Eighteen billion dollars, five million trainees, and a laptop that ships before its operating system: LEAP 2026 in full

Riyadh, 5 September 2026, from the Malham grounds. Figures are as stated by the announcing party or the Saudi news agencies.

The totals, day by day

Day 1 closed at "more than USD 15 billion." Day 2 added about USD 2.5 billion. Day 3 added USD 857 million. Day 4 came with no announced figures at all, and it may have been the most important day of the week.

Call it USD 18.4 billion across three days, against USD 22.4 billion over the first two days of 2025. The shape is roughly the same as last year, a huge opening day and then a steep drop. The opening day is the deals negotiated before the event; the rest of the week is what was agreed at the show. Read that way, Days 2 and 3 are the more honest number, and both fell by more than half.

There is a second difference, and it took me until Day 3 to see it. Last year's opening day was a row of named cheques: Groq and Aramco Digital at USD 1.5 billion, Salesforce at USD 500 million, Databricks at USD 300 million, Tencent Cloud at USD 150 million, SambaNova at USD 140 million. This year's opening day is one company's stack. The opening-day total is close to last year's, but it is now concentrated in one company, and this year the whole idea is expansion and strategic partnership, now that the build-out phase has been played.

Day 1: The stack

The minister, Abdullah Alswaha, opened with the economic figures: the digital economy up 69% since the first LEAP in 2022, from USD 118 billion to USD 199 billion; technology jobs at about 410,000, from about 150,000; nine unicorns; and first globally on digital readiness. From the hall he also gave women at 36% of the tech workforce, more than 3,000 locally manufactured servers exported, 80 km² of land and more than 12 GW of power set aside for AI data centres, and more than 20 subsea cables by 2030. He described the shift as moving from importer of technology to exporter.

Then HUMAIN took most of the oxygen. Everything it announced fits one stack.

Compute. The AMD and Cisco data centres are the only items this week with something tangible on the ground: MI355X systems with Cisco Silicon One networking, running in production as GPU-as-a-service, described as AMD's largest inference deployment outside the United States.

HUMAIN puts 13,000 AMD GPUs into service during 2026, then adds up to 250 MW from the second half of 2027, aiming at 1 GW by 2030 on the MI400 series. Alongside it, an NVIDIA HGX B300 Blackwell Ultra cloud that HUMAIN says runs above 90% cluster utilisation, heading to roughly 35 MW by mid-2027.

xAI's Saudi build, first announced at the US–Saudi Investment Forum last year, starts at 50 MW and scales to 500 MW, with roughly 18,000 NVIDIA GB300s in the first cluster and Grok deployed country-wide.
Together AI takes a 250 MW partnership with a stated target of USD 5 billion in annual revenue in its first operating year. DataVolt has broken ground on 100 MW in the Oxagon zone in NEOM, due for completion in 2028, the first phase of a campus planned at 1.5 GW.

AWS commits 50 MW in its AI Zone by 2028, with its Saudi region on track to launch in December 2026 and USD 5.3 billion behind it. stc's center3 develops capacity from 250 MW, and Olayan Alwetaid put the HUMAIN partnership at three things: data centres, connectivity and go-to-market. Microsoft's Saudi Arabia East region, with three availability zones, goes live in November.

Language models. HUMAIN M3, an Arabic model with open weights, trained on more than one trillion Arabic tokens. HUMAIN Voice, which handles the region's dialects (Saudi, Egyptian, Levantine and Maghrebi) as separate dialect targets. ALLAM, the previous generation of their model, available on the local cloud and now reachable through Amazon Bedrock, Microsoft Foundry and HUMAIN Fabric, and listed on the AWS Marketplace. And Reflection AI will run its models on HUMAIN infrastructure inside the Kingdom.

The open-weights decision for M3 is probably the best thing that happened this week for anyone building in Arabic, and its base model is MiniMax-M3, one of the best open-source models of recent months.

Applications. HUMAIN ONE comes bundled with Microsoft 365 Copilot on Azure, with ready-made Arabic agents, targeting a million users across the Middle East and Africa over the coming period.

Devices and the operating system. Horizon Ultra with Qualcomm: Snapdragon X2 Elite, 18-core Oryon CPU, Adreno GPU, Hexagon NPU. It runs Windows, enterprise orders open on 20 September, no price has been announced, and the target is one million devices across the Middle East and Africa by 2030. HUMAIN OS, which Tareq Amin calls HUMAIN's real product, turns out to be Linux with built-in AI, co-developed with KORA Systems, and ships commercially in 2027. So the flagship device ships for fifteen months running someone else's operating system.

Physical AI. Applied Intuition had a big presence and will start with autonomous trucking, aiming at thousands of vehicles on Saudi logistics supply chains by 2030.

The rest of the floor

Plenty of organisations besides HUMAIN are working on the AI map here, and the second tier had the working software.

Elm came to LEAP as business solutions partner under the line "ذكاء يمهّد الطريق", intelligence that paves the way, and its pitch is the one number on this floor that describes AI reaching citizens now: more than 30 million users across more than 16 sectors, more than 3 billion digital transactions a year, more than 350 e-services, and four international offices. Elm launched no new product. It showed AI that works out of sight and is still felt (process automation, decision support, operational efficiency), showed what it has already done and what came of it, and announced new programmes for entrepreneurs and SMEs. If any deployment in the Kingdom is going to put an Arabic model in front of thirty million people, it is more likely to look like Elm, given its large user base in the Kingdom and its growing reach in the region.

stc group, strategic partner for the fifth consecutive year, showed the infrastructure and network side of the same story, with center3 as its capacity vehicle.

Aramco Digital launched iCONNA, an industrial edge AI platform, and Aramco signed a set of agreements and MoUs on cybersecurity, industrial AI and localisation of critical technologies. Amin Nasser framed it as investment momentum that makes the Kingdom a leader in AI and digital solutions.

HP did more product than most: HP Agent One, an enterprise AI agent that handles IT service requests in Arabic; FlowMaster, for document automation and compliance; HP Garage 2.0, a startup incubator; and an R&D Center of Excellence in Dhahran working on Arabic language models, including the Najdi dialect. HP also extended "Saudi Made" with Alfanar to assemble Gen12 ProLiant servers locally and opened a Saudi Innovation Center with Intel.

Think launched Think Grid, subscription AI compute out of a Tier III facility in Riyadh, built on a SuperNode of four NVIDIA RTX PRO 6000 Blackwell accelerators with 384 GB of memory, managed by its own ILM software.

Lenovo unveiled the first laptop manufactured in Saudi Arabia, with Alat, moving that partnership from capacity-building into actual production. That is the second local laptop story of the week, and the only one with a factory behind it.

Nokia opened a Saudi R&D centre for 6G software and AI network automation. AMD, MCIT and the Digital Cooperation Organization launched a joint developer programme around ROCm, with training on AI ethics. Adobe put USD 4 billion of contribution to the Saudi creative economy over twelve months on the table, plus a year of free access to its tools for 27 million eligible residents. Mobily and BytePlus will open BytePlus's first public cloud region in the Middle East and Africa, in Riyadh. SDAIA ran DeepFest alongside the main floor: more than 50 sessions, more than 100 speakers from 17 countries, and 100 brands. And the Ministry of Human Resources and Social Development, with MCIT, Microsoft and Gulf Intelligence, launched the AI Centre of Excellence for national skills and AI adoption in the labour market.

Day 2: Capacity

Almost entirely data centres. Al Moammar Information Systems put USD 1.2 billion into expansion, taking its total capacity to 192 MW. NHC Innovation announced USD 880 million for Khuzam Digital Valley, scaling to 65 MW by 2033, with NAVER Innovation and BytePlus as its first two anchor customers. Mobily and BytePlus committed more than USD 150 million to the Riyadh cloud region. Vision Invest and the Africa Finance Corporation put USD 300 million into WIOCC for subsea and terrestrial fibre linking Saudi Arabia to East Africa.

Take those at face value and add them up, and the week put about 2.7 GW of intent against 467 MW of running capacity. That is my arithmetic, and it double-counts: AMD's 1 GW target contains the 250 MW, xAI's 500 MW contains its first 50, and the Oxagon phase is one slice of a 1.5 GW campus. Halve it and it is still nearly three times what exists.

The more useful question is when any of it arrives. Strip out everything without a stated year and the curve is much thinner than the headline:

Nothing dated lands in 2026 at all. The committed curve does not cross today's operational capacity until 2030, and more than half of what gets it there is a single AMD target four years out.

Day 3: The money gets smaller and more interesting

USD 857 million, of which about USD 293 million was venture capital across eleven deals: nine rounds worth USD 126.15 million, spread across AI, fintech, travel, gaming, HR and digital automotive services; and one new fund plus one expanded fund worth USD 166.7 million between them, being a USD 100 million Saudi–Korea deep-tech fund from Ula Capital and FuturePlay, and Impact46 and HALA taking their SME financing vehicle to USD 66.7 million. Alfanar committed USD 150 million to manufacture data-centre components locally, which is the supply chain under every megawatt above.

If a founder asked me which day to read for money, I would say this one. Day 1 belonged to the hyperscalers and the sovereign champion; Day 3 had rounds a startup can raise.

Day 4: The day nobody priced

The day came with a headcount instead of a dollar figure, and the headcount is the largest number of the week.

MCIT, the Saudi Digital Academy and Google put a Gemini programme in front of one million university students. Microsoft opened an AI Academy with one million trainees in phase one, scaling to three million by 2030, and added an Arabic Intelligence centre of excellence. ServiceNow launched Saudi AI Forward at 1.2 million beneficiaries with professional certifications. Scale AI committed to 100,000 people over four years through a programme called Alf for applied generative AI. Intel took 100,000 national cadres. Artifact Academy took more than 20,000 specialists by 2034. Schneider Electric took 10,000 by 2030. An AI Academy partnership with We Cloud Data took 500 teachers, 500 academics and 1,000 entrepreneurs in person, and 35,000 virtually. A partnership with a social media platform added 20,000 across geospatial computing, augmented reality, digital safety and marketing.

Add the headline targets and it is about 5.5 million pledged training places, in a country of roughly 35 million people. That is my arithmetic, and it certainly double-counts, since nothing stops one student appearing in the Google million, the Microsoft million and a ServiceNow certification at the same time. Discount it however you like and it is still the largest commitment of the week by reach, and the only one aimed at the demand side.

It is also the day with the weakest accountability, and plotting the pledges against their deadlines shows why.

A megawatt gets metered. A million trainees is a number nobody will ever audit, two of the three largest pledges carry no end date at all, and every one of these partners has an obvious commercial interest in the trainee learning its own platform.

What was shown

The fifth edition, themed "Into New Worlds", at the Riyadh Exhibition and Convention Centre in Malham: more than 1,000 speakers, 1,800 global brands, 600 startups, 16 stages and more than 20 tracks. Rocket Fuel took more than 3,000 applications down to 100 finalists competing for USD 1 million, equity-free. LEAP Nights ran more than 60 events across more than 15 venues. Lisa Su and Chuck Robbins were on stage; NVIDIA, Google, Meta, Uber, Roblox, Lenovo and Nokia were in the halls.

Over four days I saw no demo that looked like it was on rails. Unitree quadrupeds and humanoids, AGIBOT's A2 Ultra, COGNIXION's brain-signal interface, delivery drones, a robotaxi and a flying taxi. The one that stayed with me was Compal's POLYMEDX, an autonomous robotic wheelchair for hospitals, already running in emergency departments in Taiwan, moving patients between departments under human oversight with an emergency stop button. Physical AI that travels next to a patient is a completely different governance problem from a chatbot, and it was the only demo where somebody told me that without my asking.

Two things were not shown. Nobody ran M3 locally on a Horizon Ultra, on either day the laptop was on stage, which for a device whose whole story is on-device AI is the demo I most wanted. And the fintech and education tracks finished four days of sessions without a single sector announcement, like every previous edition, which is a little harder to explain in the week the Kingdom bought five million training places.

The week produced seventeen infrastructure announcements, fourteen skills programmes, eleven venture deals and five models. The Kingdom bought compute and students, and only a little of the software between them.

The open model

I want to be careful here, because this is the item I care most about, so it is the one I am most likely to overrate.

M3's weights being open means a team in Jeddah or Khartoum can pull them, run them, fine-tune them and publish an evaluation without asking anyone. The Arabic model market has not had many open models. It matters more this year than last, because the closed models are now everywhere: ALLAM through Bedrock and Foundry, Reflection on HUMAIN metal, Grok promised country-wide. Every one of those is a tenancy. M3 is the only one you can hold.

It matters more again after Day 4. Five million people are about to be trained on somebody's platform. If the Arabic model underneath that training is one you can only rent, the skills are portable to exactly one vendor.

Whether it delivers depends on three things HUMAIN did not publish this week: the licence terms, the quality of the weights, and the evaluation data. We will run it through our own Arabic module-generation pipeline the day it becomes downloadable, and publish the result, including what it cannot do.

What I think

The week has one pattern. The Kingdom is buying the supply side of AI at a scale and speed very few markets have matched: power, land, data centres, cloud regions, accelerators, local server and laptop assembly, and now several million training places. It is doing it through a sovereign champion, two national operators, domestic integrators like Elm that already touch tens of millions of citizens, and every hyperscaler and chip vendor with a Saudi region in its plans. The AMD cluster running in production is proof that some of this is turning into reality.

What sits between the compute and the students is thin. One open-weights model, one voice platform, one productivity bundle, a handful of enterprise agents, one device whose defining software is a year out, and eleven venture rounds. Infrastructure has to come first and skills have to come early, so I am not complaining about the order. But most of the eighteen billion is buildings with dates on them, the five million is a promise nobody will audit, and the open questions are what runs in those buildings, in what language, and whether anyone evaluates it in public.

There is one more question this year, which the schedule asked and nobody answered on stage. An event that lost its own date to a regional war spent four days committing gigawatts through 2033 and training places through 2034. Everything on the timeline below assumes the Strait stays open.

Image credits: Riyadh skyline by mo--, public domain. Strait of Hormuz map by Goran tek-en, CC BY-SA 4.0. Company logos via Wikimedia Commons, public domain or Apache 2.0 as marked, reproduced for editorial identification; all remain trademarks of their owners.