Inference

What Inference does

Inference (legal entity Inference R&D, Inc., dba Inference.net) is a San Francisco-based AI infrastructure company founded in 2023 and operating with 1-10 employees. It operates a distributed GPU cluster for LLM inference, built on NVIDIA B200 hardware (180 GiB VRAM) deployed across global data centers, with roots in the Solana ecosystem as reflected by lead investor Crypto Startup Accelerator. The company sells GPU compute, open-weight model hosting, and a full-stack LLM lifecycle platform called Catalyst, which combines monitoring (Observe), agent tracing (Trace), fine-tuning (Train), evaluation (Evaluate), and deployment (Deploy) into a single workflow that uses production traffic as supervised fine-tuning data to create self-improving specialized models. Inference also publishes proprietary open-weights models (Schematron V2 Turbo for structured output, ClipTagger-12B for video understanding) and an open-source agent optimization tool (HALO), and supports OpenAI/Anthropic-compatible APIs with a CLI-based instrumentation flow.

The commercial model is usage-based and product-led: customers access a self-serve dashboard with a $25 free credit, consume B200 GPU instances at $9.98/hr, and pay per 1K requests for optimized inference (around $0.12). Catalyst training runs are priced between $25 and several thousand dollars, with training and deployment costs currently waived during the public beta. An enterprise motion is layered on top via a "Talk to an Engineer" CTA. The platform is SOC 2 Type II certified, sells globally, and serves a mix of AI-native engineering teams, agent builders, and domain-specific application developers; named customers include Cal AI, Gravity Ads (87.3M requests/mo), Wynd Labs (Grass), Olive, and Rizz. In October 2025, Inference closed an $11.8M funding round led by Multicoin Capital and Crypto Startup Accelerator.

Name

Inference

Legal name

Inference R&D, Inc.

Website

https://inference.net

Company type

Private

Founded year

2024

Operating status

Operating

Headcount range

1–10 employees

Short description

Inference.net is a San Francisco AI infrastructure company that runs a distributed GPU cluster for LLM inference on NVIDIA B200s and offers Catalyst, a full-stack platform for serving, tracing, evaluating, and fine-tuning self-improving specialized language models for AI-native engineering teams and enterprise application builders.

Ownership category

akta.pro rank

Product category

AI Inference Infrastructure

NAICS

Computing Infrastructure Providers, Data Processing, Web Hosting, and Related Services (51821), Software Publishers (5132), Computer Systems Design Services (541512)

SIC

Services-Prepackaged Software (7372), Services-Computer Programming, Data Processing, Etc. (7370)

akta.pro primary industry

Model Deployment, Serving & Inference Platforms (HDAAABAF)

akta.pro secondary industries

Model Hosting, Serving & Inference Platforms (HDAAACAB), GPU-Accelerated & AI Training/Inference Servers (HDACABAG), Fine-Tuning, Adaptation & Custom Model Training (PEFT/LoRA/RLHF) (HDAAACAC)

Keywords

• LLM inference infrastructure
• GPU cloud hosting
• AI model deployment
• LLM fine-tuning services
• AI agent observability

Headquarters

HQ city

San Francisco

HQ country

United States

HQ region

North America

Offices1 record

Markets served

GTM type

B2B

Offering type

Software

Cost components

Technology or R&D, Infrastructure, Personnel, Marketing or Sales, Operations

Revenue model

1. Model Deployment (Usage-based): Pay-per-hour for GPU compute capacity to deploy and serve models. Based on B200 instances at $9.98/hr with 180 GiB VRAM. Revenue scales with customer usage and model deployment duration.
2. Subscription Plans: Automatically renewing subscription plans for ongoing platform access. Subscription fees billed monthly with 3-day cancellation notice requirement.
3. Catalyst Training Runs: Custom model training services ranging from $25 to several thousand dollars depending on base model size and data requirements. During public beta, training and deployment costs are covered.
4. Model Evaluation: Continuous evaluation of models against production traces, billed as part of platform usage.

Pricing tiers

Go-to-market motion1 record

Distribution channels4 records

Marketing channels6 records

Core offering

Inference operates a distributed GPU cluster (NVIDIA B200, 180 GiB VRAM) that delivers fully managed LLM inference, fine-tuning, evaluation, tracing, and observability as a unified platform. Its Catalyst workflow combines monitoring, evaluation, training, and deployment into a single product that uses production traffic as supervised fine-tuning data, enabling customers to run open-weight and custom models at 50–90% lower cost than frontier APIs.

Product overview

Inference.net is a full-stack LLM lifecycle platform providing a unified portfolio of six integrated products — Deploy (model hosting), Observe (monitoring), Trace (agent tracing), Train (fine-tuning), Evaluate (benchmarks), and HALO (agent optimization) — orchestrated through the Catalyst workflow. The platform's core value proposition is using production traffic as the training environment for supervised fine-tuning, enabling specialized language models that match or exceed frontier-quality at significantly lower cost. In addition to its platform tools, Inference.net publishes proprietary models including Schematron V2 Turbo (structured output/reasoning) and ClipTagger-12B (vision-language/video understanding).

Differentiator

Problem solved

Functional benefit

Products and services

• Deploy Fully managed, global, turn-key AI infrastructure for high-performance model hosting. Serves open-source, custom, and fine-tuned LLMs at massive scale with 99.99% uptime across public cloud, private cloud, or hybrid environments; priced at $9.98/hr per B200 instance.
• Observe Production-grade LLM monitoring with continuous benchmarking across any model and provider. Tracks latency, reliability, usage patterns, and quality signals with dashboards for cost, errors, and aggregate spend.
• Trace Traces every step AI agents take, capturing LLM calls, tool calls, and framework steps. Enables full request/response payloads, token counts, and downstream provider behavior in a unified view.
• Train Fine-tuning platform for training specialized language models tuned to quality, cost, and latency targets. Supports automatic fine-tuning workflows, curated training data generation, and continuous improvement loops; training runs range from $25 up to several thousand dollars depending on base model size and data volume.
• Evaluate Model evaluation platform using rigorous benchmarks and LLM-as-a-judge methodology. Converts production traces into evaluation datasets, scores quality across models and metrics, and enables automated regression detection.
• Catalyst Full-stack platform combining monitoring, evaluations, training, and deployment into a single product for building continuously self-improving AI models. Uses production traffic as supervised fine-tuning data, installs like an ordinary LLM tracing tool, and supports OpenAI- and Anthropic-compatible APIs.
• HALO Open-source agent optimization tool that analyzes traces, ranks failure modes, and provides actionable fixes for improving AI agent performance.
• Schematron V2 Turbo Proprietary state-of-the-art LLM for structured output generation and complex reasoning. Used for HTML-to-JSON extraction and other structured output tasks, with 150ms p50 end-to-end latency matching Opus 4.6 quality at 5% of the cost.
• ClipTagger-12B Proprietary efficient 12 billion parameter vision-language model for scalable video understanding. Processes images and video for multimodal AI tasks.
• Inference CLI Command-line interface (npm package @inference/cli) providing the 'inf' command. Supports 'inf auth login' for authentication and 'inf instrument' for automatic codebase instrumentation to route LLM calls through the Catalyst Gateway.

Quantifiable outcome

• Up to 95% cost reduction vs frontier-quality models using Catalyst training
• +4 more outcomes

Named customers5 records

Segments3 records

Ideal customer profiles2 records

Technology focussed Yes

API detail

Has API

Yes

API docs

API detail

Core technology

AI maturity

App detail

Integration6 records

AI capability10 records

Feature7 records

Scale indicators9 records

Recent moves6 records

Expansion highlights6 records

Direct peers

• Anyscale: Anyscale operates Ray-based LLM serving and inference infrastructure (Anyscale Endpoints). Targets the same AI engineering teams needing scalable, cost-efficient model hosting and fine-tuning.
• Replicate: Replicate is a cloud for running open-source ML models via API, including LLM inference. Overlaps with Inference.net on serving open-weight models and on developer/PLG distribution.
• Lambda Labs: Lambda provides GPU cloud instances (including H100/B200 clusters) and hosted inference for AI workloads. Comparable in GPU-based LLM serving and self-serve developer onboarding.
• Modal Labs: Modal offers serverless GPU compute for running LLM inference and AI workloads with developer-friendly APIs. Competes directly on developer ergonomics and per-second GPU pricing for inference.
• Together AI: Together AI runs a GPU-accelerated cloud for open-source LLM inference and fine-tuning. It is the closest direct competitor to Inference.net, with overlapping product surface (serving, fine-tuning, evaluation) and a similar target customer.
• Fireworks AI: Fireworks AI provides production LLM inference and fine-tuning infrastructure with optimizations for latency and cost. Comparable in product scope (deploy + fine-tune + observability) and target AI-native developer customers.

Broad incumbents

• Cerebras: Cerebras operates wafer-scale AI inference and training hardware integrated with its cloud. Competes in the high-performance LLM inference market that Inference targets, particularly on latency and cost-per-token.
• Hugging Face: Hugging Face hosts open-source models and offers Inference Endpoints, fine-tuning, and evaluation tooling. Broader scope than Inference but overlaps on serving open-weight LLMs and the open-science flywheel.
• Groq: Groq provides ultra-low-latency LLM inference on its LPU hardware, marketed heavily on speed and cost. Directly competitive on the inference workload side and a Catalyst Gateway integration partner.

Emerging players

• OctoAI (acquired by NVIDIA): OctoAI offered optimized model serving infrastructure for generative AI before being acquired by NVIDIA. Historically a direct comparison point on cost-efficient inference for production workloads.

Market position

Strengths5 records

Weaknesses5 records

Competitive moat5 records

Key risks6 records

Key highlights6 records

Customer concentration

• Xx.com/inference_net
• GitHubgithub.com/context-labs
• LinkedInlinkedin.com/company/inference-net

Compliance1 record

Revenue estimate

Valuation estimate

Number of profiles

Profiles3 records

Funding overview

Funding rounds1 record

Investors5 records

Funding detail is available on the Subscription and Enterprise plan.Contact sales →

M&A

Investments

M&A and investment is available on the Subscription and Enterprise plan.Contact sales →

Frequently asked questions about Inference