TL;DR

Recent European earnings point to an unexpected set of AI beneficiaries: long-established software, consulting and hosting groups rather than the model builders. SAP and OVHcloud, along with the consultancies Capgemini and Sopra Steria, each reported firmer demand or raised guidance as customers moved from trialling AI to running it.

The underlying reason is unglamorous. Getting hold of a capable model is now trivial; making it useful inside an organisation with three decades of accumulated systems is not. AI has to reach live company data while respecting permissions, leaving an audit trail and fitting workflows staff already use. That is integration work, and it is precisely what these firms were built to do long before generative AI existed. As UBS put it in a recent note, applications are where most of the value gets created.

The numbers carry the argument. SAP’s cloud backlog reached €22.9bn, up 26% once currency movements are stripped out, as finance, procurement, supply chain and HR systems moved onto platforms that increasingly host AI deployment; its purchases of Dremio and Prior Labs point the same way. Capgemini lifted its full-year growth target on bookings up 9.2%, and Sopra Steria upgraded guidance after organic growth reached 5.3%. Boston Consulting Group found upwards of 70% of investors doubtful that companies possess the operational capability to make AI work at all.

A second force favours incumbents: customers increasingly want control over where models run. Arthur Sadoun, who runs Publicis, has described clients asking for advanced models inside environments they govern themselves. Airbus picked Scaleway, an Iliad subsidiary, to carry sensitive industrial and defence work alongside Mistral-built tools, and expects roughly 70 of its critical applications running there once 2028 closes. OVHcloud’s public cloud revenue rose 20.2% in the third quarter, an early sign that appetite for infrastructure beyond the reach of the US Cloud Act converts into revenue.

Looking Forward

For UK buyers the read-across is direct, since the sovereignty and legacy-integration problems are identical this side of the Channel, and the government has been putting money behind domestic capability through its sovereign AI fund. The open question is margin: much of this work is exactly the lower-value advisory and coding effort AI itself is expected to automate. Whether these firms can hold pricing as that happens is the test the current results do not yet answer.