An aircraft has finished a flight. Somewhere, a planner must decide where it goes next. The obvious answer is wherever the timetable requires it. Then come the qualifications: maintenance, fuel consumption, turnaround time, the chance that one late arrival spoils the next departure. At Air France, changing these assignments produced an estimated saving of around 1% on fuel. The aircraft were still aircraft. The difference was in the decisions.
- The job: turn objectives and operating rules into a workable plan.
- The customers: teams building scheduling, supply-chain and allocation systems, including Air France and SAP.
- The catch: the mathematics needs an accurate description of the business.
A small number with a large denominator
Gurobi Optimization supplies the mathematical solver behind that kind of planning. Customers describe choices, restrictions and goals in a model. Its software searches for good combinations and, where the method and completed solve permit, certifies optimality within specified tolerances. You might minimize freight costs, allocate factory capacity or choose shifts. The output is a proposed action.
Air France’s operations researchers estimated aircraft fuel consumption and the costs associated with delay propagation. They built a decision-support tool that planners could adjust. Medium-haul flights made the task harder: shorter trips meant more possible assignments. Techniques that worked for long-haul schedules needed adapting. The work lay partly in translating competing concerns into mathematics, before the solver could settle anything.
Air France’s reported estimate from tail-assignment optimization. A customer result, not a promised return.
“We estimate that we are saving around 1% on fuel costs”
Solene Richard / Air France KLM
Three veterans start again
The founders were unusually familiar with this territory. Robert Bixby had co-founded CPLEX in 1987. Edward Rothberg had led the ILOG CPLEX team. Zonghao Gu had been its project architect and a lead developer. In 2008, they founded Gurobi. Its first optimizer followed in 2009.
That background matters because optimization software lives in the details: how it reduces a problem before solving it, which possibilities it investigates, how it handles numerical difficulty. Gurobi was designed to exploit modern computing architectures and multiple processor cores. Its competitive proposition combined algorithms with practical access to people who understood them.
The solver inside the software
A solver is often several layers away from the person benefiting from it. SAP began partnering with Gurobi in 2015 and expanded the arrangement in 2019. The relationship puts optimization inside planning applications. Customers can use a familiar business system while a specialist mathematical engine works underneath.
SAP describes using Gurobi both directly and within its own decomposition and heuristic techniques. Some problems need breaking into smaller pieces. Its optimization specialists meet regularly with Gurobi, and customer requirements have influenced solver features. This is an engineering relationship, with performance and numerical behavior subject to repeated scrutiny.
The market offers alternatives. IBM’s CPLEX combines mathematical programming with modeling tools and a separate constraint-programming capability. HiGHS provides open-source LP, MIP and quadratic optimization under an MIT license. Gurobi’s case rests on the combination of solver capability, deployment options and expert support. The useful comparison is a timed trial on your actual models.
Consider a demand forecast. It can estimate how many orders might arrive tomorrow. A planner must still decide which factory makes each item and which orders get scarce capacity. Optimization makes those choices explicit. It also makes disagreement more productive: if one manager prioritizes inventory and another prioritizes service, the model can expose the consequences of each preference. The business still has to choose what it values.
Pay for the decision, test the economics
Gurobi sells commercial subscriptions configured around deployment, computing requirements, concurrency and users. There are licenses for individual developers, shared access and dedicated compute servers. Instant Cloud runs managed optimization on AWS or Azure; Compute Server queues and distributes jobs on infrastructure a customer controls. Commercial pricing is quote-based.
The entry points are more accommodating. Eligible academics receive full-featured free licenses. Commercial users can evaluate the software on their own models. The Python package also includes a small, non-commercial license. Tutorials and examples reduce the distance between understanding a formulation and running it, though production work still demands data preparation and integration.
The company’s own economics have attracted private equity. Thompson Street Capital Partners invested in 2017. In September 2024, it announced a continuation transaction co-led by Hamilton Lane and StepStone funds. The investor reported that recurring revenue and employee count had each grown more than eightfold during its ownership. Those are growth multiples, not disclosed revenue totals.
A faster engine, a new front door
Gurobi 13.0, released in November 2025, added a primal-dual hybrid gradient algorithm for large linear programs and a local nonlinear barrier method. The latter offers a useful trade: a local optimum can arrive sooner, without generally guaranteeing the global best. Version 13.0.2 subsequently made GPU PDHG a fully supported feature.
Meanwhile, the company is tackling the work around the equations. Its Intelligence Hub launched in July 2026. Gurobot handles support questions; Modeler helps develop mathematical models and Python implementations; Explainer helps investigate model behavior. By October, documentation listed Modeler as generally available. Explainer and MCP integration remained experimental. Generated work still calls for independent checking.
Start with one decision
The practical lesson is to choose a recurring decision with a measurable consequence. Record the present plan. Separate the rules that cannot be broken from preferences that carry a cost. Test the model against ordinary days and awkward ones. Compare solution quality, elapsed time and the effort required to maintain it.
A useful pilot therefore includes the people who will use its recommendations. Ask them which proposed plans they would reject and why. Each objection may reveal an omitted constraint, a questionable input or a preference nobody had written down. That feedback improves the specification. Measuring against the existing process then tells you whether the extra sophistication earns its place.
A solver cannot rescue a missing maintenance rule or an objective that rewards the wrong behavior. A valid solution delivered after the dispatch deadline can be useless. Air France’s tool preserves room for planners to exercise judgment. That is a sensible destination for this technology: a decision whose trade-offs people can inspect, change and put to work.