AI compute is becoming an asset class. Its risk infrastructure hasn't caught up.
The same chips, the same issuer, seven weeks apart — priced 225 basis points and five rating notches apart, driven by what lenders could and couldn't underwrite. Eleven weeks after that, another 100 basis points wider at identical ratings, because the facility outlives its customer contracts. ComputeIQ is the independent underwriting and surveillance layer for AI compute finance: evidence-linked analytics across the full stack, from silicon and memory to power, offtake contracts, and the financing market itself.
Built on the AIMADDS evidence-enforced diligence platform. Early access for compute borrowers, lenders, insurers, and syndication participants.
Underwrite the compute, then keep watching it.
A $2B AI factory is simultaneously a technology asset, a power asset, a data-center asset, an equipment financing, and an obsolescence bet. No existing underwriting stack was designed for all of them at once.
Compute-financing deal tracker with verified comps
A living database of GPU- and HPC-backed facilities built from primary sources — SEC filings, issuer releases, rating disclosures — with every entry labeled verified or reported, never blended. The headline observable it maintains: investment-grade offtake prices at SOFR+225 across independent issuers, non-IG at SOFR+450. That 225bps differential is the market's price on underwritable information.
Full-stack technology intelligence
A registry of 38 components across 12 layers — silicon, memory, packaging, networking, cooling, electrical, power, logistics, offtake — each mapped to the credit variables rating agencies publish and monitored through citable public signals. Obsolescence, supply bottlenecks, and cost drivers surface as credit-relevant watch items, not technology trivia.
Covenant packages read from the agreement, not the summary
Whole credit agreements go through the system rather than the filing summary — reserves, cash sweeps, events of default, and the financial tests, each quoted to its section. On CoreWeave’s August 2026 facility that surfaced a Minimum Liquidity covenant in §6.12(b) that appears in no 8-K and no press coverage. Summaries lose the covenants that matter.
Post-close surveillance tied to each facility’s own covenants
Sweeps re-underwrite tracked facilities against fresh market, supply-chain and filing data, and diff against the previous sweep so what you read is what changed. Each alert names the covenant or credit variable it bears on. It never asserts a breach — that needs borrower financials we do not hold — it tells you what moved and what to re-test.
Facility teardowns with evidence-linked attribution
Side-by-side decomposition of comparable financings — same issuer, same collateral class, hundreds of basis points apart — with spread attribution expressed as ranges, every driver grounded on a quoted fact difference, and rating-notch arithmetic computed in code rather than trusted to a model.
KBRA-dimension re-leasing analysis
The re-leasing question — will anyone rent this capacity at an attractive price after the contract ends — assessed on the seven dimensions in KBRA's published framework: workload suitability, same-customer extension, third-party demand, chip fungibility, switching costs, power cost per compute, and refresh strategy.
Residual scenarios as ranges, never points
There is no liquid print for most compute residuals, so point estimates are fabricated precision. ComputeIQ produces base, stress, and severe bands across 12–60 month horizons with explicit, user-overridable driver assumptions — and widens the bands automatically when market benchmarks are absent.
The Financing-Ready Data Package
The borrower-side product: standardized, lender-ready diligence structured on KBRA's published variables — offtake counterparty credit, re-leasing, residuals, power and site — plus the data-completeness gap list: the document request a lender would issue, produced before the lender asks.
The building has a refresh cycle now.
Rack power density is rising an order of magnitude across accelerator generations — roughly 40kW to 130kW to 600kW-class. A facility engineered for air-cooled racks cannot host next-generation compute without major retrofit, so re-leasing risk now applies to the shell, not just the chips. Most of the $61B datacenter securitization market still rates this collateral on stabilized-real-estate logic. ComputeIQ puts facility obsolescence on the underwriting map.
Verified vs reported is a hard split — press-sourced facts are never presented as confirmed
No composite scores: decomposed, evidence-linked metrics mapped to rating-agency variables
Residual values ship as ranges with overridable drivers — never a confident single number
Anything a model cannot ground in provided data lands in a verification list, not in the analysis
Raising against compute? Lending against it?
Borrowers use ComputeIQ to become financeable before approaching lenders. Lenders and insurers use it to underwrite what traditional credit systems were never built to model.
Talk to us about Early Access