Key Takeaways
A year after Meta's $14.3 billion bet on Alexandr Wang, the company finally has its first proprietary AI model — but remains behind OpenAI, Anthropic and Google in a race where being second carries real costs.
Meta's decision to install Wang as its AI chief in mid-2025 was the most expensive executive hire in the technology industry's history. The payout reflected the urgency Mark Zuckerberg felt as Meta scrambled to catch up in generative AI after years of prioritizing the metaverse. Twelve months later, the company has shipped its first in-house large language model, yet the gap with frontier labs has not closed meaningfully, according to benchmark comparisons reviewed by Edgen.
"Meta is playing catch-up in a game where the leaders are accelerating, not standing still," said Sarah Guo, founder of Conviction Capital and a former Greylock partner who has tracked AI investment cycles since 2022. "The question is whether $14.3 billion buys you a seat at the table or just a ticket to the race."
The internal turbulence has been severe. Zuckerberg acknowledged in a June memo that Meta made mistakes during its transition to an AI-focused workforce, which included a 10 percent global workforce reduction and the reassignment of 7,000 employees to AI-related workflows. The company's new Applied AI Engineering unit implemented a flat management structure with ratios of up to 50 individual contributors per manager, a configuration that former employees described as chaotic. High-profile departures from rival labs that Meta had recruited aggressively — including researchers from Google Brain and DeepMind — further eroded morale.
The Model Gap Remains Wide
Meta's first proprietary model, which the company has not yet named publicly, scores in the mid-80s on the MMLU benchmark, according to people familiar with the results. That compares with scores above 90 for OpenAI's GPT-5, Anthropic's Claude 4 and Google's Gemini 2.5 Ultra. On coding benchmarks such as HumanEval, the gap is wider still, with Meta's model trailing by roughly 10 percentage points.
The performance deficit is not purely technical. OpenAI and Anthropic each raised more than $10 billion in 2025 alone, much of it earmarked for compute infrastructure. Google's DeepMind division benefits from Alphabet's $50 billion-plus data center buildout plan through 2027. Meta, despite committing more than $40 billion in annual capital expenditures for AI infrastructure, has had to split its compute allocation between training frontier models and serving inference for its social media platforms, which process billions of daily requests across Facebook, Instagram and WhatsApp.
Training costs compound the challenge. A single frontier model run can require 25,000 Nvidia H100 GPUs operating continuously for 90 days, consuming electricity equivalent to the annual usage of thousands of households. Meta's data center capacity, while expanding rapidly, has not kept pace with demand from its research teams, leading to internal queueing disputes that delayed training cycles by weeks, according to three former employees.
Zuckerberg's Pivot and the Workforce Fallout
Zuckerberg's internal memo, dated June 12, acknowledged that the complexity of restructuring around AI "inevitably led to errors." He emphasized that Meta would prioritize internal transfers over further layoffs and would scale back the expansion of manager oversight responsibilities. The company also plans to increase investment in team-building initiatives, including a large-scale hackathon in July and larger budgets for corporate events.
The workforce reductions have been particularly painful for middle managers and software developers, who bore the brunt of the 8,000 job cuts announced in early 2026. Meta has created new roles in AI-related functions to absorb displaced staff, but the transition has been uneven. Some teams lost experienced engineers while gaining junior hires with AI specialization, creating a skills mismatch that slowed project timelines.
The broader context is that Meta's AI strategy carries unusually high stakes. The company's core advertising business generated more than $160 billion in revenue last year, and Zuckerberg has staked the company's next growth phase on AI-powered features — from automated ad creation to AI-generated content recommendations. If Meta's models cannot match the quality of competitors' offerings, the risk is not just technical irrelevance but erosion of the advertising revenue that funds the entire enterprise.
What Comes Next for Meta's AI Ambitions
Zuckerberg has said Meta does not anticipate further large-scale, enterprise-wide layoffs this year, though he cautioned that the rapidly changing nature of the technology industry makes guarantees difficult. The company's next major milestone is the July hackathon, which will showcase internal AI projects and could signal which product directions Meta prioritizes.
For investors, the calculus is straightforward. Meta shares trade at roughly 23 times forward earnings, a discount to Alphabet's 26 times and a premium to the broader S&P 500. The $14.3 billion commitment to Wang represents about 9 percent of Meta's annual free cash flow — a bet that has not yet produced a competitive model but has consumed resources that could have been deployed elsewhere. Microsoft's partnership with OpenAI and Google's internal development both offer alternative models for how to compete in AI, and neither required a single executive hire costing double-digit billions.
The next six months will determine whether Wang's tenure is remembered as a bold bet that paid off or a costly detour. Meta's first proprietary model is a proof of concept, not a product. The company needs to ship something that competes on benchmarks, on inference cost per token, and on developer adoption — and it needs to do it before the next generation of models from OpenAI and Google makes the current gap look like a chasm.
This article is for informational purposes only and does not constitute investment advice.
