How AI Companies and Data Center Providers Negotiate Capacity & Revenue
As AI models scale toward trillion‑parameter architectures and GPU clusters reach industrial‑grade power densities, the commercial relationship between AI infrastructure providers and leading AI companies has become one of the most strategically important contracting domains in technology. These agreements determine how compute is delivered, how revenue is shared, how risk is allocated, and how entire data‑center campuses are financed.
This can be broken down into contract structures, the pricing logic for each, third‑party enablement strategies, and the role of hyperscale AI infrastructure companies such as OpenAI, CoreWeave, and Anthropic in structuring these agreements.
1. Capacity Reservation Contracts
Capacity reservation agreements, also referred to as Compute Tariff contracts, translate physical resources (power, racks, GPUs, cooling, network) into predictable, bankable service commitments.
Key Components (Expanded)
- Reservation Charges: fixed monthly fees secure MW‑scale power, dedicated GPU clusters, or entire liquid‑cooled halls. These charges stabilize revenue and enable project financing for massive expansions.
- Acceptance & Ramp Schedules: contracts specify delivery timelines, commissioning milestones, and ramp‑up periods. These schedules align infrastructure readiness with training roadmaps.
- Usage Charges Metered: compute (GPU‑hours, TFLOP‑hours, or token‑based equivalents) ensures AI companies pay for actual consumption, matching cost to training cycles and inference surges.
- Service Quality & Remedies: SLAs define availability, latency, thermal compliance, and network throughput. Credits or fee reductions apply if performance drops.
- Technology Refresh Governance: with rapid GPU evolution (H100 → H200 → B200 → MI300X), contracts define upgrade paths, cost allocation, and retirement of older clusters.
- Curtailment & Force Majeure: power grid constraints, cooling failures, or network outages trigger predefined curtailment rules, crucial for lender risk assessment.
- Collateral, Credit, and Step‑In Rights: provisions make AI compute contracts “bankable,” enabling debt financing for hyperscale campuses.
2. Consumption‑Based Pricing
AI companies increasingly adopt usage‑based pricing, replacing traditional SaaS licensing.
Expanded Pricing Logic:
- Token‑Based Billing: training and inference costs tied to token volume allow granular budgeting for unpredictable workloads.
- GPU‑Hour Billing: transparent pricing for training workloads, with discounts for sustained usage or liquid‑cooled deployments.
- Committed‑Spend Discounts: enterprise buyers negotiate ~30% discounts for multi‑year commitments, helping providers forecast demand.
- Budget Ceilings and Cost Controls: guardrails prevent runaway token costs, enabling CFOs to forecast AI expenses.
- Integration and Operational Overhead: total cost of ownership is often x4 the raw token cost once data pipelines, orchestration, monitoring, and model‑ops are included.
3. AI Outsourcing: New Governance and Dependency Models
As enterprises embed AI into managed services, outsourcing contracts shift from labor‑based pricing to value‑based + usage‑based models.
Expanded Contract Considerations
- Data Use and IP Rights: contracts define how training data is used, who owns derivative models, and how outputs can be commercialized.
- Model Transparency and Validation: enterprises demand visibility into model behavior, testing protocols, and validation frameworks.
- Provider Lock‑In Risks: proprietary models and integrated platforms create dependency; contracts now include portability clauses and exit rights.
- Automation‑Driven Value Pricing: pricing shifts from hourly billing to outcome‑based fees tied to business impact.
4. Revenue Sharing Models
Revenue sharing agreements are more common when AI companies want to minimize upfront capital expenditure or when infrastructure providers co‑invest in GPU clusters.
Expanded Structures
- Joint Ventures (JV): shared capex and shared revenue from compute services or model‑serving.
- Inference Revenue Sharing: providers receive a percentage of API revenue, aligning incentives with model performance.
- Token‑Economics Sharing: providers may receive a share of revenue above a threshold, creating upside for both parties.
- Performance Based Uplift: higher revenue shares if training efficiency exceeds SLA targets.
5. Developers Increase Data Center Growth
AI companies have been relying on third‑party developers to scale compute capacity faster than traditional hyperscale timelines.
Expanded Models
- Build‑to‑Suit (BTS): developers finance and construct campuses tailored to AI workloads; AI companies commit to long‑term MW‑month contracts.
- Lease + Retrofit: AI companies lease shells and add their own GPU clusters, accelerating deployment.
- Dedicated Halls in Colocation Facilities: providers build high‑density halls with liquid cooling and specialized network fabrics.
- Energy‑Integrated Development: developers build on‑site generation, integrate waste‑heat reuse, bundling energy + compute + cooling into unified tariffs.
6. The Role of AI Infrastructure & Hyperscale/Neocloud companies
A new class of AI‑focused providers (such as OpenAI, CoreWeave, Anthropic, Lambda, Voltage Park, Vantage, Digital Realty, and others) play a central role in structuring compute‑capacity agreements and enabling rapid expansion. Their expertise spans four domains:
A. Structuring Compute‑Capacity Agreements
These companies have deep experience designing contracts that balance:
- GPU cluster availability
- MW‑scale power commitments
- Liquid‑cooling requirements
- Network‑fabric guarantees
- Technology refresh cycles
- Utilization‑based discounts
AWS somewhat pioneered the Reserved Instance → Reserved GPU model; CoreWeave built the GPU Pod Reservation model; Anthropic negotiates frontier‑model training clusters with multi‑year commitments.
B. Token/API Revenue‑Sharing Models
AI companies increasingly monetize via API calls and token usage. Infrastructure providers now structure:
- Shared inference revenue
- Token‑based revenue floors and ceilings
- Performance‑based revenue uplift
- Joint commercialization agreements
CoreWeave and Anthropic have experimented with co‑monetization, where compute providers receive a percentage of API revenue in exchange for discounted GPU clusters.
C. Power Procurement & Energy Strategy
Hyperscale AI providers are now energy companies:
- They negotiate long‑term PPAs (solar, wind, nuclear microreactors).
- They build on‑site generation (batteries, gas peakers, microgrids).
- They secure grid interconnection rights years in advance.
- They design liquid‑cooling thermal envelopes tied to energy efficiency.
AWS is now again largest corporate renewable buyer in the world; CoreWeave has enabled high‑density liquid‑cooled halls; Anthropic negotiates energy‑integrated compute contracts for training.
D. Securing Powered Land & Development Rights
The most constrained resource in AI infrastructure is powered land with:
- Transmission‑level power access
- Water rights or reclaimed‑water availability
- Zoning for high‑density data centers
- Fiber routes and low‑latency proximity
- Expandable parcels for multi‑phase buildouts
Developers and companies like Vantage, QTS, Compass and CoreWeave maintain dedicated teams for:
- Land banking
- Power‑delivery negotiations
- Substation co‑development
- Liquid‑cooling permitting
- District‑heating integration
This capability is now a competitive differentiator in the AI infrastructure market.
7. AI Company contracts
Model companies like OpenAI, Anthropic, or Google DeepMind approach negotiations with clear priorities:
- Committed‑spend discounts
- Technology refresh guarantees
- Latency and throughput SLAs
- Data portability and exit rights
- Modular capacity blocks
- Integrated energy + compute pricing
These patterns shape the entire data‑center development ecosystem.
8. 2027–2035
The next phase of AI infrastructure contracting will include:
- Compute capacity traded as a commodity
- Liquid‑cooling tariffs as standalone line items
- Waste‑heat monetization embedded in compute contracts
- AI‑specific project finance instruments
- Developers co‑owning GPU clusters alongside AI companies
The commercial logic of AI infrastructure is becoming as sophisticated as the models it supports.