Meta's $100B AMD Bet Is Not a Hedge. It's a Capital Allocation Decision.

The Sequence Is the Signal

On February 17, 2026, Meta signed an expanded deal with NVIDIA for millions of chips. Seven days later, it signed a separate multiyear agreement with AMD worth up to $100 billion. That one-week gap is deliberate negotiating architecture.

Meta plans to purchase potentially up to $100 billion worth of AMD chips, enough to drive roughly six gigawatts of data center power demand. To contextualize: the agreement covers processors capable of supporting about six gigawatts of data center power demand, equivalent to the electricity required by approximately five million U.S. households for one year.

This is not hedging. Hedging looks like a pilot cluster and a vendor evaluation. A hundred billion dollars and a near-10% equity warrant is capital allocation. Meta is betting AMD can deliver at hyperscale, and structuring the relationship to profit if AMD succeeds.

What the Deal Actually Covers

Under the agreement, Meta will purchase AMD's MI540 series of GPUs and its latest generation of CPUs. Deliveries begin in the second half of this year with one gigawatt of AMD's forthcoming MI450 hardware. AMD will develop a customized version of its MI450 AI chips for Meta, primarily for inference workloads.

The warrant structure is the telling detail. AMD has issued Meta a performance-based warrant for up to 160 million shares of AMD common stock, or about 10% of the company, for $0.01 each, structured to vest alongside certain milestones. The full stock award is conditional on AMD's share price hitting $600—more than three times the closing price of $196.60 the day before the announcement.

That spread is no accident. The warrant is an alignment mechanism. Meta gets meaningful AMD equity only if AMD's stock more than triples, which implies successful execution at scale across multiple hyperscalers. Meta has made it financially rational to want AMD to win broadly.

AMD and Meta announced a six-gigawatt agreement to power Meta's next generation of AI infrastructure across multiple generations of AMD Instinct GPUs. This agreement expands on the companies' existing strategic partnership and aligns roadmaps across silicon, systems, and software to deliver AI platforms purpose-built for Meta's workloads.

Aligning silicon, systems, and software roadmaps across multiple GPU generations means Meta engineers sit in AMD's planning cycles. That integration took Google years to achieve with its TPU program.

Why Now, and Why AMD

Meta has been a close partner over multiple generations, deploying millions of AMD EPYC CPUs and significant deployments of AMD Instinct MI300 and MI350 series GPUs across their global infrastructure. This MI540 commitment builds on engineering depth that already exists.

The MI400 series has substance. The flagship MI455X packs 320 billion transistors across 12 TSMC N2 compute chiplets and three advanced 3nm chiplets, delivering up to 40 PFLOPS of FP4 performance. With 432 GB of HBM4 memory and 19.6 TB/s of memory bandwidth, it is designed for hyperscale data centers running the largest language models and multimodal AI workloads.

Discussions have also taken place between Meta and Google about using its tensor processing units for AI workloads. Meta is simultaneously courting Google's TPUs, running NVIDIA hardware, committing billions to AMD, and building in-house silicon. That is not diversification out of fear. That is procurement leverage—a credible threat against any single vendor's lock-in.

What This Does to NVIDIA

NVIDIA's position has rested on two pillars: CUDA's software ecosystem and hardware scarcity. The software moat holds and will for years. The scarcity moat is now directly tested.

The AMD partnership comes a couple of weeks after Meta struck a multiyear deal to expand its data centers with millions of NVIDIA's latest CPUs and GPUs. NVIDIA is not losing Meta as a customer. It is losing Meta's unconditional dependency. The distinction matters because the second condition is what drives pricing power and allocation priority.

If AMD delivers consistent capacity over 24 to 36 months, NVIDIA faces a structurally different negotiating dynamic with every hyperscaler. The proof point shifts from theoretical to operational. Microsoft, Google, and Amazon now have a printed deal to bring to their next NVIDIA negotiations.

Meta has pledged to invest at least $600 billion in U.S. data centers and AI infrastructure over the next several years, including a projected capital expenditure spend of $135 billion in 2026. A company spending that much in infrastructure has procurement volume sufficient to move markets. AMD's challenge is delivering yield and schedule at that volume.

The CPU Layer Is Not a Footnote

The deal is usually framed as GPUs plus CPUs, and that distinction matters. CPUs are increasingly becoming a core pillar of the AI inference compute stack because they are efficient, easier to scale, and do not tie companies solely to NVIDIA.

AMD CEO Lisa Su indicated that demand for CPUs is surging alongside the expansion of AI infrastructure and agentic systems. Inference workloads—where most production AI compute hours actually run—favor CPUs over training workloads. The AMD EPYC expansion inside Meta creates a second revenue stream structurally separate from the GPU competition with NVIDIA.

What to Watch, In Sequence

H2 2026: Whether AMD hits its first-gigawatt delivery milestone on schedule. Yield and logistics at this scale have never been tested at this pace.

Q3/Q4 2026 NVIDIA earnings: Watch gross margin guidance. If NVIDIA bundles software, expands support, or adjusts pricing to retain hyperscaler allocation priority, margin commentary will signal it before revenue does.

6 to 12 months out: Whether Microsoft, Google, or Amazon announce comparable AMD commitments. Meta's deal proves feasibility. Similar announcements from other hyperscalers would confirm structural shift rather than Meta as an outlier.

12 to 18 months out: Whether Meta's ROCm-based workload porting generalizes successfully. Meta's ML engineering team ranks among the world's most capable. Friction-free training workload migration would de-risk software portability for every other operator. Friction signals quiet retrenchment.

2027 onward: AMD's MI500 series will determine whether this relationship extends beyond the current contract. The MI540 deal establishes AMD as tier-one. The MI500 determines whether that status persists.