Meta超级智能的早期火花

Company Update North AmericaUnited States TMTInternet The Early Sparks of Meta'sSuperintelligence Valuation & Risks Benjamin Black, CFAResearch Analyst+1-212-250-9218 It is becoming increasingly clear that Meta's investments in AI are yielding everincreasing returns for advertisers, which in turn is supporting not only top-linedurability, but also advertising revenue growth acceleration. We recently hostedan ad-check call(link) with Andrew McLean, Principal and Managing Director ofInventus Media.At his agency, ad spend growth on Meta came in at +6.0% in 1Q26,exceeding the +5.5% expectation set at the quarter-ago call, driven by strong ROIfrom its Advantage+ product suite. Given the strong performance on theplatform, for 2Q26, he expects overall spend to accelerate almost 40bps (to +6.4%y/y), which is in contrast to consensus expectations that call for ~230bps ofsequential deceleration in FXN y/y growth in 2Q26 on a ~210bps tougher comp. Kunal Madhukar, CFAResearch Analyst+1-212-250-9357 Jeff Seiner, CFAResearch Associate+1-212-250-8153 Benjamin Hui, Esq.Research Associate+1-212-250-2064 Our conversations suggest that investors are expecting 1Q26 revenue of $57bn,with the high-end of the 2Q26 guide approaching $60.5-$61.0bn (up 27-28% y/y).In regards to expenses, we believe 2025's upward bias to expenses was anaberrationas the company absorbed the incremental costs of the MetaSuperintelligence Labs team (as well as the team's growing demands for cloudcompute). In 2026, we suspect the company is adopting an approach to guidancemore akin to prior years, where the bias to the expense guide was to trend flat todown throughout the year. In fact,over the last 4 years, Meta has lowered itsexpense outlook three times in the 1Q, including at 1Q25 earnings. We have alsoheard concerns that the expanded CoreWeave deal could potentially signal highertrending opex/capex, that said, we note that the MSA from the prior agreement inSept 2025 had included language that Meta had "the option to materially expandits commitment through 2032 for additional cloud computing capacity." As such,we believe this incremental spend was already contemplated in Meta's opex andcapex outlook. Raymond WongResearch Associate+1-212-250-2060 Meta also recently releasedMuse Spark, the first foundation model coming fromthe revamped Meta Superintelligence Lab (MSL), whichwe think is gearedtowards supporting Meta's core business, and the Meta AI assistant. It has beenengineered from the ground up as a "natively multimodal reasoning engine,"designed to seamlessly integrate visual data with text and tools. This differs fromsome existing models that have had multimodal capabilities added on. A keyinnovation is its "Contemplating Mode," which utilizes a multi-agent system totackle complex problems in parallel, aiming for deeper analysis without significantlatency. And even as importantly, per Meta, Muse Spark has been architected forefficiency, in that versus the Llama 4 Maverick model, it uses "an order of 15 April 2026InternetMeta magnitude" less compute resources, which is critical in an environment with severechip shortage and for deployment in smaller form factors.Based on reports fromindependent benchmarks, we believe this new model may not have entirelyleapfrogged the top tier, but it is still highly competitive, particularly in visualanalysis, reasoning, and healthcare applications. nMuse Spark is a potential software unlock that Reality Labs needed tojustify its large cash burn. Because the model is highly efficient andnatively multimodal, it appears that it is tailor-made for edge devices like theRay-Ban Meta smart glasses and future AR headsets.When an AIprocesses visualand audio inputs simultaneously and natively,itdrastically reduces latency. This means the glasses can "see" what theuser sees in real-time—overlaying virtual objects in real environments,translating live text, or offering contextual advice without the lag that ruinsthe AR illusion. nHealthcare and Wellness: Meta collaborated with over 1,000 physicians tocurate training data, enabling Muse Spark to provide more accurate andinteractivehealth insights.The model can analyze food photos fornutritional content, explain the muscle groups activated during exercises,and generate interactive health displays. This positions Muse Spark as adaily wellness companion, moving beyond generic chatbot responses.While it is not a substitute for professional medical advice, its ability tointerpret images and charts could provide more tailored guidance than text-only models.By positioning Muse Spark as a daily wellness companion(likely by integrating into wearables), Meta could create a sticky, high-frequency use-case. This strategy could yield ecosystem retention whilecollectinghigh-intent data(diet,fitness,health goals)that wouldultimately prove to be valuable for targeted advertising. nShopping: Muse Spark is set to powera new "Shopping Mode"thatleverages Meta's vast ecosystem o