Custom Silicon Pays a Different Merchant

custom-silicon-pays-a-different-merchant · conviction medium · status open · horizon 2028 · as of 2026-08-05

Hyperscalers build custom ASICs to escape NVIDIA's margin, and the escape routes the rent to Broadcom and Marvell rather than back to the buyer. The trade is a rent TRANSFER inside the accelerator layer, not a rent RECOVERY by the customer — which is why Broadcom is already the layer's second-largest name at $1.84tn.

Rests on filed figures, not on modelled shares. 3 premises (2 entity, 1 field); no derived cell is involved, so the undisclosed supply weights that put a range on other pages in this bank cannot move this one.

Exhibits

Exhibit 1Relative performance, indexed to 100How the names in this thesis have traded against SOXX.

Series available as data/custom-silicon-pays-a-different-merchant.csv

Exhibit 2Who pays CoWoS advanced-packaging capacity, and who keeps the moneyCapturers average 47.1% operating margin against payers' 44.6% — the owners of the scarce thing capture the rent, as expected.

Green/blue = model marks it as CAPTURING the rent (unbound and supplies the scarce good); faded = PAYING it (bound severe or moderate). Operating margin, live.

Exhibit 3What the conviction is actually made ofEach premise and the number it composes to. A conjunction of plausible premises is far weaker than any of them.

The rent escaped from the merchant vendor is captured by the design partner rather than by the hyperscaler

† 1 premise marked supporting — shown and arguable, but the conclusion does not depend on them, so they are not multiplied into the composed figure. Citing a filed figure should not cost conviction.

60% if the 2 gates are independent, 75% if they move together. They are claims about one industry, so the truth is between and nobody can say where. Treat this as an ordering device rather than a calibrated probability — the ranking of premises is the information, not the level.

Weakest link: HBM4 at 0.75 — A custom accelerator competes for the SAME scarce inputs as a merchant one — HBM allocation, advanced packaging, leading-edge foundry slots. Escaping

The variant

Consensus

In-house ASICs are how hyperscalers break the accelerator monopoly and take back gross margin. Every TPU or Trainium generation that ships is share taken from NVIDIA and cost recovered by the buyer, so custom silicon is a hyperscaler-margin story.

Variant

The customer does not capture the rent it escapes. Designing a competitive accelerator requires the same scarce inputs the merchant vendor competes for — leading-edge foundry slots, HBM allocation, advanced packaging — and the design partner that assembles them charges for the privilege. The rent moves ACROSS the accelerator layer from one merchant to another. Broadcom at $1.84tn against NVIDIA at $5.03tn is not a challenger discount; it is the market already pricing the transfer.

Differentiator

Consensus counts ASIC UNITS as share taken from NVIDIA. The variant watches where the MARGIN lands: hyperscaler gross margin should improve if the escape is real, and cloud-layer margin is 3.0% against an accelerator layer at 17.2%. The buyer pays a different merchant and keeps the same thin margin.

Open questions

Broadcom's AI revenue is not separable from its infrastructure-software segment in the disclosure, so the share of its $1.84tn attributable to custom accelerators cannot be isolated from the filings — the same resolution limit that memory has with dram_incl_hbm.
Marvell is materially smaller and may be the better test of the mechanism, since it has less non-AI business to hide the effect.

Reasoning chain

The rent escaped from the merchant vendor is captured by the design partner rather than by the hyperscaler VALID

premises

Broadcom Inc.0.80 strong
1.84 tn1.5 tn
Broadcom is the second-largest name in the accelerator layer at $1.84tn and is above the desk's $1.5tn size ceiling.
why
A pure design-services business does not reach that scale on services fees; it reaches it by holding a position in the flow it enables.
CoreWeave, Inc. — operating margin -2.1%0.70 moderate
69.4%
CoreWeave earns a 69.4% gross margin renting the scarcest asset in computing and a NEGATIVE operating margin.
why
A buyer of compute is not recovering rent — it is passing it through to whoever sold the fleet. If the custom-silicon escape worked differently, this is where it would show.
HBM40.75 strong
A custom accelerator competes for the SAME scarce inputs as a merchant one — HBM allocation, advanced packaging, leading-edge foundry slots.
why
Escaping a vendor does not escape the constraint that gives the vendor pricing power.

AND rather than OR: all three must hold. The design partner must be capturing scale economics, the buyer's margin must not be improving, and the underlying scarcity must be shared. If any one fails the thesis weakens sharply — a hyperscaler whose margin improves as ASIC share rises would refute it directly. THE FALSIFIER IS A MARGIN SERIES, not a unit count: cloud-layer operating margin expanding while custom-ASIC share rises kills this. That series is public and quarterly, which makes this cheap to be wrong about.

Sources

Layer market caps and median operating margins from the desk's own value-capture ladder acc 2026-08-05

Write-up

Pre-filled skeleton: custom-silicon-pays-a-different-merchant.md