Meta Chief Executive Mark Zuckerberg has acknowledged that the company’s sweeping AI-driven restructuring has not delivered the rapid gains management anticipated, offering one of his clearest admissions yet that the social media giant’s aggressive reorganization has fallen short of expectations even as it pours unprecedented sums into artificial intelligence.

Speaking during an internal town hall on Thursday, Zuckerberg told employees that Meta had overestimated how quickly AI agents would transform software development and workplace productivity, while conceding that the company’s restructuring, which included thousands of layoffs and internal transfers, had been more disruptive than executives intended.

His remarks come at a critical juncture for Meta, which has emerged as one of the biggest spenders in the global AI race. The company is projected to invest as much as $145 billion this year in AI infrastructure, contributing to a broader wave of more than $700 billion in AI spending by the world’s largest technology companies. Investors have increasingly scrutinized whether such enormous capital expenditure will translate into meaningful productivity gains and new revenue streams.

Zuckerberg said the company’s expectations for AI-powered software agents had proved overly optimistic.

“In retrospect, the trajectory of the agentic development over at least the last four months hasn’t really accelerated in the way that we expected,” Zuckerberg told employees, according to a recording heard by Reuters.

He added that the company’s strategic bets on its new organizational structure “haven’t come to fruition yet.”

The Meta chief was referring to AI agents, software systems capable of autonomously carrying out complex tasks on behalf of users with minimal human intervention. Much of the technology industry’s current investment cycle is centered on the belief that these systems will eventually automate programming, research, customer service, content creation, and many other knowledge-based tasks.

The comments amount to a rare public acknowledgment that even the world’s largest AI investors are encountering difficulties turning cutting-edge models into immediate productivity improvements.

Zuckerberg also conceded that Meta’s restructuring had not been executed as effectively as management had hoped. The company laid off roughly 10% of its global workforce in May while reassigning about 7,000 employees to AI-focused teams, moves that triggered internal criticism and renewed concerns over employee morale.

“We probably didn’t do as clean of a job on that as we could have,” Zuckerberg acknowledged, according to the recording.

The restructuring was designed to redirect resources toward AI development while reducing spending elsewhere as Meta sought to compete more aggressively against rivals including OpenAI, Anthropic, Google and xAI.

When executives began planning the overhaul earlier this year, they feared Meta was moving too slowly to capitalize on the industry’s rapid evolution.

“The conversations that I was having with our top people when we started planning all this in January and February were that they were worried that we weren’t going to move fast enough to adapt,” Zuckerberg said.

He added that executives at the time were “super optimistic” about emerging AI coding systems such as Anthropic’s Claude Code, believing that such tools would rapidly transform software engineering productivity. That optimism has since moderated as AI agents have proven less capable than many industry leaders initially anticipated.

Still, Zuckerberg maintained that Meta remains committed to its long-term AI strategy and expects the company’s massive investments to begin producing more tangible returns over the coming months.

“I expect that we’ll begin to experience more significant benefits from our AI investments within the next three to six months,” he told employees.

While AI models have demonstrated impressive capabilities in coding, reasoning and content generation, many companies are discovering that integrating those systems into real-world business workflows is proving more complex and slower than expected.

Even so, Meta executives sought to reassure employees that the company’s next generation of AI models is making rapid technical progress.

Alexandr Wang Says Meta’s Catching Up

During the same town hall, Meta’s recently appointed superintelligence chief, Alexandr Wang, said the company’s upcoming frontier model, codenamed “Watermelon,” has reached performance levels comparable to GPT-5.5, according to two people familiar with the meeting.

Wang said the assessment was based on widely followed AI benchmark tests, although he did not specify which benchmarks were used.

“Watermelon, our next model after Avocado, is currently in training,” Wang said during the town hall, according to a person familiar with the meeting.

“Watermelon uses an order of magnitude more compute than Avocado,” he added.

“Avocado” is Meta’s internal codename for Muse Spark, the first model in a new family of AI systems the company introduced in April.

The increased computing requirements illustrate the escalating AI arms race, with companies deploying ever-larger data center clusters and advanced chips to train increasingly sophisticated models.

Wang reinforced that message publicly later on Thursday in a post on X, where he indicated that an updated version of Muse Spark would be released soon.

He said the update would deliver major improvements in coding and agentic capabilities aimed at narrowing the performance gap with competing models.

When asked by an X user when Meta would produce a coding model comparable to Anthropic’s Claude Opus, Wang replied that it would be “pretty soon,” adding that users would like what the company has “cooking.”

This means Meta believes its latest generation of models can substantially improve its competitive position after earlier AI releases received mixed reviews compared with offerings from OpenAI and Anthropic.

The town hall also addressed another controversy that has generated internal criticism. Meta Chief Technology Officer Andrew Bosworth told employees that an internal investigation into the company’s mouse-tracking software found that employee information had not been incorporated into AI model training.

The software, introduced on U.S. employees’ computers in April, monitored mouse movements and digital activity to collect data intended to improve AI systems. The initiative quickly became controversial after employees objected to the level of workplace monitoring. Last month, Meta suspended the program while investigating whether sensitive information had been exposed.

Bosworth said the review concluded that employee data had not been used to train AI models.

He also announced a significant policy reversal if the program resumes.

“For people who are comfortable, that’s great, they can contribute to this kind of great human survey. To people who are not, it is not an issue,” Bosworth told employees.

Unlike the original rollout, which employees were told they could not refuse, Bosworth said any future version of the program would operate on an “opt-in” basis, allowing employees to choose whether to participate.

Meta declined to comment publicly on the town hall discussions. However, the internal admissions underscore the mounting pressure facing Meta and other technology giants as investors demand evidence that record-breaking AI spending will ultimately translate into stronger earnings growth.

China’s artificial intelligence race has entered a new phase, with Beijing-based startup Z.ai emerging as the latest company challenging the dominance of leading U.S. AI developers. Its recently launched GLM-5.2 model is winning praise from developers, technology executives and investors, bolstering the belief that China’s AI ecosystem is closing the performance gap with OpenAI and Anthropic while maintaining a significant cost advantage.

The model, launched last month, is generating growing interest across the global developer community because of its advanced coding and autonomous agent capabilities, allowing it to complete sophisticated software engineering and reasoning tasks with minimal human prompting.

According to a Reuters report, industry observers describe the enthusiasm surrounding GLM-5.2 as a “mini DeepSeek moment,” recalling the shockwaves created when DeepSeek unveiled a powerful low-cost reasoning model early last year that challenged assumptions about the enormous capital required to build frontier AI.

Unlike earlier generations of Chinese AI models, which were often viewed as cheaper but less capable alternatives to U.S. offerings, GLM-5.2 is increasingly being discussed as a genuine competitor to the latest systems from OpenAI and Anthropic.

Its rapid adoption is evident on OpenRouter, one of the world’s leading AI developer platforms, where GLM-5.2 has climbed above Anthropic’s models in usage rankings. The model has also received endorsements from influential technology leaders, including Snowflake CEO Sridhar Ramaswamy and venture capitalist Marc Andreessen, further boosting its credibility among software developers.

David Sacks, who previously served as U.S. President Donald Trump’s AI czar, said the emergence of GLM-5.2 demonstrates how rapidly China’s AI capabilities are advancing.

“We now have a Chinese open-weight model that is as good as the currently available models from OpenAI and Anthropic,” Sacks said last week, before Washington lifted restrictions on Anthropic’s Fable and Mythos models on Tuesday.

Speaking on the All-In podcast, Sacks added that GLM-5.2 is “just a tick below Opus 4.8 (from Anthropic) and right up there with GPT 5.5 (from OpenAI),” warning that “we cannot afford to do things that slow our companies down.”

Within parts of the U.S. technology industry, there is growing concern that regulatory uncertainty could weaken America’s lead in artificial intelligence just as Chinese companies are becoming more competitive.

Several analysts believe the timing has also favored Z.ai.

Washington’s temporary restrictions on Anthropic’s newest models and OpenAI’s delayed public rollout of GPT-5.6 have prompted many developers to experiment with alternative models, accelerating international interest in GLM-5.2.

Brian Tse, founder and CEO of Beijing-based AI consultancy Concordia AI, said developers are increasingly seeking alternatives to proprietary American models.

“The international developer community is increasingly aware that relying solely on proprietary, U.S.-based API models carries significant risk,” Tse said.

Cost has become another powerful advantage.

As businesses deploy increasingly sophisticated AI agents, token consumption—the units used to measure AI usage—has risen sharply, making proprietary AI services substantially more expensive.

Against that backdrop, GLM-5.2 has attracted attention by delivering performance approaching frontier models while costing roughly one-sixth as much as leading closed-source offerings from OpenAI and Anthropic. Although Z.ai has not disclosed how much it spent developing GLM-5.2, the pricing has made it particularly attractive to startups, software developers and enterprises looking to control AI infrastructure costs without sacrificing capability.

Independent benchmarks reinforce its growing reputation.

GLM-5.2 currently ranks fifth on Artificial Analysis’ large language model intelligence leaderboard, which measures overall reasoning, knowledge, and coding capabilities across numerous standardized tests. It also ranks second on Code Arena’s front-end coding leaderboard, which evaluates models’ ability to generate websites and user interfaces.

For many developers, however, the biggest attraction lies in usability rather than benchmark scores.

Tiezhen Wang, former Asia-Pacific lead at Hugging Face, said GLM-5.2 significantly lowers the technical barriers traditionally associated with deploying open-source AI.

“The shift GLM-5.2 brings is that the open-source model has become a plug-and-play, out-of-the-box product,” Wang said.

“You just deploy the model and without doing any complex fine-tuning systems, it is in a highly usable, ready-to-use state. This drastically lowers the barrier to entry for open-source adoption.”

Z.ai’s ambitions extend well beyond its current model.

In a response to Elon Musk on X last month, founder Tang Jie said the company aims to produce an AI model comparable to Anthropic’s Fable before the end of the first quarter next year, signaling its intention to compete directly with the world’s most advanced AI systems.

Even so, major challenges remain before GLM-5.2 can achieve widespread enterprise adoption outside China. Data security and geopolitical concerns continue to discourage many Western corporations, particularly banks, government agencies and cybersecurity firms, from incorporating Chinese AI models into critical systems.

Wei Sun, principal AI analyst at Counterpoint Research, said regulatory concerns remain a significant obstacle.

“I have seen some discussion among European companies about whether it could be used in enterprise settings,” Sun said.

“In the EU and U.S., some clients, partners and regulated industries may simply be unwilling to accept Chinese models in their AI stack, regardless of technical performance or price.”

Enterprise adoption also tends to move slowly because replacing AI infrastructure often requires months of testing, integration and regulatory review.

Nevertheless, some analysts argue that those concerns may be less significant than many assume. They note that companies can deploy open-weight Chinese models on their own servers or through U.S.-based cloud providers, limiting data exposure while benefiting from lower costs and greater flexibility.

Poe Zhao, founder of the Hello China Tech newsletter, said practical considerations often outweigh geopolitical ones among developers.

“Developers tend to care less about where a model comes from than whether it works, how much it costs and whether they can deploy or access it reliably,” Zhao said.

“The likely pattern is partial routing, not overnight replacement of OpenAI or Anthropic. So yes, it is a mini DeepSeek moment but in a narrower, developer-centric sense.”

Evidence suggests Chinese AI models have already been gaining international traction since DeepSeek disrupted the industry. A report published earlier this year by RAND found that Chinese large language models increased their global market share from 3% to 13% during the two months following DeepSeek’s R1 launch. The gains were particularly pronounced across developing economies and countries maintaining close economic and political ties with Beijing.

The release of DeepSeek’s low-cost reasoning model also triggered a global technology selloff by challenging the assumption that only companies spending hundreds of billions of dollars on AI infrastructure could compete at the frontier.

GLM-5.2 now appears to be extending that narrative. Rather than simply offering a low-cost alternative, Z.ai is demonstrating that Chinese AI developers are capable of producing models that approach the performance of leading American systems while remaining substantially cheaper to deploy. Although regulatory barriers and trust issues are likely to slow adoption among large Western enterprises, the model’s rapid acceptance among developers indicates that China’s AI ecosystem is becoming a more formidable competitor in the global race for artificial intelligence leadership.