Meta’s $100 Billion AMD Bet

On February 24, 2026, Meta announced a partnership with AMD to deploy 6 gigawatts of AI computing infrastructure over five years—a deal worth approximately $100 billion and structured to potentially make Meta one of AMD’s largest shareholders. Mark Zuckerberg framed it as essential to achieving “personal superintelligence.” Lisa Su positioned it as placing AMD “at the center of the global AI buildout.” But the real significance isn’t what either CEO said. It’s what Meta did two weeks earlier: signed a multiyear expansion with Nvidia for millions of Blackwell GPUs. Meta isn’t replacing Nvidia with AMD. It’s building redundancy into a supply chain where monopoly dependency has become an existential risk.

ZeitShift Intelligence | February 26, 2026

---

THE DEAL STRUCTURE: $100 BILLION, 10% OWNERSHIP, FIVE-YEAR COMMITMENT

The Meta-AMD agreement deploys 6 gigawatts of AI computing capacity through purchases of AMD’s MI450 series GPUs (customized for Meta’s workloads) and 6th generation EPYC CPUs over five years. The first gigawatt deployment begins in the second half of 2026, built on AMD’s Helios rack-scale architecture—a system AMD and Meta developed jointly through the Open Compute Project.

AMD values the transaction at “double-digit billions per gigawatt,” according to CEO Lisa Su, who declined to specify exact figures. Six gigawatts at $15-20 billion per gigawatt yields $90-120 billion. TechCrunch and multiple sources peg the total at approximately $100 billion.

The equity component is unusual. AMD issued Meta a performance-based warrant for up to 160 million shares of AMD common stock at $0.01 per share—roughly 10% of the company. The warrants vest in tranches tied to AMD meeting delivery milestones and share price targets. The final tranche requires AMD’s stock to hit $600 (it closed at $196.60 on February 24). This structure aligns incentives: Meta gets ownership upside if AMD successfully delivers at scale, and AMD gets guaranteed demand from one of the world’s largest AI infrastructure buyers.

The timeline is aggressive. Shipments supporting the first gigawatt begin late 2026. Full 6-gigawatt deployment completes over five years, with infrastructure spread across Meta’s expanding data center network—including the recently announced $10 billion, 1-gigawatt gas-powered campus in Indiana.

Meta’s 2026 capital expenditure for AI infrastructure is projected at $135 billion, up from $72 billion in 2025. The company committed to spending over $600 billion in U.S. data centers and AI infrastructure through 2028. The AMD deal represents approximately 17% of that total commitment—significant but not exclusive.

---

THE NVIDIA CONTEXT: WHY META SIGNED TWO MASSIVE DEALS IN TWO WEEKS

Two weeks before announcing the AMD partnership, Meta expanded its existing Nvidia relationship. The company committed to deploying millions of Nvidia Blackwell GPUs using NVL72 rack-scale systems, Arm-powered Grace server CPUs, and Spectrum-X networking switches with co-packaged optics.

The back-to-back announcements aren’t contradictory—they’re deliberate diversification. Meta is implementing what it calls a “platform-agnostic approach” to AI compute, building resilience through multi-vendor infrastructure. The strategy acknowledges several realities:

Supply risk: Nvidia’s Blackwell chips face production constraints and allocation pressures. Demand from OpenAI, Microsoft, Google, Amazon, and other hyperscalers exceeds available supply. Relying exclusively on Nvidia means accepting delivery timelines Nvidia controls and allocation priorities that might not align with Meta’s needs.

Pricing power: A single-vendor dependency gives that vendor extraordinary leverage in pricing negotiations. Nvidia can charge premium rates when customers have no alternatives. AMD’s entry creates competitive pressure—not just on pricing but on delivery commitments, customization willingness, and long-term roadmap alignment.

Technology risk: If Nvidia’s next-generation architecture faces delays (as Blackwell initially did), customers with no alternatives suffer those delays directly. Diversification means Meta can accelerate deployments using whichever vendor ships first, maintaining infrastructure buildout momentum regardless of individual product cycles.

Architectural flexibility: Different AI workloads benefit from different hardware characteristics. Training frontier models demands maximum raw compute—where Nvidia’s H200 and Blackwell excel. Inference (running deployed models at scale) prioritizes efficiency and throughput—where AMD’s customized MI450 architecture, optimized specifically for Meta’s workloads, might offer advantages.

The simultaneous commitments reveal Meta’s calculation: in a market where AI chips are the critical constraint on capability development, vendor lock-in is strategically unacceptable. Paying a premium for redundancy is cheaper than risking infrastructure bottlenecks that delay product launches or constrain capacity expansion.

---

PERSONAL SUPERINTELLIGENCE: WHAT META IS BUILDING

Mark Zuckerberg positions the AMD partnership as essential to achieving “personal superintelligence”—AI systems designed to deeply understand and empower individuals in their everyday lives. The phrasing signals a strategic pivot from Meta’s earlier AI positioning.

In early 2025, Meta’s AI labs competed directly with frontier model development—attempting to match or exceed GPT-4, Claude, and other cutting-edge systems. By late 2025, that positioning shifted. Meta spun up its Meta Compute organization, centralizing ownership of the entire technology stack from silicon through systems to software. The reorganization signaled focus: not competing with OpenAI and Anthropic on general-purpose frontier models, but building AI infrastructure optimized for Meta’s specific use cases at Meta’s scale.

“Personal superintelligence” in Meta’s framing means:

The computational requirements are extraordinary. Frontier model training is measured in petaflops or exaflops of compute. Deploying inference for billions of users simultaneously requires infrastructure at scales even training doesn’t approach. Meta needs compute capacity not for occasional large training runs but for continuous, real-time inference at planetary scale.

The 6 gigawatts of AMD infrastructure supports that deployment reality. A gigawatt of computing power can serve roughly 30-50 million simultaneous inference requests depending on model complexity and optimization. Six gigawatts positions Meta to handle hundreds of millions of concurrent AI interactions—still a fraction of total user base but sufficient for scaled deployment of sophisticated AI features across platforms.

The AMD chips are optimized for inference, not training. AMD and Meta customized the MI450 architecture specifically for Meta’s workload patterns—emphasizing throughput, efficiency, and the ability to serve many simultaneous requests rather than maximizing raw compute for single large jobs. This specialization differentiates the AMD deployment from the Nvidia commitment, which likely focuses more on training infrastructure.

---

THE AMD STRATEGY: FROM NVIDIA ALTERNATIVE TO STRATEGIC PARTNER

For AMD, the Meta deal represents validation of a multi-year strategy to challenge Nvidia’s AI dominance. AMD’s GPU market share in AI training and inference remains small—Nvidia controls an estimated 90%+ of the data center GPU market. But that dominance creates opportunity: hyperscalers desperate for supply are willing to work with alternative vendors if those vendors can deliver at scale.

AMD’s approach mirrors the OpenAI partnership announced in October 2025. That deal committed OpenAI to purchase AMD GPUs for multi-gigawatt deployments, with similar warrant structures tying AMD’s equity to delivery performance. The Meta agreement doubles down on the same model: secure massive, long-term commitments from AI infrastructure buyers, then issue warrants that align incentives and signal confidence in execution.

The warrants are strategic for AMD. If the company successfully delivers 6 gigawatts to Meta over five years, and if AMD’s stock appreciates to the levels the warrant structure contemplates, Meta becomes a major shareholder with vested interest in AMD’s long-term success. This creates stickiness beyond contractual commitments—Meta’s financial success becomes partially tied to AMD’s performance.

The customization strategy differentiates AMD from Nvidia. Nvidia sells standardized chips—H100, H200, Blackwell—optimized for broad AI workloads but not tailored to specific customer needs. AMD is offering bespoke architectures: MI450 chips customized for Meta, different configurations for OpenAI, roadmap alignment that lets customers influence future product development.

This approach trades gross margin (customization is expensive) for strategic positioning. If AMD becomes the vendor hyperscalers choose when they want specialized infrastructure optimized for their specific workloads, AMD carves out defensible market share even if Nvidia retains dominance in general-purpose AI chips.

Lisa Su emphasized the multi-generation collaboration: the Meta deal isn’t just MI450 chips. It includes roadmap alignment across future GPU generations, CPU integration (6th gen EPYC CPUs), rack-scale systems, and software optimization through AMD’s ROCm platform. This vertical integration across silicon, systems, and software positions AMD as a platform provider, not just a component supplier.

The Helios rack-scale architecture exemplifies this strategy. Developed jointly by AMD and Meta through the Open Compute Project, Helios integrates GPUs, CPUs, networking, and cooling into optimized rack-level systems. Meta gets infrastructure designed specifically for its deployment patterns. AMD gets reference architecture it can offer to other hyperscalers, amortizing development costs across multiple customers.

---

THE COMPETITIVE DYNAMICS: HOW NVIDIA RESPONDS

Nvidia’s position remains dominant but faces new pressures. The Meta-AMD deal doesn’t immediately threaten Nvidia’s market share—Meta is still buying millions of Blackwell GPUs. But it establishes precedent: major AI infrastructure buyers are willing to commit tens of billions to alternative vendors when supply constraints or strategic considerations warrant.

Nvidia’s response to date has been expanding production capacity and maintaining roadmap cadence. The company reported record earnings in February 2026, driven by insatiable demand for H200 and Blackwell chips. But the stock reaction was muted—investors recognized that even record results don’t address the fundamental challenge of demand exceeding supply by wide margins.

The China export control situation compounds Nvidia’s challenges. H200 sales to China remain frozen despite regulatory clearance—a situation that removes a major market from Nvidia’s addressable opportunity while doing nothing to reduce demand pressure from U.S. and allied customers. Nvidia’s record quarter occurred despite losing access to the world’s second-largest market for AI infrastructure.

This creates asymmetric opportunity for AMD. Without export control complications (AMD’s chips don’t face the same restrictions for most AI applications), AMD can potentially serve markets Nvidia cannot. The Meta deal is U.S.-focused, but AMD’s ability to operate globally without the regulatory constraints affecting Nvidia’s most advanced chips offers strategic optionality.

Nvidia’s counter-strategy appears to focus on vertical integration and ecosystem lock-in. The Grace CPU launch (included in Meta’s Nvidia expansion) represents Nvidia’s move beyond pure GPUs into full-stack infrastructure. The Spectrum-X networking switches and co-packaged optics further extend Nvidia’s reach. The strategic bet: even if customers diversify GPU suppliers, Nvidia can retain value capture through networking, CPUs, and software platforms like CUDA.

But diversification, once started, tends to accelerate. If Meta successfully deploys AMD infrastructure at gigawatt scale and demonstrates that non-Nvidia chips can handle production AI workloads, other hyperscalers will follow. Google has publicly discussed custom TPUs for years. Amazon developed Trainium and Inferentia. Microsoft announced Maia. Each of these efforts previously faced skepticism—could custom or alternative chips truly match Nvidia’s performance?

Meta’s $100 billion commitment provides an answer: at minimum, the gap is small enough and the supply constraint severe enough that diversification makes strategic sense even if AMD chips underperform Nvidia’s on pure benchmarks. And if AMD’s customized approach delivers efficiency advantages for specific workloads like inference, the case for alternative vendors strengthens further.

---

THE HYPERSCALER PLAYBOOK: EVERYONE IS DIVERSIFYING

Meta’s AMD deal isn’t isolated—it’s part of an industry-wide pattern. Every major hyperscaler is reducing Nvidia dependency through some combination of custom silicon, alternative vendors, or architectural innovation.

Amazon operates the most mature custom chip program. Trainium (for training) and Inferentia (for inference) power increasing portions of AWS’s AI infrastructure. Amazon positions these chips as cost-optimized alternatives to Nvidia for customers willing to accept slightly lower raw performance in exchange for significantly lower prices. The strategy works because many AI applications don’t require cutting-edge chips—they need sufficient performance at acceptable cost.

Google has invested billions in Tensor Processing Units (TPUs) over multiple generations. Google doesn’t sell TPUs externally (they’re Google Cloud exclusive), but the architecture demonstrates that custom silicon can match or exceed Nvidia for specific workloads. Google’s most advanced AI models train primarily on TPUs, not Nvidia GPUs.

Microsoft announced Maia—custom AI accelerators designed for Azure infrastructure. Microsoft is also a major Nvidia customer (through direct purchases and through OpenAI’s needs), but the Maia program signals the same diversification impulse driving Meta’s AMD deal.

OpenAI signed its own AMD deal in October 2025, committing to multi-gigawatt deployments structured similarly to Meta’s agreement. OpenAI remains Nvidia’s largest customer by some measures, but the AMD partnership provides supply redundancy and pricing leverage.

The pattern is consistent: hyperscalers are building multi-vendor infrastructure even when Nvidia remains their primary supplier. The AMD deals (Meta, OpenAI) are the most visible examples, but custom chip programs (Amazon, Google, Microsoft) pursue the same goal through different means.

This creates a prisoner’s dilemma dynamic for Nvidia. If Nvidia could guarantee unlimited supply at reasonable prices, hyperscalers might accept single-vendor dependency. But supply constraints are structural—semiconductor fabrication capacity takes years to build, and demand grows faster than new fabs come online. Nvidia cannot deliver all the chips all customers want simultaneously.

Faced with that reality, each hyperscaler independently concludes that diversification reduces risk. And as more hyperscalers diversify, Nvidia’s pricing power and allocation leverage diminish, creating incentives for remaining customers to diversify as well. The equilibrium shifts from Nvidia monopoly to multi-vendor competition.

---

THE INFRASTRUCTURE SCALE: WHAT 6 GIGAWATTS MEANS

Six gigawatts of computing infrastructure represents scale difficult to conceptualize. For context:

A typical nuclear power plant generates 1 gigawatt. Meta’s AMD deployment will consume power equivalent to six nuclear plants running continuously—just for the AMD infrastructure, separate from Nvidia, custom MTIA chips, and conventional data center compute.

Meta’s announced $10 billion Indiana data center campus is designed for 1 gigawatt of capacity. The AMD deal represents six such facilities—or more likely, distributed capacity across Meta’s global data center network to optimize for latency, redundancy, and regulatory compliance.

The energy requirements are enormous. Six gigawatts running continuously for a year consumes approximately 52.6 terawatt-hours of electricity—more than entire countries like Ireland or Peru consume annually. Meta’s total AI infrastructure buildout (including Nvidia, AMD, and other vendors) will rank among the world’s largest energy consumers.

This scale explains Meta’s recent $600 billion infrastructure commitment and the gas-powered Indiana facility. Renewable energy can’t scale fast enough to meet AI infrastructure demands on the timelines Meta’s roadmap requires. Natural gas provides dispatchable power that can be deployed rapidly, even if it conflicts with Meta’s longer-term sustainability commitments.

The cooling requirements are equally challenging. Six gigawatts of computing generates six gigawatts of heat (minus efficiency losses). Removing that heat requires massive cooling infrastructure—either traditional air cooling (which consumes additional power for fans and chillers) or liquid cooling (which requires complex plumbing and poses leak risks).

AMD and Meta’s Helios rack-scale architecture addresses some of these challenges through integrated design. By optimizing thermal management at the rack level rather than the individual chip level, Helios improves cooling efficiency. But even optimized systems can’t escape thermodynamic reality: six gigawatts of power becomes six gigawatts of heat that must be dissipated.

The deployment timeline—first gigawatt in late 2026, full six gigawatts over five years—reflects these physical constraints. You can’t build gigawatt-scale data centers overnight. Electrical infrastructure must be upgraded. Cooling systems must be installed. Networking must be deployed. Each gigawatt represents months of construction and billions in capital expenditure.

---

THE PERSONAL SUPERINTELLIGENCE PARADOX: CENTRALIZED INFRASTRUCTURE FOR PERSONALIZED AI

Meta’s “personal superintelligence” positioning creates an interesting tension. The AI systems Meta envisions are deeply personalized—understanding individual preferences, contexts, and needs. But the infrastructure supporting them is radically centralized—six gigawatts of computing power in Meta’s data centers, processing requests from billions of users.

This architecture raises questions about privacy, control, and the distribution of AI capability. If personal superintelligence runs in Meta’s infrastructure rather than on user devices, Meta intermediates every interaction. The AI may be “personal” in the sense that it customizes responses to individual users, but it’s not personal in the sense of being under user control.

The alternative—edge AI running on user devices—faces severe constraints. Smartphones and laptops can run inference for relatively simple models, but the “superintelligence” Meta envisions requires computational resources no personal device can provide. So the tradeoff becomes: accept centralized infrastructure controlled by Meta, or forgo access to the most capable AI systems.

This dynamic benefits Meta enormously. Users who want access to cutting-edge AI must use Meta’s platforms, generating data that improves Meta’s models, which attracts more users in a reinforcing loop. The infrastructure investment—hundreds of billions in chips, data centers, and energy—creates moats competitors struggle to match.

Even well-funded competitors face challenges. Building six gigawatts of AMD infrastructure costs approximately $100 billion. Adding equivalent Nvidia capacity costs similar amounts. The total capital requirement for competing at Meta’s scale exceeds most companies’ market capitalizations. Only Amazon, Google, Microsoft, and a handful of others can match these investment levels.

This concentration of AI infrastructure in a few hyperscalers’ hands has policy implications governments are beginning to recognize. If AI capability requires gigawatt-scale data centers that only five companies can afford, what happens to competition, innovation, and pluralism in AI development?

Meta’s answer seems to be: scale benefits accrue to users through better AI experiences, and infrastructure competition (Meta vs Google vs Microsoft) prevents monopoly even if smaller players can’t compete. Critics might argue that five-player oligopoly in foundational AI infrastructure concentrates power in ways previous technology transitions didn’t.

---

THE AMD EXECUTION QUESTION: CAN THEY DELIVER?

The Meta deal’s success depends on AMD executing at unprecedented scale. AMD has never shipped gigawatt-scale GPU deployments. The company’s total data center GPU revenue in 2025 was measured in billions, not tens of billions. Delivering $100 billion in infrastructure over five years requires roughly 10x scaling of AMD’s AI chip production and deployment capability.

AMD has faced execution challenges before. The MI300 series launch in 2024 encountered delays and supply constraints. Early customers reported software immaturity—ROCm, AMD’s CUDA competitor, lacked the polish and ecosystem breadth that made Nvidia’s platform dominant. Some AI researchers found AMD chips difficult to program effectively, limiting adoption despite competitive hardware specifications.

The MI450 series represents AMD’s answer: customization reduces software complexity by optimizing for specific workloads rather than general-purpose AI. If Meta’s inference patterns are well-defined and stable, AMD can design chips that excel at those specific tasks without needing the full flexibility Nvidia’s chips provide.

But customization creates its own risks. If Meta’s AI architecture evolves faster than AMD’s silicon roadmap, the customized MI450 chips might become obsolete before their five-year deployment window completes. Nvidia’s general-purpose approach offers more flexibility—customers can adapt workloads to leverage available hardware even as requirements change.

The warrant structure provides AMD with strong incentives to execute. If AMD delivers on schedule and Meta successfully deploys the infrastructure, both companies benefit. If AMD faces delays or performance issues, Meta could reduce orders (within contractual constraints) and AMD’s stock wouldn’t reach the levels that trigger warrant vesting.

Lisa Su’s track record provides some confidence. Under her leadership, AMD executed successful turnarounds in CPUs (Ryzen, EPYC) and GPUs (RDNA gaming chips). The company has demonstrated ability to compete with larger incumbents (Intel, Nvidia) through architectural innovation and execution discipline. But AI infrastructure at Meta’s scale tests capabilities AMD hasn’t previously demonstrated.

The supply chain is another variable. AMD doesn’t manufacture chips—it relies on TSMC for fabrication. TSMC is already running near capacity serving Nvidia, Apple, and other major customers. Can AMD secure sufficient TSMC allocation to deliver six gigawatts to Meta while also fulfilling the OpenAI commitment and serving other customers?

TSMC is expanding capacity, but new fabs take years to build. AMD’s ability to deliver on the Meta contract partly depends on TSMC’s ability to allocate sufficient advanced process capacity to AMD’s orders. This introduces dependencies outside AMD’s direct control.

---

WHAT THIS MEANS FOR THE AI INFRASTRUCTURE MARKET

The Meta-AMD deal signals several shifts in AI infrastructure economics and competitive dynamics:

Monopoly is ending. Nvidia’s 90%+ market share in AI chips is declining. Not immediately, not dramatically, but directionally. As hyperscalers commit tens of billions to alternative vendors, the market fragments. Nvidia remains dominant but no longer monopolistic.

Customization becomes competitive advantage. AMD’s willingness to tailor architectures to specific customer workloads differentiates from Nvidia’s standardized approach. Other chip designers (Broadcom, Marvell, even startups) might pursue similar strategies, offering bespoke solutions for hyperscalers’ unique requirements.

Warrants as strategic alignment. The performance-based warrant structure AMD used with Meta and OpenAI provides a model for capital-intensive technology partnerships. Customers get ownership upside, suppliers get guaranteed demand, and both parties’ incentives align around successful execution.

Infrastructure becomes moat. Meta’s $600 billion investment in data centers and chips creates competitive advantages smaller players can’t match. Concentration of AI capability in a few hyperscalers accelerates, with implications for competition policy and market structure.

Energy constraints bind. Six gigawatts for AMD chips, comparable amounts for Nvidia, custom silicon, and conventional compute—Meta’s total power consumption approaches nation-state levels. Energy availability and cost will increasingly constrain AI infrastructure expansion.

Inference optimization matters. The AMD deal focuses on inference rather than training, reflecting the reality that deployed AI at scale requires different infrastructure than frontier model development. Expect more specialization: training chips, inference chips, edge chips, each optimized for specific deployment patterns.

Multi-vendor is new normal. Every major AI infrastructure buyer will maintain relationships with multiple chip suppliers. Single-vendor dependency is strategically unacceptable when supply constraints persist and demand growth shows no signs of slowing.

The broader implication: AI infrastructure is maturing from experimental buildout to industrial-scale deployment. The hyperscalers committing hundreds of billions to chip purchases, data centers, and energy infrastructure aren’t betting on AI’s potential—they’re building for AI as the foundational compute layer for the next decade of internet services.

Meta’s simultaneous Nvidia and AMD commitments exemplify that maturity. This isn’t a startup experimenting with different vendors to find the best fit. It’s a company deploying tens of gigawatts of computing capacity using multiple suppliers because the scale of deployment requires it, the strategic risk of dependency demands it, and the market dynamics enable it.

---

CONCLUSION: THE BEGINNING OF POST-MONOPOLY AI INFRASTRUCTURE

When Nvidia reported record earnings in February 2026, the stock reaction was subdued. Investors recognized that even unprecedented revenue growth can’t solve the fundamental challenge: demand exceeds supply, customers are actively diversifying, and the monopoly that enabled Nvidia’s margins is weakening.

Meta’s $100 billion AMD deal crystallizes that shift. Two weeks after expanding its Nvidia partnership, Meta committed comparable amounts to AMD—not as replacement but as redundancy. The message to the market: in AI infrastructure, single-vendor dependency is unacceptable risk regardless of which vendor dominates today.

For AMD, the deal represents validation of Lisa Su’s multi-year strategy to challenge Nvidia in data center AI. The company still holds single-digit market share, still faces execution challenges, still lacks Nvidia’s software ecosystem maturity. But it has secured commitments from two of the world’s largest AI infrastructure buyers (Meta and OpenAI) at scales that will force AMD to execute or fail publicly.

For Meta, the deal advances “personal superintelligence” ambitions by ensuring chip supply won’t constrain infrastructure buildout. Whether AMD’s customized MI450 chips truly outperform Nvidia’s general-purpose GPUs for inference remains to be proven in production. But the diversification alone provides strategic value—Meta can’t be held hostage to any single supplier’s delivery schedule or pricing.

For the broader AI industry, the deal marks the transition from Nvidia monopoly to multi-vendor competition. Google’s TPUs, Amazon’s Trainium, Microsoft’s Maia, and now AMD’s partnerships with Meta and OpenAI all point the same direction: hyperscalers building infrastructure across multiple suppliers to mitigate concentration risk.

The irony is that this shift happens precisely when Nvidia’s technological lead seems most secure. Blackwell represents genuine architectural innovation. CUDA’s ecosystem advantages are real. Nvidia’s AI chip dominance wasn’t accidental—the company built superior products and created moats through software, developer tools, and accumulated expertise.

But monopoly in a market with explosive demand growth creates its own vulnerability. Customers with hundreds of billions to spend and existential dependency on chip supply will find alternatives even when the incumbent provides the best products. The superior technology question becomes secondary to the strategic dependency question.

Meta’s $100 billion bet on AMD isn’t a prediction that AMD will overtake Nvidia. It’s recognition that in a market where six gigawatts of computing capacity is table stakes for competing in AI, accepting single-vendor dependency is riskier than any alternative vendor’s execution challenges.

Two weeks, two massive deals, two different suppliers. The simultaneous commitments aren’t contradictory. They’re the future of AI infrastructure in a post-monopoly market.

---

Key Numbers:

MetricFigure
Meta-AMD deal total value~$100 billion
AMD infrastructure capacity6 gigawatts
Deployment timeline5 years (starting late 2026)
First gigawatt deployment2H 2026
AMD warrant shares for MetaUp to 160 million (~10% of AMD)
Warrant strike price$0.01 per share
AMD stock price (Feb 24)$196.60
Warrant full vesting trigger$600 AMD stock price
Cost per gigawatt$15-20 billion (estimated)
Meta 2026 AI capex$135 billion
Meta total infrastructure commitment (2026-2028)$600+ billion
Meta Indiana data center$10 billion, 1 gigawatt
Meta Nvidia expansionMillions of Blackwell GPUs (announced 2 weeks prior)
AMD chipsMI450 series (custom), 6th gen EPYC CPUs
Rack architectureAMD Helios (co-developed with Meta)
6 GW annual energy consumption~52.6 terawatt-hours
AMD data center GPU revenue 2025Billions (exact figures undisclosed)
Required scaling for deal~10x AMD’s current AI chip capacity
Deal announcement dateFebruary 24, 2026

---

ZeitShift Intelligence 2026

AI Infrastructure & Market Dynamics series