Meta to Spend Billions on AMD Gear, Buy Stock

Key Concepts

• Meta Compute: Meta’s ambitious initiative to acquire massive computing infrastructure (hundreds of gigawatts) to fuel AI development, aiming for superintelligence.
• AMD-Meta Deal: A multi-year agreement for AMD to supply six gigawatts worth of accelerators and other gear to Meta, including share warrants potentially making Meta a significant AMD shareholder.
• Inference vs. Training: Distinguishing between the application of AI chips for running existing models (inference) and developing new ones (training). Meta is diversifying its chip sourcing based on these different workloads.
• Hyperscaler: A company with massive data center infrastructure, like Meta, operating at a very large scale.
• Front-Loading Capacity: Meta’s strategy of aggressively acquiring computing resources now, even if it means potential overspending, to secure availability.
• Share Warrants: Financial instruments giving Meta the right to purchase AMD shares at a predetermined price, incentivizing a long-term partnership.

AMD and Meta’s Landmark Deal: A Deep Dive

The discussion centers around a significant deal between AMD and Meta, involving the supply of up to six gigawatts of computing power over time, primarily utilizing AMD accelerators, alongside other hardware. This agreement is viewed as a major endorsement for AMD, particularly as it competes with NVIDIA, which recently secured a similar deal with Meta.

The Scale of Meta’s Compute Demand

Riley highlights the “insatiable demand” from Meta, driven by Mark Zuckerberg’s “metacompute” vision – a pursuit of superintelligence through AI. Meta plans to invest approximately $135 billion in CapEx this year, with spending expected to continue at a high rate. This demand is so substantial that Meta is diversifying its chip sourcing, utilizing AMD, NVIDIA, and its own internally developed custom chips for different AI applications and workloads. Specifically, Meta is targeting projects requiring up to five gigawatts of power, contingent on regulatory approvals and energy availability.

The AMD-Meta Agreement: Structure and Incentives

The deal includes share warrants, mirroring a previous arrangement between AMD and OpenAI. These warrants allow Meta to potentially become a substantial shareholder in AMD based on operational and financial milestones. Ian explains that the warrants are structured such that Meta can acquire shares at a price significantly above the current market value (around $600 versus a recent closing price of under $200), demonstrating confidence in AMD’s future performance. Lisa Su, AMD’s CEO, emphasized that this arrangement “locks us together and tightens our relationship.” AMD has previously used similar warrant structures to solidify partnerships.

Inference Focus and Diversification

The discussion emphasizes that AMD’s role within this deal is particularly focused on inference – the process of running AI models after they’ve been trained. This contrasts with NVIDIA’s strength in training AI models. Meta’s strategy involves utilizing all three sources – AMD, NVIDIA, and its own custom chips – to support diverse applications and workloads. This diversification is a key element of Meta’s approach to managing the massive scale of its compute needs.

NVIDIA’s Position and Market Dynamics

The conversation addresses the implications for NVIDIA, whose earnings were upcoming at the time of the discussion. While Meta’s diversification is noted, the consensus is that NVIDIA remains dominant, and the overall market is expanding rapidly enough to benefit all major players. The key indicator of potential change would be if NVIDIA fails to meet or exceed Wall Street’s expectations. The discussion acknowledges the interconnectedness of chips and networking infrastructure, stating they are “now one and the same.”

Financial Context and Growth Rates

A reality check is provided regarding the relative scale of AMD and NVIDIA’s data center businesses. NVIDIA’s data center revenue is currently over $50 billion per quarter (including networking), while AMD’s total revenue across all segments is around $10 billion per quarter. The focus shifts to growth percentages as the primary indicator of AMD’s progress in catching up. Both companies currently attribute their success to the overwhelming demand, with Meta purchasing from both. Market share shifts are expected to become more apparent when market growth slows.

Meta’s Capital Expenditure Strategy

Meta’s approach to capital expenditure is described as “front-loading capacity.” Mark Zuckerberg’s strategy involves aggressively acquiring computing resources while availability is high, even if it means potentially overspending by as much as $200 billion. This is justified by the belief that securing capacity now is crucial for both AI development and maintaining its core social media business, which still generates 98% of its revenue.

AMD’s Share Structure and Validation

The discussion addresses why AMD continues to offer share warrants in its deals. Lisa Su explained that, unlike the OpenAI deal where liquidity was a factor, Meta doesn’t require capital. The warrants serve as a strong signal of Meta’s commitment to AMD and its technology. AMD, being relatively new to the high-performance accelerator market, views this as significant validation beyond a simple purchase agreement.

Conclusion

The AMD-Meta deal represents a substantial step forward for AMD in the competitive AI infrastructure market. While NVIDIA currently dominates, Meta’s diversification strategy and AMD’s focus on inference workloads create opportunities for growth. The deal’s structure, including share warrants, signifies a long-term partnership and a strong vote of confidence in AMD’s future. The overarching theme is one of massive demand for compute power, driven by the pursuit of AI and superintelligence, creating a dynamic landscape where multiple players can thrive.