Sources and checks
Noach Ark displays public observations from provider pricing pages, documented APIs, manufacturer specifications and published benchmark reports. An external collection workflow checks the sources; the application validates and renders its records. Source URLs and actual check times accompany the data. Advertised prices and order options are not checkout quotes or capacity guarantees.
Refresh attempts and failures
The server checks Vast marketplace listings hourly and OVHcloud, Scaleway and SaladCloud public rate cards daily. The daily source-review workflow covers the remaining providers. Regional catalog collection accounts for published GPU-count tiers and zones; unresolved configurations stay in a review queue. Refresh attempts do not guarantee a successful check or complete marketplace coverage. Each successful source check records its own UTC timestamp. Failed or gated requests leave the previous price and timestamp unchanged. Each offer has a freshness limit of 3, 24 or 72 hours; API observations use three hours. Missing, future and expired checks are excluded from current price summaries. Detail pages retain dated observations and mark stale information.
How summary prices are selected
For each GPU and rental term, summaries prefer listed prices with a verified GPU count, then starting prices with a verified count, then optional reference prices with a verified count, then optional rates with an unknown instance size. Within the first nonempty group, the lowest compute price for the selected whole-instance or GPU-hour basis wins; provider name and offer ID break ties. Reference fallbacks are included by default, labeled, and can be excluded. They never carry a lowest-listed-price label. Retired, unavailable, pre-order and inactive records do not contribute to summaries. The GPU can remain discoverable when it has no recent eligible rate.
Comparison results put recent listed prices with verified instance counts first, sorted by the selected whole-instance or per-GPU compute price. Other observations appear separately; reference and starting prices are included by default and can be excluded. A lowest-listed-instance-price label on the comparison requires the same GPU, GPU count, rental term and known minimum commitment. This label is a compute-price comparison, not a recommendation about reliability, performance or all-in cost. Provider buttons, provider names within offers and offer prices open the provider directly. Exact offer links are used when supplied; otherwise the action is labeled Browse. Provider navigation includes ref=noachark for attribution. Evidence and CMS links retain their exact destinations, including signed API query parameters.
On-demand and interruptible spot columns select providers independently. A provider's spot rate can exceed another provider's on-demand rate. Listing and provider counts describe recent tracked coverage, not the entire market. Provider summaries prefer hourly on-demand coverage and label other rental terms when that coverage is absent.
Marketplace sampling and unusual prices
We track identified host listings and separately labeled starting-price snapshots. Host configuration, location and quoted disk selection affect a marketplace rate. An aggregate quote is not assigned one host's resources. One listing is a sample, even when checked repeatedly. Unusual rates and material changes are checked against the precise source before importing; valid expensive listings are retained rather than hidden by comparisons with other GPUs.
Units, commitments and unknown costs
Source currencies stay separate; no exchange rate is applied. Hourly, monthly and longer-term monthly-equivalent offers remain separate. Per-GPU prices and confirmed whole-instance totals are both shown. Unknown GPU counts do not produce instance totals. Exact per-minute source rates are multiplied by 60 for a labeled hourly equivalent, retaining the native amount and billing unit. Approximate monthly calculator totals are not monthly rental plans. Taxes, storage, traffic, setup, public IP and other required charges can be additional. A known transfer rate does not establish an allowance or a complete workload charge.
History and corrections
Successful later checks append observations, including unchanged prices. No earlier history is estimated or backfilled. Changed configurations or billing identities require separate offer records. Retired offers retain their history. Source-backed corrections update current records through validation and a new check; previous observations are not silently rewritten.
Benchmarks and memory estimates
Benchmark evidence retains the model/checkpoint, GPU variant and count, engine, precision, input/output lengths, system and publication or measurement date. Aggregate maximum-load output throughput is not single-user generation speed. Token-cost scenarios combine matched rental configuration prices with published throughput; they do not establish measured rental performance, utilization, latency or an all-in bill.
Memory scenarios use pinned checkpoint weights and supported cache configurations for an explicit token count and number of resident sequences. Runtime overhead is additional. A subtotal below a GPU's nominal memory is not a guarantee that the workload fits or that memory is pooled across devices.
Finder scope and empty results
The V1 finder sizes the entire model and cache on one GPU. It prices confirmed multi-GPU instances in full, but does not evaluate splitting the model across them. Empty results distinguish no eligible observed offers from an unevaluated distributed deployment strategy; neither means the model cannot run. Memory fit, performance evidence, pricing completeness and availability are separate checks.
Deployment-aware finder and V2 API
The website and V2 API share one evaluation service. Original BF16 precision is the default. Smaller-memory FP8 alternatives require opt-in and retain their own pinned checkpoint, exact resident tensor-byte accounting, source evidence and quality qualifications. They appear separately from original-precision deployments. The original single-GPU model-page calculator and V1 API retain their existing sizing and cost semantics.
Supported deployment profiles pin the engine version, GPU variant/count, parallelism, runtime and required instance topology. Only single-GPU or single-node tensor parallelism with exactly 2, 4 or 8 GPUs is curated. Dense matrices and vocabulary are sharded according to the pinned engine; norms and quantization metadata are replicated. K/V projection allocation follows the KV-head sharding/replication rules. BF16 KV cache and the editable runtime reserve are calculated per GPU. No offload, pipeline parallelism, expert parallelism, multi-node deployment or invented combination of listings is assumed. All MoE weights remain resident even when only some experts are active.
Recommendations require exact instance configuration and topology evidence checked within the preceding 72 hours. GPU specifications alone cannot prove rental topology. Deployment checks are refreshed only after actually rechecking the cited sources. Missing or stale instance evidence excludes a recommendation. Profiles are curated estimates; driver/CUDA compatibility, checkpoint access, actual peak memory, buffers and allocator overhead remain runtime checks.
Response targets and decision evidence
First-token delay and per-active-user output speed targets require explicit input/output lengths, whose sum becomes the cached-token assumption. Conflicting totals are rejected. Evidence must match the pinned checkpoint, precision, engine version, GPU variant/count/topology, lengths, concurrency and measurement definitions. Interactive targets use p95 first-token delay and p5 per-user output tokens per second. Aggregate maximum-load capacity benchmarks cannot satisfy an interactive target. Missing measurements produce Performance target not evaluated; those options remain visible for memory and price comparison.
Result states distinguish matching estimated deployments, no eligible observed offers, an unsupported deployment strategy and missing sizing evidence. Exclusion summaries give one primary reason per observed listing, not a statement about all providers. Memory feasibility, performance evidence, pricing completeness and availability are separate. Ranking uses full-instance compute cost within each precision group. Priced extras, unknown charges and billing qualifications remain beside a partial subtotal; it is never a complete quote. Scenario links preserve the workload and chosen profile and re-evaluate current evidence.
Limited market coverage
Single tracked listing means one eligible recent observation for the GPU and rental term under the selected summary policy. It is a provider quote, not a market average or evidence that no other rentals exist. Repeated checks do not turn one listing into broad coverage. Unusual prices are checked against their sources and retained when valid.
Benchmarked-GPU summaries match the benchmark's exact GPU variant and instance count. Derived token-cost scenarios use three significant figures and show the compute price's check time and age; source precision remains stored. Regional rows combine only when configuration, price, terms, charges and qualifications match, retaining each offer's evidence and history link. Country and provider zone filters are separate from broad coverage labels, which remain visible in source details.
Published price confidence and historical coverage
Listed price means an exact published price for the stated GPU configuration and billing basis, including official rate cards. Unknown stock or minimum commitment is qualified separately; neither alone makes a confirmed published price reference-only. Unknown commitments cannot establish a qualifying no-commitment rental for the finder or Product markup.
Every active GPU remains discoverable. When no current USD hourly offer qualifies, the directory may show its last dated quote with source age, price basis and listing status. Such historical quotes never enter current price rankings, benchmark cost scenarios or Product offers. A current-quotes-only filter is available. A one-listing median is that listing's price, not a market benchmark.
The directory and comparison can rank whole-instance prices or GPU-hour rates. Required bundle sizes and complete instance costs remain visible. Display formatting uses three significant digits below one dollar and two decimals at or above one dollar. Source Decimal values, calculation precision and numeric sorting remain unchanged.
Saved evidence and rental eligibility
Saved pricing evidence records the exact marketplace contract or rate-card configuration, check time, native units and source fingerprint for that observation. Live provider links are separate and may change. Historical aliases never enter current results. Verified terms can fill a previously unknown commitment on the same configuration without changing its offer URL; the earlier unknown value remains in history.
Finder country counts describe eligible listed rentals for the selected currency, billing term and commitment before memory and deployment-fit checks. A saved country filter remains valid when coverage is zero. Provider-wide terms must be refreshed within 72 hours; unknown or expired commitment evidence cannot qualify a rental. Catalog and public-page audits check evidence, identities, units and API agreement independently of collection.