Meta's very bad 2026 (so far)

Meta has just agreed to a settlement worth up to $17.1 billion over allegations that Facebook and Instagram were deliberately designed to keep children hooked on its platforms.

The case had been running since 2023 and finally reached a federal trial on 18 August. The states accused Meta of knowingly designing addictive features to maximise the amount of time children and teenagers spent on Facebook and Instagram, while misleading parents and the public about the risks. Meta denied the allegations and fought the case for years. New York Attorney General: start of the Meta trial

Then the trial actually started.

Eight days later, Meta agreed to pay at least $12.1 billion, potentially rising to $17.1 billion, while accepting legally enforceable restrictions covering how minors use Facebook and Instagram. Those include a two-hour daily limit, overnight restrictions, limits on notifications, stronger age assurance and the option to avoid algorithmic feeds. New York Attorney General: full $17.1bn settlement terms

Meta hasn't admitted the allegations were true, but after spending nearly three years fighting the entire premise, paying up to $17.1 billion and accepting restrictions aimed directly at the behaviour being challenged is one hell of a way to end the argument.

Especially because Meta was having another wonderfully timed problem at almost exactly the same time.

Please stop filming people with the Facebook glasses:

Ray-Ban Meta glasses are probably the most successful new hardware product Meta has produced in years. Around seven million pairs were reportedly sold during 2025, meaning this isn't another failed gadget nobody cared enough to buy.

Instead, people bought them and demonstrated the fairly obvious social problem with putting a discreet camera inside something deliberately designed to look like ordinary glasses.

Reports of people covertly filming strangers, particularly women, helped turn "pervert glasses" from an internet insult into a mainstream nickname. Influencers promoting them reported losing followers, businesses started banning them, and courts in England and Wales banned them over concerns about unauthorised recording. The Guardian: backlash against Meta's "pervert glasses"The Guardian: Meta glasses banned from courts

US law enforcement has since started worrying about exactly the same thing, with agencies warning that the glasses could be used to covertly record police facilities and other sensitive locations. The Guardian: US police concerns over Meta smart glasses

For almost any other company, having your most promising new gadget culturally rebranded as voyeur hardware would be an impressive branding disaster. For Meta, after decades of privacy scandals, it feels almost precision-engineered to remind everyone why they already distrusted the company.

So within weeks Meta had agreed to one of the largest social-media settlements ever while its big new hardware success was getting called pervert glasses.

And neither is the biggest hole Meta is digging this year.

$145 billion for fucking what?

Meta now expects to spend $130 billion to $145 billion on capital expenditure during 2026, up from $72.22 billion in 2025, with infrastructure for its AI ambitions accounting for much of the increase. Meta: Q2 2026 results and capex guidance

And one of the flagship results we're apparently supposed to admire is Muse Spark.

Meta launched the original Spark in April and claimed Meta Superintelligence Labs had rebuilt its AI stack "from the ground up" in nine months. After the embarrassing Llama 4 generation and a huge reorganisation of Meta's AI operation, it suddenly had a closed-weight model producing highly competitive benchmark scores against OpenAI, Anthropic and Google. Meta: original Muse Spark announcement

The timeline inevitably raised questions about how much of that extraordinary jump came from synthetic data and distillation from existing frontier systems. I wrote about this back in May.

At the same time this supposedly ground-up competitor appeared, Meta employees were using Claude so heavily that somebody inside the company had built a leaderboard called Claudeonomics, ranking employees by how many Anthropic tokens they consumed. Meta shut it down after the data escaped outside the company. Fortune: Meta's Claudeonomics token leaderboard

Nine months from rebuilding the stack to a closed model suddenly trading benchmark blows with the frontier, while your own workforce is hammering Claude hard enough to turn token consumption into an office competition.

Sure.

More importantly, even if you accept every benchmark number Meta prints, there is still a much simpler problem:

What is the fucking business?

Meta says Spark is built for its products. It can improve Meta AI, recommendations, shopping, Facebook, Instagram, WhatsApp, Messenger and the smart glasses. Meta's own launch material describes exactly that integration. Meta: where Muse Spark is actually being deployed

Fine. Meta already owns one of the largest and most optimised advertising systems ever created.

Facebook and Instagram already rank content, predict engagement, automate campaigns, target advertising and squeeze money from billions of people's attention. AI can obviously improve those systems further. Better recommendations might keep people scrolling longer. Better targeting might squeeze another percentage point from conversions. Generative tools can make producing advertisements cheaper.

But advertising is already the fucking company.

After spending up to $145 billion in a single year, is the revolutionary new economic opportunity really that Instagram can serve you a slightly better Temu advert?

Meta has now launched the broader Muse personal agent as well, capable of sending emails, shopping, making payments, booking travel and controlling other apps. In other words, another general-purpose AI agent entering a market already full of OpenAI, Anthropic, Google and everyone else trying to build exactly this sort of thing. Reuters: Meta launches the Muse personal agent

There are subscriptions and API revenue attached to this stuff, so Meta technically has ways of charging for it. What it does not have is anything resembling a demonstrated business proportional to $130 billion to $145 billion of annual capex.

Another chatbot isn't that business. Slightly better adverts aren't that business. A shopping agent isn't that business. The numbers are simply on completely different planets.

The money is already disappearing:

I've already covered Meta's ridiculous Q2 results separately, so there's little reason to go through the entire quarter again.

Revenue grew 28% to $60.8 billion.

Costs and expenses grew 55% to $42 billion.

Operating income fell 8%.

Meta generated $31.86 billion of operating cash flow, then spent $31.08 billion on capex and finance leases, leaving only $784 million of free cash flow, down from $8.55 billion a year earlier. Meta: Q2 2026 financial results

Even Family of Apps managed to grow revenue 28% while its operating profit fell 6%, with costs rising 67%. Meta's own filing lists data centres, technical infrastructure, cloud services and third-party AI token costs among the things driving expenses higher. Meta Q2 10-Q: Family of Apps and Reality Labs results

That's the absurdity of Meta's current position. Its advertising machine added more than $13 billion of quarterly revenue year on year and the company still ended up with lower operating profit and almost no free cash flow after investment.

The core business is running faster and faster while Meta shovels the extra money into an AI buildout whose eventual economics still amount to trust us, it'll be massive.

Meta tried the AI future on itself:

Meta even tried demonstrating those productivity gains internally.

Zuckerberg's Project OT, short for Organization Transformation, explored reorganising the company around AI agents and shrinking some teams by as much as 60%.

Then Meta got a rather useful preview of its own AI economics.

Reuters reported that major site emergencies increased by around 40%, while employee time spent dealing with those emergencies increased by around 70%. Zuckerberg ultimately abandoned plans for another company-wide layoff wave later in the year. Reuters investigation: how Meta's Project OT imploded

So Meta is spending historic amounts of money partly on the promise that AI will produce enormous productivity gains, while its own aggressive attempt to reorganise the company around that idea created more technical emergencies and required more humans to clean them up.

That's considerably more informative than another benchmark chart.

The old hole is still there:

None of this means Meta has finished paying for its previous corporate obsession.

Reality Labs generated $431 million of revenue in Q2 and lost $4.62 billion from operations. Meta Q2 10-Q: Reality Labs segment results

Horizon Worlds has meanwhile produced one of the better corporate strategy sequences of the year.

In February, Meta announced that Worlds was shifting to be "almost exclusively mobile", openly acknowledging that VR had not grown as much or as quickly as expected. Then it went further and announced Horizon Worlds would actually be removed from Quest completely in June, making Meta's flagship metaverse a mobile-only product. Meta: renewed mobile focus for Horizon Worlds

A month later, CTO Andrew Bosworth suddenly announced Meta had changed its mind and would keep existing Horizon Worlds VR experiences alive after users complained. New development would still focus overwhelmingly on mobile. TechCrunch: Meta reverses Horizon Worlds VR shutdown

Meta spent tens of billions trying to manufacture a VR metaverse important enough to rename the entire company after it, announced it was taking that metaverse out of VR, then reversed the decision almost immediately and left the existing VR version on life support while focusing on phones.

You genuinely couldn't plan the symbolism better.

Some other highlights from Meta's excellent year:

The bigger disasters haven't stopped Meta finding time for smaller ones.

In June, attackers managed to manipulate Meta's AI customer-support chatbot into handing over high-profile Instagram accounts, including an inactive Obama White House account, Sephora and an account belonging to a senior US Space Force official. The attackers exploited the automated support process and Meta later fixed the vulnerability. Reuters: Meta AI support chatbot account breach

It's a beautifully compact version of the whole strategy: automate a sensitive job with AI, remove more human involvement, then discover the AI has been given enough authority for somebody to persuade it to hand over the keys.

Then there was the timing masterpiece.

While Meta was literally being taken to court over alleged harms to children, researchers were discovering that Meta's own advertising systems had approved paid ads containing AI-generated child sexual abuse material.

The Tech Transparency Project eventually identified more than 350 abusive video ads published across Facebook, Instagram and Threads since late 2025. Some used images of real children, while many directed users towards AI "nudify" applications. Meta removed the ads after they were reported and said it had earned less than $5,000 from them. WIRED: more than 350 abusive ads found on Meta platforms

The amount Meta earned is almost beside the point. One of the largest advertising companies on Earth, spending tens of billions on AI and constantly talking about automated safety systems, was accepting money for AI-generated child sexual abuse ads while defending itself in a trial about the safety of its platforms for children.

You can't parody that timing.

Europe has been keeping Meta occupied as well. In July, the European Commission preliminarily found that the addictive design of Facebook and Instagram breached the Digital Services Act, specifically examining infinite scroll, autoplay, notifications and personalised recommendation systems. European Commission: preliminary addictive-design finding against Meta

That came after another preliminary finding in April that Meta was failing to adequately stop children under 13 from accessing Facebook and Instagram, including because they could simply enter a false birth date without an effective check. European Commission: Meta's under-13 age-assurance failure

Apparently one enormous regulatory fight about children wasn't enough.