Why the Metaverse Failed and What That Means for AI's Optimistic Scenarios

Between 2021 and 2023, Meta spent approximately $46 billion on its Reality Labs division — the unit responsible for VR hardware and the metaverse. This is not a rounding error. $46 billion is more than most countries’ annual defense budgets. It’s more than the GDP of Bolivia. It funded thousands of engineers, researchers, industrial designers, and content creators working on what Mark Zuckerberg publicly described as the successor to the mobile internet.

In Q4 2022, Reality Labs had 21,000 employees and was losing $4.3 billion per quarter. By the end of 2023, the Quest 3 headset was genuinely impressive hardware — comfortable, high-resolution, capable of convincing mixed reality. And nobody was using the metaverse.

Horizon Worlds, Meta’s flagship virtual social platform, had approximately 200,000 monthly active users in early 2023, down from 300,000 in late 2022. For context: MySpace, at its most embarrassing decline, still had millions of users. A company that in the same period had 3 billion monthly active users across Facebook, Instagram, and WhatsApp could not get 300,000 of them to spend meaningful time in its flagship metaverse product.

The question worth asking is: why? The failure was not inevitable. The technology worked. The investment was real. The company had the largest social platform on the planet. If any organization could have made a virtual social world work, Meta was it. What went wrong, and what does the answer mean for AI?

The failure mode has a name in product strategy: the jobs-to-be-done mismatch.

Clayton Christensen’s jobs-to-be-done framework asks a simple question about any product: what job is the user hiring this product to do? Not “what does the product do,” but “what task was the user trying to complete when they turned to this product?” The distinction matters because products that are technically impressive sometimes don’t get hired because something else already does the job adequately.

The metaverse was trying to get hired for “virtual social connection.” This is a real job — people want to feel connected to people they can’t physically be with. But the metaverse competed with a set of alternatives that were deeply established, extremely convenient, and good enough: phone calls, video calls, text messages, social media. The pandemic, far from having “trained a billion people to live online,” mostly confirmed that people preferred simple, low-friction video calls (Zoom, FaceTime) over elaborate virtual environments. They chose the technology that got the job done adequately, not the technology that was most technically ambitious.

The metaverse’s specific problem was friction. Not just technical friction — the headset discomfort, the setup process, the limited battery life — but social friction. Organizing a virtual reality meetup requires all participants to own a headset, install the same application, navigate an unfamiliar interface, and tolerate the limitations of avatar-based representation that most people find slightly uncanny. Organizing a Zoom call requires a link.

This friction asymmetry is fatal in consumer social products. Social platforms win on network effects — the platform everyone is on is the platform you use, because the people you want to talk to are there. But network effects compound for low-friction products and barely work for high-friction ones. Every additional person who joins your WhatsApp group makes WhatsApp more valuable. Every additional person you’d need to convince to buy a Quest headset and set it up to join your virtual hangout makes the virtual hangout less likely to happen.

The 2006 virtual world Second Life peaked at around one million monthly active users and never grew beyond that in any sustained way, despite being technically functional and having active corporate investment. The lesson was available in 2021. Meta chose not to take it.

There’s a version of this analysis that is too cynical — that says Meta’s metaverse was always theater, a pivot narrative to distract from Facebook’s stagnating growth and regulatory pressure. The cynical version has something to it (the October 2021 rebrand to Meta happened exactly as Facebook whistleblower Frances Haugen’s congressional testimony was dominating the news), but it’s incomplete. The engineering investment was real. The hardware got meaningfully better. The company genuinely believed, or had a large faction that genuinely believed, that VR was the next major computing platform.

The sincere believers were wrong because they confused capability improvement with inevitability. VR hardware got better every year from 2016 to 2024 — resolution, latency, weight, comfort, tracking — and at no point did “better hardware” translate into “product that people chose to use for daily social interaction.” The capability improvements didn’t address the friction problem or the jobs-to-be-done mismatch. They made the product better at the thing people had already decided they didn’t want.

The analogy from consumer electronics is 3D TV. 3D television was real technology, shipped by real companies (Sony, LG, Samsung), with real content investments from Hollywood. The picture quality was impressive. Adoption was essentially zero. By 2016, every major TV manufacturer had stopped making 3D TVs. The technology worked. It solved a problem (flat images look flat) that consumers had already decided wasn’t a problem worth solving with glasses and headaches.

What does this mean for AI?

The optimistic AI scenarios involve conversational AI becoming the primary interface for information, creative work, coding, customer service, healthcare, education — essentially, the replacement of most existing software interfaces with natural language. Some of this is already happening and the evidence for it is strong: GitHub Copilot demonstrably speeds up programming, AI tutoring systems improve learning outcomes in controlled studies, customer service automation reduces handling time on routine queries.

But there are adjacent optimistic scenarios where the jobs-to-be-done analysis should make you skeptical.

AI companions — services like Replika, Character.ai, and various successors that offer AI-powered emotional relationships and friendship — are getting a lot of investment and press coverage. The technology works, in the sense that people do engage with them. The jobs-to-be-done question is: what job are users hiring an AI companion for? If it’s “manage acute loneliness in the absence of other options,” that’s a job the technology can do, but it’s a job you might not want to optimize for because “give lonely people a better AI substitute for human connection” has complicated social effects that aren’t obviously good.

If the job is “provide a consistent, low-judgment conversational partner for people developing social skills,” that’s a better job — and there’s evidence some users hire AI companions for exactly this purpose. But whether this is a large enough and stable enough use case to build businesses on is a different question.

The more direct AI parallel to the metaverse failure is AI in the enterprise. The story being told in 2025 is that every enterprise will become an AI-first organization, that AI will transform knowledge work, that companies not adopting AI will be outcompeted. The investment numbers are credible. The executive attention is real. The consulting industry has built a cottage economy around the transformation narrative.

What’s less clear is whether enterprise AI is being hired for jobs that the technology is actually good at, or whether a significant portion of the investment is hiring AI for the job of “demonstrating to investors and boards that we are taking AI seriously.” The latter job doesn’t require AI to actually improve productivity — it requires AI to be visibly deployed. This is a jobs-to-be-done mismatch, and it’s one that produces adoption statistics (look, we have 50,000 Copilot seats!) that don’t reflect actual utilization or value creation.

The metaverse lesson isn’t that transformative technology can’t succeed. It’s that capability does not produce adoption, adoption requires something specific from users (jobs to be done), and enthusiasm from investors and executives is not evidence that users are enthusiastically hiring the product. Meta had the investor enthusiasm and the executive commitment. The users, given the choice, kept using Zoom.

AI has something the metaverse never had: clear, demonstrable utility in specific narrow tasks. Code completion, document summarization, research assistance — these are real jobs where AI is genuinely better than the alternatives in ways that matter. The risk is that the AI field uses the success of narrow-but-real use cases as evidence for the viability of broad-and-speculative ones. That inference doesn’t follow. The metaverse might have been excellent at making VR games, and it was — but “excellent VR games” didn’t validate “replace the social internet.”

The AI use cases that have shown genuine sustained adoption share a specific characteristic: they reduce friction in tasks that users were already doing and already valued. GitHub Copilot autocompletion reduces the friction of typing repetitive code. ChatGPT reduces the friction of starting a blank-page writing task. Perplexity reduces the friction of search-synthesize-summarize research tasks. These are improvements to existing workflows. The friction reduction is real and measurable.

The AI use cases that are struggling to find sustained adoption tend to involve asking users to do new things rather than do existing things more easily. AI virtual meetings (nobody asked for this). AI-driven social networking features (users find them uncanny rather than useful). Autonomous AI agents that handle complex multi-step tasks (users are nervous about handing over control and the error rates remain too high for most contexts). These are capability demonstrations in search of a use case.

The metaverse lesson isn’t that transformative technology can’t succeed. It’s that capability does not produce adoption, adoption requires something specific from users (jobs to be done), and enthusiasm from investors and executives is not evidence that users are enthusiastically hiring the product. Meta had the investor enthusiasm and the executive commitment. The users, given the choice, kept using Zoom.

The deeper question the metaverse poses for AI is about the narrative arc. Every major tech company is now, to varying degrees, betting its future on AI in the way Meta bet on the metaverse. The bet is less obviously wrong — AI has more demonstrated utility across more domains than VR social spaces ever did. But the scale of investment, the certainty of the predictions, and the competitive pressure to keep investing regardless of evidence are the same dynamics that produced the $46 billion writedown.

What specific human need does this AI application meet better than the alternative? Answer it concretely, with evidence. If you can’t, you’re building a metaverse.