IDC’s FutureScape 2027 forecast gives hosting companies something to chew on, and not all of it is revenue. Three predictions stand out: more AI work moving to open models, more contracts assigning liability for what agents do, and more regional rules for AI. Each one adds an obligation along with the opportunity.
Take open models first. IDC predicts enterprises will move 60% of AI workloads from proprietary large language models toward open-source ones by 2028, using routing to cut costs and lower reliance on one supplier. Who hosts them is a separate question, since a company could use its own hardware, a specialist or a big cloud. And hosting an open model means somebody owns the capacity planning, updates, tuning and recovery. That labor costs money, so a low price per token tells buyers little about their real bill. Routing a request to a cheaper model can save compute, but retries or a hand-off to a larger model can eat the savings.
Liability is where it gets uncomfortable. IDC expects 75% of Global 2000 vendor contracts to cover agent-initiated actions by 2028, and 70% of those enterprises to verify agent decisions continuously. These are forecasts, not rules. Even so, picture the position of a managed provider. It may not control the permissions a customer grants or the business rules a developer writes, yet the buyer may still want one party to blame. That boundary could end up as important as the price of compute.
A kill switch shows the problem. Stopping an agent is easy to promise. Someone still has to decide when to pull it, keep the records and work out whether past actions need fixing.
Regional controls add cost too. IDC predicts 35% of Global 2000 companies will run jurisdiction-specific AI model registries by 2029. Supporting different approved models and update schedules for similar customers gets expensive quickly.
A caution about the research: the headline predictions cover large enterprises, so nobody should use them to decide how many accelerators to buy. Providers need signed commitments first. If a customer wants spending caps, continuous verification and responsibility for agent mistakes, each demand needs a price and a named owner. Otherwise the service picks up three extra costs before the first invoice.
