Feb 25, 2026 · 24-Hour AI Briefing: Meta’s $100B AMD Bet, and the NAND Cycle Fight Over “AI Storage” Valuations
Two stories in the last 24 hours pulled attention back to hardware reality. One is Meta reportedly locking in long-term compute supply with AMD, pushing into custom silicon and even potential equity alignment. The other is Citron shorting SanDisk on the view that the NAND cycle is near a peak, while Samsung intensifies competition in high-end SSDs. One debate is about compute certainty; the other is about whether storage can earn a structural premium in an AI world.
1. Meta reportedly signs a five-year, $100B strategic deal with AMD, targeting 6GW of compute and deep MI450 customization, potentially ending up with around 10% ownership via warrants
Commentary:
Reportedly, Meta has historically leaned heavily on NVIDIA, but as AI demand scales exponentially and supply-chain concentration risk rises, a single-vendor posture becomes increasingly fragile. The “6GW” framing signals long-horizon locking of training and inference capacity, turning compute procurement from “buying GPUs” into “securing power-plant-scale throughput.”
If MI450 is truly optimized for a single customer, the real story isn’t just specs. It’s Meta encoding its requirements for low power, high throughput, and specific workloads into a tighter hardware-software loop. The objective is less about peak benchmarks and more about $/token, supply stability, and a reproducible systems playbook.
The most consequential element is the procurement-for-warrants structure. If the reported 160M-share warrant package is tied to shipment volume, price, and technical milestones, and if it ultimately implies roughly 10% ownership, that would elevate the relationship into risk-sharing and incentive alignment. For AMD, it’s durable demand plus ecosystem pull-through. For Meta, it’s leverage over roadmap and allocation priority, plus partial hedging against future pricing pressure.
NVIDIA’s moat remains CUDA and its entrenched developer ecosystem. If Meta, as a top-tier practitioner, can migrate meaningful production workloads to ROCm with strong reliability and economics, it would materially boost AMD platform credibility and encourage other hyperscale buyers to re-evaluate single-ecosystem lock-in. Who will be the next Meta-style mover?
2. Citron Research shorts SanDisk: NAND cycle near a peak, “not NVIDIA,” and Samsung ramps high-end SSD competition
Commentary:
Storage is a deeply cyclical business, and valuation ceilings tend to depend on whether a company can earn a premium beyond the cycle. Citron’s thesis is straightforward: as NAND approaches a peak, pricing, margins, and sentiment can roll over quickly. Without a platform-style moat like NVIDIA’s, SanDisk is easier to price as a cycle-driven name. If Samsung pushes harder in premium SSDs, competition shifts from capacity to execution across product tiers and customer qualification.
But the “no moat” claim may be overly one-dimensional. SanDisk’s shared Yokkaichi NAND manufacturing footprint with Kioxia can support cost-competitive access to high-quality 3D NAND, and enterprise QLC SSD capabilities align well with the capacity-per-dollar needs of genAI inference. If it’s already qualified across major hyperscaler lists and holds meaningful OEM channel positions, the market’s re-rating case becomes less about the NAND cycle and more about “AI storage infrastructure.”
The key question is durability: do these advantages hold through the downcycle with stable orders and resilient margins, or do they fade with pricing pressure? That’s the line between “infrastructure premium” and “classic cyclicals.” Do you agree with Citron’s short here?