The Epoch Brief - May 22, 2026

Key Takeaways

• Memory now 63% of AI chip component costs, up from 52%
• High‑bandwidth memory spending grew to $32 B in 2025 across major vendors
• Frontier labs could capture most global AI compute headroom within years
• Further AI scaling will be limited by chip supply without economic shift
• FrontierMath workshops launch globally to surface verifiable math problems

Pulse Analysis

Memory has become the single largest cost driver in AI chips, now representing roughly two‑thirds of component spending. High‑bandwidth memory (HBM) outlays surged from $12 billion in 2024 to $32 billion in 2025 across Nvidia, AMD, Google and Amazon designs, outpacing all other parts. This rapid escalation pressures chip manufacturers to innovate cheaper, higher‑density memory stacks and reshapes supply‑chain negotiations, as memory shortages could directly inflate AI system prices.

At the same time, frontier AI labs are poised to dominate compute consumption. Josh You’s analysis suggests that while these labs currently use under 50% of global AI compute, they could absorb the bulk of the remaining capacity within a few years. Once that threshold is reached, further model scaling will be constrained by the physical availability of chips, despite AI capital expenditures nearing $1 trillion annually. Overcoming this barrier will likely require an unprecedented acceleration in semiconductor production and possibly new economic models to fund the expansion.

Beyond hardware, Epoch AI is fostering a collaborative research ecosystem. The FrontierMath: Open Problems workshops, launching in New York and rolling out to London, Berkeley, Boston, Los Angeles and Toronto, aim to surface mathematically rigorous, verifiable challenges that can accelerate AI breakthroughs. Coupled with a hiring push for designers, researchers and data scientists, the initiative underscores the firm’s commitment to building the talent pipeline needed to tackle the next generation of AI problems.