The challenges of sourcing GPUs as AI demand surges

As businesses across sectors look to AI to make productivity gains and grow, demand for graphics processing units (GPUs) has surged. GPUs are critical to the operation of AI as they enable data processing at a level of complexity and scale that AI systems need which computer processing units (CPUs) and other types of computer chips cannot match.

However, with manufacturers struggling to produce enough GPUs to go around, distributors are being forced to take steps to ensure they are not left out of pocket when meeting customer orders, while businesses that source from those distributors are having to adapt to changes to traditional technology procurement and decide what value they place on obtaining access to available GPU supplies in the context of their AI projects.

The impact of high demand for GPUs

When demand exceeds available supply, manufacturers allocate production volumes among the participants of the downstream supply chain – including distributors, original equipment manufacturers, cloud providers and strategic customers. Distributors then determine how scarce inventory is distributed among end customers. In practice, priority is commonly given to:

• customers with established purchasing relationships;
• hyperscale and enterprise buyers;
• customers willing to enter into multi-year commitments;
• projects already underway; and
• customers purchasing complete infrastructure solutions rather than standalone GPUs.

As a result, procurement strategies that rely on spot purchasing can become ineffective during periods of acute supply constraint. Businesses seeking access to AI infrastructure increasingly need to secure supply windows in advance through contractual commitments and strategic supplier relationships.

Long-term commitments are becoming a prerequisite for supply

Distributors are increasingly requiring customers to place orders well in advance, enter framework agreements, commit to minimum purchase volumes and provide deposits or prepayments. These commitments enable distributors to negotiate future manufacturing allocations with upstream suppliers and GPU vendors.

From a legal perspective, such arrangements shift commercial relationships away from traditional product sales towards capacity reservation models. Customers are often asked to assume greater commitment risk in exchange for improved certainty of supply. This development also creates new drafting challenges around cancellation rights, deposit treatment, termination provisions and the consequences of reduced upstream allocation.

Capacity reservation is replacing conventional sales models

One of the most significant developments in the AI infrastructure market is the emergence of what could be described as "future capacity sales". High-end AI accelerators are frequently sold before production occurs.

Under this model, manufacturers forecast future output and allocate anticipated production to distributors and strategic partners. Distributors then offer customers delivery rights relating to future production slots.

In this market, the customer is not purchasing goods per se but rather:

• access to future manufacturing capacity;
• priority within an allocation queue; and
• contractual rights to receive products once production is completed.

To support these arrangements, distributors often require binding purchase orders, deposits, advance payments, minimum volume commitments and cancellation penalties.

The scarce asset is no longer necessarily the GPU itself. Rather, it is the right to receive the GPU at a specified future date. This resembles established practices in industries such as shipbuilding, aircraft manufacturing and energy infrastructure, where future production capacity is routinely reserved years before delivery.

Pricing mechanisms are evolving to address volatility

Severe shortages have also altered pricing practices. Distributors are often unwilling to provide long-term fixed pricing commitments where replacement inventory is uncertain and market prices remain volatile. Instead, contracts increasingly rely on alternative pricing structures. Some common models have emerged.

Fixed pricing

The customer agrees a fixed unit price and future delivery date. Under this approach, the distributor bears the risk of market price increases. Consequently, fixed-price arrangements are relatively uncommon during periods of significant shortage.

Price-at-shipment

Customers reserve future capacity, but final pricing is determined closer to delivery based on manufacturer pricing, component costs, market conditions and allocation status. This structure transfers a significant portion of pricing risk to the customer.

Reservation fee plus final purchase price

Customers pay a non-refundable reservation fee to secure future capacity and then pay the final purchase price upon shipment. This model provides distributors with demand certainty while reducing inventory and working capital risk.

For many AI projects, the financial impact of delayed deployment significantly exceeds the cost of the hardware itself. As a result, customers are often willing to accept pricing uncertainty in exchange for delivery certainty.

Alternative supply strategies are becoming increasingly important

To preserve customer relationships and maintain project momentum, distributors are also directing customers towards substitute solutions. These may include alternative accelerator platforms, earlier-generation products or cloud-based consumption models.

At the same time, scarce GPUs are increasingly bundled into larger infrastructure offerings, including AI clusters, enterprise server configurations, managed AI infrastructure and integrated systems.

Bundling allows distributors to allocate limited resources to higher-value opportunities while enabling customers to procure a complete operational environment rather than individual hardware components.

Legal considerations

The legal analysis should follow the commercial reality of the transaction. In many cases, the most important questions relate to allocation rights, delivery uncertainty and pricing mechanisms rather than the sale of physical goods.

When drafting agreements, distributors and their customers should focus on:

• the precise subject matter being acquired, including whether the customer is purchasing hardware, configured systems, cloud capacity or future allocation rights;
• the treatment of allocation risk where upstream supply is delayed, reduced or reprioritised;
• pricing mechanisms, escalation rights and quotation validity periods;
• deposits, reservation fees, prepayments and cancellation rights;
• substitution rights and bundling arrangements; and
• regulatory issues including export controls, sanctions compliance, end-user restrictions and audit rights.

What businesses should do now

Businesses procuring AI infrastructure should ensure that legal and procurement teams recognise that the market operates fundamentally differently from traditional technology supply chains. The commercial substance of many transactions is not the purchase of existing products but the acquisition of future access to constrained computing capacity.

Accordingly, contractual risk allocation becomes critical. Organisations should scrutinise allocation provisions, delivery commitments, pricing adjustment mechanisms and cancellation rights when committing significant capital to AI infrastructure projects.

In the current GPU market, certainty of access has become a commodity in its own right. Contracts should expressly address who bears the risk when that certainty cannot be delivered.