By George Titus · 14 August 2026 · 4 min read
Announcements have been dominated by capacity and capital. A run of recent deals shows who is actually buying, and on what terms. On 20 July, e& UAE and Core42 launched a platform called Sovereign AI Compute. It gives enterprises and government bodies in the UAE on-demand access to GPU capacity inside the country. Core42, part of the G42 group, supplies the accelerators and implementation services. e& UAE supplies the network and customer relationships.
The commercial terms are the interesting part. No upfront capital cost. Zero egress fees, meaning customers are not charged to move their data out. Jaafar Al Hashmi of Core42 said the partnership moves sovereign AI from a concept into an operational capability. That is a product launch, not an infrastructure announcement. It is a useful marker of where this market has got to.
What has been built
The capacity numbers in the region are large and mostly verified. Stargate UAE is a 1GW compute cluster in Abu Dhabi, built by Khazna and operated by OpenAI and Oracle, with G42, NVIDIA, Cisco and SoftBank as partners. The first 200MW phase is expected live in 2026. It sits inside the UAE-US AI Campus, a site targeting 5GW.
Microsoft and G42 announced a separate 200MW expansion through Khazna, part of Microsoft’s $15.2 billion UAE commitment through 2029. It is expected to start coming online before the end of 2026.
In Saudi Arabia, HUMAIN agreed a financing framework with the National Infrastructure Fund worth up to $1.2 billion, covering up to 250MW of hyperscale capacity. Those terms are non-binding. HUMAIN’s stated roadmap runs to 6.6GW over a decade. Google Cloud and PIF are funding a $10 billion AI hub in the Kingdom, launched with HUMAIN.
And on 1 July, MGX closed its first fund at $49 billion, above a $45 billion target, drawing investors from the Gulf, North America, Asia and Europe. It has invested in 14 companies across semiconductors, AI infrastructure and platforms.
Almost all of the above is denominated in megawatts and dollars. Very few disclose who has agreed to pay for the output. That is normal for infrastructure at this stage, and it is not evidence of a problem. But it does mean the announcements tell you what is being built and not what it is for. Which is why two recent items are worth more attention than their size suggests.
Cohere is the worked example
On 9 July, HUMAIN and Cohere announced a partnership. HUMAIN will designate at least 50 megawatts of dedicated compute to Cohere’s next-generation foundation models, expandable over five years, expected live in the fourth quarter of 2027. Fifty megawatts is small next to a gigawatt. But it is contracted, dated and attached to a named customer.
Cohere is Canadian. Semafor reported that this is its first major deployment outside North America. The company could have placed the workload in North America or Europe. Aidan Gomez, Cohere’s chief executive, framed the decision around compute access, saying the partnership gives the company the scale and flexibility it needs for future generations of enterprise models. That is what demand looks like when it shows up: a specific customer, a specific quantity, a specific date.
What changes for a company in the region
The e& and Core42 launch answers a different question, and a more immediate one for most businesses here. Until recently, a regulated company in the UAE had two unattractive options. Use offshore GPU capacity and deal with data residency problems. Or build capacity in-house, which takes capital and time.
The new platform is aimed squarely at that gap. Core42 positions it as covering the full AI lifecycle rather than being a general cloud with GPUs attached, with data residency and access controls built through the stack. For a bank, insurer or healthcare group, the practical effect is that in-country compute becomes something you rent rather than something you build.
Two commercial details are worth noting because they are unusual. No capital expenditure requirement removes the main budgeting obstacle. Zero egress fees remove the main lock-in objection, since customers can leave without paying to take their data with them.
What to watch
The build phase in the Gulf is well documented. The next phase is less visible and more important: who signs up, for how much, and for how long. Three things would tell you the market is maturing. Named customers rather than named partners. Contracted quantities rather than site capacity. And repeat business, meaning the second and third deals with the same buyer.
The Cohere agreement is the clearest example so far of the first two. The e& and Core42 platform is a bet on volume from ordinary enterprises rather than from frontier labs. Both are the beginning of the demand side becoming legible. That is the part of this story worth following over the next few quarters, because it is what will determine whether the capacity now under construction gets used at the rate it is being built.