AI Hardware & Chip Architecture · Frontier labs building custom chips to own infrastructure stack and cut compute costs

episode briefing

extracted intelligence

Bloomberg Tech

OpenAI Unveils First Custom AI Chip With Broadcom | Bloomberg Tech 6/24/2026

now viewing▶ 1:33

Frontier labs building custom chips to own infrastructure stack and cut compute costs

2026-06-24 · 5 company · 6 thematic

sentiment

5 bull0 bear0 neu

speakers

david george

David George leads the growth investing team at Andreessen Horowitz, managing $22 billion across five funds; he focuses on late-stage venture investments in founder-led companies like SpaceX and articulates the thesis for late-stage venture as a distinct asset class.

andrew feldman

Leads Cerebras, a public company (CBRS) building wafer-scale AI inference chips; previously co-founded SeaMicro (acquired by AMD); describes $25B backlog and global data center buildout at unprecedented scale.

hock tan

celine

Celine manages an actively managed ETF focused on robotics and automation, with SK Hynix as a top holding; she specializes in AI infrastructure, memory chips, and the intersection of physical AI with compute demand.

now playing · AI Hardware & Chip Architecture

AI Hardware & Chip Architecturetailwindscore 9/10hock tan

Frontier labs building custom chips to own infrastructure stack and cut compute costs

Every major frontier AI lab (OpenAI, Anthropic, etc.) will develop custom silicon via partners like Broadcom to control their infrastructure stack, optimize for their specific workloads, an…

Memory & Storagetailwindscore 9/10celine

Structural memory supply deficit driven by AI inference demand and rising manufacturing complexity

Memory demand has shifted from cyclical to structural as cloud providers continuously add memory to maximize inference performance despite price increases; supply cannot keep pace due to es…

Data Center Infrastructuretailwindscore 8/10brody

Hyperscalers committed to $850B in future data center leases despite ROI uncertainty

Meta, Microsoft, and other hyperscalers have signed $850B in future data center leases — most not yet built — reflecting a consensus that they must spend aggressively to secure AI capacity,…

Space Economytailwindscore 8/10david george

SpaceX Starship reusability could enable orbital GPU data centers with superior economics

Once Starship achieves rapid reusability, SpaceX can deploy 'airplane-sized GPU racks in space' with solar arrays, providing incremental compute capacity that avoids terrestrial power/land…

Semiconductorstailwindscore 8/10andrew feldman

Heterogeneous compute architectures will fragment the AI hardware market beyond GPUs

The AI compute market will not consolidate around GPUs; instead, a heterogeneous mix of custom ASICs (hyperscalers, frontier labs), wafer-scale engines (Cerebras), and other novel architect…

Private Markets & Venture Capitaltailwindscore 7/10david george

Late-stage venture emerges as $5T founder-centric asset class distinct from public markets

Late-stage private investing has grown into a $5T asset class where returns are driven by backing exceptional founders making long-term bets, not capital markets timing; top funds like a16z…