Mubadala and Together AI explore UAE infrastructure; Abu Dhabi office confirmed, but no data centre yet

The business so what: this alters your vendor roadmap, not your production architecture this quarter. Mubadala put $100 million into Together AI's Series C and has attached an exploratory partnership to it, which provides a strong signal that open-model inference capacity will eventually arrive in the Gulf. However, the confirmed scope remains an Abu Dhabi office and an agreement to identify opportunities. There is no named data centre, no allocated megawattage, no UAE region in the API catalog, and no operational timetable. A bank, healthcare provider or government entity planning 2027 workloads should track the partnership, but any supplier claiming this agreement satisfies UAE data residency requirements today is confusing a press release with tenancy architecture.

Why it matters anyway: the unit economics of efficiently served open models are low enough to reshape build-versus-buy calculations across the GCC. Together AI's published serverless model catalog lists DeepSeek-V4-Flash-0731 at $0.14 input and $0.28 output per million tokens with a 1,048,576-token context window in FP4, openai/gpt-oss-120b at $0.15 and $0.60 in MXFP4, and frontier-tier moonshotai/Kimi-K3 at $3.00 and $15.00. Worked example: an enterprise assistant processing 200 million input tokens and 50 million output tokens monthly costs roughly $42 on DeepSeek-V4-Flash, compared to $1,350 on Kimi-K3. That economic divergence explains why Gulf sovereign capital is backing both sides of this market. Aramco Ventures led the $800 million round and Mubadala committed $100 million, while round investors pledged over 500 MW of compute capacity capitalised separately. Note that this 500 MW figure represents global capacity, not domestic UAE power.

Where engineering teams risk making mistakes: conflating sovereign investment branding with legal jurisdiction, and coupling code to a single API. Nothing in the announcements indicates where inference tokens are processed. What teams can inspect today is standard commercial cloud pricing: Together AI's dedicated model inference pricing charges per minute for each dedicated replica, billing $3.99 hourly for a 1xnvidia-h100-80gb and $8.99 hourly for a 1xnvidia-b200-180gb, with H200 141GB, B300 280GB and GB300 280GB available on custom quotation. An H100 replica running continuously across a 30-day month incurs approximately $2,873. That is a standard global cloud contract without domestic data residency guarantees. The secondary hazard is architectural lock-in: embedding proprietary SDK bindings and provider-specific model strings directly into services makes future migration to local UAE capacity unnecessarily costly.

How Azrty approaches this technically: deploy an OpenAI-compatible gateway in front of all workloads so provider selection becomes a configuration change rather than a code rewrite. Our FastLLM Proxy sits before internal model servers and external hosted providers, consolidating routing, budgets, and access control. On-premises or sovereign GPU fleets run on Kubernetes via our Kuvryn and Kuvryn AI platforms. Routing logic should enforce data classification: restricted records remain entirely on local GPUs, while non-sensitive workloads can burst to external serverless endpoints with exact model strings pinned. Illustrative routing policy:

routes:
  - name: restricted-data
    match: { data_class: restricted }
    upstream: own-gpu-cluster        # internal infrastructure, models served locally
    models: ["qwen3.6-35b-a3b-fp8"]
  - name: general-assist
    match: { data_class: internal }
    upstream: together-serverless
    models: ["deepseek-ai/DeepSeek-V4-Flash-0731", "openai/gpt-oss-120b"]
    fallback: own-gpu-cluster
budgets:
  team-platform: { monthly_usd: 5000, action: throttle }

The strategic upside for the region is substantial: the provider that first delivers production open-model inference in the GCC captures regional latency advantages, sovereign compliance, and enterprise developer volume simultaneously. Gulf sovereign wealth is backing the probable leaders of that shift. The operational risk lies in mistaking financial announcements for production infrastructure. Insist on three verification criteria before routing regulated workloads: an identifiable physical facility, verified local capacity metrics (MW or GPU count), and a UAE endpoint within the provider API. Until then, maintain model-agnostic routing, audit serverless costs against dedicated infrastructure, and treat local capacity milestones as procurement options rather than immediate migration triggers.