By Djellal Djouad
This is the inaugural CrossVol Research note. Welcome.
TL;DR — five things to remember:
DeepSeek V4 Pro inference cost is $1.10/M output tokens. The US frontier sells the same intelligence at $25-30/M. Three falsifiable tests show this is not a subsidy.
Huawei Atlas 800 (Ascend 910C) is now delivering H100-class inference at 30% of NVIDIA system cost. Chinese AI Capex is converging on US Capex without buying NVIDIA.
PJM cleared $329.17/MW-day vs $28.92 two years prior. A ~300 GW data-center demand gap by 2030 against ~60 GW of realistic generation supply.
Hyperscalers carry $145B+ of IG debt with 8-40Y duration, funding 18-24 month commoditizing token revenue. The duration mismatch is structural, not cyclical.
Two dated catalysts in the next ~6 months: Nov 10 (tariff truce expiry), Nov 27 (export-control suspension expiry). Together they remove the consensus assumption that the bilateral relationship stays containable.
Sell-side AI infrastructure equity targets implicitly assume none of these. We argue they will all bind by Q2 2026 → Q1 2027. Expected rerate: 25-40% on the AI infrastructure equity complex.
Why this is non-consensus
The sell-side is six months late on three transmission mechanisms most analysts treat as separate stories:
Token economics (a software story)
Hardware parity (a geopolitics story)
Power and capital markets (an infrastructure story)
We argue these are one story. The same balance sheet question — what is a dollar of AI Capex worth if token revenue commoditizes in 18-24 months? — runs through all three.
The five-vector framework gives you the lens to read it.
The five vectors
1. Token commoditization
DeepSeek V4 Pro: $1.10/M output. GPT-4o-class: $25-30/M. A 20-25x gap.
There are three explanations for this gap, and only one of them is true:
(a) Subsidy from the Chinese state
(b) DeepSeek is using cheap, depreciated hardware (Ascend, not H100)
(c) The frontier price is a margin, not a cost
The note walks through the three falsifiable tests. Spoiler: (c) is closest. The frontier sells convenience and reliability, not raw inference.
2. Chinese hardware parity
Ascend 910C / Atlas 800 numbers from the field: H100-class inference, 30% of NVIDIA system cost on a like-for-like deployment. Chinese hyperscalers now have a domestic substitute that does not require US export licenses.
The note details the deployment pipeline through Q1 2027 and the implications for NVIDIA's China Total Available Market (TAM).
3. US power grid bottleneck
PJM 2026/27 capacity auction cleared at $329.17/MW-day, up from $28.92 in 2024. The cause is structural: ~300 GW of new data-center demand by 2030 against ~60 GW of realistic new supply (PJM queue, ISO interconnection times, transmission build-out).
Power, not silicon, is the binding constraint on US AI deployment.
4. Trump-admin response: physical, not regulatory
The note argues the next phase of US-China AI policy is physical (export controls on equipment, rare-earth restrictions, tariff truce expirations) rather than regulatory (chip-design controls, software restrictions). Two dated catalysts in November 2026 test this.
5. Hyperscaler bond wall
$145B+ of IG debt issued by hyperscalers in 2024-2025, with 8-40Y duration, funding AI Capex with 18-24 month commoditizing revenue.
The duration mismatch is the cleanest single transmission mechanism between the four vectors above and equity valuations. Credit spreads will reprice first; equity follows on a 60-90 day lag.
What I'm publishing here
CrossVol Research is the research arm of the CrossVol derivatives desk. We publish:
Non-consensus thesis notes like this one (~4-6 per year, 5-10k words each, fully sourced)
Trade construction notes (dispersion, calendar, vol-of-vol) tied to thesis catalysts
Tactical weekly briefs on macro vol, positioning, term structure
The framework is the framework I use on the desk every day. The differentiator: we read derivatives positioning the way most analysts read price.
The desk publishes timestamped public calls on @DerivativeProFR since 2024. Five of them are referenced in the long-form version of this thesis.
Full thesis with 39 sources, 4 satellite notes (dispersion trade construction, bond-wall transmission, PJM grid bottleneck, token economics): crossvol.com
What's next
This week:
Satellite #1 — Dispersion trade construction on long single-name vol / short index vol for the AI infra rerate window
Satellite #2 — The hyperscaler bond wall: which credit spreads reprice first
Satellite #3 — PJM grid bottleneck mechanics for equity investors
Satellite #4 — DeepSeek token economics: three falsifiable tests
Subscribe to receive these as they drop.
— CrossVol Research Two decades on derivatives desks across FXCM LLC, Knight Capital, JP Funds, Allianz, BCFX, Trafigura, VTB Capital. Five timestamped public calls on @DerivativeProFR since 2024.
CrossVol Team & Djellal Djouad
Disclaimer: This material is for informational purposes only and does not constitute investment advice, an offer, or a solicitation. CrossVol Research does not make trade recommendations. The views expressed are the analytical opinion of the authors. Readers should consult their own advisors. CrossVol Research may hold positions in entities mentioned. Past performance does not guarantee future returns.
Further reading