US greenlights sale of 35,000 AI chips to Gulf firms G42 and Humain

On November 20, 2025 the U.S. Commerce Department lifted a major export‑control hurdle, authorizing the sale of up to 35,000

Nvidia Blackwell GB300 GPUs

to two flagship Gulf technology firms: G42 in the United Arab Emirates and Humain in Saudi Arabia. The move, valued at roughly $1 billion, is more than a headline‑worthy hardware transfer; it signals a strategic realignment of AI leadership in the Middle East, reshapes supply‑chain dynamics for high‑performance compute, and sets a precedent for future U.S.–Gulf tech cooperation.

• The deal authorizes equal allocations of 35,000 GB300 GPUs to G42 and Humain, marking the first time a single U.S. export decision empowers both UAE and Saudi Arabia with identical cutting‑edge AI hardware.

• Blackwell’s ~200 TFLOPS FP32 performance and 4.5× efficiency over Nvidia’s previous H100 positions Gulf data centres to train large language models (LLMs) and other demanding workloads at lower energy costs.

• The transaction is wrapped in strict EAR/ITAR reporting requirements, underscoring the U.S.’s intent to balance alliance support with containment of high‑risk technology transfer.

• For hardware vendors, data‑centre operators, and enterprise buyers, the deal signals an immediate opportunity to capitalize on a new market segment while preparing for supply‑chain scaling and competitive pressure from AMD and other players.

Enterprise architects who are scouting next‑generation GPU platforms must now factor in the Gulf’s accelerated access to Blackwell. The implications unfold across three axes:

capability expansion, cost optimization, and geopolitical risk management.

The GB300’s 200 TFLOPS FP32 throughput is roughly 4.5× the H100’s peak, enabling Gulf data centres to train or fine‑tune models with 175 billion parameters in a fraction of the time and energy. For an enterprise that relies on LLMs for customer service, fraud detection, or product recommendation, this translates into:

• Up to 30% faster model convergence when using mixed‑precision pipelines.

• A 25–35% reduction in per‑epoch energy consumption , given the GB300’s ~600 W/TFLOP efficiency.

• The ability to run parallel inference workloads for multiple tenants without overprovisioning.

Data‑centre operators are increasingly pressured to meet carbon‑neutral targets. The GB300’s improved power efficiency directly impacts the total cost of ownership (TCO). Assuming a typical 20 kW per GPU rack, a shift from H100 to GB300 can cut data‑centre cooling and power budgets by ~15–20%. Over a three‑year horizon, this equates to:

• Approximately $1.2 million in avoided energy costs for a 5,000‑GPU deployment.

• Lower capital expenditures (CapEx) on cooling infrastructure due to reduced thermal density.

• A faster return on investment (ROI) , especially when coupled with the increased throughput per watt.

The U.S. has embedded rigorous security reporting requirements into the deal, a move that serves two purposes:

• It ensures that downstream use of the GPUs remains within U.S. policy bounds, preventing spill‑over to adversarial entities.

• It gives enterprises confidence that their hardware investments are protected by a stable regulatory environment, reducing the risk of sudden revocation or compliance penalties.

For businesses with global supply chains, this also means that any integration of Blackwell GPUs must be accompanied by robust audit trails and adherence to EAR/ITAR protocols—an operational consideration that can influence procurement timelines and contract structures.

Deploying 35,000 GB300 GPUs is a logistical challenge. Below is a pragmatic roadmap that balances hardware acquisition, infrastructure scaling, and software stack readiness.

• Engage with Nvidia’s sales team to secure early access to the 4 nm TSMC process line, which will be critical for meeting delivery windows.

• Coordinate with cooling vendors (e.g., liquid‑cooling providers) to design racks that accommodate the GB300’s higher thermal density without compromising airflow.

• Secure power contracts that can support a projected 500‑MW data‑centre footprint—typical for a Gulf‑scale deployment.

• Power Distribution: Allocate 20 kW per rack with a 15% headroom to accommodate future upgrades.

• Cooling Strategy: Deploy rear‑door exhaust liquid cooling or in‑rack immersion systems to maintain < 35°C ambient temperatures.

• Networking: Implement InfiniBand HDR (200 Gbps) interconnects to minimize latency between GPU nodes, essential for distributed training workloads.

• Upgrade to Nvidia CUDA 12.5 , which introduces enhanced kernel fusion and memory‑overlap features tailored to GB300.

• Adopt TensorRT 9 for inference acceleration, leveraging the GPU’s FP16/INT8 throughput improvements.

• Utilize Hugging Face Accelerate or DeepSpeed ZeRO‑3 to scale LLM training across thousands of GPUs efficiently.

• Implement a dedicated compliance officer role focused on EAR/ITAR reporting, ensuring that all data transfers, software licenses, and hardware movements are logged.

• Integrate automated audit tools (e.g., ComplianceTrack ) to generate quarterly reports required by the U.S. Commerce Department.

The Gulf’s acquisition of Blackwell chips reshapes the competitive landscape for AI hardware in 2025. While Nvidia maintains a dominant share, AMD and Intel are poised to capture niche segments if supply constraints emerge.

• Blackwell’s performance leap cements Nvidia’s position as the sole provider of high‑throughput GPUs capable of training 175B+ parameter models at scale.

• The deal’s value—$1 billion—highlights the commercial viability of selling GPUs to sovereign entities, a model that could be replicated in other emerging markets.

While AMD chips were included in broader export approvals, they did not feature as the flagship line. However, AMD’s

MI300X

offers competitive FP16 performance and lower TDP for certain workloads. Enterprises that prioritize cost over raw throughput may consider a hybrid approach.

Intel’s upcoming

Lakefield‑AI

ASIC, targeting edge inference, remains a distant competitor for data‑centre scale training but could become relevant for multi‑tenant inference services in the Gulf region.

For an enterprise planning to deploy 5,000 GB300 GPUs across two data centres, the financial upside is substantial. The following simplified model assumes a 30% throughput gain and a 20% energy cost reduction over H100.

Metric

Baseline (H100)

Post‑Blackwell

Total Training Time (per epoch)

10 hours

7 hours

Energy Consumption per Epoch (kWh)

15,000

12,000

Annual Energy Cost ($)

$3.6 million

$2.9 million

CapEx for Cooling Infrastructure (per 1k GPUs)

$500 k

$400 k

Total CapEx (5k GPUs)

$2.5 million

$2.0 million

Net Savings Over 3 Years

-

$1.8 million

These figures suggest that a well‑executed Blackwell deployment can deliver a

~15% cost reduction** in operating expenses (OpEx) and a significant upfront CapEx saving, translating into an accelerated ROI timeline.

• Secure Early Access: Engage Nvidia’s channel partners to lock in supply contracts before the 4 nm fab ramp-up completes.

• Integrate Compliance from Day One: Build EAR/ITAR reporting workflows into procurement and deployment pipelines to avoid costly delays.

• Adopt a Hybrid Hardware Strategy: Combine Blackwell GPUs for training with AMD or Intel ASICs for inference, optimizing cost versus performance.

• Leverage Gulf Partnerships: Explore joint‑venture models with G42 or Humain to share infrastructure costs and accelerate AI product launches in the MENA market.

• Plan for Future Upgrades: Anticipate Nvidia’s next‑generation GPUs (e.g., Blackwell successor) by designing modular racks that can accommodate future ASICs without major redesign.

The 2025 Gulf deal is a catalyst for several emerging trends:

• Regional AI Ecosystem Maturation: With Blackwell in place, UAE and Saudi Arabia can move from model training to commercial AI services, attracting talent and investment.

• Supply‑Chain Resilience: The demand spike may prompt Nvidia and TSMC to expand 4 nm capacity, potentially reducing lead times for other customers.

• Export‑Control Evolution: The U.S. will likely refine EAR/ITAR requirements, balancing alliance support with geopolitical risk mitigation, which could affect future hardware deals.

• Competitive Diversification: AMD and Intel may accelerate their AI chip roadmaps to capture mid‑tier market segments if Nvidia’s supply chain faces constraints.

The U.S. Commerce Department’s approval of 35,000 Nvidia Blackwell GPUs for G42 and Humain is more than a hardware transaction; it is a strategic endorsement of Gulf AI leadership underpinned by stringent export controls. For enterprises, the deal offers an immediate pathway to high‑performance training capabilities, lower operating costs, and a robust compliance framework. It also signals that the next wave of AI infrastructure will be shaped not just by technological advances but by geopolitical alignments and supply‑chain resilience.

Business leaders who act now—securing early access, integrating rigorous compliance measures, and adopting a hybrid hardware strategy—will position themselves at the forefront of an evolving AI landscape that is set to redefine data‑centre economics and regional technology leadership in 2025 and beyond.