Oracle will deploy HPE Juniper networking across its AI data centres

Oracle will deploy HPE Juniper routing and switching equipment across its AI data centres globally under an expanded multi-year agreement with Hewlett Packard Enterprise, announced alongside HPE's third-quarter results. The deal covers HPE Juniper PTX and MX routing platforms and QFX and EX switching platforms, together with multi-year support services and financing. The detail that separates this from an ordinary vendor renewal is that HPE issued Oracle warrants to purchase shares of HPE common stock. Warrants given to a customer are a way of tying that customer's economics to the supplier's performance, and they show up in deals where the supplier wants a very large commitment and is prepared to give away future equity to secure it. It is a structure that has appeared more often in AI infrastructure over the past two years, and it says something about the bargaining position of a buyer with Oracle's current build plans. Those plans are the reason the deal is large. Oracle has been expanding AI supercluster capacity aggressively, and networking is the part of that build that scales worst if it is got wrong. When tens of thousands of GPUs are training a single model, the interconnect determines how much of the theoretical compute is actually usable. Bandwidth per node, tail latency, congestion behaviour under all-to-all traffic patterns and how quickly the fabric recovers from a failed link all translate directly into training time. A cluster with excellent accelerators and mediocre networking is an expensive way to keep GPUs idle. For HPE the agreement is the clearest validation so far of its acquisition of Juniper Networks. That deal was expensive and contested, and the argument for it was that networking would become the constraint in AI infrastructure while HPE's traditional server business faced commodity pressure. Winning a global, multi-year deployment across the AI data centres of a company building at Oracle's current pace is the kind of proof point that argument required, and HPE is describing the deployment as among the industry's largest AI infrastructure buildouts. The collaboration is not new. Oracle and Juniper have worked together for more than a decade, which matters because networking migrations are painful and incumbency is worth a great deal. What has changed is the scale and the fact that the incumbent is now inside HPE. The competitive frame is worth stating. Networking for AI clusters is contested between Ethernet-based approaches and NVIDIA's InfiniBand, and between the major Ethernet suppliers, principally Arista, Cisco and now HPE Juniper. A commitment of this size from a hyperscale-class buyer is a data point in favour of merchant Ethernet at the top end of AI infrastructure, in a segment where InfiniBand held an early advantage. There is a Gulf angle. Oracle has been building out cloud regions in Saudi Arabia and the wider Middle East, and AI data centre capacity in the region is expanding quickly across several vendors, with AMD, Cisco and HUMAIN having brought production infrastructure live in the Kingdom and AWS preparing to open a Saudi region in December. If Oracle standardises its AI data centre networking globally on HPE Juniper, that standard applies to whatever it builds regionally, and it puts HPE deeper into an infrastructure stack the company is separately trying to localise through its Saudi Made programme. The companies have not disclosed the contract value, the number of sites, the deployment schedule, the strike price or size of the warrant issuance, or which Oracle facilities are covered first. HPE has described the arrangement as multi-year and global without attaching a figure to it, which is common in networking agreements and leaves the commercial significance hard to size from outside. The next place this becomes measurable is HPE's networking segment revenue over the coming quarters, and any disclosure of the warrant terms in the company's filings.