Inside Meta Platforms’ AI teams, engineers have adopted a phrase that now circulates in group chats and all-hands meetings: they describe the internal environment as a “soul-crushing gulag,” according to TechCrunch, which reported the characterization this week. The phrase is extreme, but it captures a mood that has spread through the company’s model-building efforts in recent months, as deadlines slip and the quality gap with rivals refuses to close.
The frustration comes at a delicate moment. One year ago, Chief Executive Mark Zuckerberg brought in Alexandr Wang, the founder of data-labeling company Scale AI, to take charge of a new model effort. The mandate was simple: build something that could compete with the best models from OpenAI and Google, using the enormous compute budget Meta has assembled. That project has now reached the stage where it is supposed to be sold to customers, according to CNBC, and the reception has been lukewarm.
The core problem is not compute, which Meta has in abundance, but data. The company has spent heavily on the graphics processors needed to train large models, but the text that powers them has become the industry’s scarcest resource. Most of the high-quality writing on the open web has already been absorbed into the training runs of the leading labs, and publishers have grown aggressive about licensing their archives, people familiar with the industry said.
Zuckerberg acknowledged the problem in an interview this week, saying Meta had made mistakes in its AI data strategy. He did not specify which decisions he meant, but people close to the company point to two: an early reliance on public web data at a moment when competitors were locking up exclusive archives, and a slow start on the partnerships with publishers and content platforms that now underpin rival models.
The stakes are large. Meta has committed roughly $14.3 billion to its AI push this year, a budget that covers data centers, chips and the people who build the models. That spending was supposed to produce a model that could anchor the company’s advertising tools and its consumer apps. Instead, it has produced a series of releases that reviewers have judged competent but not leading, according to third-party evaluations cited by analysts.
Money, it turns out, does not buy everything. Scale AI’s role was supposed to fix the data problem with human-annotated examples, and Wang’s team has built pipelines that Meta lacks. But the fundamental constraint remains: the best models are trained on data that competitors already control, from Google’s access to YouTube transcripts and books to OpenAI’s licensing deals with news organizations.
The internal mood has not helped. Engineers have described long hours, shifting priorities and a management structure that rewards visibility over results, according to people who have left the teams. The “gulag” phrase, however crude, reflects a workplace where attrition has picked up and recruiting has become harder, headhunters said.
Meta’s options are narrowing. The company can license more content, as it has begun to do with news publishers, or it can lean on synthetic data, generated by other models, a technique that has produced mixed results. Both approaches cost money and neither guarantees the quality jump Zuckerberg wants, analysts said.
The broader question is whether Meta’s spending will eventually close the gap or simply keep it in place. The company has shown it can catch up in features, as it did with short-form video, but AI models do not reward fast followers the way social features do. A model trained on inferior data does not catch up by spending more; it needs something its rivals do not have.
The data question has already reshaped the industry’s economics. Licensing deals that were once minor line items have become strategic assets: OpenAI has paid hundreds of millions of dollars for archives, and Google’s control over YouTube transcripts and scanned books gives its models a corpus no rival can replicate. Publishers, burned by years of scraping disputes, now negotiate as gatekeepers, and the prices they command keep rising, according to people who broker such deals.
Meta’s own assets are not worthless. The company’s apps generate an ocean of user content, and its researchers have argued that behavioral signals from its social graph can substitute for traditional training text. But converting engagement data into a competitive model has proved harder than the theory suggests, and the company’s public benchmarks still trail those of its rivals, analysts said.
The financial stakes are visible in Meta’s stock. The company’s AI spending has weighed on margins, and investors have shown they will punish the shares if the outlays do not produce a payoff. A strong model would justify the budget; a mediocre one would revive the questions that followed the company’s earlier metaverse spending, which consumed tens of billions of dollars before being wound down.
For now, Meta is betting that its distribution advantage matters more than model quality. The company can put its AI in front of three billion users through its apps, a reach no lab can match, and it plans to use that reach to sell the model to businesses. Whether customers will pay for a model that trails the frontier is the question Zuckerberg has not yet answered.
