Meta’s Superintelligence Gamble

Meta has rebuilt part of its Menlo Park operation into something closer to a national research laboratory. The Superintelligence Labs and a smaller group known as TBD Lab, a deliberately provisional name for its most confidential work, are organised around one objective, an AI able to adapt, learn, and improve itself while it runs. The spending matches the ambition. Meta expects capital expenditure of up to 72 billion dollars in 2025, most of it directed at AI infrastructure. The recruitment has been more striking still. One twenty-four-year-old researcher was reportedly offered 250 million dollars across four years, with as much as 100 million in the first. Several senior researchers have been approached with packages beyond 300 million dollars.

The target is larger than a chatbot or a productivity tool. Zuckerberg describes it as personal superintelligence, an assistant that knows the user, anticipates them, and acts on their behalf, present across every Meta application and device and, in principle, across much of daily life for billions of people. Recruitment on this scale aims past advertising performance, toward the point where billions of people reach information and make decisions each day.

The pitch is simple enough: a single assistant that is always available, learns from continuous use, and fits into the texture of an ordinary day. It can book a meeting, order a gift, check a flight, read a contract, or surface one fact at the moment it is needed, whether the user is at a desk, on a call, or looking through a pair of Ray-Ban Meta glasses. Simple is not the same as easy.

Trust is the first barrier, and Meta's data history works against it

The first obstacle is trust, and not the kind involved in handing a platform a few photographs. A personal assistant becomes useful only when it can read across the joined-up detail of someone's life, the messages, the calendar, the locations, the purchases. That cross-contextual memory is the feature, and it is also the first thing a hesitant user turns off. Meta approaches this from a difficult position. The European Union fined it 1.3 billion dollars for data-protection breaches in 2023, and Apple has spent years marketing on-device processing as a signal of restraint.

The data that would make the assistant useful is the same data that drives Meta's advertising revenue. Privacy-preserving methods such as on-device processing and federated learning can narrow the gap, and Meta could commit to opt-in consent and independent audits. None of that removes the underlying tension. An assistant built to know everything about a person and a business built on monetising what is known about that person pull in opposite directions. Meta cannot fully be both at once.

Competitors have already installed the habits the assistant would need to displace

Trust would only open the door. Daily use is won or lost on whether the assistant does something the alternatives do not, and the alternatives are already installed. Microsoft's Copilot ships inside Office and Windows, present by default for more than a billion users. Google has placed Gemini across Gmail and Search. OpenAI's ChatGPT carries the recognition that comes with roughly 800 million monthly users. Perplexity has built a following on cited, checkable answers, and Claude has become the preferred tool for many writers and engineers. Each of these occupies a habit Meta would have to displace.

Meta's distinct advantage is its open release of models such as Llama, which has drawn a large developer base and given the company influence over how AI is built across the industry, beyond its own products. Its infrastructure adds weight, with multi-gigawatt clusters named Prometheus and Hyperion supporting a rate of iteration few competitors can match. Where rivals control the desktop or the search bar, Meta controls the glasses, the headset, and the social graph, surfaces the others cannot reach.

The assistant has to feel continuous across platforms Meta does not control

Embedding AI across Meta's properties is the simpler half of the problem. Making it behave as one continuous presence, instead of a set of separate features, is what tests the design. A request made to the glasses to identify a restaurant, shared through WhatsApp, checked against reviews on Instagram, and booked through Facebook, should read as one conversation that remembers each step. Continuity of that kind runs straight into surfaces Meta does not own. Microsoft controls the desktop, Google shapes the habits formed around search, and Apple decides which assistants may run on its devices. These constraints are written into the architecture of rival platforms, and distribution alone does not remove them.

The routes around these constraints are limited. Meta can carry context across devices through secure tokens instead of deep access, and it can embed with services such as Shopify and Notion so the assistant stays useful where it is kept out of the operating system. Regulation may help here, since the European Union's Digital Markets Act pushes toward the interoperability that Apple and Google would prefer to avoid. The strongest pressure comes from demand itself. An assistant that becomes the way groups plan and coordinate through WhatsApp builds that pressure from below.

A rival assistant could intercept the engagement Meta's business depends on

For Meta the bet is defensive before it is expansionary. Its advertising income depends on people coming to Facebook, Instagram, WhatsApp, and Messenger to share and discover and buy. An assistant installed at the operating-system level, owned by someone else, would meet many of those intentions before they reached a Meta application. A person who asks a rival assistant to plan a weekend, find a restaurant, or share a set of photographs completes that task somewhere other than Meta, and each completion elsewhere is attention Meta does not capture.

The loss would be quiet. Meta's applications would be reduced to features inside a rival's service, the direct relationship with users would thin, and a new layer of revenue, transaction fees from commerce, paid AI subscriptions, and advertising placed inside a context-aware assistant, would form around whoever owns that assistant.

Win the same contest and the position inverts. Meta would stand at the point where billions of people look for information and decide what to do, directing that attention into its own services before rivals saw it. Microsoft, Google, and Apple would be left adjusting to Meta's choices instead of setting their own. Adoption on a large enough scale might even quiet the privacy history, since proven usefulness tends to dull old objections.

For Zuckerberg the personal stakes match the corporate ones. He changed how the world connected once, and owning the layer beneath the next phase of computing would place him among the figures who defined it. His record of hard pivots, mobile and then Reels, suggests he reads this as survival. What differs this time is the nature of the threat, which is the possibility of another company owning the ground beneath everything he has built.

Success would turn Meta into the target its rivals coordinate against

Reaching the top would not bring quiet. A lead in personal AI would draw a coordinated response from companies with deep reserves and their own reasons to resist a single dominant assistant. Apple could combine tighter platform rules with a renewed privacy campaign, timed to the moment Meta's assistant scales, and use its work in augmented and virtual reality to make Meta's hardware look dated before that hardware becomes essential. Google could secure exclusive data and content arrangements for Gemini while encouraging regulators to examine how Meta uses its data, slowing the climb at its steepest point. Microsoft tends to flank, and a partnership with a large messaging or entertainment service could move billions of interactions toward Copilot. Beyond any single move, rivals could form alliances and press for standards aimed at concentrated AI, casting Meta as a giant in need of constraint.

Scale that signals victory would also make Meta the obvious target, and each misstep would carry further once competitors found common cause against it.

All of it depends on whether the technology can be built at all

Every part of this rests on one assumption, that a personal superintelligence of the kind Zuckerberg describes can be built at all. Current large language models perform well inside controlled conditions and become unreliable as tasks grow long and interdependent. The product Meta describes would need steady reasoning across many simultaneous conversations, accurate recall over months of interaction, and dependable execution of tasks with several steps. Present models still invent details when asked to complete something as ordinary as a restaurant booking.

The infrastructure carries its own limits. Serving billions of people with responses fast enough to feel human demands computing resources that strain physical, environmental, and financial boundaries at once. An assistant that pauses noticeably to think reads as broken. Data adds a further constraint, since more training material brings smaller gains while amplifying the biases already present, and Meta's social corpus, however large, may lack the structured knowledge that reliable reasoning needs.

Should progress slow, the competitive argument weakens quickly. A rival could move ahead through a different approach, whether neuromorphic hardware, hybrid symbolic methods, or some advance that leaves today's architectures behind. The state of the underlying science will show the direction of the contest earlier than any product launch.

Whether the bet is working will show in a few concrete measures

The bet will become legible within roughly two years, against a few plain measures. One is whether a meaningful share of Meta's daily users, above thirty per cent, come to rely on the assistant for everyday tasks instead of testing it once. Another is whether large outside services such as Shopify or Uber build native connections to it. The last is whether continuity across devices becomes reliable enough to trust without thinking about it. These separate a novelty from a habit.

One dependency runs underneath all of it. The bet reflects Zuckerberg's own appetite for risk and his willingness to spend at a rate few boards would tolerate. A successor, or a shift in the company's priorities, could bring pressure to monetise these features early or to slow the spending before the science is settled. Whether Meta's reinvention outlasts the intensity of its founder is a question the company has not yet had to answer.

What Meta is attempting is a change in what it is for, from a company that grew the world's social networks into one that wants to be the layer through which people reach everything else. The spending assumes that change is possible and that the technology will deliver what the plan requires. Both remain open, and the company has chosen to commit tens of billions of dollars before either is known.