Special Analysis
For most of the past year, the central question in artificial intelligence was what the technology can do. Over the past week, the focus shifted to what it costs and who carries the risk. The most significant developments did not concern new capabilities but their consequences: rising prices for the hardware AI depends on, growing pressure on the capacity and financing needed to build it, and a quiet change in what leading developers are actually selling.
Taken together, competitive advantage appears to depend less and less on access to the most capable model, and more and more on the ability to manage costs, secure capacity, limit exposure to financial risk and reduce dependence on individual suppliers. As AI infrastructure expands, its costs and risks are increasingly leaving the technology sector and passing through to the companies and consumers that rely on it.
The cost of building AI infrastructure reaches the consumer
Google has confirmed that its next generation of Pixel devices will be more expensive, citing the sharp rise in memory chip prices driven by demand from AI data centres. This is not an isolated decision by one manufacturer but a reflection of a broader shift in the electronics market.
According to research the company cites, the price of a gigabyte of the memory used in smartphones rose from around 2.80 dollars in 2025 to approximately 12 dollars in 2026, more than a fourfold increase. For a premium device with sixteen gigabytes of memory, that increase alone lifts the cost of the component from roughly 45 to nearly 190 dollars, before the processor, screen or camera are counted.
The cause is structural. Memory manufacturers are redirecting production capacity towards high-bandwidth memory, which is essential for AI servers, reducing the supply available to consumer electronics manufacturers. Analysts expect memory to account for a growing share of AI infrastructure investment in the coming years and do not anticipate any significant improvement in supply before 2028.
For companies, this is a reminder that the cost of building AI infrastructure is no longer confined to those developing AI systems. The consequences will be felt by every organisation buying servers, laptops or other equipment, and by every organisation planning its own on-premises infrastructure. Hardware budgets and procurement decisions increasingly have to account for a cost driven by demand that has nothing to do with how the buyer itself uses AI.
Even the largest providers are rationing compute
The shortage is not limited to memory. Across the industry, the compute capacity needed to run AI systems has become a constrained resource, and even the largest cloud providers are struggling to keep up with demand.
Microsoft, which reports its most recent quarterly results at the end of July, has stated for several consecutive quarters that demand for its cloud services exceeds supply and that this constraint is expected to persist at least until the end of 2026. Company executives say what they lack is the space and electrical power needed to install new equipment, rather than the hardware itself. The value of contracted but not yet delivered revenue has risen into the hundreds of billions of dollars, as customers commit in advance, often at a higher price, simply to secure future access to capacity.
In practice, this means companies can no longer treat cloud computing as an all but unlimited service available on demand. A company may price an AI service at today's rates and then, several months later, find that reserved capacity is harder to secure, that memory-intensive systems carry additional cost, or that on-premises equipment has become more expensive. The model has not changed; the economics underneath it have.
For organisations, capacity planning therefore becomes a strategic question rather than a procurement detail. Capacity contracted in advance, the ability to use more than one provider, and designing systems to use smaller or more efficient models wherever possible are becoming part of a well-constructed AI strategy. Not because any particular model requires it, but because the availability of computing power can no longer be taken for granted.
When a chipmaker guarantees its own customer
According to the Wall Street Journal, subsequently confirmed by other agencies, the chipmaker NVIDIA is negotiating financial guarantees worth around 250 billion dollars to help OpenAI secure a large data centre complex in the US state of Ohio. A separate arrangement, estimated at 350 billion dollars, would help finance OpenAI's purchase of chips, with the total value of the project expected to exceed 500 billion dollars.
The structure of the arrangement is unusual. A guarantee of this kind means NVIDIA would cover OpenAI's obligations if the company proved unable to meet its lease and financing commitments. According to available information, the arrangement was necessary because traditional lenders, given the company's financial profile, were not prepared to lend to OpenAI directly. In practice, a company's principal chip supplier would simultaneously be guaranteeing its ability to buy those chips.
The scale is exceptional. The proposed guarantee is several times NVIDIA's cash reserves and exceeds its total annual revenue. Some investors have described the arrangement as a form of "circular financing", in which a supplier guarantees a customer's spending on the supplier's own products. Negotiations have not concluded and the terms may still change.
For the wider market, the significance lies less in the specific agreement than in what it signals. When a project of this scale depends on a supplier's balance sheet rather than on conventional lending, it suggests that financing the expansion of AI infrastructure is becoming increasingly strained. That can affect the price and availability of compute for any organisation relying on it, including those far removed from the companies involved in this arrangement.
Competition moves from models to platforms
Over the same period, leading developers signalled a change in what they are selling. OpenAI introduced Presence, a platform for deploying AI agents across voice and text channels in large organisations, with the banking group BBVA, SoftBank and the insurer IAG among its first customers. This shift came even as the leading laboratories continued to release new versions of their foundation models, which shows that the competitive contest is no longer decided on model capability alone.
The centre of gravity is moving up the stack, from the model itself to the platform that deploys, monitors and governs whole populations of agents inside a company. As the differences between leading models narrow, competition is increasingly about who controls the environment in which those models run.
For organisations, this changes the nature of the decision. The question shifts from which model to adopt to whose platform to build on. Unlike a model, a platform creates a greater degree of dependency: it determines how agents operate, how identity and access are controlled, how data is handled and how the service is billed. Choosing a platform is a longer commitment than choosing a model, and considerably harder to reverse.
Take away
Seen separately, these developments look like four different stories: one about consumer prices, one about compute capacity, one about financial markets and one about enterprise software. Together, however, they describe a single common theme: the growing concentration of dependency in the AI market. Companies increasingly rely on a small number of memory manufacturers, on the constrained capacity of a handful of cloud providers, on the financial strength of individual suppliers, and on the platforms of a small number of companies.
Companies cannot avoid these dependencies entirely, but they can prevent any one of them from becoming a single point of failure whose interruption would jeopardise the whole business. Understanding real cost exposure before prices move, securing capacity in advance, preserving the ability to move models and data, and treating supplier concentration as a strategic risk are becoming as important as the choice of technology itself. As the market's focus shifts from questions of capability to questions of cost and control, resilience becomes the decisive advantage.