• Home
  • Technology
  • Gaming
  • Entertainment
  • World & Business
  • Science
  • Sports
  • AI
HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
Technology
Reaction

Meta may buy a large allocation of AMD MI450 GPUs in 2026

A user argues that Meta buys AMD's Instinct GPUs to hedge against Nvidia's pricing and supply risks.

MikeMI
1 Source, 365d ago, first seen 365d ago

TLDR

In an October 2025 post, a user claims Meta is already a major buyer of AMD Instinct GPUs and expects a large MI450 allocation in 2026. The user frames AMD as a way for Meta to diversify beyond Nvidia, citing cost and supply concerns, AI inference needs and open-source software.

Combined views

69.1K

1 Source, first seen 365d ago

310 likes15 comments67 saves29 reposts

Combined views

69.1K

1 Source, first seen 365d ago

310 likes15 comments67 saves29 reposts

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

Featured Source

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

1 Source

Mike@MikeLongTerm$AMD 🤝$META is going to be MASSIVE 🔥🔥 @Meta is a major direct buyer of AMD's Instinct series GPUs (e.g., MI300X, MI350/MI355X) for its own AI infrastructure.and serves as a key partner in the broader AI ecosystem. Meta is also going to be a massive buyer of MI450(2026) from $AMD. So far, Meta or Mark Zuckerberg @finkd allocated 42% on $AMD GPUs and 58% on $NVDA GPUs. It wouldn't be wrong to say AMD entire AI Accelerator year of supply could not supply Meta demand alone. A big distinction between Meta and other Hyperscalers: Meta builds and operates its own massive data centers (e.g., a planned 2.2 GW facility in Louisiana), unlike pure cloud providers. However, Meta has explored cloud partnerships, including rumored $20B multi-year deals with Oracle for Nvidia GPUs (not AMD-specific). The "unlike other hyperscalers" angle highlights Meta's aggressive diversification into AMD as a Nvidia alternative, a strategy less pronounced among peers like Amazon (AWS) or Alphabet (Google), who prioritize custom ASICs or Nvidia exclusivity. Meta is going to be a big customer for $AMD "Secret weapon" as well. Making AI workload in mobile/smaller devices for future products. Meta has been one of AMD's largest customers since 2023, scaling rapidly amid its $600B+ AI infrastructure commitment through 2028. Key milestones: 1. 2023-2024 Purchases: Meta acquired ~173,000 AMD Instinct MI300X GPUs (per Omdia data), equivalent to ~600,000 Nvidia H100s in AI horsepower by end-2024. These were deployed for Llama model training/inference (e.g., Llama 3 on MI300X). 2. 2025 Ramp-Up: Meta is allocating 15-20% of its AI capex (~$90-120B through 2028) to AMD's MI350/MI355X series. Early adopters include Meta for inference workloads like AI stickers, image editing, and its Meta AI assistant. By Q3 2025, Meta confirmed testing MI355X for distributed inference, with volume deployments expected in H2 2025. 3. 2026 Pipeline & Beyond: Meta is evaluating MI400 (2026 launch) for rack-scale "Helios" systems (72-GPU clusters). Total AMD allocation: 300,000-500,000 units annually by 2026, per SemiAnalysis and other estimates. The Juicy Part: Why Meta chose AMD: Meta's decision stems from a "Nvidia hedge" philosophy, driven by cost, supply risks, and open ecosystems. Unlike peers locked into Nvidia (80-90% of their GPU spend), Meta treats hardware as a commodity for its open-source Llama models 1. Cost Efficiency (Tokens/$ Edge): ~AMD GPUs offer 2-4x better price-performance for inference (Meta's bread-and-butter, ~70% of workloads). MI300X delivers 35x faster LLM inference than MI250 at ~40% lower cost per token vs. Nvidia H100. ~Example: Meta's Llama 2 70B inference on MI300X achieves 1.5s latency with linear scaling—critical for real-time features like Instagram AI. Nvidia B200 wins on raw training FLOPs but costs 2-3x more ($60K/unit vs. AMD's $25-30K for MI355X). ~Meta's capex: $35-40B in 2025 alone; AMD saves 20-30% ($7-12B) vs. all-Nvidia. 2. Supply Chain Diversification: ~ $TSM capacity for AMD (10-15% of CoWoS) is less contested than Nvidia's 70%. And TSMC will have to allocate more for AMD in H2 2025 and 2026. ~ @elonmusk @xai confirmed this: AMD GPUs "work very well" for small-medium models, aligning with Meta's Llama focus. Meta avoids Nvidia's monopoly pricing (up 20-50% for locked-in buyers). 3. Open-Source Ecosystem Alignment: ~Meta co-developed PyTorch (open-source) and pushes ROCm (AMD's open CUDA alternative). ROCm 7 (Q3 2025) and ROCm 8 (2026) enable seamless porting—unlike Nvidia's proprietary CUDA lock-in. ~Partners: Meta collaborates with AMD on Llama 3 optimization; Oracle's OCI Supercluster (AMD-powered) could host Meta's overflow tests. 4. Inference-First Workloads(Best Inference): ~Meta's 3.2B daily users generate massive inference demand (e.g., Reels recommendations). AMD excels here: MI355X's 288GB HBM3E handles larger batches at lower power (1,400W TDP vs. Nvidia's 1,000W+ but higher $/FLOP). ~Training (20-30% of spend) stays Nvidia-heavy, but inference shifts to AMD for ROI. Meta's AMD bet is a calculated play for cost-leadership and resilience in an under-invested $TSM world—more aggressive than peers due to its consumer-scale inference needs and open-source ethos. META will be AMD largest customer for a long time, especially a large allocation of MI450 is expected for @Meta in 2026.365d
    • Home
    • Technology
    • Gaming
    • Entertainment
    • World & Business
    • Science
    • Sports
    • AI

    1 Source

    Mike@MikeLongTerm$AMD 🤝$META is going to be MASSIVE 🔥🔥 @Meta is a major direct buyer of AMD's Instinct series GPUs (e.g., MI300X, MI350/MI355X) for its own AI infrastructure.and serves as a key partner in the broader AI ecosystem. Meta is also going to be a massive buyer of MI450(2026) from $AMD. So far, Meta or Mark Zuckerberg @finkd allocated 42% on $AMD GPUs and 58% on $NVDA GPUs. It wouldn't be wrong to say AMD entire AI Accelerator year of supply could not supply Meta demand alone. A big distinction between Meta and other Hyperscalers: Meta builds and operates its own massive data centers (e.g., a planned 2.2 GW facility in Louisiana), unlike pure cloud providers. However, Meta has explored cloud partnerships, including rumored $20B multi-year deals with Oracle for Nvidia GPUs (not AMD-specific). The "unlike other hyperscalers" angle highlights Meta's aggressive diversification into AMD as a Nvidia alternative, a strategy less pronounced among peers like Amazon (AWS) or Alphabet (Google), who prioritize custom ASICs or Nvidia exclusivity. Meta is going to be a big customer for $AMD "Secret weapon" as well. Making AI workload in mobile/smaller devices for future products. Meta has been one of AMD's largest customers since 2023, scaling rapidly amid its $600B+ AI infrastructure commitment through 2028. Key milestones: 1. 2023-2024 Purchases: Meta acquired ~173,000 AMD Instinct MI300X GPUs (per Omdia data), equivalent to ~600,000 Nvidia H100s in AI horsepower by end-2024. These were deployed for Llama model training/inference (e.g., Llama 3 on MI300X). 2. 2025 Ramp-Up: Meta is allocating 15-20% of its AI capex (~$90-120B through 2028) to AMD's MI350/MI355X series. Early adopters include Meta for inference workloads like AI stickers, image editing, and its Meta AI assistant. By Q3 2025, Meta confirmed testing MI355X for distributed inference, with volume deployments expected in H2 2025. 3. 2026 Pipeline & Beyond: Meta is evaluating MI400 (2026 launch) for rack-scale "Helios" systems (72-GPU clusters). Total AMD allocation: 300,000-500,000 units annually by 2026, per SemiAnalysis and other estimates. The Juicy Part: Why Meta chose AMD: Meta's decision stems from a "Nvidia hedge" philosophy, driven by cost, supply risks, and open ecosystems. Unlike peers locked into Nvidia (80-90% of their GPU spend), Meta treats hardware as a commodity for its open-source Llama models 1. Cost Efficiency (Tokens/$ Edge): ~AMD GPUs offer 2-4x better price-performance for inference (Meta's bread-and-butter, ~70% of workloads). MI300X delivers 35x faster LLM inference than MI250 at ~40% lower cost per token vs. Nvidia H100. ~Example: Meta's Llama 2 70B inference on MI300X achieves 1.5s latency with linear scaling—critical for real-time features like Instagram AI. Nvidia B200 wins on raw training FLOPs but costs 2-3x more ($60K/unit vs. AMD's $25-30K for MI355X). ~Meta's capex: $35-40B in 2025 alone; AMD saves 20-30% ($7-12B) vs. all-Nvidia. 2. Supply Chain Diversification: ~ $TSM capacity for AMD (10-15% of CoWoS) is less contested than Nvidia's 70%. And TSMC will have to allocate more for AMD in H2 2025 and 2026. ~ @elonmusk @xai confirmed this: AMD GPUs "work very well" for small-medium models, aligning with Meta's Llama focus. Meta avoids Nvidia's monopoly pricing (up 20-50% for locked-in buyers). 3. Open-Source Ecosystem Alignment: ~Meta co-developed PyTorch (open-source) and pushes ROCm (AMD's open CUDA alternative). ROCm 7 (Q3 2025) and ROCm 8 (2026) enable seamless porting—unlike Nvidia's proprietary CUDA lock-in. ~Partners: Meta collaborates with AMD on Llama 3 optimization; Oracle's OCI Supercluster (AMD-powered) could host Meta's overflow tests. 4. Inference-First Workloads(Best Inference): ~Meta's 3.2B daily users generate massive inference demand (e.g., Reels recommendations). AMD excels here: MI355X's 288GB HBM3E handles larger batches at lower power (1,400W TDP vs. Nvidia's 1,000W+ but higher $/FLOP). ~Training (20-30% of spend) stays Nvidia-heavy, but inference shifts to AMD for ROI. Meta's AMD bet is a calculated play for cost-leadership and resilience in an under-invested $TSM world—more aggressive than peers due to its consumer-scale inference needs and open-source ethos. META will be AMD largest customer for a long time, especially a large allocation of MI450 is expected for @Meta in 2026.365d
    Today's Rank

    —

    Not ranked yet

    Today's Rank

    —

    Not ranked yet