router head to head and measure total emissions directly.

Why not just buy cloud AI, as the SWOT study prescribes? The prescription in Strategy of use Artificial Intelligence for optimization and stability cake business seasonal “Case study on small-medium scale model BR company” was never implemented. Per-call billing scales with exactly the 150× surges that small live sellers face, and financing constraints (Kuang et al.) make such variable costs hard to absorb. We model a pay-per-call tier in the provisioning comparison, using token counts priced under stated assumptions, rather than dismissing it.

Why not just rerank by product greenness? That is the approach in Towards Carbon Footprint-Aware Recommender Systems for Greener Item Recommendation. It answers a different question, about the carbon of the goods sold, and it could be layered on top of any configuration we study. We exclude it so that the compute effect is not confounded.

Proposed Experiments

The first step checks data feasibility. The file does not say whether KuaiLive exposes per-room content such as titles, chat text or frames. If only interaction logs exist, the content tiers will use the textual and metadata signals that are available. We will state that LARM-level multimodality is approximated, and we will not claim to replicate LARM's reported numbers.

The main experiment trains all tiers and routers on KuaiLive's first 18 days and tests on the final three. Evaluation uses Hit Rate@K and a stratified long-tail Hit Rate. Rooms are bucketed by prior interaction count, with the fewer-than-five bucket as the primary target. Energy per refresh and per ranking call is measured on two hardware classes, following FUEL's H100-versus-L40 comparison, with and without W8A8 quantization. Embodied carbon comes from SCARIF-style whole-server estimates. PCAM-style distributions over grid regions and yields are propagated through the whole calculation.

The first decisive check is the value-concentration test. If more than half of the content tier's Hit Rate gain over ID-only comes from rooms outside the cold bucket, our core hypothesis fails. In that case the router reduces to fixed-rate refresh.

The second decisive check is the provisioning flip. It asks whether, once refresh volume is cut by the router, the lifecycle-optimal hardware under surge-sized provisioning moves from the new accelerator to the legacy one. We would compare this against uniform refresh, which we expect to favor the new accelerator.

The third check is the sunk-carbon test on live traffic. We measure total emissions from the SCI-style router against the operational-only router, across utilization levels.

The final check is the SWOT concordance. We score the operator scenario with a standard factor matrix, as in the micro-business study, and ask whether its recommended posture matches the measured carbon-and-cost optimum. A mismatch would be concrete, quantitative evidence of where static situational analysis misleads small live-commerce adopters.