Meta Bets Billions on AWS Chips While Cutting 8,000 Jobs
What happened
Meta Platforms signed a multiyear agreement with Amazon Web Services on April 24 to deploy tens of millions of AWS Graviton5 CPU cores, beginning a significant diversification away from Nvidia GPUs for AI workloads. The deal is valued in the billions. The Graviton5 chip uses a 3-nanometer process with 192 cores and offers up to 25% better performance than the previous generation. The agreement is specifically targeted at agentic AI: systems that reason, plan, and execute multi-step tasks in real time, which are CPU-intensive rather than GPU-intensive. Meta simultaneously confirmed it will cut approximately 8,000 employees and close 6,000 open positions effective May 20, while projecting 2026 capital expenditures of $115-135 billion, up from $72 billion in 2025.
The shift from GPU to CPU for inference and agentic AI is real, and it undermines Nvidia's structural position: the moat Nvidia built on training is less relevant for the deployment phase that now matters commercially.
The Hidden Bet
Agentic AI workloads are fundamentally CPU-driven
The current consensus that inference and agents are CPU-intensive may not hold as model architectures evolve. If future agentic systems require more complex reasoning that benefits from GPU acceleration, the Meta-AWS bet could become a strategic mistake within 18-24 months. Nvidia is not standing still on inference efficiency.
Meta's $125B+ capex in 2026 represents productive investment
S&P Global projects Meta's Reality Labs will post roughly $5 billion in losses in 2026. The AI investment thesis is that it will eventually translate into revenue. Analysts have no clear model for how or when. The capex expansion is happening while simultaneously cutting the workforce that would build and use those systems.
AWS winning Meta's CPU business is a straightforward Amazon win
Meta is using AWS as a compute vendor while also building its own data centers and custom silicon with Arm Holdings. AWS gets revenue but Meta retains architectural control. If AWS's chips are commoditized relative to Meta's proprietary silicon, Meta can exit the deal. The dependency runs the wrong direction from Amazon's perspective.
The Real Disagreement
The real fork in this story is not about chips. It is about whether hyperscale cloud vendors or vertically integrated AI companies will control the infrastructure layer of the AI economy. Meta's strategy is to use AWS as a cost-efficient supplier while building proprietary infrastructure in parallel. Amazon's strategy is to become a core partner that Meta cannot easily replace. Both cannot be right about who holds the leverage. If agentic AI systems become the primary commercial product of the AI era, the company that controls the CPU layer for inference controls the economics of deployment. Amazon is betting it can own that layer for companies that cannot build their own. Meta is betting it can build its own while using AWS as a flexible overflow.
What No One Is Saying
The layoffs and the capex increase happening simultaneously are not contradictory from a capital allocation perspective: Meta is replacing human labor with AI capital. The 8,000 jobs cut are, in many cases, being substituted by the compute the company is buying. The story being told publicly is 'efficiency.' The story being executed is 'the marginal return on human labor in our AI-optimized workflow is now below the cost of capital.' That is either visionary or a bet on AI capabilities that do not yet exist at commercial reliability.
Who Pays
Meta's 8,000 laid-off employees
Effective May 20, 2026
Job cuts effective May 20. These are predominantly technical and operational roles that the company's own internal AI is being positioned to replace. The layoffs come while the company is spending more than $100B on AI infrastructure.
Nvidia
Gradual; structural shift over 12-24 months as inference workloads scale
If the CPU-for-inference thesis holds and other hyperscale customers follow Meta's example, Nvidia's growth story shifts from a multi-year supercycle to a market that is maturing in its dependence on GPUs. Nvidia is not displaced but its pricing power in inference diminishes.
Smaller AI companies
Already visible in cost-per-token comparisons; worsening as proprietary silicon scales
Meta, Google, Microsoft, and Amazon are building proprietary infrastructure that gives them structural cost advantages over companies that rely entirely on GPU cloud rental. The marginal cost of running AI models falls for companies with custom silicon and rises in relative terms for everyone else.
Scenarios
CPU-era thesis validates: agentic AI scales on Graviton
Meta's agentic AI products (reasoning, coding, multi-step task execution) achieve commercial scale running on Graviton cores. AWS becomes the default CPU infrastructure for agentic AI. Nvidia's share of the AI infrastructure market falls as inference displaces training as the primary workload.
Signal Meta reports Q2 or Q3 2026 results showing AI-driven revenue acceleration; AWS reports CPU utilization growth outpacing GPU growth
Agentic AI requires GPU: Nvidia retains dominance
Agentic AI products require more sophisticated reasoning than current CPU architectures efficiently support. Nvidia releases inference-optimized GPUs that outperform custom CPUs at comparable cost. The Meta-AWS bet is not wrong but proves to be an optimization rather than a structural shift.
Signal Nvidia announces Blackwell-derived inference chips with pricing that undercuts ARM-based alternatives at scale
Meta builds proprietary silicon, exits AWS
Meta's collaboration with Arm Holdings produces proprietary chips optimized for its specific agentic AI workloads. Within 2-3 years, Meta reduces its AWS dependency significantly, using Graviton as a transition technology. Amazon's bet on becoming a core partner fails.
Signal Meta announces a proprietary inference chip in 2027-2028; executives refer to AWS as a 'bridge' technology
What Would Change This
If Meta's Q1 2026 earnings (releasing April 29) show AI-driven revenue growth that is clearly accelerating and attributable to agentic products, the capex strategy is no longer speculative: there is a revenue line being built. Conversely, if Reality Labs losses accelerate or advertising revenue growth slows, the $125B capex number becomes a liability rather than an investment.
Related
Anthropic Is Paying Amazon $100 Billion to Not Run Out of Computers
powerDeepSeek Is Abandoning Nvidia. Jensen Huang Says That's Catastrophic. He's Right for the Wrong Reasons.
powerNVIDIA Just Released the First AI Models Built to Fix Quantum Computing's Core Problem
powerThe Export Control Bottleneck