Meta: Can Afford Trillion-Dollar Computing Power, But Can't Keep Key People
Meta's AI Ambition: A $135 Billion Bet on Chips, But Losing Key Talent
In July 2025, Meta recruited top AI infrastructure engineer Ruoming Pang from Apple with a compensation package worth over $200 million. However, just seven months later, he left for OpenAI, forfeiting much of his unvested equity. This high-profile departure is part of a broader trend of key talent leaving Meta's AI division, including Chief AI Scientist Yann LeCun and other senior figures.
The exodus is largely attributed to the fallout from the Llama 4 model's release in April 2025. The model was later revealed to have been benchmarked unethically, using different model versions to optimize scores on different tests, severely damaging trust within the developer community. This scandal led CEO Mark Zuckerberg to lose confidence in the team, resulting in a major reorganization. He appointed 28-year-old Scale AI CEO Alexandr Wang as Chief AI Officer, who now oversees the new Meta Superintelligent Lab (MSL). The planned flagship model, Llama 4 Behemoth, was indefinitely delayed.
Compounding these software issues, Meta also canceled its most advanced in-house AI training chip project, a critical part of its plan to reduce reliance on Nvidia. This failure has triggered a panic-buying spree. In February 2026, Meta announced a capital expenditure budget of $115-$135 billion, nearly double the previous year's. Within ten days, it signed massive, multi-year chip deals: a multi-billion dollar agreement with Nvidia for Blackwell and Vera Rubin GPUs, a $60-$100 billion deal with AMD for MI450 GPUs (which included warrants for a 10% stake), and a multi-billion dollar deal to rent Google's TPU chips.
This strategy of acquiring immense, diverse hardware from three different architectures (CUDA, ROCm, XLA/JAX) creates immense engineering complexity. Ironically, Meta is spending hundreds of billions on the world's most complex hardware while losing the rare engineers, like Pang, capable of building the cross-platform frameworks needed to make it all work.
Zuckerberg's gamble mirrors his all-in bet on the metaverse: see a trend, spend heavily, reorganize frequently. The difference is that AI is a more tangible opportunity, and Meta's core advertising business generates massive cash flow to fund it. However, the article concludes that money can buy chips and算力, but it cannot guarantee the retention of top talent or the development of a winning model. If Meta's next model, codenamed "Avocado," fails to compete with GPT-5 and Gemini 3 Ultra, its massive expenditure will have only built expensive data centers full of underutilized hardware. The AI race is won by those who can build transformative models, not just those who can write the biggest checks.
marsbit02/28 09:41
