Pipeshift pricing, plans and limits

Choose Pipeshift when a working open-weight model has to hit a hard latency budget, especially voice agents with a hard first-token budget, and you accept a sales-led start. With no rate card and a small balance sheet, treat it as a scoped engagement and price Baseten or Fireworks AI first.

Pipeshift is a managed inference cloud founded in 2024 that serves open-weight, custom and fine-tuned models in single-tenant clusters or inside a customer VPC. Its homepage claims 99.999% cluster uptime across 10+ regions, and partners including Neysa in India and Armada's Bridge Marketplace resell or host the service.

About Pipeshift

Pipeshift runs production inference for teams whose workloads are latency-bound rather than throughput-bound. The pitch is narrow on purpose: real-time agents, particularly voice, where a slow first token is the difference between a usable product and an unusable one. What it sells is managed clusters and a serving runtime, not raw capacity you tune yourself.

The differentiator is the runtime layer rather than the metal. Its in-house framework, MAGIC (Modular Architecture for GPU Inference Clusters), compiles a serving pipeline per workload, with custom kernels, speculative decoding, KV caching and GPU scheduling, then autoscales clusters and releases them when idle. That puts it between raw GPU rental, the model CoreWeave sells, and a shared hosted API such as the one at Fireworks AI. The homepage quotes a NetApp engineering director crediting over 500 tokens per second without compression or quantization, and sets a first-token SLA for its voice-agent page.

Deployment comes in two forms: dedicated single-tenant clusters on Pipeshift Cloud, or self-hosted inside the customer's own VPC. Three partner announcements widen where it runs. Neysa (27 May 2026) gives it an India-based cloud so Indian teams can keep inference in-country. Armada's Bridge Marketplace (19 May 2026) lists it for Armada customers. Atlas TTS (28 September 2026), a 1.7B-parameter speech model from runAtlas, is offered as a managed endpoint or a VPC install.

The newest evidence is a benchmark Pipeshift published on 26 October 2026. On a single Hopper-generation GPU serving Qwen3-TTS, it reports p95 time to first audio rising from 24.6 ms at 1 request a second to 33.7 ms at 40, with 0 of 41,337 requests failing, while SGLang-Omni, vLLM-Omni and Nari Labs' engine slowed and then failed under load. Treat that as a vendor-run result on one model and one GPU, not an independent comparison.

The other side is straightforward. If you want a rate card, a card-on-file signup and prices you can model today, Published rates and self-serve signup are what Baseten adds as the nearest managed rival. Modal and Replicate cover bursty custom models, Groq sells fast hosted inference on its own chips, and Crusoe rents the GPUs outright. A team that only needs speech output can skip the serving stack entirely with ElevenLabs as a finished API.

The company was founded in 2024, with Arko C as CEO and Enrique Ferrao as CTO, and the site footer names Infercloud Inc. as the operating entity. Reported funding is roughly $3.5M, which is early-stage money for infrastructure and the honest frame for anyone weighing it against an established provider.

Plans and pricing

Pipeshift publishes no rate card: pipeshift.com/pricing returned a 404 on 28 October 2026, and every call to action on the site is a Talk to an engineer booking link. Atlas TTS shows the two commercial shapes the vendor describes: Pipeshift bills a managed endpoint, while a VPC install uses an annual runAtlas licence plus GPUs you provide, with no per-character meter. Plan on a scoping call before you can model spend, and check the published per-token rates at Together AI before you book that call.

Key Features

• MAGIC serving runtime: Compiles a workload-specific pipeline with custom kernels, speculative decoding and KV caching; the homepage quotes a NetApp director citing over 500 tokens per second without quantization.
• Latency SLA for voice agents: Lists a voice SLA of under 100 milliseconds to first token and runs speech-to-text, LLM and text-to-speech together on one pod.
• Scale-to-zero autoscaling: Scales on a customer-defined SLA, with fast cold starts and GPU bin-packing so idle accelerators are released between bursts.
• Single-tenant or VPC deployment: Runs dedicated clusters on Pipeshift Cloud across 10+ regions, or self-hosted inside the buyer's own VPC.
• Observability and access control: Tracks model API metrics, cost and GPU and CPU use in one console, with RBAC team settings.
• Forward deployed engineers: Puts engineers on the customer's serving stack to build, optimise and scale it, a hands-on motion suited to a few large accounts.
• Partner channels: Runs on Neysa's India-based cloud (May 2026), is listed on Armada's Bridge Marketplace (May 2026) and hosts runAtlas's Atlas TTS (September 2026).

Strengths

• Assigns forward deployed engineers to the account, a level of hands-on serving optimisation that self-serve inference providers do not offer at any tier.
• Runs your chosen open-weight or fine-tuned model rather than a proprietary endpoint, so the serving layer can be changed without rewriting against a new model API.
• Offers a VPC route as well as managed clusters, and for Atlas TTS keeps one request shape across both, so a pilot can move into your own account unchanged.

Weaknesses

• No published pricing anywhere on the site, so total cost cannot be modelled before a sales conversation and a scoping call.
• Reported funding of about $3.5M makes it a small vendor to place production inference on, next to providers with far deeper balance sheets.
• SOC 2 is the only attestation named, with no HIPAA or ISO 27001 claim, which rules it out for regulated healthcare workloads without further diligence.
• Its headline latency result is self-published on one model and one GPU, and the homepage gives 99.999% cluster uptime in one place and 99.99% in another.