How long should AI compute investments and contracts be underwritten given rapid hardware depreciation?

Andrew Feldman · Oct 6, 2025

The real question in AI hardware depreciation is how much faster future chip generations are than the current one, not accounting convention

Depreciation schedules only matter to the extent successor generations obsolete current hardware

0:00 20VC: Cerebras CEO on Why Raise $1BN and Delay the IPO | NVIDIA Showing Signs They Are Worried About Growth | Concentration of Value in Mag7: Will the AI Train Come to a Halt | Can the US Supply the Energy for AI with Andrew Feldman

Andrew Feldman · Oct 6, 2025

The true driver of chip depreciation is how much faster each new generation is than the last, not elapsed time — you retire a fully paid-off chip only when a new part is fast and power-efficient enough that its data-centre footprint is worth more.

In a fixed-megawatt data centre, chips are displaced when new parts deliver more dollars per unit of power and space, even at zero remaining depreciated cost.

15:17 20VC: Cerebras CEO on Why Raise $1BN and Delay the IPO | NVIDIA Showing Signs They Are Worried About Growth | Concentration of Value in Mag7: Will the AI Train Come to a Halt | Can the US Supply the Energy for AI with Andrew Feldman

Eiso Kant · Oct 7, 2024

Nvidia's generational gains for training have consistently been about 2x every two years, so Blackwell's training gains are largely marketing

Every two-year Nvidia generation has delivered roughly a 2x training performance increase

Scope: training specifically

34:30 20VC: Raising $500M To Compete in the Race for AGI | Will Scaling Laws Continue: Is Access to Compute Everything | Will Nvidia Continue To Dominate | The Biggest Bottlenecks in the Race for AGI with Eiso Kant, CTO @ Poolside

Andrew Feldman · Oct 6, 2025

Real generational chip performance gains are only about two to two and a half times per meaningful generation, far less than vendor marketing implies, once you compare like for like.

Engineering digging on apples-to-apples comparisons (eight bit to eight bit, four bit to four bit) shows memory bandwidth doesn't improve more than 2x even if flops do.

Scope: comparing equivalent precision formats

16:52 20VC: Cerebras CEO on Why Raise $1BN and Delay the IPO | NVIDIA Showing Signs They Are Worried About Growth | Concentration of Value in Mag7: Will the AI Train Come to a Halt | Can the US Supply the Energy for AI with Andrew Feldman

Jonathan Ross · Sep 29, 2025

The industry is amortizing chips over too long a period; a five-year chip life makes no sense because new chips will drive old chips' value below their electricity and data center operating costs

There are two phases to a chip's value — covering CapEx at deployment and merely beating OpEx to stay running — and the bet everyone is making is that new chips won't push old chips below OpEx, which he doesn't believe holds at five years

Scope: Groq uses a more conservative internal number and upgrades chips about once a year

20:07 20VC: OpenAI and Anthropic Will Build Their Own Chips | NVIDIA Will Be Worth $10TRN | How to Solve the Energy Required for AI... Nuclear | Why China is Behind the US in the Race for AGI with Jonathan Ross, Groq Founder

David Cahn · Oct 27, 2025

The core risk of the buildout is chip vintage: today's spend is on H100s and B100s, and if the breakthrough only arrives with Rubin or the 2028 Feynman chips, investors are left with warehouses of legacy hardware

AI is physical — you can't just upgrade a chip like a finger; the hardware sits in a giant warehouse and may become legacy if the timeline stretches to ten years instead of two

28:02 20VC: Sequoia's David Cahn on The Winners and Losers in AI | The $0-$100M Revenue Club: Is Triple, Triple, Double, Double Dead? | The Future of Defence: Who Wins and Who Loses | How to Analyse Margins and Growth Rates in a World of AI

David Cahn · Aug 5, 2024

No one will ever train a frontier model on the same data center twice

By the time a training run finishes, the GPUs are outdated and the data center is too small for the next model

0:00 Training runs outpace their own data centers becoming obsolete before completion

Steeve Morin · Feb 24, 2025

Preemptively bought compute does risk becoming outdated, but buyers have been spared because Blackwell is late and other chips are being cancelled, keeping the H-series current

delays in next-generation silicon extend the useful life of what people already bought

23:30 Delayed next gen chip releases extend the useful life of existing hardware

Jonathan Ross · Sep 29, 2025

The shorter the payback period you underwrite, the clearer the outcome, which is why Groq avoids long-dated compute bets

Long contracts add a third calculation — whether breaking the contract is cheaper than running the chip at a loss — and he can't predict how that resolves

21:21 Shorter underwriting payback periods yield clearer predictable outcomes

Eiso Kant · Oct 7, 2024

Frontier labs are not forced to upgrade to each new chip generation, because a 2x more efficient chip unlocks no new capability, only more of the same math

The operations are the same matrix multiplications and additions; Blackwell just means more can be done per chip, which devalues his H200s without obliging him to replace them

35:20 Chip upgrades add throughput not new capability so adoption isnt forced

Andrew Feldman · Mar 24, 2025

Ten years is an unreasonably long planning horizon in the chip and AI industry — three to five years is the meaningful frame

NVIDIA was worth only $10 billion ten years ago, showing how much can change in that span

58:28 Three to five year horizon is the meaningful planning frame not ten years