Meta's over $100 billion AI spend, $600 billion planned by 2030, has no clear path to recoupment; Muse's 5 million downloads in 22 days came from promotion to existing users, few will pay, and profit must come from undisclosed retailer commissions. Agent errors make calling it superintelligent an exaggeration.

sharply critical of Meta's strategy in the field of artificial intelligence on this channel in the past. We have been wondering why the company has an AI strategy in the first place.

They have already spent well over $100 billion on AI data centers, and plan to spend $600 billion by 2030. That's a huge amount, even for Meta.

Until recently, it was not at all clear how they planned to recoup this investment. During the first two years, they developed a family of open-weight "Llama" models. Despite being widely used by developers around the world, it can be downloaded and used for free without any licensing fees. Therefore, "Meta" did not generate any revenue.

Zuckerberg also made the unusual decision to appoint 28-year-old startup founder Alexander Wang to be in charge of the social media giant's artificial intelligence efforts.

In July, reports surfaced that Meta was planning to lease some of its unused data center capacity to other AI companies. If so, this would represent a humiliating admission that the company’s internal ambitions in artificial intelligence are faltering.

But it seems that Zuckerberg has changed his mind and now wants to keep all the data center capacity for himself.

They launched a personal assistant called "Muse". Initial adoption figures are promising, with the "Muse" app garnering 5 million downloads in just 22 days. According to some estimates, this makes it the fastest-growing app ever.

"Muse" is a personal assistant that runs almost continuously in the background on a cloud computer in one of Meta's data centers. Computing costs are likely to be enormous, and the vast majority of users will likely not pay anything at all.

In this video, we'll delve into the depths of "Muse," the reason "Meta" created it, and most importantly, how they plan to eventually profit from it.

First, let's differentiate between "Mews" from "Meta" and traditional chatbots based on large language models such as "ChatGPT" and "Clude".

Each Mews user gets their own virtual machine within Meta's data center. This virtual machine has its own web browser, storage space, processing power, and memory. It also stores user files, its "Mews" memory, and login data for any services the user connects to.

The AI model itself does not work inside the virtual machine. When the virtual machine needs to think, it uses the "Muse Spark" language model, which is located in a different part of the "Meta" data center.

The linguistic model provides the necessary intelligence. It essentially tells the virtual machine what it should do. The virtual machine stores all the relevant data that makes your "Muse" proxy unique to you.

Since the device is in the cloud and not on the user's phone, "Muse" can continue working on tasks after the application is closed. The main selling point of "Muse" is that it always runs in the background.

Therefore, it performs tasks over a long period of time.

For example, you could ask her to read your email inbox every morning and filter out the important messages. If there is a particular product you want to buy, she can search for it online daily and notify you as soon as there is a discount on it.

Meta also unveiled a device called the "Muse charm". It is essentially a small smart speaker that allows you to communicate with your "Muse" agent via voice commands. It is supposed to connect to your smartphone via Bluetooth.

The Muse device is not yet available for sale and its price has not been announced, but Zuckerberg says shipments will begin in time for the upcoming holiday season.

The nature of Muse's work, which is constantly running, means that it likely consumes a huge amount of computing resources. Despite varying estimates, Meta will likely need data center infrastructure worth several billion dollars for every 100 million Muse users.

However, Meta is believed to already have enormous amounts of data center capacity that are not fully utilized. Therefore, the only real cost is the marginal cost of running the electricity.

In theory, the launch of Muse was a huge success with over 5 million app downloads in the first 22 days. However, Meta heavily promoted Muse to its existing base of Facebook, Instagram and WhatsApp users.

We do not know how much these people actually use it.

So, how does Meta plan to generate revenue from Muse? Technically, there are paid options available. The free version gives you 100 million tokens per week. For $16 a month, you can increase the limit to 500 million.

For $80, you can get 3 billion tokens per week. For the vast majority of people, the 100 million tokens in the free plan are more than enough. Realistically, a very small number of people would pay for Muse outright.

The real profits will come from shopping through smart agents. The idea is to ask your Muse agent to find deals online, perhaps for a particular item of clothing. Muse will display the product list, and you simply need to press a button to complete the order.

Ultimately, Meta may make deals with various e-commerce sites. Meta will program Muse to route a larger share of shopping orders to partner retailers in exchange for a commission.

The commission will be paid by the retailer, not the consumer. Therefore, you as a user will not even know about it.

While shopping through agents may work in theory, the bigger question is whether consumers will trust it or not. Muse has an incentive to direct you toward products where the retailer has agreed to pay commissions to Meta.

These will not necessarily be the best deals. In reality, these are unlikely to be the best deals, as the retailer has to recoup the commission cost in the form of higher prices.

This leads us to broader concerns regarding data privacy in relation to Muse. Meta does not have a successful track record when it comes to handling user data. Having an AI agent that is always on and has access to vast amounts of sensitive information about you could lead to serious risks.

To the extent that Muse grants access to your passwords or credit card numbers, this information is stored entirely within your virtual machine, which is theoretically isolated.

Even Meta itself cannot access the data on your virtual machine. Meta claims that it also does not use your Muse chat history to target you with ads on Facebook or Instagram.

We know that Meta strongly encourages its employees to use AI proxy programming tools. This has led to a number of internal security failures. Meta is so concerned about potential security vulnerabilities in Muse that it is offering rewards of up to $300,000 if anyone can find a way to hack into their Muse servers.

At the launch event, Zuckerberg promised that you would be able to make money using Muse. Muse also has an innovative business model. We believe that Muse will make you money. We support this by making Muse free for a large number of tokens, with the expectation that we will generate a profit over time by taking small transaction fees.

One obvious place to make money is Facebook Marketplace. By default, Muse is already connected to your Facebook account.

In September 2026, a Canadian YouTube user named Matt Robb asked Muse to sell a keyboard on Facebook Marketplace. Muse took over the management of the listing, including communication with potential buyers.

Muse gave Matt's address to a potential buyer and told him to come anytime between 10 a.m. and 10 p.m. He himself had no idea about it and was not at home at the time. Muse never asked Matt for approval of the sale or a delivery date.

If you buy something from Facebook Marketplace, you can't tell if you're talking to an AI agent from Muse. You'll think you're talking to a real person. A few hours later, the buyer arrived at Matt's house.

Matt wasn't home, so there was no one to open the door. While the buyer waited outside, Muse repeatedly assured him that Matt was on his way home. None of this was true. Throughout this time, Matt had no idea what was going on.

The problem is that large language models are merely predictive text generators. She is adept at writing texts that sound convincing based on her training data, but she lacks any real understanding of the situation.

Just a few days later, Matt Robb faced another failure for Muse. Muse was asked to put his Google Pixel smartphone up for sale. While Muse could negotiate, the minimum price should have been $700.

A potential buyer offered $600. Muse accepted the $600 offer despite receiving explicit instructions regarding the minimum price.

These mistakes made by the agent have repercussions in the real world. In Matt Robb's case, he didn't lose any money, but his reputation on Facebook Marketplace was badly damaged by very angry buyers.

This creates a kind of dilemma. Muse is only useful if it can perform long-term proxy tasks on your behalf. To perform these tasks, he needs access to things like your calendar, your email, and possibly even your payment information.

But the more freedom you give him, the more likely he is to make costly mistakes.

Zuckerberg has been talking about personal superintelligence for some time. We now know that this is what he was referring to. I think it's an exaggeration to call Muse super-intelligent when it can't even sell a keyboard on Facebook Marketplace.