The Ripple Effect of the AI Boom

A few weeks ago, we ran into an unexpected infrastructure problem at our organisation.
Several RAM modules in our infrastructure started failing in quick succession. We suddenly found ourselves having to source replacement RAM modules more frequently than usual. That turned out to be a surprisingly difficult exercise. RAM was expensive, availability was limited, and some of the suppliers who had stock were offering modules without proper replacement or return warranties. We had little choice but to keep looking and keep buying.
This got me curious. Was this just a temporary supply issue, or was something larger happening in the semiconductor industry?
As I started looking into it, I found an interesting connection with the enormous demand being created by the AI infrastructure boom. And that led me to three developments that, when viewed together, raise some interesting questions about the economics of the AI industry.
1. The AI boom is changing what semiconductor manufacturers choose to produce
Not all memory is the same.
The RAM we use in ordinary computers and servers is generally conventional DRAM. AI data centres, particularly those running powerful GPUs, need something more specialised called High Bandwidth Memory (HBM).
HBM allows GPUs to access enormous amounts of data at very high speeds. The demand for HBM has grown rapidly with the expansion of AI infrastructure. Memory manufacturers such as Samsung, SK Hynix and Micron are therefore investing heavily in HBM production. There is a strong commercial incentive to prioritise these higher-value products.
But there is a catch.
Manufacturing capacity is not unlimited. What it means is that a significant portion of manufacturing capacity is being allocated to a relatively specialised category of memory. This creates a crowding-out effect. Less capacity is available for conventional memory, putting pressure on supply and prices.
And this is not just an isolated observation. The Center for Strategic and International Studies (CSIS), in its September 2026 analysis, describes the current memory shortage as a structural supply-demand problem, with the shortage potentially persisting through 2027 and beyond.
So, the AI infrastructure boom is beginning to influence the economics of components used in everyday computing infrastructure.
2. NVIDIA is not just selling the picks and shovels. It is helping finance the gold rush.
Another development that caught my attention was the recent reporting around NVIDIA's financial arrangements with companies building AI infrastructure.
NVIDIA is, of course, the dominant supplier of GPUs used in much of today's AI infrastructure. But its role is becoming more complex. It has invested in some AI infrastructure companies, entered into arrangements to support their growth, and made commitments to purchase computing services from selected AI cloud providers.
Consider how this can work.
An infrastructure company needs capital to build a data centre. Financial support or a loan guarantee helps it raise that capital. It uses the money to purchase NVIDIA GPUs. NVIDIA records the chip sale as revenue. NVIDIA has also entered into cloud service agreements with selected AI cloud providers. Under these arrangements, the providers can sell computing capacity to third-party customers, and NVIDIA's commitments reduce as that capacity gets used.
This creates an interesting financial relationship.
NVIDIA is simultaneously a technology supplier, an investor, a source of financial support and, in some arrangements, a potential customer.
Its July 2026 financial filing disclosed $279 billion in supply and capacity commitments, up from $119 billion in the previous quarter. It also disclosed $36 billion in AI cloud service commitments and a conditional guarantee of up to $105 billion for the OpenAI-related Ohio data centre project.
These numbers represent different types of commitments and financial exposure. They are not all money already spent or debt owed.
NVIDIA has also announced plans to mobilise more than $500 billion of third-party capital through financing platforms. This is capital to be independently underwritten and provided by financial institutions, not $500 billion of NVIDIA's own investment.
There are legitimate strategic reasons for these arrangements. NVIDIA wants to ensure that the infrastructure required to run its technology gets built.
But they raise an important question: how much of the current infrastructure demand is coming from independent customer economics, and how much is being enabled by financial relationships within the ecosystem?
This does not make the reported GPU sales unreal. It does mean that we should look beyond revenue figures to understand the underlying economics of demand.
3. Are we building infrastructure faster than AI can monetise it?
AI computing requirements are undoubtedly growing. But projected demand and actual commercial returns are two different things.
Consider the scale of investment. S&P Global Ratings estimates that capital expenditure by the six largest hyperscalers could exceed $1.3 trillion by 2027. This is an enormous bet on future AI demand.
Building data centres involves substantial investments in GPUs, memory, networking, land, electricity and cooling. These investments are made today, while the returns are expected to materialise over several years.
If AI adoption and the revenue generated from AI applications grow more slowly than anticipated, we could see underutilised data centres, financial pressure on infrastructure providers and excess capacity.
This isn't a prediction that AI demand will collapse or that AI is a bubble. It simply highlights the financial risks of investing at a pace that future demand may not be able to sustain.
A founder's perspective
What started as a problem of sourcing replacement RAM got me looking at the AI infrastructure boom a little differently. The impact is no longer confined to companies building AI. It is beginning to affect the wider technology ecosystem, from memory availability and component prices to energy requirements.
As a co-founder of an AI company, I am optimistic about AI's potential. At Cere Labs, we are building in this space ourselves and see its value first-hand.
But perhaps we should pay as much attention to the economics of building AI as we do to the excitement around what AI can do. After all, infrastructure investments eventually need customers, utilisation and returns to justify them.



Comments