Infrastructure Analysis — AWS just made Graviton 5 generally available. 192 cores on 3nm. Purpose-built for agentic AI. Meta is deploying tens of millions of cores. This is not a chip announcement. It is Amazon building its own Layer 2 to escape the Nvidia Tax.
Meta deploying tens of millions of Graviton cores is the most important data point. Meta just cut 8,000 jobs to fund $145B in AI capex. Part of that capex is going to Amazon’s custom silicon instead of Nvidia’s GPUs.
The Nvidia Tax Escape
In the Map of AI, Layer 2 (Compute) has been a near-monopoly. Nvidia holds 75% of AI compute spend at 75% gross margins. Every hyperscaler pays the Nvidia Tax.
But the five biggest Nvidia customers are also the five biggest Nvidia competitors:
Custom Silicon vs Nvidia — The Layer 2 Diversification
MetaMTIA + now Graviton — dual-sourcing away from Nvidia
AppleApple Silicon — on-device AI, zero Nvidia dependency
Custom silicon growing 3x faster than Nvidia GPUs. The Nvidia Tax has an expiry date.
Graviton 5 is not competing with Nvidia’s H100/B200 for training frontier models. It is competing for everything else: inference, agentic workloads, databases, web applications, EDA tools. These workloads represent the majority of cloud compute spend — and they don’t need Nvidia’s margins.
The Agentic AI Angle
AWS explicitly designed Graviton 5 for agentic AI — “real-time reasoning, code generation, multi-step task orchestration.” This connects directly to Apple’s Agent OS Bet and the broader thesis that the agent is becoming the computer.
The structural insight: AI agents don’t need GPUs. They need CPUs with massive cache, fast memory, and low latency — exactly what Graviton 5 provides. The agent orchestration layer (deciding which model to call, routing queries, managing context) runs on CPUs, not GPUs.
This is why Amazon built Graviton 5 with 5x larger cache and 33% lower inter-core latency. The agent doesn’t train models — it coordinates them. And coordination is a CPU workload.
The Goldman $7.6T Read
Goldman’s $7.6 trillion AI capex projection assumes Nvidia at 75% of compute spend. If custom silicon (Graviton, TPU, Maia, MTIA) captures even 10% of that over six years, that is $380 billion that doesn’t flow through Nvidia.
Graviton 5 going GA with Meta as a flagship customer is the most concrete signal yet that the Nvidia Tax is being eroded from within — by the very customers who pay it.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
