✨HellasQCI at the Q-PrEP (Quantum Preparedness) Workshop ⏰ 5 Feb 2026 | 📍 Athens (NCSR “Demokritos”) Experts collaborate to address post-quantum security challenges & advance quantum-safe communications. 🔗Details: shorturl.at/TeglV 📝Registrer: shorturl.at/2OnWP

$AMD 40-50% Growth & $META CapEx Strategy 🧵 @Meta is aggressively scaling its AI infrastructure to power Llama models, recommendation systems, and consumer-facing features like AI-generated content and assistants. CapEx refers to the cash Meta spends on long-term assets, with AI-focused CapEx dominating covering data centers, servers, GPUs, CPUs, networking, and power/cooling systems. After Q3 ER, Meta's full-year 2025 CapEx guidance is $70–72 billion (up from an initial $60–65B in January and $66–72B in July), a ~81% YoY increase from 2024's ~$39B. This funds ~1 GW of new compute capacity, bringing total GPUs to >1.3 million by year-end. How to understand Meta CapEx Allocation: ~Hardware (GPUs/CPUs/Servers: 60–70%, ~$42–50B): Dominates for AI accelerators and compute nodes. ~Data Centers/Facilities (20–25%, ~$14–18B): Includes a planned 2.2 GW Louisiana site (equivalent to two nuclear plants) and global expansions. ~Networking/Power/Cooling (10–15%, ~$7–11B): Rack-scale designs like Open Rack Wide (ORW) for dense AI clusters. ~Software/Other (5%, ~$3–4B): ROCm optimizations and in-house MTIA chips (but <10% of spend). Meta is AMD's largest hyperscaler customer, hedging NVIDIA (58–70% of GPU spend) with AMD for diversification, cost (20–30% savings), and open ecosystems (ROCm vs. CUDA). This includes MI-series GPUs for inference/training and EPYC CPUs for servers. AMD Instinct MI Series (GPUs: ~$7–10B, 70–75% of AMD allocation) ~Current (MI300X/MI325X): Meta deployed ~250K MI300X units (equiv. 600K NVIDIA H100s in perf), powering all Llama 405B inference. Handles 70% of workloads ( AI stickers, image editing). ~(MI350/MI355X, H2 2025): 250–300K units allocated (~$6–8B at $25–30K/unit vol. discount). Excels in inference (35x gen-over-gen uplift, 288GB HBM3E). Meta co-optimized for ORW racks; testing shows 1.5x faster Llama inference vs. NVIDIA H200 at 40% lower cost/token ~Future (MI450, 2026): Lead partner for Helios racks (72-GPU clusters, 6.4 EFLOPS). Expected 800K units (~$24B at $30–35K/unit), part of 1–2 GW ramp. Zuckerberg hinted at "massive" 2026 buys post-OpenAI's 6GW AMD deal. ~Why MI series? 2–4x better price/perf for inference; ROCm maturity (80% parity with CUDA) and a reliable supplier. Total 2025–2028: 800k–1m units/year. ~Meta runs >1.5M EPYC units globally (5th Gen "Turin" for Grand Teton platform). Powers deep learning recs and memory-intensive tasks. ~Expecting massive ramp up on 2026 EPYC "Venice" due to chiplet breakthrough design. But things changed after Creative Financing with Blue Owl Meta's headline $70–72B 2025 CapEx (Q3 guidance, up ~81% YoY) captures on-balance-sheet spend, but total effective AI infrastructure investment balloons to $90–100B when factoring in these creative vehicles. The Blue Owl JV alone adds ~$27B in "ghost CapEx"—funded externally but enabling Meta's 1 GW compute ramp (total GPUs >1.3M by YE). This hybrid model (PE bonds leasebacks) lets Meta retain operational control while Blue Owl's funds ( anchored by PIMCO's $18B debt tranche) shoulder 80% of the risk/ownership. ~Structure: Meta contributes 20% equity ($5.4B), gets a $3B upfront payout, and leases the asset back on 4-year terms (not long-term liabilities under GAAP). Debt stays in the SPV, not Meta's books. ~Scale Acceleration: Enables 2–3x faster builds Hyperion (Richland Parish, LA) will house 500K GPUs by 2027, equivalent to two nuclear plants' output. Without this, Meta's grid/power constraints (permitting delays) would cap growth at 500–700MW. Basically due to this creative financing structure, $META can spend 1.5-2x more on $AMD CPU/GPU in 2026-2030. AMD stands at the precipice of a decade-defining growth explosion, poised to deliver 40–50% compound annual total revenue growth from 2026 through 2030, catapulting its top line from ~$34 billion in 2025 to $180–220 billion by 2030, a 5–6x expansion. This isn’t speculative hype; it’s grounded in Meta’s revolutionary $27 billion Blue Owl JV, which acts as a force multiplier for AMD’s Data Center dominance, unlocking unprecedented scale in AI accelerators (MI series) and server CPUs (EPYC) without the traditional constraints of CapEx, power, or grid bottlenecks. Meta’s off-balance-sheet financing model structured via SPVs, leasebacks, and private credit at 6–7% yields removes the ceiling on compute deployment. Hyperion, the 2.2 GW AI supercluster in Louisiana, will host over 1 million AMD GPUs and 2 million EPYC cores by 2027-2028, with MI450X-powered Helios racks delivering 6.4 ExaFLOPS per cluster at 30% lower power density than NVIDIA equivalents. $META can easily scale 5GW with AMD CPU/GPU into 2030 with higher AI CapEx allocation in the coming years. This isn’t just efficiency, it’s strategic lock-in: Meta’s open-architecture ORW standard and ROCm co-optimization make AMD the default inference engine for Llama 4/5, Reels AI, and next-gen agents.

💡How can #HPC help deliver safer, personalised pediatric dosimetry? #EuroCCGreece, #GRNET & #Bioemtech share the answer in a #FFplus interview with P. Papadimitroulas. ▶️Watch👉 shorturl.at/5YGvz 📘Read👉 shorturl.at/sKkLf 🔬Success Story👉 shorturl.at/yVz0P