Dynamic Model of Circular Financing in the AI Compute Market: Stability with Neobanks in the Loop
Unpublished manuscript, 2026
Abstract
Neobanks have become a new kind of participant in the AI compute market: they own large behavioural datasets and train proprietary models, yet own no hardware and increasingly receive capital from the chip vendor itself. We ask how such a participant changes the stability of the market. The market is modelled in control-theory terms as a ninth-order block diagram linking NVIDIA with hyperscalers (Microsoft, Meta, Alphabet, Amazon, Oracle), AI labs connected to NVIDIA by cross-investments (OpenAI, Anthropic, xAI, Mistral), GPU-rental neoclouds (CoreWeave, Nebius, Nscale) and end consumers of compute, with Revolut as the neobank case. The model contains a positive feedback loop (circular vendor financing) and a negative one (substitution by in-house chips). Its fast market link is identified from NVIDIA’s share-price decline of October 6 – November 10, 2026 (−29.4%), driven by an October background shift, the November 3 US midterm election and a hypothetical risk specific to NVIDIA (from November 9): T = 0.70 days, ξ = 0.40, RMS error 1.8 pp, R² = 0.96. A coupling parameter θ describes how much of the vendor’s financing reaches the neobank channel. With θ = 0, Revolut’s demand is paid for by its own customers: its pole stays outside the loops and the stability margin with respect to circular financing is 8.8. In the hypothetical scenario θ = 1, the neobank becomes part of the positive loop: the critical financing gain falls from 1.31 to 0.79, the margin from 8.8 to 5.2 (−40%), shock amplification rises by 9% and the response slows by 22%. Only the four lowest-order coefficients of the characteristic polynomial change, so the coupling affects the slow fundamental dynamics while the market reaction itself remains intact. Data on Revolut published by Pavel Nesterov add a second effect: its blended LLM inference cost fell from $1.49 to $0.49 per million tokens between April and November 2026, and its AI coding inference costs fell by about 90% after it moved from commercial APIs to open-weight models self-hosted on 56 NVIDIA B200 GPUs. Treated as a slow drift of the neobank channel gain, falling token cost leaves the margin unchanged when compute is customer-paid; when compute is vendor-financed, the margin ranges from 7.2 to 2.9 depending on the price elasticity of token demand. Neobanks thus act either as a stabilising feedforward channel or as an additional amplifier, depending on who pays for their compute. The bank’s own compute channel remains stable for any coupling, but vendor financing transmits NVIDIA-specific shocks to it: together, the election shock and the hypothetical risk cut its compute purchases by about 3.7%, with the effect building up over about half a year.
Keywords: neobanks; Revolut; AI compute market; stability margin; positive feedback; circular financing; characteristic equation; root locus; block diagram; system identification; token cost; price elasticity; parametric drift; robust stability; NVIDIA.
BibTeX
@misc{desyatirikov2026circular,
title = {Dynamic Model of Circular Financing in the AI Compute Market: Stability with Neobanks in the Loop},
author = {Desyatirikov, Felix},
year = {2026},
note = {Manuscript}
}