H100$6.39/hr▼ 1.2% 7d
A100 80GB$2.45/hr▼ 0.5% 7d
H200$10.29/hr▼ 0.8% 7d
L40S$1.28/hr▲ 0.3% 7d
T4$0.24/hr▼ 0.6% 7d
L4$0.45/hr▼ 1.1% 7d
H100$6.39/hr▼ 1.2% 7d
A100 80GB$2.45/hr▼ 0.5% 7d
H200$10.29/hr▼ 0.8% 7d
L40S$1.28/hr▲ 0.3% 7d
T4$0.24/hr▼ 0.6% 7d
L4$0.45/hr▼ 1.1% 7d
NewMENA AI Factsheet Q3 2026Read the report
Middle East & North Africa · Q1 2026

MENA AI Factsheet Q1 2026

Data as at 31 March 2026. Period 1 Jan to 31 Mar 2026. Free to read and download.

This factsheet takes a holistic look at AI infrastructure in MENA in Q1 2026. Live AI data centre capacity (section 1) and projects due to go live (section 2) give an overview of AI supply in the region. Compute pricing (section 3) gives insight into demand for AI compute, for training and inference.

Our analysis puts this data into context. Section 4 estimates the returns operators and investors could expect from a new GPU in the region at Q1 prices. For CTOs, CFOs and AI developers, section 5 compares the cost of deployment options and which suits each workload.

1

Live AI data centre footprint

Powered by Signwl Capacity Atlas™

Estimated data centre capacity that was live at the end of the quarter in each market, for all workloads and for AI, with the number of live facilities. Powered by Signwl Capacity Atlas™: project-level pipeline, with each project's status checked against construction evidence.

The UAE and Israel lead MENA's live AI capacity, with 39% and 27% of the region. All of the AI capacity under construction from projects due by 31 Mar is in Saudi Arabia (ICS Arabia Desert Dragon Riyadh phase 1, section 2); none of it is confirmed live. If it comes online, the region's AI capacity would rise about 1.4 times, with Saudi Arabia holding 46% and the UAE 28%. Two of three AWS availability zones in the UAE and one of three in Bahrain lost power on 1-2 Mar 2026; restoration by 31 Mar is not confirmed in AWS statements, so these buildings are not counted as live facilities. Live IT is not adjusted because their sizes are not published.

CountryLive
facilities
Live IT,
MW
Live AI IT,
MW
AI share
of live IT
AI watts
per person
Share of
MENA AI
Live + under construction AI IT, MWShare of MENA AI, live + under construction
UAE49375–40522–418%2.0–3.839%22–4128%
Israel39180–23012–3110%1.2–3.127%12–3119%
Saudi Arabia47135–24514–2310%0.4–0.723%47–5646%
Qatar1244–461–23%0.3–0.82%1–21%
Bahrain87–230–13%0.1–0.61%0–1<1%
Oman1012–240–38%0.1–0.62%0–32%
Egypt3029–731–34%0.03%1–32%
Other MENA†6743–1251–32%0.03%1–32%
MENA262825–1,17151–1088%0.1–0.3100%84–140100%

Sources and methodology. Signwl Capacity Atlas™ combines Signwl's data centre register, operator and regulator disclosures, market reports and AI server import data to estimate live and under-construction IT and AI capacity.
†Morocco, Kuwait, Jordan, Iraq, Lebanon, Tunisia and Algeria.

2

Status of projects due to go live by 31 Mar 2026

Powered by Signwl Capacity Atlas™

Publicly announced data centre projects that were due to be live by the end of the quarter, and whether each has been confirmed live. Powered by Signwl Capacity Atlas™: project-level pipeline, with each project's status checked against construction evidence.

Of 173 MW publicly announced for completion by 31 Mar 2026, none is confirmed live; 65 MW is under construction and 108 MW shows no evidence of construction. 113 MW of the total is in Saudi Arabia; a further 590 MW is due in Q2 on announced dates.

DeliveredUnder construction, not confirmed liveNo evidence of construction
ICS Arabia Desert Dragon Riyadh phase 11 Saudi Arabia
65 MW
Naver / Nvidia Nouaceur phase 1 Morocco
40 MW
Ezditek RUH01 (Riyadh) Saudi Arabia
24 MW
Saudi Data Centre Fund / Sahayeb expansion Saudi Arabia
24 MW
SoleCrypt Bizerte2 Tunisia
20 MW

1.Now expected Jun 2026 (ICS Arabia, Jan 2026); part of a 187 MW programme across Al-Kharj, Jeddah, Dammam and NEOM.

2.Now expected 2027 (SoleCrypt, Feb 2026).

Sources and methodology. Signwl Capacity Atlas™ tracks publicly announced projects with a stated MW and a due date, and checks their status against operator, contractor and regulator evidence dated on or before the quarter-end. Under construction means physical works have been reported within the previous 12 months but the capacity is not confirmed live. Projects are shown at their originally announced capacity and date; any project not confirmed live by the quarter-end is behind its announced schedule, and announced revised dates are footnoted.

3

Compute pricing: on-demand and spot

Powered by Signwl GPU Price Index™

Hourly GPU rental prices from the major cloud providers in MENA, on demand and on spot (Q1 2026 averages of daily prices). Spot is spare capacity sold at a discount, so its gap to on-demand signals demand. No discount points to a tight market, a narrowing one to tightening, and a widening one to growing spare capacity. Powered by Signwl GPU Price Index™: daily on-demand, reserved and spot prices for 20+ GPUs across 40+ metros.

GB200 and H200 spot averaged above on-demand in Q1, so they show no spot discount; Signwl analysis rates demand for them as High. L4 also shows high demand, with the lowest discount in the Dubai market, where spot averaged $0.56 in Q1, 24% below on-demand, and spot rose to an average of $0.71 per GPU-hour in March. Between 12 Jan and 31 Mar, spot discounts narrowed for A10G by 11 points, and L4 from 59% to 31%, which Signwl analysis reads as rising demand. Spot discounts widened for A100 80GB from 47% to 75%, which Signwl analysis reads as more spare capacity. Demand is low for MI300X, V100, L40S and A10, where spot sold at more than 70% below on-demand on average in Q1 (70.5%, 71%, 72% and 77%).

Training GPUs

GPUOn-demand
$/GPU-hr
Spot
$/GPU-hr
Spot discount,
Q1 average
Daily spot
discount, Q1
12 Jan → 31 MarMENA
demand
GB20024.19above on-demand†none0% → 0%High
H20014.40above on-demand†none0% → 0%High
H1009.052.8868%68% → 68%Medium
MI300X4.681.3870.5%71% → 71%Low
A100 80GB3.621.1867%47% → 75%Medium
V1002.410.7171%71% → 71%Low

Inference GPUs

GPUOn-demand
$/GPU-hr
Spot
$/GPU-hr
Spot discount,
Q1 average
Daily spot
discount, Q1
12 Jan → 31 MarMENA
demand
L40S2.010.5572%––Low
A101.370.3277%78% → 77%Low
A10G0.960.4653%51% → 40%Medium
L40.820.4839%59% → 31%High
T40.350.1069%70% → 72%Medium

Sources and methodology. Signwl GPU Price Index™: Q1 2026 averages of daily list prices from the major cloud providers in MENA (price data from 8 Jan 2026; daily spot discount line from 12 Jan). Spot discount compares spot with on-demand in the same market. MENA demand from the Q1 average spot discount: the smaller the discount, the less spare capacity providers have to sell. High: under 50%; Medium: 50–70%; Low: over 70%.
†Spot listed above the on-demand price; no discount. L40S has no daily series: priced on too few days for the daily panel.

4

GPU economics in MENA

Powered by Signwl Returns Model™ · Signwl Component Tracker™

What one GPU in a new MENA build earns over three years when rented out at reserved list prices, after power, facility and maintenance costs. Powered by Signwl Returns Model™: GPU and data centre returns on your own assumptions; Signwl Component Tracker™: RAM, CPU and storage costs bundled into GPU pricing.

Signwl analysis suggests the H100 pays back fastest, after 9.0 months, followed by the GB200, after 10 months. At reserved list prices and 80% utilisation, H100, GB200 and H200 repay their cost within 3 years (300%, 246%, 132%). GB200 and H100 make money at half of list price; the others do not. A100 80GB returns 10% over 3 years, and Signwl analysis suggests MI300X, L40S and L4 do not pay back within a 3-year horizon.

GPUAll-in cost
$ per GPU
Rent, 3-yr
$/GPU-hr
Annual price
decline
Other costs
$ a year
Payback
months
Year-1
return
3-year
ROI
GB20054,20010.5225%11,1851091%246%
H20045,7076.2624%8,5731653%132%
H10031,2506.9116%6,8389.0117%300%
MI300X48,7873.3615%8,5083916%-8%
A100 80GB28,7832.2814%5,4083322%10%
L40S15,9640.9025%3,06259-5%-39%
L46,4680.3621%1,138551%-35%

Sources and methodology. Signwl Returns Model™ estimates the return on one new GPU in MENA, bought for cash and rented at Q1 2026 reserved list prices, using Signwl price and depreciation data (at 31 Mar 2026) with published hardware, power and facility costs; hardware prices and the power tariff are those published on or before 31 Mar 2026. Hardware is priced at US dollar list prices, the market the region buys from.

5

AI deployment cost comparison

Powered by Signwl Compute Outlook™

The monthly cost of running two typical AI workloads on models that keep processing inside MENA: an API, or an open model on rented MENA GPUs. Only APIs and GPU instances available in the region at 31 Mar 2026 are included. Powered by Signwl Compute Outlook™.

For high-volume simple tasks at 1bn tokens a month, Amazon Nova Pro costs $1.0k, $303 more than running an open model on frontier GB200 GPUs, and Amazon Nova Lite costs $78, $137 less than running an open model on the cheapest suitable GPU, MI300X.

For complex reasoning and agents at 1bn tokens a month, Amazon Nova Pro costs $1.3k, $976 less than running an open model on frontier GB200 GPUs and $621 more than running an open model on the cheapest suitable GPU, MI300X.

Monthly cost at 1bn tokens a month (bars scaled within each workload; all options processed in MENA)

  • Ready-made service (in-region API)
  • Open model, self-hosted on rented MENA GPUs

Best available

High-volume simple tasks
Amazon Nova Pro$1.04 per 1M tokens
$1.0k
gpt-oss-120b on 1x GB200$0.74 per 1M tokens
$737
Complex reasoning and agents
Amazon Nova Pro$1.28 per 1M tokens
$1.3k
Leading open model on 8x GB200$2.26 per 1M tokens
$2.3k

Lowest cost

High-volume simple tasks
Amazon Nova Lite$0.08 per 1M tokens
$78
gpt-oss-120b on 1x MI300X$0.22 per 1M tokens
$215
Complex reasoning and agents
Amazon Nova Pro$1.28 per 1M tokens
$1.3k
Leading open model on 8x MI300X$0.66 per 1M tokens
$659

Sources and methodology. Signwl Compute Outlook™ compares in-region API list prices with self-hosting open models on rented MENA GPUs, using published benchmark throughput. Only models processed in a MENA cloud region at 31 Mar 2026 are included, with availability and Nova list prices from AWS documentation and price lists dated on or before that date; for complex work one per-token API met this test, so the best available and lowest-cost services are the same. Quality scores are not used for Q1; Best available is the higher-tier in-region API. Self-hosted cost scales with the GPU-hours used. Excludes start-up time and engineering and operations costs, and assumes GPUs are available when needed.

To discuss this factsheet or Signwl data, contact us at signwl.com/contact

Signwl, GPU Price Index, Component Tracker, Capacity Atlas, Returns Model and Compute Outlook are trademarks of Signwl.© 2026 Signwl