HOLDco

AI hardware margins are historically high — and historically temporary. This episode breaks down the forces already compressing AI silicon economics and what long-term holders should actually own as the cycle turns.

Show Notes

Nvidia's data-center gross margins recently touched 75%. Memory suppliers are posting similar numbers. By conventional metrics, AI hardware looks like the most lucrative infrastructure buildout in a generation — but HOLD.co's analysis of AI hardware margin collapse and its implications argues those numbers are scarcity rents, not durable economics. History suggests they end. The only open question is how fast.

This episode walks through the structural forces already pushing on AI silicon pricing, what that means for adjacent industries, and where long-term value actually accumulates when compute gets cheap. Key points covered include:

  • Scarcity rents always normalize: IBM mainframes, long-haul bandwidth, and solar modules all followed the same arc — high margins held until fab capacity caught up, buyers got sophisticated, and second sources arrived. AI silicon has no credible reason to escape this pattern.
  • Inference pricing is the early warning system: The cost of running a GPT-4-class model has fallen from roughly $30 per million tokens in early 2023 to well under $0.50 today — a decline Andreessen Horowitz tracked at approximately 10x per year for a fixed quality bar.
  • Hyperscalers are building their own chips: Google, Amazon, Microsoft, and Meta are all shipping custom ASICs on leading-edge nodes, with some analysts projecting Nvidia's inference market share could fall from above 90% today to the 20–30% range by 2028 — eroding pricing power from the very top of the demand curve.
  • Adjacent manufacturing is mispriced: Cooling systems, power distribution, and cable assemblies are being valued as though current demand is permanent; the underlying work is durable, but the pricing power is not, and the right frame is cash flow across a full cycle.
  • Land with power is the asymmetric position: Grid interconnects, water rights, and permitted substation capacity will remain scarce long after the chips running inside those buildings have commoditized — a durable scarcity that silicon cannot replicate.
  • Falling inference costs are an operating tailwind: A capability that cost $20 per million tokens three years ago and now costs under $0.40 is no longer an experiment — it's a straightforward input for agentic AI and automated workflows across the portfolio today.

The episode draws a sharp line between owning businesses whose economics survive compute normalization and owning multiples that assume normalization never comes. For more on how these dynamics intersect with acquisitions and portfolio construction, see Hold.co's mergers and acquisitions advisory services. Also from the show: What a Sale-Leaseback Actually Nets After Taxes and Rent examines a related question about how asset structure affects long-term returns.

Hold.co

VDR.ai

What is HOLDco?

An operator-led view of holding company work: acquiring, building and running durable, cash-producing businesses in the real economy. Deal criteria, diligence, integration, capital allocation, and the management questions that arrive the day after a close.

Each episode takes one decision — what to pay, what to fix first, when to keep the seller and when not to, how to fund the next deal — and reasons it through from an operator's chair rather than a spreadsheet. Written for people buying and running businesses, not spectating on them. Five or six minutes an episode.

Topics include deal criteria and screening, diligence that finds the real risk, deal structure and seller financing, integration priorities after close, capital allocation, management transitions, and running several businesses at once.

Produced by HOLD.co, an operator-led holding company. Full details, services and further reading at https://hold.co