Qualcomm and AWS have announced a multi-generation collaboration to jointly design silicon dedicated to inference and very high-speed optical interconnects. The deal comes with warrants allowing Amazon to acquire up to 25 million Qualcomm shares. The stock rose by around 10% on the announcement.
It is the fourth move in three weeks in the same direction, and the picture is becoming legible.
The sequence
Let us recap what we have covered. Nvidia puts an accelerator dedicated to inference into production. OpenAI publishes the first results of its in-house chip. Nvidia buys the open-model platform, then spends 3.5 billion on MediaTek. And a leading Chinese model runs on domestic silicon.
Add this new deal, and the movement is clear: everyone is building an alternative, and Nvidia is spending to hold its position.
Why inference specifically
We have explained this several times and it is worth repeating, because it is the key to all these moves.
Training and serving a model do not call for the same qualities. Training demands raw power and tolerates slowness. Inference demands producing each token quickly, one after another, and runs up above all against the speed at which data circulates between memory and processor.
Yet inference has become the dominant cost. We documented this in our article on the two costs of AI: when you are serving hundreds of millions of users and agents are running in loops, every point of efficiency per token translates into millions.
A general-purpose chip spends part of its time waiting. A chip designed for this specific task organises the flow differently. It is a market large enough to justify dedicated design, and recent enough that Nvidia's position in it is less established.
The deal mentions very high-speed optical connections, and this point is more important than it appears. When a model is too big to fit on a single accelerator, it has to be split up, and the components must then exchange data constantly. The speed of those links becomes the limiting factor, not the computing power. This is what we noted about pooling ordinary machines: a slow network cancels out the benefit of fast machines. A good part of Nvidia's current advantage rests precisely on its internal interconnects.
What the structure of the deal reveals
The warrant mechanism deserves a note. It allows Amazon to acquire a significant stake in Qualcomm under conditions tied to the execution of the partnership.
This is not a simple supply contract: it is an alignment of interests over time. Amazon needs not to depend on a single supplier, Qualcomm needs a guaranteed customer to amortise a multi-generation development.
It is the classic structure when a buyer wants to create an alternative to a dominant supplier: it does not settle for buying, it funds the existence of the competitor.
What to take away
This deal will not weaken Nvidia in the short term: demand exceeds supply and the software ecosystem built over twenty years remains a considerable advantage.
What it signals is that the market is recomposing around a segmentation. Training and serving are becoming two businesses with two kinds of hardware, and the second attracts all those who cannot compete in the first.
For a user, the consequence is indirect and welcome: more competition in inference means a cost per token that keeps falling. This is what explains why a model like Images 2.5 halves its latency at an unchanged price. Hardware gains always end up reaching the bill.