Signals Inbox·August 27, 2026·AI Infrastructure

What startups have Meta bought recently?

Meta has bought PlayAI, WaveForms, Rivos, Limitless, Moltbook and Assured Robot Intelligence in its recent startup-buying wave, plus Manus, whose acquisition is now being unwound. The pattern behind those deals is even clearer: Meta is buying compressed R&D time across voice, chips, memory, agents and robotics.

Q1What startups has Meta actually bought recently?

Meta has made seven clear outright startup acquisitions in its current buying wave, although only six still meaningfully count as Meta-owned today because the Manus deal is being unwound.

The run started with PlayAI in July 2025, followed by WaveForms, Rivos, Limitless and Manus before the end of that year. Meta then bought Moltbook and Assured Robot Intelligence, or ARI, in 2026. We found no later confirmed outright startup acquisition after ARI. Meta has kept doing startup deals since then, but the more recent ones have mostly been acqui-hires rather than purchases of the whole company.

That distinction matters with Meta because several headlines make the acquisition list look longer than it really is. Dreamer, Gizmo and Virtue AI all lost important people to Meta, but Meta did not simply buy those companies. Scale AI is another special case: Meta spent $14.3 billion for a 49% stake and recruited CEO Alexandr Wang, but Scale remained a separate company.

Meta's recent outright startup acquisitions

Q2Why did Meta suddenly start buying startups again?

Meta started buying startups again because its AI push created technical gaps that the company wanted to close much faster than normal internal hiring could manage.

The change is easy to see in Meta's own accounts. Its 2024 annual report showed no goodwill added from acquisitions. One year later, Meta reported $4.09 billion of cash consideration and another $450 million of shares used for several business acquisitions. Meta said the resulting goodwill was primarily tied to advancing its AI work, acquiring talent, creating synergies and finding future ways to make money from the technology.

The timing lines up almost perfectly with the creation of Meta Superintelligence Labs and Zuckerberg's aggressive attempt to strengthen Meta's position against OpenAI, Google and Anthropic. PlayAI arrived first. WaveForms followed only weeks later. Rivos came next, then Limitless and Manus.

Meta these days is using acquisitions almost like an accelerated recruiting and R&D program. Instead of waiting for a new internal group to spend two years building expertise in speech, chips, agents or robotics, Meta can buy a team that has already spent those two years solving the problem.

Q3Why did Meta buy both PlayAI and WaveForms within a month?

Meta bought PlayAI and WaveForms almost back-to-back because voice had become important enough to justify collecting several strong teams rather than choosing one winner.

PlayAI already had a working platform for generating and cloning realistic voices. The startup had raised about $23.5 million, and Bloomberg reported that its entire team would join Meta. An internal Meta memo specifically connected PlayAI's technology with Meta AI, AI characters, wearables and audio creation.

WaveForms was a much younger company. It had existed for only about eight months, yet Andreessen Horowitz had already invested $40 million at a reported $160 million pre-money valuation. Co-founder Alexis Conneau had previously worked at OpenAI, where he helped build the neural-network technology behind GPT-4o's Advanced Voice Mode.

The acquisitions were slightly different. PlayAI brought a mature voice-generation platform and a full operating team. WaveForms brought frontier speech researchers working on more natural and emotionally expressive conversation.

Buying both within roughly four weeks suggests Meta expects voice to become a major interface for AI, especially on devices where typing makes little sense. Smart glasses are the obvious example. AI characters and assistants that people speak with all day are another.

Q4Is Rivos really Meta's way out of Nvidia?

Rivos gives Meta more control over its own AI chips, but currently it has done very little to free Meta from Nvidia.

Meta bought Rivos after already becoming one of the startup's largest customers. Rivos was developing processors around the open RISC-V architecture and had reportedly been seeking funding at a valuation above $2 billion. Meta already had its own MTIA chip program, so buying Rivos added hundreds of engineers, chip architecture and physical-design expertise that Meta had previously relied more heavily on outside partners to provide.

The integration has since gone badly enough to weaken the original thesis. The Information reported that Meta cut more than a quarter of roughly 450 former Rivos employees only months after the acquisition. One major chip project built around Rivos technology, code-named Olympus, was halted, while another project had its timeline pushed back.

Meanwhile, Nvidia remains central to Meta's giant AI infrastructure buildout. Meta is still spending enormous amounts on external accelerators even while developing its own chips.

Rivos may eventually lower Meta's cost per AI query or give it more bargaining power with suppliers. For now, the acquisition looks more like a difficult attempt to build a serious internal chip operation than an Nvidia escape hatch.

Q5Why did Meta buy Limitless when it already sells smart glasses?

Meta bought Limitless because the startup had spent years working on one of the hardest problems for AI glasses: giving an assistant a useful memory of what happens around you.

Limitless started as Rewind, software that recorded activity on a computer and made it searchable. The company later moved into hardware with a $99 pendant that could record conversations, transcribe them and let users search through what they had heard.

By the time Meta acquired the company, Limitless had raised more than $33 million from investors including Andreessen Horowitz, First Round Capital and NEA. Meta immediately connected the deal to its AI-wearables effort. Limitless then stopped selling new Pendants and began winding down parts of the standalone product.

That sequence tells us more than the acquisition announcement itself. Meta did not need another necklace to compete with Ray-Ban Meta glasses. It wanted a team that already understood continuous audio capture, personal context, searchable memory and the awkward privacy problems that appear when a device is always listening.

Those capabilities become much more valuable when the device is already sitting on someone's face all day.

Q6Why did Meta pay more than $2 billion for Manus?

Meta paid more than $2 billion for Manus because Manus had already shown something most AI-agent startups were still trying to prove: people would actually pay meaningful money for autonomous agents.

Manus launched its general-purpose AI agent in 2025 and said it crossed $100 million in annual recurring revenue within roughly eight months. The product could browse websites, research topics, write code and carry out multi-step tasks inside a virtual computer.

The price also shows how quickly expectations changed. Benchmark had led a $75 million funding round earlier in 2025 that valued Manus at around $500 million post-money. Meta's reported purchase price was roughly four times that level only months later.

Even using Manus's $100 million-plus ARR figure as a rough floor, Meta was paying something around 20 times annualized revenue. That would be expensive for a normal software company. Manus was growing at a speed that made normal software comparisons less useful, and Meta was also buying technology, users, engineers and a position in one of the hottest parts of AI.

This was the closest Meta came in its recent buying wave to purchasing a real independent AI business rather than primarily absorbing technical talent.

Q7Does Meta still own Manus today?

No. Manus is currently separating from Meta and returning to independent operations after Chinese regulators forced the acquisition to be unwound.

China's National Development and Reform Commission blocked the transaction in the spring and ordered the companies to reverse it. Manus had been founded by Chinese engineers before moving its headquarters to Singapore, and Beijing treated the transfer of its technology to a U.S. company as a foreign-investment issue.

Meta subsequently separated Manus from its internal systems and stopped data sharing between the companies. Manus has now told users that it is returning to independent operations and that some data created after Meta's acquisition will have to be deleted as part of the separation.

Reuters also reported that Tencent had been discussing becoming Manus's largest shareholder in the new ownership structure.

Manus should remain in any historical list of startups Meta bought recently because the transaction genuinely happened. As of now, though, listing Manus as a startup Meta owns would be wrong.

Q8Why did Meta buy Moltbook, the social network for AI agents?

Meta bought Moltbook because its most interesting feature was the network connecting AI agents, not the viral spectacle of bots posting to each other.

Moltbook appeared in early 2026 as a Reddit-like network where agents running through OpenClaw could create profiles, post, comment and interact. The site became famous almost immediately, partly because people were fascinated by apparently autonomous bots discussing their own lives and partly because questions emerged about how autonomous some of the viral activity really was.

Meta acquired Moltbook only weeks after its launch and brought founders Matt Schlicht and Ben Parr into Meta Superintelligence Labs.

Meta's explanation was unusually specific. The company highlighted Moltbook's "always-on directory" for connecting agents and said the approach could help agents work for people and businesses.

That fits where Meta is going now. The company has started pushing business agents for customer service and other commercial tasks, while Zuckerberg has talked about a broader enterprise opportunity around APIs, agents and compute. If millions of agents eventually need to discover each other, verify identities or coordinate tasks, Moltbook was exploring that infrastructure unusually early.

Q9What does Meta buying Assured Robot Intelligence say about its humanoid robot plans?

Meta's purchase of Assured Robot Intelligence shows that humanoid robotics is now a real part of its AI research agenda, even if a Meta-branded robot is still far from guaranteed.

ARI was founded by robotics researchers Lerrel Pinto and Xiaolong Wang and was building foundation models that help robots understand people, predict what happens in physical environments and learn physical tasks. Meta bought the startup and moved the team into Meta Superintelligence Labs.

That fits work Meta was already doing. The company has built world models such as V-JEPA that learn how physical scenes change and has a dedicated robotics effort focused on the software intelligence behind humanoid machines.

ARI adds researchers working specifically on whole-body robot control and learning from human behavior. Meta therefore has several pieces of the same puzzle now: vision models that understand the physical world, wearable devices that observe the world from a human point of view, and robotics researchers trying to convert that understanding into physical action.

We still have very little evidence that Meta wants to manufacture humanoid robots at scale itself. The stronger case today is that Meta wants to own part of the intelligence layer that could eventually run inside many robots.

Q10Did Meta actually buy Dreamer, Gizmo or Virtue AI?

Meta did not buy Dreamer, Gizmo or Virtue AI outright, even though all three belong in the story of how Meta has been absorbing startup teams lately.

Dreamer's founders and team joined Meta Superintelligence Labs after building a system that let users create personalized AI agents. Dreamer remained a separate legal entity, while Meta received a non-exclusive license to use its technology.

The Gizmo deal was similar. Meta hired the engineering team behind Atma Sciences, the company that built the AI-generated mini-app platform Gizmo, and licensed its technology. The startup itself was not acquired in the conventional sense.

Virtue AI is the newest example. Meta recently hired co-founders Dawn Song, Bo Li and Sanmi Koyejo along with other team members from the AI-security startup. Axios reported that the group would work on security and safety around increasingly capable AI agents. Virtue AI itself continued operating and was later acquired by Fortinet.

Calling all three "Meta acquisitions" therefore inflates the real list. Calling them ordinary hiring also misses how coordinated these transactions were.

Meta startup deals that were not outright acquisitions

Q11Did Meta buy Scale AI?

Meta did not buy Scale AI. Meta paid $14.3 billion for a 49% stake and recruited Scale CEO Alexandr Wang, while Scale remained a separate company.

The size of the deal makes the confusion understandable. Meta spent more on its Scale stake than it has publicly spent on any startup acquisition in this recent wave, and Wang immediately became one of the most important people inside Meta's AI organization.

The structure was still very different from a takeover. Meta stayed below majority ownership, Scale continued as an independent company and Scale appointed a new CEO after Wang left.

Scale is still important to understanding the broader pattern. The transaction came just before Meta began its run of AI startup acquisitions, and Wang now runs Meta's superintelligence effort. It shows how aggressively Meta was willing to spend for talent and strategic access when an outright acquisition was either unnecessary or impractical.

But for a literal answer to "what startups has Meta bought?", Scale AI should stay off the acquisition list.

Q12Is Meta mostly buying startup products or startup teams?

Meta is mostly buying teams and technical capability, with the standalone startup product often becoming secondary once the deal closes.

PlayAI's entire team moved into Meta. WaveForms' founders joined Meta and its standalone web presence disappeared. Rivos was folded into Meta's chip organization. Limitless stopped selling new Pendants. ARI's founders and team moved into Meta Superintelligence Labs. Moltbook's creators joined the same AI organization.

The acqui-hires make the pattern even stronger. Dreamer, Gizmo and Virtue AI all involved Meta taking people without buying the whole company.

Manus was the unusual case because Meta originally planned to keep operating and expanding the consumer product. Manus already had real revenue and millions of users, so shutting it down would have destroyed a major part of what Meta was buying.

Across the wider set, the behavior is pretty clear. Meta seems willing to buy a very young startup once a small group of people has demonstrated unusually fast technical progress. WaveForms had existed for months. Moltbook had existed for weeks. Dreamer had barely opened its beta.

For Meta right now, a functioning team can be valuable long before it becomes a functioning business.

Q13How much money is Meta actually spending on these startup acquisitions?

Meta's startup acquisition push is already a multibillion-dollar program, even though most individual purchase prices remain secret.

Meta's 2025 annual report gives us the cleanest aggregate. The company completed several business acquisitions with $4.09 billion in cash consideration and another $450 million in Meta shares, bringing total purchase consideration to roughly $4.54 billion.

Meta did not disclose exactly how that amount was split between PlayAI, WaveForms, Rivos, Limitless, Manus and any smaller transactions included in the accounting. We therefore cannot responsibly assign made-up prices to each startup.

The accounting also shows how abrupt the change was. Meta recorded no acquisition-related increase in goodwill in 2024. In 2025, acquisitions added about $3.8 billion of goodwill, and Meta explicitly linked most of that value to its AI efforts, workforce and expected synergies.

Spending slowed in cash terms during the first half of 2026, when Meta reported $474 million of cash used for acquisitions of businesses and intangible assets. The comparison is imperfect because deal structures and accounting timing differ, but the order of magnitude is still useful.

Meta went from essentially no visible acquisition contribution to goodwill in 2024 to several billion dollars of acquisition consideration one year later. That is a genuine strategic change, not a handful of opportunistic deals.

Q14Have Meta's startup deals already turned into real products?

Yes. Meta's Gizmo acqui-hire has already turned into a very visible consumer product called Pocket.

Meta hired the Atma Sciences engineering team behind Gizmo earlier this year and obtained a non-exclusive license to its technology. Gizmo let people describe simple interactive experiences with text prompts, then publish those mini-games or toy apps into a social feed.

Meta has now rolled Pocket out across the United States. Pocket works in almost exactly that format: users describe an interactive "gizmo," AI generates it, and other users can play, remix, like or share it. The app is powered by Meta's Muse Spark model.

The speed is notable. We are seeing the output of the Gizmo team inside a Meta-owned product only months after the team joined. Gizmo itself is being wound down.

We have much less visibility into the other deals. PlayAI and WaveForms technology can disappear inside Meta AI without a public label saying where it came from. Limitless expertise could show up gradually in future wearables. ARI may take years to produce anything consumer-facing.

Rivos provides the cautionary example. As pointed out above, reporting about layoffs and cancelled chip plans suggests that absorbing a strong startup team does not automatically produce faster execution.

Q15So which startups has Meta bought recently?

As of now, the clean list is PlayAI, WaveForms, Rivos, Limitless, Moltbook and Assured Robot Intelligence, plus Manus as a completed acquisition that is currently being unwound.

Dreamer, Gizmo and Virtue AI belong in a separate acqui-hire category. Scale AI belongs in another category again because Meta bought a large minority stake rather than the company.

The more interesting conclusion is how concentrated the buying has become. Meta has collected technology around realistic voice, AI memory, agents, agent networks, custom chips and humanoid robotics in a relatively short period. Most of those companies were tiny compared with Meta, and several were barely old enough to have built normal businesses.

Meta's current startup strategy is much closer to buying compressed R&D time than recreating the Instagram and WhatsApp playbook. It is paying to bring small groups of unusually fast builders inside the company before their technology becomes strategically important somewhere else.