Research · 2026-07-18

Research

Asset Economics

Earning power per spec-sheet unit, how it ages, and how it compares to what the hardware cost.

Chip Economics

The daily rate re-denominated per GB, per TB/s, per petaFLOP (PFLOP, 10^15 floating-point operations per second), per kW. One lens makes the cross-section flat.

Open →

Rent and MSRP: Five Generations of Posted Prices Every generation’s gross posted-rate revenue passed its manufacturer’s suggested retail price (MSRP) anchor within 11–17 months; acquisition cost per unit of bandwidth fell 2.5× while rent per unit did not. GPU Rental Rates by Generation: Nine Years of Posted Prices Per-generation posted-rate histories, 2017–2026: declines arrive in steps set by the repricers, interrupted twice by demand — in 2021–23, and again from late 2025. Rent and Age: Five Generations The July 6, 2026 cross-section: ≈−26%/yr in dollars (R² 0.96) with age, flat per unit of memory bandwidth, banded per watt.

Term & Price Structure

On-demand is one point on a surface — the rate broken out by tenor and by where it actually transacts.

Monitor · updated daily

Committed Term Structure

Posted committed asks by tenor — guaranteed capacity from one month to a year, per chip.

Open → Monitor · updated daily

On-Demand vs Clearing Basis

Posted on-demand asks against marketplace clearing — where list rates sit above what transacts.

Open →

The Commitment Discount: Three Years of H100 Posted Prices Neocloud H100 posted prices, Oct 2023 – Jun 2026: a persistent committed discount, a $2.80 on-demand floor through 2025, a turn upward since December.

Demand & Utilization

Rate cards are asks. These read demand composition and realized economics from filings and from the fleet itself.

Monitor · updated daily

Realized vs Rate Card

Per-GPU-hour realization estimated from public filings — well below posted rate cards.

Open → Monitor · updated daily

Utilization & Availability

How much of the tracked fleet is rented versus idle, and how many venues have capacity.

Open →

Training, Inference, and What Compute Earns A primer for credit readers: training demand behaves like construction backlog, inference like utility load — with the observables for each.

Contracts & Credit Structure

The paper behind the prices — what a compute contract binds, how lenders ring-fence the fleet, and the vocabulary both are written in. Read alongside the credit ledger.

Monitor · updated daily

GPU-Backed Credit

The debt stack behind AI compute, as filed — every instrument traced to a public document.

Open → Monitor · updated daily

Credit Reference Series

The derived layer: spread at signing, refi velocity, and lender-filed fair-value marks — arithmetic on filed numbers only.

Open →

A GPU Lease Rate Enters the Rating Case: Fitch’s BB+ on CoreWeave’s DDTL 5.5 Fitch’s July 16 rating case for the proposed $2.6B facility leads with a new assumption: customer contracts that expire before the debt matures are renewed or replaced at favorable GPU lease rates — language absent from the April and May actions. Same Facility, Different Marks: Compute Credit in Lender Filings Four advisers file 99.0 / 100.0 / 102.0 / 103.1 on one GPU-backed facility at the same date. A 16-tranche cross-sector control shows performing credit normally agrees within a quarter point — compute paper is marked like performing credit and dispersed like stressed credit. The GPU SPV: How Lenders Take Compute as Collateral The filed credit agreements behind the largest GPU special-purpose vehicle (SPV) financings: serial-number covenants, Uniform Commercial Code (UCC) account control, contract-realization tests — and a rating that travels with the offtake, six notches above the parent. The Compute Contract: From On-Demand Terms to Take-or-Pay How GPU-hours are actually sold, from published terms of service to U.S. Securities and Exchange Commission (SEC)-filed master services agreements — including the termination formula that computes damages in the on-demand spread. A Compute Glossary for Finance Professionals FLOPS, bandwidth, TDP, PUE, MFU, goodput, NVL72, training vs. inference vs. edge — the vocabulary of compute paper, defined in finance terms.

Precedents & History

Prior markets that financed expensive, rapidly improving machines — documented from the contemporaneous record.

The Company That Insured Itself Against Technological Change Itel, 1979: the largest independent computer lessor in the world, residual-dependent leases insured at Lloyd’s, and the IBM 4300 — a $433M loss ($2B today). That era’s structures were far more aggressive than today’s: GPU credit amortizes its collateral to zero and references no market residual.