Glossary
Speak AI compute.
- Accelerator
- Any specialized chip (GPU, TPU, AI accelerator) used for training or inference.
- Availability
- Percentage of a provider's advertised capacity that is currently deployable.
- B200
- NVIDIA's Blackwell-generation datacenter GPU, ~2x H100 training throughput.
- Batch inference
- Serving many inference requests together to increase throughput and reduce cost per token.
- Blended spot
- Volume-weighted average of spot prices across all listed providers for a given accelerator.
- Compute reservation
- Pre-purchased commitment for a defined quantity of accelerators over a defined term.
- Cost per token
- Normalized inference cost expressed per million output tokens for a target model.
- GDPR
- EU regulation governing personal data processing; central to EU-sovereign compute procurement.
- H100
- NVIDIA's Hopper-generation datacenter GPU; workhorse of 2023–2025 training.
- H200
- H100 refresh with 141GB HBM3e memory, improving long-context and inference throughput.
- Hyperscaler
- Very large cloud operator (AWS, Azure, GCP) with global footprint and enterprise contracts.
- Interconnect
- Network fabric linking accelerators within and across nodes (NVLink, InfiniBand, EFA, OCS).
- Neocloud
- Post-2020 GPU-specialist cloud built for AI-first workloads.
- NVLink
- NVIDIA's high-bandwidth GPU-to-GPU interconnect, up to 3.2 Tb/s aggregate on H100 platforms.
- Reserved capacity
- Committed inventory guaranteed by contract; discounted vs on-demand and spot.
- RFQ
- Request for quote — structured procurement flow used for large or enterprise compute purchases.
- SLA
- Service-level agreement covering uptime, response time, and remediation credits.
- Sovereign cloud
- A cloud whose ownership, operation, and jurisdiction all sit within a target legal region.
- Spot
- Short-term, discounted capacity that can be reclaimed by the provider on short notice.
- Tier 1
- Established GPU cloud (CoreWeave, Lambda, Nebius) with enterprise-grade operations and SLAs.
- TPU v5p
- Google's fifth-generation Tensor Processing Unit for large-scale training and inference.
- Training
- The compute-intensive phase of building or fine-tuning a model.
- Uptime
- Percentage of time a service is available; typically expressed as 99.9%–99.999%.
- Volatility
- Rolling standard deviation of price used to measure market stability.