Who actually builds each part

Market default; its API shape became the industry’s de facto interface.

Strong enterprise reputation for coding, long context, and safety posture.

Multimodal from the ground up, bundled with Google’s cloud and TPU hardware.

Challenger scaling aggressively on its own colossal GPU clusters.

Frontier-quality models with an open-weights heritage; the sovereignty option for Europe.

The family that made open weights a serious enterprise option; enormous ecosystem.

Prolific family across sizes; often the benchmark leader among open weights.

Proved frontier-adjacent reasoning at a fraction of assumed cost; reset price expectations.

Google’s open line, distilled from Gemini research; strong at small sizes.

Compact, efficient weights with permissive licensing — a self-hosting favorite.

PagedAttention made GPU memory manageable; the engine under most self-hosted serving.

Co-designed with its RadixAttention cache sharing; a favorite for structured agent workloads.

Squeezes maximum throughput from NVIDIA hardware; the vendor’s own path.

Splits reading (prefill) and writing (decode) onto different GPUs at datacenter scale.

Runs models on a laptop in one command; where most developers first touch local AI.

The earlier generation of open serving; still deployed, now largely legacy beside vLLM.

Spills cache from GPU memory to CPU, disk, and network so prefixes survive across requests.

A distributed KV-cache pool proven at Kimi’s production scale; now open source.

NVIDIA’s plumbing for moving cache between GPUs and storage at wire speed.

The physical substrate: scarce on-GPU memory where the cache lives. Its supply is an industry bottleneck.

Frontier and open models behind AWS identity, billing, and compliance.

Model catalog plus agent tooling inside the Azure enterprise estate.

Gemini plus third-party models on Google’s TPU/GPU infrastructure.

Fast serving of open weights; also rents dedicated GPU clusters.

Performance-obsessed inference platform; compound-AI positioning.

Production serving with your own autoscaling GPU capacity, minus the platform team.