Internet congestion control can be viewed as a resource allocation mechanism implemented via a distributed computation by end-systems. It provides a concrete, measurable example of an economic tatonnement process, with TCP's congestion avoidance algorithms playing the role of a `Walrasian auctioneer' searching for market clearing allocations. What are the theoretical limits upon the scalability of such a distributed computation? In this talk we review recent progress, and discuss the constraints imposed by: limited information available to end-systems; the finite speed of light; inherent randomness; and incentive compatibility.