Oluwadi
Investigating the Overlooked
Region
What
Who For
Analysis
Oluwadi
Investigating the Overlooked
HUMAIN is Saudi Arabia's national AI company, owned outright by the Public Investment Fund (PIF), the kingdom's sovereign wealth fund — launched in May 2026 as an explicit national champion, the same institutional move the PIF has already made in sports, gaming, and entertainment.[1] Its infrastructure ambitions are large on their own terms: a $3 billion data-center partnership with Blackstone's AirTrunk, a target of 1.9 gigawatts of AI compute capacity by 2030, and a stated goal of becoming the world's third-largest AI infrastructure provider.[2] Saudi Aramco — the state oil company that has funded the kingdom for a century — is separately negotiating a "significant minority stake" in HUMAIN, meaning the oil economy is directly capitalizing what Saudi leadership hopes replaces it.[2]
Humain-m3 is not the first serious Arabic-language model out of the Gulf. In August 2023, Inception (G42's AI arm), Cerebras, and MBZUAI — the Mohamed bin Zayed University of Artificial Intelligence, a real research university in Abu Dhabi, not a corporate lab wearing an academic name — jointly released Jais, a 13-billion-parameter Arabic model trained from scratch on 116 billion Arabic tokens, at the time billed as the world's most advanced Arabic LLM. A second generation, Jais 2, has already followed.[3] MBZUAI didn't stop there: in May 2024, in partnership with Petuum and LLM360, it released K2-65B, a 65-billion-parameter open-source model — this time general-purpose, not Arabic-specific, and with MBZUAI in the lead role rather than one partner among three.[3] The UAE got to Arabic AI first, built it natively with a real university as a genuine technical partner rather than a funder, and has kept producing independent frontier research since. Saudi Arabia's own earlier attempt, ALLaM, came from SDAIA (the Saudi Data and AI Authority, a government body, not a university) and never reached comparable scale or attention.[3] HUMAIN isn't just late to the UAE's research. It's Saudi Arabia's second attempt at its own.
The reason this lineage — Jais, Falcon, ALLaM, Humain-m3 — is worth tracking at all isn't primarily geopolitical. It's linguistic. A model trained by an institution for whom Arabic is the native language, not one slice of an English-centric training mix, has a real chance of understanding register, dialect, and cultural context in ways a model where Arabic is a minority share of the data structurally can't. That's not a hedge or a hunch. The 9/11 Commission found that the FBI "lacked sufficient translators proficient in Arabic," leaving 35% of all Arabic-language national-security wiretaps untranslated — and, in the starkest single case, NSA intercepts of Al-Qaeda's own words the day before the attack, "Tomorrow is zero hour," went untranslated for days after it.[4] The intelligence community's own later assessment of itself: an apparatus "laden with Russian speakers and disastrously short of Arabic speakers" — built for the last war's language, not the live one.[4] Troy's own early work at In-Q-Tel, CIA's strategic investment arm — public record, as is the fact of his employment there — was natural-language processing focused on Arabic specifically, from 2003 to 2007, years before "AI" was the term anyone used for it: a direct, technical answer to the exact capability gap the country had just watched cost it. The lesson from that era wasn't abstract. Language capability that actually understands the language, not a passable approximation of it, is the difference between a warning translated in time and one translated after.
On September 3, 2026, HUMAIN unveiled Humain-m3, its own answer — and its foundation is MiniMax's open M3 base, a Chinese company's architecture, trained further on more than a trillion Arabic-language tokens.[5] The model uses a mixture-of-experts design with roughly 428 billion total parameters and 23 billion active per token — a real, substantial system, not a demo.[5] The contrast with Jais is the actual story: the UAE spent three years building its own Arabic model from the ground up with a research university. Saudi Arabia skipped that step entirely, adapting an existing Chinese company's weights to get a comparable result in a fraction of the time. At the same time, HUMAIN's infrastructure partnerships run through Nvidia, AMD, and Mistral, and AWS is separately building $5.3 billion of data centers inside the kingdom.[6] The same company is deepening ties with Chinese model architecture and American chip and cloud infrastructure simultaneously, in public, without treating the two as in tension.
Most countries building sovereign AI have had to pick a supply chain. Saudi Arabia is testing whether a state with enough capital doesn't. The kingdom holds real, deep US chip access — Nvidia and AMD partnerships, an AWS build-out — the kind of access the US has denied or restricted for other countries entirely. It's spending that access on infrastructure while sourcing its actual model architecture from China, betting that speed to a working Arabic-language model matters more than which country's technology stack it's built on. That's not indecision. It's a wager that HUMAIN's real advantage isn't its allegiance, it's the capital to run both tracks in parallel that a smaller or less-capitalized state couldn't afford to run at all.
The UAE's TII (Technology Innovation Institute) has run the identical open-weights strategy on its own model, Falcon, since 2023 — a government research institute, not a university, but the same underlying bet: publish the weights, let outside adoption do the validation work. Troy ran Falcon locally, on his own machine (a 4090-equipped rig he calls Dagger), and found it held up better than most models available at the time — he hasn't run it recently, so this is a dated, firsthand assessment, not a current claim. His own conclusion from actually deploying it, not just reading about it: a strong base model is roughly half of what makes a system useful. The other half is the harness built around it — the tooling, the scaffolding, the product engineering that turns raw weights into something that actually works for a task. Adapting someone else's open model, the way HUMAIN adapted MiniMax's, gets a state the first half fast. It says nothing about whether the second half — the harness — gets built at all.
Why does this matter? The dominant framing for AI infrastructure right now is a binary — US-aligned or China-aligned, one stack or the other. HUMAIN is a real, current, well-funded case that doesn't fit that binary, run by a state with the sovereign capital to make "both" an actual option rather than a hedge. Whether that position holds — whether the US or China eventually forces a real choice — is exactly why this is worth tracking as it unfolds, not writing up as settled.