The Investment Request You Can't Refuse
Half your revenue comes from two clients. If those two clients demand you buy new equipment and hire more staff, can you say no? Refuse, and half your revenue is at risk. Agree, and you've handed over control of your own investment decisions. That was exactly the position Amazon found itself in last week. As Nvidia GPU orders ballooned from 2 million units to three times that figure within a matter of days, it wasn't Amazon calling the shots on that scale — it was Amazon's biggest customers.
What Happened
Start with the numbers. AWS posted $39.8 billion in operating income, accounting for 58% of Amazon's total operating profit. A 39% operating margin shows the cloud business is still solidly profitable. But 44% of that profit comes from just four customers. Analysts also noted that the new-customer pipeline has been stalled for six months. There's growth here, but it's growth from existing customers spending more — not from the customer base expanding.
What happened over a few days in late August, on top of this structure, is the crux of this week's story. Reports came in quick succession: 2 million additional Nvidia GPUs ordered on the 27th, 3 million more on the 28th, and by the 30th, an order three times the original size. Amazon already has its own chip, called Trainium, and its CEO had described demand for it as very strong. Yet Amazon kept buying more Nvidia chips anyway. Over the same stretch, reports surfaced that free cash flow had turned negative, and Amazon moved to lock down power and network capacity directly too, signing a 200MW wind power deal in Sweden and a fiber-optic supply deal with Corning.
Why Buy Someone Else's Chip When You Have Your Own?
The reason lies with the customers. Amazon's top clients have already built their software stacks on top of Nvidia's CUDA ecosystem. Telling them to migrate to Trainium is effectively asking them to abandon their own investment. Add memory chip supply constraints on top of that, and there's a hard limit on how fast Amazon's own chip can substitute in. In the end, Amazon chose to buy the chips its customers wanted, in the quantities they wanted.
This is where the real shape of customer concentration risk reveals itself. We usually think of this risk purely as a revenue line item: how much revenue drops if that customer leaves. But Amazon's week shows that dependence runs in both directions. The bigger a customer's share grows, the larger the investment needed to meet their demands — and the less freedom you have to choose which supplier fills that investment. Mapped out as a chain, it looks like this.
Customer concentration swallowed supplier choice, and at the end of that chain, cash flow turned negative. Even the big-picture goal of $1 trillion in AWS revenue ultimately rests on the premise that this small group of customers keeps spending more.
The Lesson for Solo Businesses
The scale is different, but the logic is the same. If two clients make up half your revenue, the tools, platforms, and delivery specs they use become your investment list by default. If a client requires files in a specific software format, you have to buy that license. If they specify particular equipment, you have to acquire it. It doesn't matter if you know a cheaper, better alternative — that's the same reason Amazon bought Nvidia chips instead of using its own Trainium.
That's why you need to put two numbers on the same table. The first row is your top clients' share of revenue. The second is how replaceable the key suppliers are that those clients have locked you into. Replaceability asks a simple question: if this supplier raises prices or cuts off supply, can you switch to another one while keeping your contract with the client intact? If the answer is no, that supplier's cost isn't a separate line item — it's part of your customer concentration risk. Track the two rows side by side every quarter, and you can watch exactly how far replaceability drops as customer concentration rises.
The way to reduce this risk comes from the same table. Filling the new-customer pipeline — the very thing Amazon has failed to do for six months — is how you bring down the first row. Negotiating delivery requirements around output specs rather than specific tools is how you protect the second.
One Thing to Try This Week
This week, try building just one table. In the left column, list your top three clients and their share of your revenue. In the right column, list the suppliers and tools you use because of each client, and whether they're replaceable. If a row has "no" stacked up in the right column, that's the client you can't refuse, whatever they demand next quarter. Amazon was standing in exactly that spot when it tripled its GPU order. Knowing where that spot is in your own business is where you start.
