In March 2024 Shanghai published the Implementation Plan for High-Quality Development of Intelligent Compute Infrastructure, known as the “Compute Pujiang” action plan for 2024–2025. It sets out arrangements on intelligent compute capacity, AI data-centre construction, compute scheduling and the industrial ecosystem. The word “infrastructure” in the title matters: the policy reaches the bottom of the supply chain rather than stopping at application subsidies.
Four strands of the plan
Capacity refers to total supply, construction to spatial layout, scheduling to matching resources across operators, and ecosystem to coordination among chips, frameworks, models and applications. Together they turn compute from equipment that a company buys into a public capability that a city can schedule.
For AI companies, the immediate effect is more sourcing options: build, lease, buy cloud services or connect to a scheduling platform, each with different cost and compliance profiles. The more options there are, the more the contract matters. Compute is not office space; delivery quality, availability and exit mechanics are hard to fix after the fact.
Clauses that are easy to miss in a compute lease
- What “capacity” means. Contracted capacity and usable capacity are rarely the same. Specify card type, interconnect bandwidth, storage ratio, fault response and remedies.
- Cost pass-through. Who bears electricity price movements, curtailment arrangements and maintenance fees should be explicit, so that “market adjustment” cannot justify unilateral changes.
- Racking and exit. Rack location, migration assistance, data deletion and equipment disposal after termination should be settled at signing.
- Sub-leasing limits. Compute providers often prohibit resale or subleasing. A model company that supplies compute to its own customers needs written consent in advance.
- Force majeure. Allocation of responsibility for power, network and facility incidents cannot rely on boilerplate alone.
Energy and facility approvals
AI data centres run at higher power density than conventional facilities, so energy quotas, energy-saving review and electricity connection are preconditions. The builder usually carries these obligations, but a tenant is not entirely detached: insufficient power capacity or constrained retrofits eventually surface as service unavailability. Before choosing a site or signing, ask for the status of approvals and expansion plans rather than relying on commercial assurances.
Fire safety, environmental assessment and land-use approvals affect delivery timing in the same way. This information is not a trade secret, yet it is often omitted. Putting it on a diligence list costs little.
Data security and cross-border compute
Where training data sits in a third-party facility, access control, encryption and audit responsibilities have to be allocated in the contract. Multi-tenant isolation, operator access rights and data exposure during troubleshooting are real risk points.
Calling compute across borders adds another layer: processing data abroad may constitute an outbound transfer, and even a domestic facility accessed remotely by an overseas team can raise the same question. Map the data flow first, then decide where compute sits, rather than signing a compute contract and fixing compliance afterwards.
Why the local template matters
Shanghai’s plan is not unique; competition on compute supply will continue across regions. For companies, local policy differences are both a siting variable and negotiating leverage: electricity prices, energy quotas and scheduling-platform access conditions affect the real unit cost of compute.
Treating compute supply as a compliance and contract matter rather than a pure purchase is the most practical change these policies bring.
For security and access control at compute facilities, see cybersecurity compliance; for compute lease and service terms, see commercial contracts; for data storage and outbound transfer analysis, see data compliance.
This article is based on the Shanghai Compute Pujiang intelligent-compute action plan for 2024–2025, published in March 2024, and on public information. It is for reference only and does not constitute legal advice.