Saudi Arabia Is Buying Its Way Into the AI Model Layer

HUMAIN has bought compute, chips and equity. It has built one thing at the model layer, and it did not build the model. That is the honest summary of Saudi Arabia’s fourteen-month push into artificial-intelligence models, and the HUMAIN Cohere partnership announced on 9 July 2026 is the clearest example of the pattern: HUMAIN supplies at least 50 megawatts of dedicated compute to Cohere, and in exchange gets access to a model developer and a promise of joint Arabic-language work. The capacity is expected to be live in the fourth quarter of 2027 [S1].

That is seventeen months from announcement to first electron. Nothing obliges either party to ship before then, and neither has published a contract value, a binding instrument, or an interim milestone.

The Cohere deal is not an outlier. Across ten announced model-layer relationships — Cohere, Adobe, xAI, Groq, Qualcomm, AMD, Cisco, NVIDIA, AWS and stc — the recurring instrument is a press release describing an intention. One is a non-binding term sheet. One is a memorandum of understanding extended rather than executed. Exactly one product is in public hands: HUMAIN Chat, launched 25 August 2025, running ALLaM 34B — a model developed not by HUMAIN but by the Saudi Data and AI Authority, and derived from Meta’s open-weight Llama-2 [S4][S12].

This page is the ledger. For HUMAIN’s company profile and infrastructure footprint, see HUMAIN, Saudi Arabia’s PIF-owned AI company; this piece covers only the model layer — who HUMAIN partnered with to obtain models, what instrument each relationship rests on, and what has verifiably shipped.

The HUMAIN Model-Layer Ledger

Last verified: 31 July 2026.

Two columns carry the argument. The instrument column is dominated by announcements rather than agreements; the shipped column is almost empty.

Is HUMAIN Partnered With Cohere?

Yes — since 9 July 2026, on terms that begin in late 2027. The announcement landed during Canadian Prime Minister Mark Carney’s visit to Riyadh, framed by both sides as Cohere’s first major deployment outside North America [S3].

What the HUMAIN Cohere partnership actually commits

HUMAIN designates at least 50 megawatts of AI-dedicated compute to support Cohere’s next-generation foundation models, expandable over five years in line with demand, live in Q4 2027 [S1][S2]. The parties also said they would jointly develop sovereign Arabic-language models, sector-specific models for financial services, healthcare and energy, and enterprise applications keeping Saudi customer data in-Kingdom.

Tareq Amin, HUMAIN’s chief executive, framed it in supply terms: “Access to computing capacity will be the defining factor in the future of artificial intelligence.” Aidan Gomez, Cohere’s co-founder and chief executive, said the same from the other side: “Developing new generations of AI models requires sustained access to high-performance computing” [S2]. Both quotes describe a compute transaction. Neither describes a model.

What the partnership does not do

No contract value has been published by either party. No instrument type has been named — the releases say “strategic partnership” and “collaborate”, not “definitive agreement” or “binding”. No delivery milestone exists before Q4 2027. And Cohere’s capital structure carries no Saudi money: its August 2025 round of $500m at a $6.8bn valuation was led by Radical Ventures and Inovia Capital, with AMD Ventures, NVIDIA, PSP Investments and Salesforce Ventures participating, and its 2026 Series E was led by Germany’s Schwarz Group. HUMAIN is not a Cohere shareholder. The relationship is a customer-supplier arrangement in the language of alliance — and the supplier is HUMAIN.

What Is ALLaM?

ALLaM is a family of Arabic and English language models built by SDAIA’s National Center for AI (NCAI) — and the flagship variant HUMAIN Chat runs on is documented nowhere.

The peer-reviewed record is solid and worth crediting. ALLaM: Large Language Models for Arabic and English was accepted as a poster at ICLR 2025; all 25 authors, led by M. Saiful Bari, are at NCAI inside the Saudi Data and Artificial Intelligence Authority [S4]. The method is documented: merge an Arabic tokenizer into Llama-2’s vocabulary, initialise the new embeddings from averaged existing ones, and continue pretraining. The 7B and 13B variants saw 1.2 trillion mixed tokens, split roughly 540bn Arabic to 660bn English — and half the Arabic, some 270bn tokens, is machine-translated, a detail absent from every press account. One 7B was trained from scratch, and it is the only variant ever released as open weights.

ALLaM 34B is a different thing. It appeared in August 2025 under HUMAIN, not in the paper, which covers 7B, 13B and 70B. There is no technical report, no open weights and no public API; the only way to use it is HUMAIN Chat, launched 25 August 2025 for web, iOS and Android and marketed at roughly 400 million Arabic speakers [S12]. The Hugging Face organisation ALLaM-AI has been renamed humain-ai, while the model card still credits SDAIA’s NCAI. The brand transferred to HUMAIN; the technical record did not follow it.

Deployment has also contracted rather than expanded. ALLaM reached Microsoft’s Azure AI model catalogue in September 2024 and IBM’s watsonx on 21 May 2024 — but IBM deprecated the watsonx listing on 8 May 2026 and withdrew it on 8 June 2026, leaving it available only to customers who provision it themselves. A model losing a hyperscaler distribution channel is not the trajectory the announcements describe.

Is ALLaM competitive on Arabic benchmarks?

Independent evaluations do exist — and they place ALLaM in the middle of the field, not at the front. This is the section most coverage gets wrong in both directions: the boosters cite only SDAIA’s own tables, and the sceptics claim no third-party evidence exists. Three arm’s-length results are public.

Open Arabic LLM Leaderboard v2 — the community leaderboard run with Hugging Face. ALLaM-7B-Instruct-preview was evaluated on 19 February 2025 and scores 60.49% overall (AraTrust 86.93, AlGhafa 69.49, ArabicMMLU 64.90, MadinahQA 54.24, Arabic EXAMS 51.58) [S21]. On the same pipeline, Google’s gemma-3-27b-it reaches 68.70% and Qatar’s Fanar-1-9B-Instruct 64.57%. So does Yehia-7B-preview at 60.92% — a community fine-tune of ALLaM itself, built by Navid-AI. The national model is narrowly beaten by a hobbyist derivative of the national model.

IslamicMMLU — 10,013 questions across Quran, Hadith and Fiqh, 26 models, published by researchers at the University of Edinburgh in March 2026 [S22]. This is the sharpest test available, because Islamic religious knowledge is precisely the domain in which a Saudi Arabic-first sovereign model should have an unassailable edge.

ALLaM ties a two-generation-old GPT-4, sits 22 points behind Qatar’s Fanar-Sadiq and 34 points behind Gemini 3 Flash, and scores 43.2% on the Quran track — where Fanar-Sadiq scores 94.4%. A third study, ALPS (February 2026), reaches the compatible conclusion that commercial models beat every Arabic-native system, with the best of them, Jais-2-70B, at 83.6% against a human baseline of 84.6% [S23].

And then the finding nobody has reported: Saudi Arabia’s own national Arabic benchmark does not evaluate ALLaM. BALSAM, published 30 July 2025 and hosted by the King Salman Global Academy for Arabic Language, spans 78 tasks and 52,000 examples and ranks 22 models. ALLaM is not among them. Yehia-7B preview — the fine-tune of ALLaM — is. And four BALSAM co-authors (Norah Alzahrani, Ahmed Abdelali, Nora Altwairesh and Abdulmohsen Al-Thubaity) are also ALLaM co-authors [S24]. The same researchers built both, and left the national model out of the national benchmark.

Two caveats cut the other way. Every independent result above measures ALLaM 7B, an early, small, open-weights model — and no independent comparative evaluation of ALLaM 34B exists anywhere. The only dedicated 34B study is a single-author, UI-level assessment of 23 prompts judged by GPT-5, Gemini 2.5 Pro and Claude Sonnet 4, with no baseline models at all [S5]. The most-repeated endorsement — that Cohere ranked ALLaM 34B the most advanced Arabic LLM built in the Arab world on MMLU — appears in no Cohere publication and carries a qualifier (“built in the Arab world”) that excludes every frontier model.

Has Saudi Arabia Built Its Own AI Model?

Yes — but not a frontier model, and not at HUMAIN. The distinction is the whole argument of this page.

Saudi Arabia has built a capable, deployed, Arabic-first model derived from an American open-weight base, inside a government authority, and put it in front of consumers through a PIF-owned company. That is a genuine national capability. It is also downstream adaptation: the recipe, the architecture and the base weights came from Meta.

Amin has been candid about the gap. Describing HUMAIN’s first year, he said he wanted the science team “to learn how difficult it is to build frontier models” — a formulation conceding the destination has not been reached. HUMAIN positions itself across the full value chain, but the frontier-model rung is aspiration, not inventory.

Against the comparators, Saudi Arabia is mid-pack in a crowded regional field: the UAE has Falcon (TII) and Jais (G42/Inception/MBZUAI), Qatar has Fanar (QCRI), and all three are — like ALLaM — adaptations or mid-scale trains rather than frontier runs. Nobody in the Gulf has trained at the compute scale of GPT-5-class systems, and the constraint is not ambition.

The Hardware Partners Are Model-Layer Partners Too

The chip announcements are usually filed as infrastructure. Several are actually about model access, which is why they belong here.

xAI is the largest. HUMAIN took $3bn in xAI’s Series E, announced 19 February 2026, on top of a previously announced 500MW infrastructure partnership, with the arrangement contemplating Grok deployment across Saudi Arabia [S13]. This is the one relationship where HUMAIN holds equity in a frontier lab, and the one exposed to a moving mark: xAI merged with SpaceX in a deal announced 3 February 2026 valuing the combination at about $1.25tn, and the merged entity listed on 3 June 2026 at $135 a share, roughly $1.77tn. The stock traded near $116 on 25 July 2026, about 13% below issue.

Groq is the one that quietly works. HUMAIN hosts OpenAI’s open-weight models on Groq’s inference hardware inside the Kingdom, sold as sovereign, low-latency inference. It is a live endpoint rather than a memorandum — and also the purest expression of the renting thesis, because the models are OpenAI’s and the silicon is Groq’s.

Adobe signed the most model-shaped announcement. On 19 November 2025 at the US–Saudi Investment Forum, the two said they would build Arabic-first multimodal generative AI spanning audio, image, video, 3D and digital twins, pairing ALLaM and HUMAIN Cloud with Adobe Firefly Foundry, and that Adobe intends to become HUMAIN’s first global AI data center customer [S14]. Eight months on, no joint model, no product and no capacity figure has been announced.

Qualcomm committed to 200MW of AI200 and AI250 rack-scale inference systems starting in 2026, with ALLaM integration and a Qualcomm AI engineering centre at HUMAIN in Riyadh [S10]. AMD and Cisco announced on 19 November 2025 that they plan to form a joint venture for up to 1GW by 2030, phase one 100MW on AMD Instinct MI450s [S7]; Reuters reported the first customer would be generative-video startup Luma AI [S16]. As of 31 July 2026 the entity has not been publicly named or confirmed as incorporated.

NVIDIA anchors the hardware story and illustrates its ceiling. The US Commerce Department authorised exports to HUMAIN and the UAE’s G42 on 19 November 2025, each up to the equivalent of 35,000 Blackwell GB300s, worth roughly $1bn in total, under security and reporting requirements monitored by the Bureau of Industry and Security [S6]. Set that against the 600,000-GPU ambition attached to HUMAIN in 2025 and the arithmetic is unforgiving. NVIDIA does not break out Saudi or HUMAIN revenue in its quarterly reporting. For the licence mechanics, see NVIDIA GPUs, Saudi AI and export controls.

AWS is the row that fails outright. The $5bn-plus “AI Zone” was announced in May 2025 and an in-Kingdom AWS region has been promised since 2024. AWS’s own regional reference documentation, checked on 31 July 2026, lists exactly two Middle East regions — me-south-1 (Bahrain) and me-central-1 (UAE) — with Saudi Arabia appearing only among announced future regions [S11]. Secondary reports of a January 2026 general availability are not supported by AWS’s published table. For what the two did build, see HUMAIN ONE and AWS.

Chinese model relationships could not be stood up. No verifiable HUMAIN agreement with Alibaba’s Qwen, DeepSeek, Zhipu or Moonshot exists in the public record as of 31 July 2026, and no Huawei Ascend deployment at HUMAIN has been documented. The exposure runs the other way: on 7 July 2026 Reuters reported that China’s commerce ministry and NDRC had held closed-door meetings with Alibaba, ByteDance and Z.ai about restricting overseas access to advanced Chinese models. A Gulf operator hedging American models with Chinese ones would now face export controls from both directions. Context: Saudi–China relations.

The stc Joint Venture Was Extended, Not Signed

HUMAIN and stc signed a memorandum of understanding on 18 December 2025 to establish a data center joint venture through stc’s Center3 subsidiary — HUMAIN 51%, stc 49% — targeting up to 1GW of AI-focused capacity, with an initial phase of up to 250MW subject to customer commitments.

On 18 June 2026, stc disclosed to Tadawul that the MoU had been extended by six months to complete regulatory, operational and commercial requirements, and stated there was no material financial impact [S8]. That is a single six-month extension of the original MoU — not, as has been reported elsewhere, a second extension. Six months after signature, the joint venture was not incorporated.

The correct reading is not that the deal is dead. It is that a listed Saudi telecom operator with a disclosure obligation declined to convert an AI data center MoU into a company on the original timetable, and told the exchange the extension carried no material financial impact. A listed counterparty’s own characterisation of materiality is the most reliable evidence available about a private company’s pipeline.

Who Owns HUMAIN and How Much Money Has Actually Arrived

HUMAIN was established on 12 May 2025, is majority-owned by the Public Investment Fund, and is chaired by Crown Prince Mohammed bin Salman personally. Tareq Amin, previously chief executive of Aramco Digital, has led it since launch.

In October 2025, PIF and Saudi Aramco announced a non-binding term sheet for Aramco to acquire a significant minority stake, with both shareholders contributing AI assets, capabilities and talent [S9]. Nine months later, no stake percentage, no valuation, no completion date and no definitive agreement have been published. The transaction remains subject to definitive documentation and regulatory approvals.

On disbursement, the record is thin in a specific way. HUMAIN has never published revenue, paid-in capital or a balance sheet. The announced figures in circulation — $10bn for AMD, $23bn of aggregate technology agreements, $77bn, $100bn — are commitment or ambition numbers, not cash. The verifiable outflow is the $3bn xAI investment [S13]. A $10bn venture fund, Humain Ventures, was announced on 28 May 2025 [S17]; no portfolio has been published. Reuters reported on 19 May 2026 that HUMAIN had picked Goldman Sachs to advise on roughly SAR20bn ($5.3bn) of data center debt financing [S15]; no closing has been announced — see HUMAIN’s Goldman Sachs mandate. Money in, as opposed to announced, comes from two named sources: PIF, in undisclosed amounts, and Blackstone-backed AirTrunk, in an approximately $3bn data center campus partnership announced 28 October 2025 [S19].

Where Saudi Arabia Actually Sits on the AI Stack

Strip the announcements away and the position is legible, layer by layer.

The honest verdict is neither the boosterish one nor the dismissive one. Saudi Arabia is plausibly competitive at the energy and data layers, where its advantages are structural and hard to replicate. It is a buyer at the compute layer, where a foreign regulator sets the ceiling. It is a renter at the model layer, where every frontier system it deploys belongs to somebody else. And it is an early but genuine builder at the application layer — the layer with the least defensible moat. That configuration describes most countries. What is unusual is the capital being deployed to change it, and the confidence with which the model layer is described as owned rather than leased.

Can Saudi Arabia Compete in AI?

At some layers, yes. At the model layer, not on the current evidence — and three separate constraints explain why.

Power is binding before chips are

Saudi Arabia operated 467MW of data center capacity in Q1 2026 on MCIT’s own figures, against HUMAIN’s pipeline of 6.6GW by 2034 — a fourteen-fold gap that no export licence closes. Grid interconnection, substations and gas-turbine lead times set the timetable, not accelerators. This is the site’s pillar finding and it is set out in full in Saudi Arabia’s AI data center power constraint. It is also why the Cohere compute lands in Q4 2027 rather than Q4 2026, and why HUMAIN’s Riyadh and Dammam data centers had not been confirmed operational as of 31 July 2026 despite a Q2 2026 commitment.

Chip access is a foreign policy variable

The November 2025 authorisation is capped, conditioned and revocable, granted to a named company under BIS monitoring, and it treats HUMAIN and G42 as parallel cases while the underlying country-tier framework keeps moving. A model-layer strategy whose central input is licensed by another government has a foreign veto inside it.

The law binds less than people assume — except where it binds a lot

There is no Saudi AI statute. What constrains a deployment is the Personal Data Protection Law, SDAIA guidance and the National Cybersecurity Authority’s controls — see how Saudi Arabia regulates AI and the Personal Data Protection Law. SDAIA’s AI ethics principles are non-binding; the PDPL is enforceable. That combination is permissive for model training and restrictive for personal-data transfer — precisely the shape that favours training in-Kingdom on in-Kingdom data.

Article 26(4) Is the Advantage Nobody Has Priced

Here is the structural asset that no competing analysis has connected to the model-layer strategy.

From 12 August 2026, Saudi Arabia’s new Copyright Law contains an express, standalone artificial-intelligence training exception. Article 26(4) permits copying a work “for the purposes of developing artificial intelligence products and algorithms” — under a chapeau that disposes of both permission and payment: without the author’s permission and without compensation. Three conditions attach: the work must have been lawfully published, the original copy lawfully acquired, and the copying limited to what the purpose requires. There is no rights-holder opt-out and no attribution duty [S18]. The full statutory analysis, including the contested 1 August versus 12 August commencement date, is in Saudi Arabia’s Copyright Law and Article 26.

Compare the regimes a model developer would otherwise train under. The EU’s text-and-data-mining exception under Article 4 of the DSM Directive collapses the moment a rights holder inserts a machine-readable reservation — exactly the infrastructure European publishers spent 2024 and 2025 building. The UK abandoned its proposed broad TDM exception after a publishing backlash, leaving commercial training in limbo. The United States is resolving the question case by case in court. Japan and Singapore are permissive, but neither offers subsidised gigawatt-scale power alongside.

Saudi Arabia is now the jurisdiction that offers cheap firm power, a licensed chip supply, a large curated Arabic corpus, and statutory certainty that training on lawfully obtained copyrighted material is permitted without payment. No other country assembles all four. For a developer whose training-data exposure is a line item in its risk register, that combination is worth real money — and it is worth more to a foreign lab with a licensing problem than to a Saudi one.

The catch is who captures the value. An exception that makes in-Kingdom training cheap is only an advantage if somebody trains in-Kingdom. On the current ledger, the entity most likely to exploit Article 26(4) at scale is not HUMAIN — which has no frontier pretraining programme — but a foreign lab renting HUMAIN’s megawatts. Cohere is contracted to do exactly that from Q4 2027, for the stated purpose of developing next-generation foundation models. The legal advantage is Saudi. The models may not be. And the advantage has a shelf life: Article 37(2) of the same law adds an open-ended non-commercial exception, neither provision has been tested against the Berne three-step test, and rights-holder groups have not mobilised largely because the law has not yet commenced. For the wider framework, see intellectual property in Saudi Arabia.

Why This Matters for Vision 2030

The AI programme is the largest non-oil bet in the current Vision 2030 portfolio, and its public case rests on Saudi Arabia becoming a producer rather than a consumer of artificial intelligence. HUMAIN’s stated target — processing 7% of global AI workloads by 2030 — is a compute-hosting metric, not a model metric. That is the metric the organisation is optimising for, and it is achievable without owning a single frontier model.

If the compute-hosting thesis works, Saudi Arabia becomes a well-paid landlord in an industry whose margins accrue mostly to model owners and application layers. That is a real business — roughly the business Ireland built in cloud — but it is not technological sovereignty, and both the national AI strategy and the digital sovereignty framing promise more than landlording. The gap between the rhetoric and the leasing reality is measurable: count the models trained in-Kingdom, not the megawatts announced.

Risks, Contradictions and Open Questions

The instrument problem is systemic, not incidental. Ten relationships, and the strongest instruments in the set are an equity cheque (xAI), a non-binding term sheet (Aramco) and an extended MoU (stc). A portfolio whose binding documents are the ones nobody has signed has an execution question at its centre.

Partner silence is under-reported. NVIDIA does not break out Saudi revenue. Cohere’s funding communications carry no Saudi participation. stc told its own exchange the JV extension had no material financial impact. Adobe has announced nothing further in eight months. None of that proves failure; all of it is inconsistent with the scale the announcements imply, and it deserves the most weight precisely because it comes from the counterparties rather than from Riyadh.

The AWS region conflict is unresolved in the secondary literature. Multiple outlets report a January 2026 general availability for a Saudi AWS region. AWS’s own published region table on 31 July 2026 does not list one [S11]. We follow the primary source, and will correct this page if AWS publishes a launch.

ALLaM’s competitive standing is genuinely unknown. Absent an independent head-to-head, nobody outside SDAIA can say where ALLaM 34B sits against Falcon-H1 Arabic, Jais, Fanar or the Arabic performance of GPT-5-class systems. Until somebody runs it, treat every superlative as vendor copy.

HUMAIN’s finances are opaque by construction. A private PIF company with no disclosure obligation, publishing no revenue or capital figures, is not auditable from outside. Every disbursement number here is inferred from counterparty announcements and could be materially wrong in either direction.

The Article 26(4) advantage is untested. It commences on 12 August 2026, the implementing regulations were unpublished as of 31 July 2026, and no model developer has yet relied on it publicly.

What to Watch Next

• 12 August 2026 — the Copyright Law commences. Watch whether SAIP’s implementing regulations narrow Article 26(4), and whether any developer publicly cites it as a siting reason.
• Mid-December 2026 — the stc/HUMAIN MoU’s extended term expires. Incorporation, a second extension, or lapse. All three are informative.
• Q3 2026 reporting season — whether NVIDIA, AMD, Qualcomm, Cisco or Adobe name HUMAIN in filings or on calls. Continued silence from five listed counterparties is itself a finding.
• Any HUMAIN Riyadh or Dammam energisation announcement — the Q2 2026 commitment lapsed on 30 June unconfirmed.
• The AMD/Cisco JV’s legal name — a registered entity converts a plan into a company.
• Q4 2027 — the Cohere compute delivery date, and the first hard test of whether HUMAIN can hand a foreign model developer working megawatts on schedule.
• Any published independent Arabic benchmark including ALLaM 34B — the single piece of evidence that would settle the model-quality question.

Related Vision 2030 Context

• HUMAIN, Saudi Arabia’s PIF-owned AI company — the full company profile and capitalisation history
• Saudi Arabia’s AI data center power constraint — why the grid, not chips, sets the timetable
• HUMAIN’s Riyadh and Dammam data centers: status — the dated verification tracker
• How Saudi Arabia regulates AI: SDAIA’s framework — what actually binds an AI deployment
• Saudi Copyright Law 2026 and Article 26 — the AI-training exception in full
• NVIDIA GPUs, Saudi AI and export controls — the licensing regime
• Saudi Arabia’s data center sector — the capacity base
• Cloud computing in Saudi Arabia — hyperscaler regions and data residency
• SDAIA: the Saudi Data and AI Authority — ALLaM’s owner
• The Public Investment Fund — HUMAIN’s majority shareholder
• HUMAIN’s Goldman Sachs data center financing mandate — the debt raise
• HUMAIN ONE and AWS — the operating-layer ambition
• Saudi–China relations — the alternative-supplier question
• Digital sovereignty — the doctrine behind sovereign AI

Sources

• [S1] Reuters, Saudi Arabia’s Humain, Canada’s Cohere to Work Together on AI Compute, Sovereign AI Models, news report, 9 July 2026. https://finance.yahoo.com/technology/ai/articles/saudi-arabias-humain-canadas-cohere-131646877.html
• [S2] Asharq Al-Awsat, Humain, Cohere Launch Strategic Partnership to Expand AI Infrastructure in Saudi Arabia, news report, 10 July 2026. https://english.aawsat.com/business/5294414-humain-cohere-launch-strategic-partnership-expand-ai-infrastructure-saudi-arabia
• [S3] Semafor, Saudi AI Champion HUMAIN to Supply Compute to Canada Startup Cohere, news analysis, 14 July 2026. https://www.semafor.com/article/07/14/2026/saudi-ai-champion-humain-to-supply-compute-to-canada-startup-cohere
• [S4] M. S. Bari et al. (SDAIA), ALLaM: Large Language Models for Arabic and English, technical paper, arXiv:2407.15390. https://arxiv.org/abs/2407.15390
• [S5] O. Nacar, UI-Level Evaluation of ALLaM 34B: Measuring an Arabic-Centric LLM via HUMAIN Chat, preprint, 24 August 2025. https://arxiv.org/abs/2508.17378
• [S6] US Department of Commerce, Statement on UAE and Saudi Chip Exports, press release, 19 November 2025. https://www.commerce.gov/news/press-releases/2025/11/statement-uae-and-saudi-chip-exports
• [S7] AMD Newsroom, AMD, Cisco and HUMAIN to Form Joint Venture to Deliver World-Leading AI Infrastructure, press release, 19 November 2025. https://newsroom.amd.com/news/amd-cisco-and-humain-to-form-joint-venture/
• [S8] Argaam, stc Extends HUMAIN AI Data Center MoU by 6 Months, Tadawul disclosure report, 18 June 2026. https://www.argaam.com/en/article/articledetail/id/1913648
• [S9] Public Investment Fund, PIF and Aramco Agree for Aramco to Acquire a Significant Minority Stake in HUMAIN, With PIF Retaining Majority Ownership, press release, October 2025. https://www.pif.gov.sa/en/news-and-insights/press-releases/2025/pif-and-aramco-agree-for-aramco-to-acquire-a-significant-minority-stake-in-humain-with-pif-retaining-majority-ownership/
• [S10] Qualcomm, Qualcomm to Establish AI Engineering Center at HUMAIN, press release, 19 November 2025. https://www.qualcomm.com/news/releases/2025/11/qualcomm-to-establish-ai-engineering-center-at-humain
• [S11] Amazon Web Services, AWS Regions, global infrastructure reference documentation, retrieved 31 July 2026. https://docs.aws.amazon.com/global-infrastructure/latest/regions/aws-regions.html
• [S12] Bloomberg, Saudi’s Humain Launches Arabic Chatbot With ‘Islamic’ Values, news report, 25 August 2025. https://www.bloomberg.com/news/articles/2025-08-25/saudi-s-humain-launches-arabic-chatbot-with-islamic-values
• [S13] CNBC, Saudi’s Humain Invested $3 Billion in xAI’s Series E Funding Round, news report, 19 February 2026. https://www.cnbc.com/2026/02/19/saudis-humain-invested-3-billion-in-xais-series-e-funding-round.html
• [S14] Adobe Newsroom, HUMAIN and Adobe Announce Global Strategic Partnership to Build AI Models and AI-Powered Applications Tuned for the Arab World, press release, 19 November 2025. https://news.adobe.com/news/2025/11/humain-adobe-announce-global-strategic-partnership
• [S15] Reuters, Saudi Arabia’s Humain Picks Goldman Sachs to Advise on Data Centre Financing, Sources Say, news report, 19 May 2026. https://money.usnews.com/investing/news/articles/2026-05-19/saudi-arabias-humain-picks-goldman-sachs-to-advise-on-data-centre-financing-sources-say
• [S16] Reuters, AMD, Cisco and Saudi’s Humain Launch AI Joint Venture, Land First Major Customer, exclusive report, 19 November 2025. https://kfgo.com/2025/11/19/exclusive-amd-cisco-and-saudis-humain-launch-ai-joint-venture-land-first-major-customer/
• [S17] TechCrunch, Humain Planning $10B VC Fund to Invest in US, European, and Asian Startups, news report, 28 May 2025. https://techcrunch.com/2025/05/28/humain-planning-10b-vc-fund-to-invest-in-us-european-and-asian-startups/
• [S18] Copyright Law of Saudi Arabia, Royal Decree No. M/169, Article 26 and Article 61, gazetted in Umm al-Qura, 13 February 2026; in force 12 August 2026.
• [S19] Blackstone, HUMAIN and Blackstone-Backed AirTrunk Announce Partnership to Build State-of-the-Art Data Centers in the Kingdom of Saudi Arabia, press release, 28 October 2025. https://www.blackstone.com/news/press/humain-and-blackstone-backed-airtrunk-announce-partnership-to-build-state-of-the-art-data-centers-in-the-kingdom-of-saudi-arabia/
• [S20] Amazon Web Services / Hugging Face, Open Arabic LLM Leaderboard v2, community leaderboard maintained with Hugging Face, 2A2I and TII. https://huggingface.co/spaces/OALL/Open-Arabic-LLM-Leaderboard
• [S21] Open Arabic LLM Leaderboard v2, detailed evaluation results for ALLaM-7B-Instruct-preview, run 19 February 2025. https://huggingface.co/datasets/OALL/details_ALLaM-AI__ALLaM-7B-Instruct-preview_v2
• [S22] A. Abdelaal, M. N. Al Haffar, M. Fawzi and W. Magdy (University of Edinburgh), IslamicMMLU: A Benchmark for Evaluating LLMs on Islamic Knowledge, preprint, 24 March 2026 (rev. 11 July 2026). https://arxiv.org/abs/2603.23750
• [S23] H. S. Al-Olimat and A. Alshareef, ALPS: A Diagnostic Challenge Set for Arabic Linguistic and Pragmatic Reasoning, preprint, 19 February 2026. https://arxiv.org/abs/2602.17054
• [S24] R. Al-Matham et al., BALSAM: A Platform for Benchmarking Arabic Large Language Models, preprint, 30 July 2025; platform at benchmarks.ksaa.gov.sa. https://arxiv.org/abs/2507.22603