ARTICLE · 1225581
Muse登顶App Store,钱却难赚 | BARRON'S
01 文章选材
【Para. 1】Meta新发布AI产品Muse广受好评
【Para. 2】但作者分析Muse目前仍处于亏损状态，且难以收回本
【Para. 3】Meta原有的广告战略有所改变，难以通过这一渠道变现
【Para. 4-5】而寄望于电商和订阅费则显得更加不明确
【Para. 6-8】Muse的用户大部分仍属于免费层级，Meta试图通过低价来抢占市场份额
【Para. 9-11】Meta的计划和套餐设置和目前市场情况存在出入
【Para. 12-13】重度用户的比例不高，用户总数的增多反而不利于利润增长
【Para. 14-15】AI市场目前都处于高成本投入阶段，如何稳定推高利润率会是各公司的重点
906 words ★★☆☆☆（难度评级）
【Para. 1】Meta Platforms’ Muse personal AI agent is an early hit, holding the iOS App Store’s top spot since soon after its Sept. 8 release. According to research firm Human, it was already the No. 1 agent last month, responsible for 40% of agent-based traffic on the internet. Largely because of its release, agent traffic tripled in a month. Meta stock is up 18% since Muse’s release.
【Para. 2】But Meta is probably losing money on Muse now, and it will be very hard for it to defray the costs down the line.
【Para. 3】As Meta demonstrated with its Threads microblogging platform, its strategy is to scale a new product and then monetize it with ads. But Muse may be different, because it’s promising to bring a level of privacy and security to it that Meta isn’t known for. Soon, interactions with Muse will be opaque to everyone aside from the user, including Meta. The kind of highly effective ad targeting that Meta’s data collection enables in Facebook and Instagram may not work as well in Muse. If user behavior in Muse is anonymized, it won’t provide Meta with data to further refine its targeting.
【Para. 4】In any case, Meta isn’t yet talking about ads in Muse. For now, it’s looking to e-commerce and premium subscriptions to subsidize the costs.
【Para. 5】And those costs may be very high. At least to begin with, Meta is being very generous with the Muse free tier, allocating up to 100 million tokens a week for each user. Tokens are the basic unit of artificial-intelligence text; for example, this sentence is made up of about 30 tokens when measured by common methods. Though we don’t know exactly what the computing costs of 100 million tokens are to Meta, we have a decent idea of what it would charge if someone were using the underlying model, Muse Spark, to do other things, like writing software code using the company’s application programming interface, or API.
【Para. 6】On the API, there are different prices for tokens depending on how they’re being used, with the cheapest being “cache” tokens. Even if 97% of a Muse user’s tokens are for cache, free-tier users are getting the equivalent of a $1,100 annual API credit. In the past 12 months, Meta has generated only $63.54 per each of its 3.6 billion daily app users, almost entirely through ads. The tough job ahead of Meta is bridging that gap for Muse.
【Para. 7】It’s unclear what Meta’s profit margin is on the Muse Spark API, but we can deduce that it isn’t as high as the top models from OpenAI and Anthropic. In the “cost per task” metric of research firm Artificial Analysis, Muse Spark is closer to inexpensive Chinese models than the ones from the U.S. frontier AI labs. The Muse Spark API is a great deal for users, and Meta may well be underpricing it to gain market share.
【Para. 8】Revenue per user from the Muse agent is likely going to have to be much higher than what Meta earns from its apps to recoup the costs. If Meta were able to capture all of Visa’s revenue with its Muse e-commerce fees—$45 billion in the past 12 months—that could be enough, but scaling to the size of Visa won’t be easy.
【Para. 9】This year, Meta intends to spend up to $145 billion on capital expenditure for new AI data centers.
【Para. 10】The Muse math gets even tougher for Meta in the paid tiers. The $240-a-year plan offers the equivalent of $5,500 in annual API credit, and the $1,200-a-year tier comes with a staggering $33,000 of annual credit. On the free tier, where most of the casual users of Muse will be, Meta can count on many of them not coming close to their free token allocation, bringing average token use down. But people on the paid plans will probably be power users who will be using Muse to run their lives like a full-time human personal assistant would. They may come close to using up their entire token budget every week.
【Para. 11】Meta didn’t respond to a request for comment on the wide gulf between its average revenue per user and Muse token allocations.
【Para. 12】The company is hoping that everyone becomes a power user, and that they make Muse a daily habit like its social-media apps. But the more successful Muse is, and the more people use it, the more difficult profitability will become.
【Para. 13】This dovetails with a recent report from research firm SemiAnalysis, which showed that the API costs of the token allocations for OpenAI and Anthropic subscription plans are well in excess of their subscription prices. In effect, at all three companies, API customers are subsidizing free and subscription users.
【Para. 14】The situation is indicative of where we are right now in AI. The costs are evident in the hundreds of billions of dollars being spent for new AI data centers. Cloud providers like Amazon.com, Microsoft, and Alphabet are earning revenue from renting out these servers, but companies with no cloud service like Meta are fishing around for income, and it isn’t clear exactly where the profits come from when the expenses are so high.
【Para. 15】Meanwhile, Meta shareholders have to endure reduced operating profit margins, down to 37% in the latest quarter (after excluding some one-time charges) from 43% last year. Making Muse a mass-market service is going to be no mean feat, but making it a profitable one may prove the bigger hill to climb.
05 检验题
Meta's Muse personal AI agent is an early hit, holding the iOS App Store's top spot since its release. Largely because of its release, agent traffic ____【11】____ in a month. But Meta is losing money on Muse now, and it will be very hard to ____【12】____ the costs down the line.
Meta's usual strategy is to scale a new product and then monetize it with ads. But Muse may be different, because it promises privacy and security. Soon Muse interactions will be ____【13】____ to everyone except the user, including Meta. If user behaviour is anonymized, it will not provide Meta with data to refine its advertising ____【14】____. For now, it is looking to e-commerce and premium subscriptions to ____【15】____ the costs.
The costs may be very high. The free tier allocates up to 100 million tokens a week, the equivalent of a $1,100 annual API credit. Meta generated only $63.54 per daily app user last year. The tough job ahead is ____【16】____ that gap. Revenue per user may need to exceed Meta's app earnings to ____【17】____ the costs.
Paid tiers make the math even tougher. The $1,200-a-year plan comes with $33,000 of annual credit. Casual users on the free tier may not come close to their allocation, but paid users are likely to be power users who may use their entire ____【18】____ every week. Meta hopes Muse becomes a daily ____【19】____, but the more successful Muse is, and the more people use it, the more difficult ____【20】____ will become.
Multiple Choice Questions
11. [A] doubled [B] fell [C] tripled [D] paused
12. [A] create [B] defray [C] advertise [D] reveal
13. [A] transparent [B] visible [C] public [D] opaque
14. [A] targeting [B] avoidance [C] language [D] storage
15. [A] increase [B] measure [C] subsidise [D] document
16. [A] widening [B] bridging [C] ignoring [D] reporting
17. [A] recoup [B] expand [C] conceal [D] borrow
18. [A] sentence [B] profile [C] password [D] budget
19. [A] rival [B] mistake [C] habit [D] law
20. [A] recruitment [B] profitability [C] moderation [D] translation
看更多单词注释，全文翻译，长难句解析，写作积累以及答疑服务；
扫码下方报名训练营获取👇👇👇
注：在线课程报名之后不接受退款，报名之前有疑问请先加助教微信（ttzhujiao888）查看例文免费试学和训练营介绍，考虑清楚再报名
