The State of AI Compute

Key Takeaways

• AI compute capacity grew 8.5× in two years.
• Five to six firms dominate global AI hardware holdings.
• Google, Microsoft, Amazon, Meta, Oracle, xAI lead the race.
• Compute ownership becomes a strategic moat for the next decade.

Pulse Analysis

The pace of AI hardware expansion has outstripped most forecasts, with the H100‑equivalent metric serving as a proxy for raw model‑training power. Moving from 2.5 million units in early 2024 to over 21 million by late 2025 signals not just demand for larger models but a concerted effort to lock in the physical substrate that fuels them. This acceleration is driven by falling costs of high‑bandwidth memory, advances in chip packaging, and the strategic push by cloud providers to bundle compute with proprietary services, creating a virtuous cycle of demand and supply.

A deeper look reveals that the surge is concentrated among a tight cluster of tech giants—Google, Microsoft, Amazon, Meta, Oracle, and the emerging xAI venture. Their dominance stems from massive capital outlays, exclusive access to NVIDIA’s flagship H100 chips, and the ability to integrate compute into broader ecosystems of software, data, and talent. This concentration creates a formidable barrier for smaller players, as owning the compute layer translates directly into pricing power, faster model iteration, and the capacity to offer premium AI products that competitors cannot match.

Looking ahead, the trajectory suggests that compute will become the new strategic asset class, akin to data or bandwidth in the early 2010s. New entrants may seek alternative architectures—such as custom silicon, optical processors, or federated edge compute—to bypass the entrenched incumbents. Meanwhile, regulators may scrutinize the emerging duopoly for anticompetitive risks, especially as AI services permeate critical sectors. Companies that can secure long‑term access to large‑scale compute, either through partnerships or innovative hardware strategies, will likely dictate the pace of AI innovation for the next decade.