How it started
My first operations research class in mechanical engineering.
Seventeen years ago I came across Hamdy Taha's Operations Research: An Introduction. I was a mechanical engineer who had always been good at maths — calculus and probability especially — and it had never occurred to me that maths could be used to see a supply chain, let alone estimate what it would do next.
Taha's argument is the one that stuck: build a model of the real system, and the quality of every decision that follows depends on how honestly that model represents reality. Everything I have worked on since has been a version of that problem.
A decade of experience
Four industries,
one common bottleneck.
Long-lead industrial equipment
A centrifugal compressor is closer to a lithography machine than most people assume — twelve to sixteen months from order to the end of the assembly line. The fan is a critical component. It cannot be batch-made and held in inventory, so a single machining or casting error moves the whole date.
I was on techno-commercial sales, which meant I was the person giving those dates to customers in oil, gas and automotive. When a customer's machine went down we had to find them a spare so the factory did not stop. A Bajaj two-wheeler plant once lost a compressor and the line stood still for hours — and the entire conversation was about one thing: by when. Nobody could answer it from a system. We answered it from relationships and guesswork.
Building an edge compute ADAS
For the next few years I built an edge compute ADAS device — a unit that runs an inference model locally to read driver behaviour and improve safety in commercial vehicles. We designed the PCB and fabricated the whole prototype in house.
The delays were constant and almost never ours. A shipping issue, or something at a supplier — and the supplier could rarely tell us what the delay actually was, only that there was one.
Scale that up and it is the same position a hyperscaler is in today, with a much larger number attached. A coolant distribution unit slips, nobody can give an exact date, and the entire working capital committed to that build sits idle until the part lands.
It's a commodity world
At eFeed we worked with maize, soya and a long tail of other commodities. After the geopolitical events of that period, prices moved fast enough that holding a margin became the hardest part of the business — harder than sourcing, harder than demand.
The same mechanism sits inside every large manufacturer, only with more zeros: an input whose price is set by events thousands of kilometres away, a contract priced weeks earlier, and no system anywhere connecting the two.
Across these three: lead time delays, supplier dependency, and geopolitical events. Three unrelated industries, one common bottleneck.
Same pattern emerging
About six months ago I started tracking the largest hyperscalers and their suppliers, and the same pattern appeared. Delivery dates going well past what was committed, with no assurance attached to the new ones.
Why it matters: most of the capital going into AI is getting stuck in working capital. If hyperscalers cannot operationalise on the timeline they have underwritten, the return on that capital takes a serious hit.
First principles
The replies were the surprise.
The Dark Supply Chain began as a blog about commodity constraints, read from the supply chain side rather than the price side.
Then the replies started. Futures desks on one side. CFOs and commodity managers at manufacturers on the other. Both asking for the same thing: detailed research into what sits between the commodity and the company.
"How do we make the supply chains more visible?"
The research
Top academic group
across the world.
I partnered with professors at Arizona State, Glasgow and INSEAD to understand how supply chain models have evolved through the decade of AI. We have collaborated on a paper, to be published shortly, on resource competitiveness between AI and agriculture — a supply chain perspective.
Out of countless hours of brainstorming and writing came the first version of the product: DSC Intelligence.
Now
DSC Intelligence is made for
hyperscalers and their listed suppliers.
The newsletter still covers critical components across AI hyperscalers and their suppliers, and builds indicators for studying the delays and bottlenecks in those chains.
The research is our go-to-market engine. Serious readers subscribe, and then reach out directly with a problem they think we can solve.
So if you recognise yourself in any of the above — please do reach out, and share the word.