Reckon Research builds the intelligence that decides where intelligence runs.
The world spends its largest capital line on silicon, and workloads land on that silicon by default, habit, and vendor lock. Compute is traded as a commodity, dollars per GPU-hour, but consumed as a service: useful work under latency, cost, and energy constraints. The gap between those two framings is large and unmeasured. The same workload can cost multiples more in one placement than another while meeting the same service target.
Frontier labs train models. Providers rent chips. We build the allocator: a learned world model of computational performance, and the policies that act on it.
The corpus
Workload-conditional performance data across silicon, software stacks, providers, prices, and time. Hyperscalers hold fragments they cannot publish. A neutral lab is the only actor that can assemble it with open provenance.
The environment
Live compute markets are a decision environment with measurable consequences. Placements, sizings, and migrations produce realized costs, met or missed targets, and counterfactuals.
The verifier
Deterministic ground truth: did this placement meet its p99 at this cost on this silicon. Our verification apparatus is built before our claims are.
The calibrated foundation: auditable analytical models fitted from measured traces, and the placement-premium index, published with confidence intervals and raw data.
Performance surrogates: learned models that predict latency, cost, and energy for unseen combinations of workload, silicon, and stack, with uncertainty honest enough to defer to measurement.
Allocation policies: trained in corpus-calibrated simulation against the verifier, deployed first in shadow, placing and migrating under uncertainty better than static rules.
Allocation as a market: mechanism design for heterogeneous compute, and assay and settlement research for financialized compute markets.
Our first instrument: placement primitives for inference engineers. Profile a workload, estimate it across every chip and provider, place and migrate it under a policy you write in plain Python. Apache-2.0; the core is standard-library only so every estimate is auditable by hand.
- Every published number ships with its raw traces.
- Methodology and validators are open; disputes with traces attached outrank everything else.
- Unfavorable results publish. Kill criteria are set before experiments run.
- Hardware access never buys editorial input; vendor data is labeled and independently replicated.
- Serve open-weight models: use berth, and tell us where its numbers are wrong, with traces.
- Operate hardware or a cloud: contribute measurement access under our published governance rules.
- Research performance modeling, RL for systems, or market design: a genuine open problem, a deterministic verifier, a proprietary corpus, and a live environment.
hello@reckonresearch.com