In order to assure Quality of Service (QoS) connectivity, Network Service Providers (NSPs) negotiate Service Level Agreements (SLAs). However, a committed SLA might fail to respect its QoS promises. In such a case, the customer is refunded. To maximize their revenues, the NSPs must deal with risks of SLA violations, which are correlated to their network capacities. Due to the complexity of the problem, we rst study a system with one NSP provider and give a method to compute its risk-aware optimal strategy using (max; +)-algebras. Using the same method, we study the case where two NSPs collaborate and the case where they compete, and we derive the Price of Anarchy. This method provides optimal negotiation strategies but, when modeling customers' reaction to SLA failure, analytical results do not hold. Hence, we propose a learning framework that chooses the NSP risk-aware optimal strategy under failures capturing the impact of reputation. Finally, by simulation, we observe how the NSP can bene t from such a framework.