Meta booked an 8.3% lift in ad clicks and a 15.7% lift in conversions last quarter from AI that no user agreed to. The vision everyone is arguing about is free, optional, and needs a graphics card selling for $4,699.
Zuckerberg published 6,500 words promising personal superintelligence for everyone, and the reaction was hostile enough that TechCrunch spent a podcast episode on it. None of that reaches Meta’s income statement. Meta’s AI already earns its return inside the ad auction, where advertisers pay and users get no vote, and the free half of the program runs on hardware Meta never has to buy.
Rebecca Bellan of TechCrunch decided to try Muse Glimmer, the open model Meta shipped the same day as the essay. She could not put it on her MacBook. The weights need a machine with 24GB of memory to spare, which her laptop did not have. That gap between the title of the document and the specification of the artifact is the whole story, and it is not a marketing failure. It is the design.
What Happened
On August 10, Meta published “The Future Is for Everyone” at meta.com, a 14-page essay in which Zuckerberg argued that concentrating AI in a few hands is itself the safety risk, and promised that everyone will get a capable personal agent that understands their goals. The same day, Meta Superintelligence Labs released Muse Glimmer, a 29.6-billion-parameter model under Apache 2.0, built for always-on local agents and distilled from the proprietary Muse Spark.
The reception went badly. TechCrunch AI editor Russell Brandom argued on August 10 that the manifesto explains why people dislike AI. On the August 16 Equity podcast, Anthony Ha, Kirsten Korosec and Bellan picked over the same ground: the messenger problem, the Pollyanna tone, the gap between a promised future and anything a reader would want. Bellan read Zuckerberg’s posture as “almost like an anti-Dario”, a deliberate counterweight to Anthropic’s Dario Amodei, who spent that same weekend arguing the AI backlash is a crisis of trust.
The polling supports them. In the Bentley University-Gallup Business in Society survey of 3,270 US adults conducted May 4 to 11 and published July 28, the share of Americans saying AI does more harm than good climbed from 31% to 39% in a year, while the share saying it does more good fell to 9%. Only 27% said they trust businesses to use AI responsibly, down from 31%. Among adults aged 18 to 29, the share with no trust at all in business use of AI rose from 29% to 41%.
The Backstory
Meta has been here before, and it has the receipts. Zuckerberg renamed the company in 2021 around a future consumers were asked to buy a headset to enter. They mostly declined. Reality Labs has now run operating losses in every quarter since Meta began reporting the segment in late 2020, reaching roughly $88 billion against about $12 billion of cumulative revenue. In January, after a $6.02 billion quarter, Meta cut hundreds of jobs from the division and moved Horizon Worlds into maintenance mode. Q2 2026 still lost $4.62 billion on $431 million of revenue.
That is what it costs Meta when a consumer purchase decision sits between the vision and the user. The AI program removes the purchase decision. Glimmer is free, permissively licensed and hosted on Hugging Face. Meta AI arrives inside WhatsApp, Instagram and Facebook, apps 3.6 billion people already open daily. Nobody has to be persuaded to install anything, and nobody has to be persuaded to like it either.
The license carried its own signal, which we covered when Glimmer landed: Meta dropped every restriction it had attached to Llama, because terms are a price you can only charge when the buyer has no substitute.
The Plan
Zuckerberg laid out three revenue lanes on the July 29 earnings call, and consumer affection is required for none of them.
The first is the core business. LLMs now sit inside Instagram and Facebook ranking. Every public Reels and Feed post on Instagram runs through a language model that scores it across topic and tone before anyone sees it. On the ads side, Meta shipped its Generative Recommender this quarter, which reasons about ad content and user preference together rather than scoring each ad alone.
The second is enterprise: the Muse Spark API, now on OpenRouter for US developers, plus business agents. More than a million businesses already run Meta Business Agents weekly on WhatsApp and Messenger. Movida, a Brazilian car rental chain with close to 400 locations, put one on the whole booking flow and reported 44% more daily bookings through the channel, with 85% of conversations closed without a human.
The third is compute itself. Zuckerberg told analysts Meta is fielding offers to sell compute at a premium to what it paid, while arguing the margin on selling intelligence beats the margin on selling capacity. Meta guides 2026 capital expenditure of $130 billion to $145 billion, and it is moving the largest projects off its own balance sheet through joint ventures.
The personal agent, the thing the essay is about, has no price, no ship date and no revenue line. Zuckerberg told analysts it has to work out of the box for billions of people, and that Meta has not shipped it yet.
The Business Model Angle
Meta’s Q2 ad revenue grew 27% to $59.4 billion. Split that into its parts and the arithmetic is tidy: ad impressions rose 14%, average price per ad rose 12%, and 1.14 multiplied by 1.12 gives 27.7%. Impressions grew because AI recommendations held people in the feed longer. Price grew because AI made the ads convert better. Susan Li put a figure on the second half: Meta’s GEM ranking model and sequence learning, combined with new user-understanding models, delivered an 8.3% increase in ad clicks and a 15.7% uplift in conversions on Facebook.
No user was asked. There is no toggle for LLM-scored ranking, no consent screen for generative ad retrieval, no version of Instagram that skips it. The AI that pays for Meta’s buildout operates on the far side of a decision the user never gets to make. Our earlier piece on Meta’s automated account bans made the structural point: advertisers pay, and users are the inventory. The AI program follows the same map. Sentiment is a cost only where sentiment gates a transaction, and in the ranking stack nothing is transacted with the person being ranked.
Now look at the half people are shouting about. Meta compressed Glimmer’s weights to 4-bit precision, landing under 20GB, so the model plus its KV cache, perception encoder and speculative-decoding drafter fit a 24GB or 32GB envelope. That decision moves inference cost off Meta’s servers and onto your desk. You buy the silicon, you pay the electricity, and Meta pays nothing per token.
The Steam Hardware Survey is the friendliest sample available, since PC gamers own far more discrete GPU memory than the general public. In July 2026, 16GB cards became the most common configuration at 25.90%, 8GB sat at 25.32%, and the 24GB tier that Glimmer targets reached 5.43%. Roughly one machine in eighteen, in the most GPU-rich consumer population on earth, can hold the model Meta built for everyone.
The price of clearing that bar has gone the wrong way, and Meta helped push it. Meta published Glimmer’s throughput figures on an RTX 5090. That card launched at $1,999 in January 2025 and carried a median US listing price of $4,699.99 in August 2026 per Tom’s Hardware, up from $4,299.99 in June, about 135% above list. The cause is memory: Samsung, SK hynix and Micron have been steering wafer and packaging capacity toward high-bandwidth memory for AI data centers, leaving GDDR7 short. Consumer device makers have been paying that bill all year. When Meta raised its own capex guidance in April, Li named higher memory-chip pricing as a reason. The buildout that makes Meta’s free model possible is inside the price of the card you need to run it.
One line in the expense disclosure deserves more attention than it got. Meta’s 55% expense growth was driven in part by third-party AI token costs. Meta is buying inference from other labs while giving its own weights away.
The Risk
The honest counterargument is that Meta is asking consumers to buy something after all, and it is working. Reality Labs revenue grew 16% on AI glasses even as Quest sales fell. Zuckerberg said early sales of the new Meta Glasses line, built with EssilorLuxottica, beat internal expectations, and he calls glasses the ideal form factor for an assistant. Connect on September 23 will show whether the glasses become the agent surface. If they do, preference starts mattering at the till.
Trust is also an input cost, not decoration. Zuckerberg told analysts that personal superintelligence needs a clear mental model of everything happening in a user’s life. Calendar, health, money, relationships. Meta cannot default anyone into that grant the way it defaults them into a ranked feed, and Gallup says 27% of Americans trust businesses with AI. Meta One, the new subscription, has the same problem: someone has to choose to pay.
The consent-free layer carries the regulatory exposure. Meta took $2.4 billion in legal charges in Q2 and told investors that youth-related trials scheduled this year may produce a material loss. Ranking and recommendation systems are precisely what those cases are about. Distribution defaults are a licence, not a property right, and the EU has spent two years demonstrating it.
Then there is the model itself. Meta’s own comparison table shows Qwen 3.6 27B ahead of Glimmer on OSWorld-Verified, TerminalBench 2.1, GDPval-AA and most multimodal tests. Free is not the same as best, and the open-weight tier is already crowded with cheaper Chinese options.
Quick Questions
Does the backlash cost Meta anything today? Not in the ad business, which produced 98% of Family of Apps revenue last quarter and grew 27% on levers users do not control. It costs Meta later, in the subscription, agent and glasses lines that need someone to say yes.
Why give a capable model away? Meta has no model P&L to protect. Free weights cost Meta nothing on the income statement and cost rivals pricing power on theirs.
What does on-device inference do for Meta’s economics? It converts a variable server cost into a customer-owned fixed asset. Meta ships the file once and never pays for a token.
Can most people run Muse Glimmer? No. The 4-bit build needs just under 20GB of free memory, and only 5.43% of surveyed Steam machines carry a 24GB card. Meta’s cloud-hosted assistant in WhatsApp remains the version almost everyone touches.
What settles the question? Muse Spark 1.2’s weights, promised without a date, license or hardware spec. That release tells you whether Meta is distributing intelligence or distributing the models it no longer needs.
The Business Model Analyst Take
The essay reads as a request for permission that Meta’s plan does not require. Zuckerberg has built an AI business with two payers, and the user is neither of them. Advertisers fund the ranking stack through an auction they cannot opt out of on behalf of the people being ranked. Consumers fund the local agent through a graphics card whose price Meta’s own data centers helped double. In between sits a manifesto asking for enthusiasm about the only layer with no invoice attached.
That structure is the lesson Meta took from the metaverse. Eighty-eight billion dollars of Reality Labs losses bought a simple insight: never put a consumer purchase decision between your vision and your distribution. So the model is free, the license has no terms, and the assistant appears in a chat thread people already have open. The public gets a vote on the part that does not matter and no vote on the part that does.
Watch three things this quarter. Whether average price per ad keeps climbing, which tells you the ranking gains are compounding rather than a one-off. Whether Meta One reports a subscriber number, which would be the first honest test of whether anyone will pay Meta for AI. And whether the glasses at Connect are priced as hardware or as a subsidised doorway to an agent, because that choice reveals which side of the ledger Meta thinks the consumer belongs on.
