Inference that is
ApexMarket splits batch AI workloads across spot providers, verifies every shred against a trusted Nebius Nemotron reference anchor, and releases payment only when proof succeeds.
Proof Before Payment with Reference Anchor Security
Built on unquantized reference models, deterministic cryptographic verification, and sub-cent double-entry ledger balance conservation.
Nebius Nemotron-3-Nano Anchor
Unquantized reference anchor (nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B) running at temperature 0.0 with 5% secret-seeded CSPRNG sampling and logprob drift detection to catch silent provider model substitutions.
5-Tier Proof Before Payment
Every batch shred passes offline tokenizer billing, rotated canary probes, reference anchor sampling, cross-entropy logprob scoring, and a Nemotron semantic arbiter before payment escrow is released.
32.18% Cost Arbitrage
Empirically validated multi-provider cost arbitrage drops per-query costs from $0.001361 to $0.000923, delivering 32.18% net savings even after accounting for full 9% verification audit overhead.
32.18% Verified Cost Savings Without Quality Loss
Measured across empirical benchmark runs on identical NVIDIA Nemotron open-weight models. High throughput and 52ms TTFT with zero vendor lock-in.
Arbitrage Cost Matrix vs Hyperscaler Menus
Calculates empirical savings across open-weight architectures, accounting for the 9% verification audit overhead while delivering real-time streaming instead of 24-hour cold batch queues.
On-Demand Single Provider
Standard hyperscaler pay-as-you-go API rate. Immediate execution with zero verification.
- •Single provider lock-in
- •0% Verification / No canary audit
- •100% retail menu cost
Batch API Baseline
Standard 50% discount offered by commercial providers, constrained by 24-hour SLA delays.
- 24-hour turnaround delay
- •Offline queue with zero latency SLA
- •Silent quantization risks unmitigated
ApexMarket Verified
Empirical cost arbitrage combining RunPod spot nodes and Nebius reference anchor with full 9% audit overhead included.
- Real-Time Execution (<120ms TTFT)
- 5-Tier Proof Before Payment Audit
- Nebius Nemotron-3-Nano Anchor
Saves $0.24 / month while eliminating the 24-hour turnaround delay of Batch APIs.
Interactive 5-Tier Protected Pipeline Simulator
Step through how ApexMarket protects enterprise batch workloads: capacity-aware greedy routing, active 4% HMAC canary challenges, and INT4 drift failover to the Nebius reference anchor.
Stage 1: Batch Ingestion & Shredding
Batch Ingestion Queue: 50 Prompts
Offline HuggingFace AutoTokenizer · PII Redaction Active
Prompts 00-09
Prompts 10-19
Prompts 20-29
Prompts 30-39
Prompts 40-49
Stage 1 Details: 50 raw enterprise prompts ingested, PII sanitized, and partitioned into deterministic micro-batches via offline HuggingFace AutoTokenizer.
Nominal Production FlowCheap capacity exists. Trusting it is the hard part.
Routing by price alone is easy. Knowing the answer is right, and paying fairly for it, is what ApexMarket adds.
Same model, different price
Providers publish different per-token rates for identical open-weight models. A fixed default provider leaves that gap unclaimed.
Same name, different quality
Quantization and serving differences can change outputs with no flag in the API. Cheap capacity is only useful if it is checked.
Fees eat tiny payments
A payout fee can exceed the work itself. ApexMarket keeps its own ledger and settles in batches.
Four steps from raw batch to verified payout.
Shred
Your batch is split into small shreds. Secrets and personal data are redacted before anything leaves, and jobs default to Confidential, which stays on the trusted Nebius anchor.
Five checks stand between a node and its money.
Open a tier to see what it catches. Try the demo to see what a failed node earns.
Token counts come from a local tokenizer, never the provider's own usage numbers. Outputs are scanned for injection payloads and unsafe artifacts.
Hidden prompts with known answers are mixed into each shred. Two or more failures fail the shred and slash its earnings.
A secret-seeded 5% of real prompts also run on the trusted Nebius anchor at temperature 0, then get compared.
The anchor scores each candidate output. A spike in negative log-likelihood points to heavy quantization or model substitution.
A Nemotron model arbitrates borderline shreds that the other checks cannot settle.