Five-Layer Architecture

LinkCompute is built as a five-layer pyramid, bottom to top. Each layer addresses a specific pain point in the industry.

Layers at a glance

L1 — Global Compute Resource Layer

The utility layer — compute as water, power and gas.

• Compute-center aggregation. Already connected to 10+ compute centers at home and abroad, with 10+ more under-construction centers in early partnership to lock in future low-cost, scalable supply.
• Compute absorption service. Idle resources at partner centers come onto the platform, get activated by deploying open-source models, and the platform takes a service fee.
• Shared colocation and hosting. Dedicated rack zones at partner centers provide hosting to customers who have compute but lack space or operations.
• Cross-border channel. Low-cost domestic compute is packaged through a compliant overseas pipeline — from scheduling and data prep to training and intelligent app delivery.

Serves: compute centers (absorption), GPU holders (monetization), overseas customers (low-cost compute)

L2 — All-Category Model Layer

One-stop global model supply.

Closed-source aggregation

• Multi-channel access to the same model — e.g. GPT-4o via OpenAI official, AWS, Microsoft Azure, etc.
• Users pick the channel that fits their needs
• Deep discounts already negotiated with multiple vendors give us a price edge

Open-source deployment

• Deploy DeepSeek, Qwen, Llama and other popular open models on partner compute centers' idle GPUs
• Solves consumption for the centers, and gives platform users cheap access to open models

Serves: developers and enterprises calling models; compute centers absorbing idle capacity

L3 — Data & Evaluation Layer

This is what sets LinkCompute apart.

"Where does compute come from" price-comparison

For any open-source model (e.g. DeepSeek-V3), the platform shows real-time prices, available concurrency, and average latency across different compute centers.

Example: a Xinjiang site at 0.25 RMB / 1k tokens, an Inner Mongolia site at 0.30, a Tibet site at 0.40 — the differences are visible at a glance.

The same comparison applies to closed-source models across cloud providers and channels.

"AI Barometer" rankings

Built from real global call volume on the platform:

• Model call-volume charts (daily / weekly / monthly)
• Compute-center heat charts (most active, lowest latency)
• Tool / Agent charts (most-called AI tools)

OpenRouter has already proved the value of call-volume data — analyzing AI trends from 100 trillion tokens of real consumption rather than static benchmarks. LinkCompute goes further by fusing call volume with price, latency, and compute-source data into a more complete decision layer.

Serves: everyone picking a model or a compute source; research and media tracking AI trends

L4 — Tool & Agent Distribution Layer

The developers' App Store.

• Tool aggregation. One-stop showcase, invocation, billing and settlement for AI tools and agents.
• Built-in promotion engine. Solves the "I built it, nobody uses it" problem for indie developers.
• Unified integration. A single standard API — integrate once, reach global users.
• Revenue share. The platform takes a call-based cut; developers focus on building.

Why L4, not L1? Tool marketplaces only reach critical mass when there's enough underlying model usage and developer base. The lower three layers must succeed first.

Serves: indie developers and small teams (monetization); enterprise customers (discovery and use)

L5 — AI Orchestration Layer

Using AI to orchestrate AI.

• Natural-language driven. Users type "I want to fine-tune a 7B model" or "run a high-concurrency support agent" and the system parses the requirement.
• Smart matching. AI picks the optimal compute configuration, best-fit model, best-value plan, estimates cost and time, and ships a ready-to-use environment in one click.
• Continuous improvement. The orchestration engine self-tunes from platform-wide call data, improving match quality over time.

Keep the complexity on our side; give the user something simple.

Serves: everyone who wants to use AI without managing configuration

Two usage modes

On top of the five layers, the platform exposes two interaction modes:

• Simple mode (driven by L5) — for newcomers and non-technical users. Describe the goal, get a plan and a working environment.
• Pro mode (L1–L4 exposed) — for experienced developers. Pick model, channel, compute center, price range, latency target — all data transparent.