HUMAIN Is Building the Distribution Layer, Not the Lab

Between August 31st and September 3rd, 2026, at the LEAP conference in Riyadh, HUMAIN made three key announcements. One, that its NVIDIA cloud had gone live^[1]. Two, that its AMD cluster was serving customers^[2]. And three, that it had a frontier Arabic language model^[3]. These were promoted across Twelve press releases in four days. It was, by volume, the most productive week in the company's sixteen-month life.

The releases themselves rather than their coverage convey a different type of week. The cloud that went live is 1.1 megawatts. The AMD cluster's capacity is not disclosed. And the frontier Arabic model was, in HUMAIN's own headline, "Developed by MiniMax" (a Shanghai lab) with the body text adding that it was "commissioned by HUMAIN and delivered by MiniMax," built on MiniMax's own model, and destined for release under MiniMax's own license.

Saudi Arabia's frontier-AI strategy centers on compute, routing, enterprise software and national alignment, while its model intelligence comes from external vendors. Of the six Western labs HUMAIN has announced model partnerships with, none has shipped anything. The one vendor that has actually delivered a model is Chinese.

Sovereignty, on this evidence, is being redefined from we make it to we route it. Whether that is a retreat or a strategy is a genuinely open question.

Nearly every fact here comes from HUMAIN's own publications. Though not because HUMAIN is unusually candid; the real value of its disclosures is actually that they conflict with one another. The 1.1 megawatt figure sits four paragraphs from a plan for five hundred. The benchmark losses are printed on its own product page. It has described its chat model's authorship three different ways on three of its own surfaces. Taken together, HUMAIN's publications often establish less than the company's broader presentation suggests.

The Company That Is Not Called HUMAIN

HUMAIN was announced on May 12, 2025, by Crown Prince Mohammed bin Salman, as a company wholly owned by the Public Investment Fund. The launch release is unusually direct about the governance: HUMAIN is "Chaired by HRH the Crown Prince."^[4] He also chairs PIF, which owns it, and the board of SDAIA, the state authority that regulates data and AI.

The legal entity is not called HUMAIN. Its own Terms of Use name it Al-Mustaqbal Lil-Thaka Al-Istinai Company, also rendered Future Artificial Intelligence Company, a joint stock company registered under commercial registration number 1009089438^[5], which the US Patent and Trademark Office independently corroborates on HUMAIN's American trademark application.^[6] That application, filed in September 2025, was refused: a final office action issued on August 19, 2026. The reason is not in the record.

The company also predates its own announcement. Saudi trademark priority filings for the HUMAIN mark are dated March 25, 2025 - seven weeks before the launch - and HUMAIN's own year-in-review says it existed as a working group from October 2024 with "No office. No employees." The actual incorporation date has never been published.

Then there is the leadership page.^[7] It renders at least eleven directors plus the Chairman: three sitting cabinet ministers, two Royal Court advisers, four PIF-side seats, the president of SDAIA, and the chief executive of Saudi Aramco. There are no independent directors among them. Against those at least twelve names, the page renders exactly one executive: Tareq Amin, the chief executive.

No chief financial officer, chief operating officer, chief technology officer or general counsel has ever been named publicly. Across at least sixty-six items in its newsroom archive, HUMAIN has never announced a single personnel change - no hire, no departure, no promotion. The absence of named executives and personnel announcements leaves an unusually thin public record for a company now preparing to assemble an IPO team.

One Point One Megawatts

On August 31, 2026, HUMAIN announced the availability of its AI cloud, "powered by NVIDIA HGX B300 accelerated computing systems and already supporting government and enterprise AI workloads across the Kingdom." Four paragraphs in, it discloses the size.

"Today, HUMAIN's AI cloud provides 1.1 MW of operational capacity expanding to 35.1 MW by end of 2027 and is supporting private and public sector with agentic AI and GPU-accelerated production workloads with more than 90% utilization of current cluster."^[1]

The same release itemizes it further: 57 NVIDIA HGX B300 systems deployed, 1.1 MW of operational capacity, six government and enterprise customers, sixteen models in production. NVIDIA's own specification for the HGX B300 gives the form factor as "8x NVIDIA Blackwell Ultra SXM,"^[8] which makes 57 systems approximately 456 GPUs.

And two paragraphs above the disclosure, the same document restates the ambition: "HUMAIN and NVIDIA previously announced plans to build AI factories with a projected capacity of up to 500 megawatts, powered by several hundred thousand NVIDIA GPUs over five years."

The comparison is HUMAIN's own. Both figures are in the same release, four paragraphs apart, and the company printed them together.

First, 1.1 MW describes the NVIDIA cloud specifically, not HUMAIN in total. The AMD Instinct MI355X deployment announced the same day is described as "now live... and serving HUMAIN customers" with no megawatts, no GPU count and no site named. The missing AMD figures leave HUMAIN's total live capacity unknown. AMD's own release, which carries the same wire text, adds nothing: the same two forward numbers - 250 MW from 2027, 1 GW by 2030 - and none for what is live.^[2] Second, the 500 MW is a five-year ceiling governed by "up to," and mocking month-sixteen actuals against a year-five ceiling is a cheap trick. Third, and cutting the other way: the cluster is running at more than 90% utilization. So it's small, and it's already full.

The number that actually matters is the one HUMAIN committed to in the same sentence: 35.1 megawatts by the end of 2027. That is a thirty-two-fold scale-up on a stated deadline, and it's the figure against which the company can be fairly judged. Everything else is a ceiling with a "hundreds of thousands" attached.

For scale: a single NVIDIA GB300 NVL72 rack draws on the order of 130 kilowatts. HUMAIN's entire live AI cloud is about eight racks' worth of power. Amazon's capital expenditure for calendar 2025 was $131.8 billion; Microsoft's for fiscal 2026 was $115.9 billion; Alphabet's $91.4 billion; Meta's $69.7 billion.^[9] Against that, the AMD collaboration's "up to $10 billion" is a joint figure across two parties over five years.

Announcement as a Product

HUMAIN's newsroom holds at least sixty-six unique items between May 12th, 2025 and September 3rd, 2026. Roughly a third of them were published on three days: the founding week in May 2025, a single day in November 2025 that produced eight releases all tagged "at the U.S.-Saudi Investment Forum," and the four days of LEAP 2026.^[10]

The legal force of the verbs matters more than the number of announcements. Read the instrument each release actually names:

Qualcomm is the only counterparty in the entire archive that named its instrument precisely. Everyone else chose a word - "collaboration," "partnership," "initiative" - that carries no legal content at all.

There's a specific hazard buried in this table, and this publication walked straight into it. Five separate HUMAIN deals carry a "250 MW" figure: the Infra framework, the center3 joint venture, the AMD/Cisco second phase, the unsigned Al Moammar letter of award, and Together AI. Three of the five place a dollar figure in the same release without asserting any relationship between the two numbers.

On September 4th, 2026, our own newsdesk reported a "$2.5 billion fund for 250MW." That was two facts welded together: the $2.5 billion attaches to the Al Moammar letter of award, and the 250 MW we paired it with came from the older, separate Infra framework. The item was refused before publication, but the lesson took a second pass to state properly. Numbers should be paired only when a source explicitly states their relationship, ideally in the same sentence.

Whether the AMD/Cisco, Infra, Al Moammar and Together AI megawatts are the same physical megawatts is not established by any primary document. The disclosures do not establish whether these projects overlap, and HUMAIN has never published a total that resolves the ambiguity.

The Money That Is Not There

HUMAIN has never disclosed revenue, profit, EBITDA, backlog or valuation. Unremarkable for a private company. How far the silence extends, though, is remarkable.

The founding release announces a PIF-owned company and contains no capital figure at all - no commitment, no funding round, no balance sheet. PIF's own annual report, published August 2026, gives HUMAIN narrative treatment across five of eighty pages and does not break it out in the financial statements.

In October 2025, PIF and Aramco announced the signing of "a non-binding term sheet" under which Aramco would take a "significant minority" stake. No percentage was ever disclosed by anyone. Aramco's own full-year 2025 results release, published the following March, mentions HUMAIN at least twice: once in a forward-looking bullet on plans to acquire a significant minority interest in HUMAIN, and once in the glossary, which defines it as a "PIF portfolio company specialized in artificial intelligence technologies."^[11] Aramco's filing still treated the transaction as pending. Its half-year 2026 results and its 2026 interim report mention HUMAIN not once. And PIF's own annual report, published August 2026, still describes the arrangement as a signed term sheet. The honest formulation: as of the most recent annual report, ten months on, the transaction remained a term sheet.

Then the sharpest negative in the file. No US-listed counterparty has ever disclosed a HUMAIN contract value, equity stake or commitment in a periodic report. Not NVIDIA, not AMD, not Cisco, not Qualcomm. A full-text search of SEC filings returns no 10-K or 10-Q disclosure. NVIDIA's most recent quarterly names HUMAIN exactly once, and it is under robotaxi collaborations, alongside Foxconn and Uber. AMD's most recent names it once, in a list of Helios rack customers. The "up to $10 billion" is booked as a commitment in no audited filing anywhere.^[12]

The direction of travel that is documented runs the other way. HUMAIN put $3 billion into xAI's Series E in February 2026^[13] and led Luma AI's $900 million Series C in November 2025^[14] - after which Luma announced a 2 GW supercluster with HUMAIN, making the company HUMAIN funded into the customer of the capacity HUMAIN is building. The arrangement resembles vendor financing if Luma is a paying customer of the capacity being developed with HUMAIN.

PIF's annual report contains the single most quotable sentence anyone has written about HUMAIN, and it's the owner's own assessment. HUMAIN, it says, "evolved from primarily reselling third-party technologies to becoming an intellectual property creator."

Who Really Built ALLaM?

HUMAIN's consumer product, HUMAIN Chat, launched in August 2025 on "ALLAM 34B, HUMAIN's flagship Arabic large language model."^[15] The model family is presented throughout HUMAIN's material as its own: "the ALLAM family of Arabic models," "ALLAM, our Arabic-first large language model."

The only ALLaM anyone can download is a 7-billion-parameter preview, and the model card - hosted in HUMAIN's own Hugging Face namespace - says this:

"ALLaM is a series of powerful language models designed to advance Arabic Language Technology (ALT) developed by the National Center for Artificial Intelligence (NCAI) at the Saudi Data and AI Authority (SDAIA)."^[16]

Its structured metadata repeats it - "Developed by: National Center for Artificial Intelligence at SDAIA" - and its acknowledgements add, "This release is funded by Saudi Data & AI Authority." The card text still refers to the model by its old path, ALLaM-AI/ALLaM-7B-Instruct-preview. That organization no longer exists: huggingface.co/ALLaM-AI returns 404, and the repository now lives at huggingface.co/humain-ai, which holds exactly one model.

The academic record agrees. The ALLaM paper was submitted in July 2024 - ten months before HUMAIN was announced - with twenty-five authors credited to NCAI.^[17] The Hugging Face repository was created on February 13, 2025, three months before HUMAIN existed, and has not been updated since July 14, 2025.

What happened, on the documentary record, is a change of custody with no published change of ownership. And the mechanism appears to have been a person rather than a contract. The ALLaM model card thanks "Mishari Almishari and Yaser Alonaizan for leading the ALLaM project." Dr. Yaser Al-Onaizan is today HUMAIN's Deputy CEO and President of AI Products.^[18] Both of those facts sit on HUMAIN's own web properties.

The continuity of personnel makes this an attribution problem: the ALLaM lead moved into HUMAIN, while the public record of model ownership did not move with him. HUMAIN's own launch release makes the tension explicit within a single paragraph. First:

"HUMAIN's ALLAM 34B builds on the previous released models developed by the National Center for AI under The Saudi Data & AI Authority (SDAIA) national mandate."

Then, two sentences later:

"Over 120 AI specialists, including 35 PhD holders, with a 50/50 gender ratio, have built this model from the ground up - hosted in Saudi, by Saudis, with global talent alongside them."^[15]

These can be reconciled - a fresh pretrain following an inherited recipe would satisfy both. The point is that nobody outside HUMAIN can tell, because ALLaM 34B has no published weights, no model card, no technical report and no benchmark table of its own. It has been a live consumer product for over a year and has never been documented by its owner. Because HUMAIN has published no weights or technical documentation for ALLaM 34B, the public record cannot establish which account of its development is accurate.

Commissioned and Delivered

Which brings us to September 3rd, 2026, and a headline that gives the game away without apology:

"HUMAIN Unveils humain-m3, a Frontier Arabic Language Model Developed by MiniMax, in Research Preview on HUMAIN Node"^[3]

The body is equally direct. humain-m3 is "a frontier Arabic-language model commissioned by HUMAIN and delivered by MiniMax." It is "built on the MiniMax-M3 lineage" - a 428-billion-parameter mixture-of-experts model, 23 billion active per token - "further pre-trained on more than one trillion tokens of Arabic-native content."

MiniMax is a Shanghai AI lab, listed in Hong Kong. Its own MiniMax-M3 model card gives the same architecture, unchanged: approximately 428 billion total parameters, approximately 23 billion activated, one-million-token context, natively multimodal.^[19] HUMAIN's contribution is the Arabic continued pre-training layer. Nothing else is claimed.

HUMAIN "expects the model weights to be released under the MiniMax Community License following completion of safety training and alignment, currently targeted for next month." Saudi Arabia's sovereign frontier model is expected to ship under a Chinese lab's license.

As of today the weights do not exist publicly; the model is a gated research preview. And ALLaM 34B's weights have been promised and unpublished for over a year. The October target remains an uncompleted plan.

One more asymmetry worth noting: the disclosure ran one way. HUMAIN named MiniMax in its headline, while MiniMax made no corresponding announcement. Its Hong Kong exchange record for August and the first week of September is eleven routine filings - monthly returns, an EGM notice, board changes - and its interim results for the half-year to June 30th, released on August 26th, do not mention HUMAIN, Saudi Arabia or an Arabic model at all, which is consistent with a deal signed after the period, or with one too small to disclose.^[20] The commercial terms (the fee, the exclusivity, who owns the Arabic checkpoint) are not public in any jurisdiction.

Seven Benchmarks, Self-Run

The press release announcing humain-m3 claims it "achieved the highest average score among the frontier models tested" across "seven public Arabic benchmarks." It prints no numbers, no benchmark names and no competitor names. There's a heading reading "Arabic Benchmark Performance" with prose underneath and no table.

The table exists, on a different property. node.humain.com (HUMAIN Node, a separate, server-rendered site that nothing in the press release links to) publishes it in full.^[21]

Start with what's good about it, because it's the part that gets least attention. These seven are exactly the scored suite of the Open Arabic LLM Leaderboard, version 2 - a public, independently maintained, machine-translation-free benchmark set.^[22] HUMAIN did not invent a favorable evaluation, and it did not drop an inconvenient one. It used the standard set, unmodified. It also prints its own losses: humain-m3 leads five of seven and loses ALRAGE to GPT-5.6 Sol and Translated MMLU to Opus 5. Node says "leads five of seven" in its own copy.

The result shows a margin of about two points over the tested frontier models and 9.03 points over MiniMax-M3, its own base model. Node states: "HUMAIN's Arabic post-training adds nine points on average." HUMAIN has demonstrated useful Arabic post-training of an existing base, without demonstrating frontier pretraining.

There's one legitimate external comparison available. The Open Arabic LLM Leaderboard's results repository holds the raw evaluation file for an independent run of ALLaM-7B from February 19th, 2025, scored by the leaderboard's own harness rather than by HUMAIN.^[23] Regrouping that file's per-category scores into HUMAIN's own seven groupings, equally weighted, gives ALLaM-7B a score of 65.25 - a calculation this piece performed directly against that raw file, not a figure OALL or HUMAIN has published as such. Set against HUMAIN's figures:

• 65.25 - ALLaM-7B, Saudi Arabia's own 2025 sovereign model, scored independently

• 80.34 - MiniMax-M3, untuned, before any Arabic work

• 89.37 - humain-m3

Because the evaluator and model generation differ, the numbers show only how the systems compare on a similarly grouped task set. But the shape is instructive: the distance from Saudi Arabia's own model to the Chinese base it commissioned on top of (+15.09) is larger than the distance HUMAIN's Arabic post-training then added (+9.03).

And then the limits, which are severe. The scores are HUMAIN's own - its caption says so: "Scores are HUMAIN's evaluation of the previewed humain-m3 checkpoint." The competitor scores for GPT-5.6 Sol and Opus 5 were also produced by HUMAIN. No harness is named, no shot count, no decoding configuration, no evaluator. Nobody outside the company can reproduce any of it: there are no weights, and API access is gated behind an approval queue. The independent leaderboard cannot settle it either - it has never scored a 2026 frontier model, so there is no published Arabic evaluation anywhere that these numbers can be laid against.

What Is the "Saudi Alignment Guardrail"?

Node's access page describes three steps, and the second is the most interesting one.

"Once approved, you can use HUMAIN M3 with the Saudi alignment guardrail - made for everyday use and testing. The guardrail adds a little latency, with thinking and streaming off."

Above it sits a third tier: "Once approved, you can use the full checkpoint to push the model to its limits," with "Thinking, streaming, and lower latency."

So the configuration in the benchmark table is tier three. The configuration most approved users will actually get is tier two, with reasoning disabled. Readers of our Fable 5.1 review will recognize the structure exactly: a model name that no longer answers the question of what you actually get.^[24] Over the past year, frontier labs have increasingly used routing policies to determine which model configuration customers receive. HUMAIN has now arrived at the same place from a third direction.

HUMAIN's distinctive contribution is a national policy layer applied as a product tier. That phrase - "the Saudi alignment guardrail" - is HUMAIN's, not ours. It's offered alongside a hosting menu that lets a customer choose "Global, in-Kingdom, or sovereign hosting" per workload.

HUMAIN's more distinctive claim lies in policy control over model behavior, because its compute and base models remain dependent on foreign suppliers. Node describes itself as carrying "100+" models from "the world's leading providers" behind one API. In HUMAIN's product design, the national guardrail supplies the Saudi policy layer across models sourced from multiple providers.

The Rule That Was Never Rescinded

The Biden administration's "Framework for Artificial Intelligence Diffusion" was published in January 2025. Despite Commerce's headline, the framework remains codified while enforcement is suspended. License Exception AIA at 15 CFR 740.27 is live; Supplement No. 5 still lists its nineteen authorized destinations; the country allocation cap of 790,000,000 total processing performance for 2025 to 2027 is still in the text.^[25]

What happened on May 12th, 2025 was a Commerce Department press release. Its headline says the department "Rescinds" the rule. Its body says something materially weaker: Commerce "initiated a rescission," "plans to publish a Federal Register notice formalizing the rescission," and the Under Secretary "instructed BIS enforcement officials not to enforce" it.^[26]

That Federal Register notice has still not been published. The regulatory agenda entry has slipped from a projected July 2025 to a projected July 2026, and both dates have passed. Sixteen months on, the gap between what the law says and what the agency does is the regime.

The rule never uses the words "Tier 1," "Tier 2" or "Tier 3" - that framing is journalistic shorthand and should not be quoted or attributed to the government. And the rule never mentions Saudi Arabia at all.

How Does Saudi Arabia's Chip Access Compare With the UAE's?

Saudi Arabia sits in Country Groups D:3 and D:4 in the Export Administration Regulations, and in no A-group. The consequence is concrete: every advanced AI chip shipment to Saudi Arabia requires an individual export license, reviewed under a presumption of approval up to that 790-million-TPP cumulative cap, after which the policy becomes denial.^[27] License Exception AIA is not available to it.

The United Arab Emirates took a different route. On July 10th, 2026, a final rule moved the UAE into Country Group A:5, removed it from D:3 and D:4, and created a new Supplement No. 8 naming the entities authorized to receive advanced computing items license-free - G42 and Core42 by name, alongside eight US companies. The rule publishes an accession procedure: an advisory-opinion request, a thirty-day clock, and named decision-makers. It also sets an expiry: the G42 and Core42 authorization "shall automatically expire on April 6th, 2027."^[28]

Neither "Saudi Arabia" nor "HUMAIN" appears anywhere in that rule.

The UAE's access is rule-based, entity-named, time-limited, and has a published door that others can walk through. Saudi Arabia's is discretionary, unnamed, open-ended, and has no published door at all. Saudi Arabia ended up with the weaker legal footing and the lighter published scrutiny simultaneously.

What HUMAIN actually received is four paragraphs on a Commerce webpage, dated November 19, 2025: Commerce "has authorized the export of advanced American semiconductors to G42... and Humain," with "both companies... receiving approvals to purchase the equivalent of up to 35,000 Nvidia Blackwell chips (GB300s)," conditioned on "rigorous security and reporting requirements" - about which the very next sentence concedes that BIS "is engaging with the companies regarding these requirements."^[29]

That is the entire public disclosure. The statement disclosed no dollar figure, license number, validity period, expiry, operator requirement, China clause or audit regime. The conditions were still being negotiated when the approval was announced, and they have never been published. The wording does not establish whether the 35,000-chip limit applies to each company or to both combined.

In September 2025, while the deal was still under negotiation, the Center for a New American Security reported that HUMAIN "could purchase hundreds of thousands of advanced NVIDIA and AMD chips to build a 1 gigawatt data center."^[30] What Commerce authorized two months later was up to 35,000.

Supplement No. 10 to Part 748 - the Data Center Validated End User regime - requires a recipient to demonstrate "a credible plan to or demonstrated track record of respecting human rights"; to be free of ties to military or military-intelligence end users; to have "eliminated supply chain dependencies on advanced semiconductors... produced by any entities headquartered in Macau or destinations specified in Country Group D:5"; and to "notify the U.S. Government... of all cooperative activities, such as joint ventures," with such entities.^[31]

The Data Center VEU regime has no designated entities in Saudi Arabia, the UAE or any other country. The human-rights criterion, the China-divestment clause and the audit architecture sit in the Code of Federal Regulations, fully drafted, applying to nobody. Both Gulf arrangements route around them - Saudi Arabia through individual licenses with unpublished conditions, the UAE through a bespoke supplement carrying none of these criteria.

The GAIN AI Act excludes Saudi shipments under its country definition; the destination-agnostic Chip Security Act would cover the chip classes licensed to HUMAIN. The GAIN AI Act's text defines "country of concern" as Country Group D:5 or E, plus Macau and Hong Kong.^[32] Saudi Arabia is D:3 and D:4. The Chip Security Act would require location verification on exactly the chip classes licensed to HUMAIN, regardless of destination.^[33] Both bills remain introduced and referred to committee; the GAIN AI Act's provisions have separately advanced as part of the Senate's fiscal year 2026 defense authorization bill, but that vehicle does not alter the underlying country-of-concern definition.

The Regulator That Shares a Chairman

Saudi Arabia has no AI law. SDAIA's AI Ethics Principles, first published in September 2023, are still labeled a draft on the live page three years later. Compliance runs through "Optional Registration," which "aims to motivate target entities," with follow-up via "optional reports" and the possibility of "Motivational Badges." No penalty, fine, sanction, revocation or appeal appears anywhere in the document.^[34]

This is worth contrasting with the Personal Data Protection Law, which is real and does bite: Royal Decree M/19 carries imprisonment of up to two years and fines to SAR 3 million for disclosing sensitive data with intent to harm, and administrative fines to SAR 5 million, doubling on repeat.^[35] On the same national platform, PDPL controller registration is mandatory and AI-ethics registration is optional. The distinction is deliberate, and those PDPL penalties should never be cited as though they're attached to the AI framework.

Nor would enforcement be visible if it happened. The rules of procedure for the adjudicating committee state that "all deliberations of the Committee and its decisions shall be confidential unless the decision specifically provides for the publication of a summary." Publication is opt-in, decided by the committee itself.

Mohammed bin Salman chairs SDAIA's board, PIF, and HUMAIN. SDAIA's president, Dr. Abdullah Alghamdi, sits on HUMAIN's board - and previously chaired SCAI, the PIF artificial-intelligence vehicle whose people were folded into HUMAIN. And HUMAIN's own About page lists SDAIA under the heading "OUR STRATEGIC ALLIES," first in an array that continues Qualcomm, NVIDIA and Cisco.

A fair-minded Saudi official has an answer to most of this, and it's a good one: of course the Crown Prince chairs the national AI champion - he chairs the sovereign wealth fund and much of its portfolio, and every sovereign AI program on earth involves state coordination. That answer is correct, and it disposes of the org chart as such.

The entity sharing a chairman with HUMAIN is not merely a state body; it's the regulator - the authority that writes Saudi data and AI rules and enforces the PDPL under which HUMAIN Chat is hosted. A regulator that shares a chairman with the operator it regulates, seats its president on that operator's board, and is presented on that operator's website at vendor tier beside its chip suppliers is a structure with a name. The question it raises isn't rhetorical: what happens when HUMAIN needs a PDPL ruling?

And there's one documented consequence rather than a hypothetical one. A state-developed asset - ALLaM - moved from SDAIA's namespace into a PIF commercial company's, with the paperwork still bearing the previous owner's name, and no instrument anywhere recording the transfer. The overlapping leadership provides a plausible institutional explanation for the absence of a visible transfer process, though the documents do not establish causation.

Those facts are reported here because a company chaired by the Crown Prince, partnered with the Interior Ministry, and selling a national alignment layer is operating inside that context. They are not offered as an allegation against HUMAIN, because on the public record no such allegation exists.

The Case for the Defense

If frontier intelligence is becoming commoditized, HUMAIN's emphasis on distribution may be an efficient strategy. Open-weight models now trail the frontier by months rather than years. Six Western labs have queued up to partner. In that world, spending tens of billions to train a model that will be matched by an open release within a year is the worst available use of capital - and owning the compute, the routing layer, the enterprise integration and the alignment chokepoint is the best. Those are the layers where capital and location compound. The strategy directs capital toward infrastructure and distribution assets that may retain value while sourcing models that depreciate quickly.

Open weights, once downloaded, cannot be revoked by Beijing. Every NVIDIA and AMD rack in Riyadh runs under US export permissions that can be withdrawn at any time, under a regime whose central rule was suspended by a press release and could be unsuspended the same way. Under that definition, HUMAIN's exposure to revocable US chip permissions may exceed its exposure to a downloaded model checkpoint.

Three honest answers, none of them complete. First, the license terms on the weights are not public, and "cannot be revoked" is an assumption about a document nobody outside the parties has read. Second, the guardrail and the router are themselves dependencies - for everyone downstream of HUMAIN, sovereignty is being relocated rather than distributed. Third, and most damning for the strategy as executed rather than as conceived: on the ALLaM record, "we route it" has so far meant "we announce it." A distribution layer is only a strategy if the distribution works, and two of its four components - the enterprise agent OS and the model marketplace - remain future tense ten months after their stage demos.

A note on the explanation this piece deliberately does not offer. The obvious answer to "why build here" is cheap energy, and we cannot support it. We could not verify the Saudi industrial tariff instrument to a standard that would let us print a number, so no cents-per-kilowatt-hour figure appears in this piece and no comparison to American wholesale power is drawn. What can be said is narrower and comes from HUMAIN: it claims its data centers are "about 30% cheaper for compute customers than elsewhere in the world"^[42] - HUMAIN's figure concerns customer compute prices, which can reflect subsidies, land, tax treatment and expected returns as well as electricity costs.

HUMAIN's strongest case rests on access to firm megawatts on a state-backed timeline.

What Should We Watch Next?

Everything above is a snapshot of a company that publishes in bursts and is measured in years. These are the falsifiable things, with dates attached, and a reader can check every one themselves.

35.1 megawatts by the end of 2027. HUMAIN's own commitment, in the same sentence as the 1.1. Its next test is scaling the NVIDIA cloud thirty-two-fold to 35.1 megawatts by the end of 2027.

The humain-m3 weights, promised for October 2026, under the MiniMax Community License. If they appear, HUMAIN's Arabic post-training becomes independently checkable for the first time and the benchmark table stops being unfalsifiable. If they slip the way ALLaM 34B's have slipped for a year, that is its own answer.

The Al Moammar contract. A letter of award for 250 MW, explicitly unsigned as of August 27th, 2026, with the parties expected to sign within two weeks. That deadline has now passed.

Whether Aramco's term sheet ever closes, and at what percentage. It has been non-binding and unsized since October 2025.

Whether a chief financial officer is ever named - particularly now that the company has begun assembling an IPO preparation team, with a previously stated timeframe extending to 2029 for listings in Riyadh and New York.^[43]

Whether the Federal Register notice rescinding the AI Diffusion Rule ever appears. Sixteen months, two missed agenda deadlines, and a rule still on the books that no one is enforcing.

HUMAIN has energized about eight racks of power over those same sixteen months, shipped one consumer product, inherited one model family and commissioned another, and convened most of the American semiconductor industry in Riyadh. Its owner says it has become an intellectual property creator, although its newest model still depends on MiniMax's license, Washington's chip approvals and a guardrail written in Riyadh.

For HUMAIN, sovereignty will be measured by whether control of compute, routing and alignment can compensate for dependence on foreign models and American chips.

Sources

1. HUMAIN: NVIDIA cloud launch and operational capacity Inline ↗

   5 passages checked · read September 12, 2026

   …Today, HUMAIN's AI cloud provides 1.1 MW of operational capacity expanding to 35.1 MW by end of 2027 and is supporting private and public sector with agentic AI and…

   …private and public sector with agentic AI and GPU-accelerated production workloads with more than 90% utilization of current cluster.…

   …HUMAIN's infrastructure currently includes: 57 NVIDIA HGX B300 systems deployed and available fully operated by HUMAIN running HUMAIN Sovereign Cloud Stack offering comprehensive AI…

   …offering comprehensive AI Cloud Services 1.1 MW of operational AI infrastructure capacity 6 government and enterprise customers using the platform 16 leading AI models, multiple applications and production workloads running on the…

   …and NVIDIA previously announced plans to build AI factories with a projected capacity of up to 500 megawatts, powered by several hundred thousand NVIDIA GPUs over five years.…

2. AMD: HUMAIN Instinct systems deployment Inline ↗

   3 passages checked · read September 12, 2026

   …AMD, Cisco and HUMAIN plan to deploy up to 250 MW of AI infrastructure as part of the next phase of the buildout beginning in 2027.…

   …MI355X GPU-based AI infrastructure built on Cisco Silicon One-based AI networking is now live in Saudi Arabia and serving HUMAIN customers in the Kingdom and beyond.…

   …customer demand, the AMD, Cisco and HUMAIN joint venture remains on track to deploy up to 1 GW of AI infrastructure by 2030.…

3. HUMAIN: humain-m3 research preview Inline ↗

   4 passages checked · read September 12, 2026

   Commissioned by HUMAIN and delivered by MiniMax, humain-m3 advances Arabic AI performance across seven public benchmarks and is available in research preview today through HUMAIN Node RIYADH, Saudi Arabia — 3 September, 2026 — HUMAIN, a…

   …Built on the MiniMax-M3 lineage, humain-m3 is a 428-billion-parameter mixture-of-experts model further pre-trained on more than one trillion tokens of Arabic-native content.…

   …HUMAIN expects the model weights to be released under the MiniMax Community License following completion of safety training and alignment, currently targeted for next month.…

   …In evaluations across seven public Arabic benchmarks, humain-m3 achieved the highest average score among the frontier models tested, demonstrating strong performance across Arabic language understanding and reasoning.…

4. PIF: HUMAIN launch announcement Inline ↗

   3 passages checked · read September 12, 2026

   …Chaired by HRH the Crown Prince, HUMAIN will provide a comprehensive range of AI services, products and tools, including next-generation…

   …HUMAIN will operate and invest across the artificial intelligence (AI) value chain as a unified operating company.…

   …The company will also offer one of the world’s most powerful multimodal Arabic large language models (LLMs).…

5. HUMAIN: Terms of Use and legal entity Inline ↗

   3 passages checked · read September 12, 2026

   …May 2025 INTRODUCTION Our website, https://www.humain.ai (the “ Site ”) is provided by Al-Mustaqbal Lil-Thaka Al-Istinai Company, also known as Future Artificial Intelligence Company (“ HUMAIN ”).…

   …We are a joint stock company registered in the Kingdom of Saudi Arabia (“ KSA ”) under commercial registration number 1009089438, with our…

   …registered in the Kingdom of Saudi Arabia (“ KSA ”) under commercial registration number 1009089438, with our office located at Al Aqeeq District 7956, Al Intessar Street 3229, Post Code 13511, Riyadh, KSA.…

6. USPTO: HUMAIN trademark application Inline ↗

   2 passages checked · read September 12, 2026

   …No Filed No Basis: No Currently No Basis: No Current Owner(s) Information Owner Name: Al-Mustaqbal Lil-Thaka Al-Istinai Company Owner Address: King Fahad District 3254 King Fahad Branch Road 7295, Post Code 12262 Riyadh SAUDI…

   …Foreign Application Filing Date: Mar. 25, 2025 Foreign Application/Registration Country: Saudi Arabia Priority Claimed: Yes Foreign Application Number: TM01001256925 Foreign Application Filing Date: Mar. 25, 2025…

7. HUMAIN: leadership page Inline ↗

   3 passages checked · read September 12, 2026

   …Alhumied Mr. Turqi bin Abdulrahman Alnowaiser Mr. Abdullah bin Fahad Alabduljabbar Tareq Amin is a technologist, business leader and entrepreneur with over two decades of leadership experience in technology…

   …bin Abdullah Toonsi H.E. Mr. Mohammed bin Mazyad Al-Tuwaijri H.E. Dr. Abdullah bin Sharaf Alghamdi H.E. Dr. Hamad Al Al-Sheikh Eng.…

   …Amin bin Hassan Nasser Mr. Yazeed bin Abdulrahman Alhumied Mr. Turqi bin Abdulrahman Alnowaiser Mr. Abdullah bin Fahad Alabduljabbar Tareq…

8. NVIDIA: HGX platform specifications Inline ↗

   1 passage checked · read September 12, 2026

   …class=" third-column odd-column odd-row"> HGX B200<sup>4</sup> Form Factor 8x NVIDIA Blackwell Ultra SXM 8x NVIDIA Blackwell SXM 1</sup>" data-col-index="0" data-row-index="2" class=" first-column odd-column…

9. SEC: capital expenditure company-concept data Inline ↗

10. HUMAIN: Media Centre archive Inline ↗

    2 passages checked · read September 12, 2026

    …More HUMAIN and KORA Partner to Build Operating System at LEAP 2026 Read More HUMAIN Unveils humain-m3, a Frontier Arabic Language Model Developed by MiniMax, in Research Preview on HUMAIN Node Read More Microsoft and HUMAIN Expand…

    …Personal Computing Unveiled by CEO Tareq Amin at Qualcomm’s Snapdragon Summit Read more HRH Crown Prince launches HUMAIN as global AI powerhouse Read More HUMAIN Horizon Pro: The Future of Agentic AI Personal Computing Unveiled by CEO…

11. Saudi Aramco: results and HUMAIN term sheet Inline ↗

12. SEC EDGAR: HUMAIN filing search Inline ↗

13. HUMAIN: xAI Series E investment Inline ↗

    3 passages checked · read September 12, 2026

    …full-stack artificial intelligence capabilities globally, today announced a $3 billion strategic investment in xAI as part of the company’s Series E financing round.…

    …· As a result of the transaction, HUMAIN became a significant minority shareholder, with its xAI holdings converted into SpaceX shares.…

    …HUMAIN became a significant minority shareholder in xAI, with its holdings subsequently converted into shares in SpaceX.…

14. HUMAIN: Luma AI Series C and supercluster Inline ↗

    3 passages checked · read September 12, 2026

    …frontier artificial intelligence company building multimodal AGI, today announced it has raised $900 million in Series C funding (including SAFEs and two tranches of funding), led by HUMAIN , a PIF company delivering global…

    …$900 million in Series C funding (including SAFEs and two tranches of funding), led by HUMAIN , a PIF company delivering global full-stack AI solutions, with significant participation from AMD, Amazon, and existing investors Andreessen…

    …HUMAIN to build Project Halo, one of the world’s largest AI compute clusters that will include the training of World Models, the next…

15. HUMAIN: HUMAIN Chat and ALLAM 34B launch Inline ↗

    4 passages checked · read September 12, 2026

    …HUMAIN’s ALLAM 34B builds on the previous released models developed by the National Center for AI under The Saudi Data & AI Authority (SDAIA) national mandate.…

    …Over 120 AI specialists, including 35 PhD holders, with a 50/50 gender ratio, have built this model from the ground up - hosted in Saudi, by Saudis, with global talent alongside them.…

    …announced the launch of HUMAIN Chat, a next-generation Arabic conversational AI app powered by ALLAM 34B, HUMAIN’s flagship Arabic large language model.…

    …ALLAM 34B, its most advanced model, has been independently verified by Cohere on MMLU as the world’s most advanced Arabic LLM built in the Arab world and will continue to improve with consumer engagement post launch.…

16. Hugging Face: ALLaM-7B model card Inline ↗

    4 passages checked · read September 12, 2026

    …models designed to advance Arabic Language Technology (ALT) developed by the National Center for Artificial Intelligence (NCAI) at the Saudi Data and AI Authority (SDAIA) .…

    …Acknowledgment This release is funded by Saudi Data & AI Authority .…

    …We thank Mishari Almishari and Yaser Alonaizan for leading the ALLaM project.…

    …Our pretraining from scratch recipe consists of two steps: training on 4T English tokens followed by training on 1.2T mixed Arabic/English tokens.…

17. arXiv: ALLaM research paper Inline ↗

    1 passage checked · read September 12, 2026

    [2407.15390] ALLaM: Large Language Models for Arabic and English Skip to main content Search Submit Donate Log in Search arXiv Press Enter to search · Advanced search --> Computer Science > Computation and Language arXiv:2407.15390…

18. HUMAIN: Yaser Al-Onaizan byline and role Inline ↗

    1 passage checked · read September 12, 2026

    Dr. Yaser Al-Onaizan Deputy CEO & President of AI Products The key to enterprise AI adoption is not task automation In 2025, an MIT report found that 95% of generative AI initiatives at large companies have produced no measurable impact on…

19. Hugging Face: MiniMax-M3 model card Inline ↗

    2 passages checked · read September 12, 2026

    …It has ~428B parameters and ~23B activated parameters.…

    …agent coding video conversational custom_code Eval Results arxiv: 2606.13392 License: minimax-community Model card Files Files and versions xet Community 29 Deploy Copy to bucket new Use this model Instructions to…

20. MiniMax: interim results announcement Inline ↗

    2 passages checked · read September 12, 2026

    …Code: 0100) INTERIM RESULTS ANNOUNCEMENT FOR THE SIX MONTHS ENDED JUNE 30, 2026 FINANCIAL HIGHLIGHTS Six months ended June 30,…

    …Revenue generated outside Chinese mainland contributed approximately 60.8% of our total revenue for the six months ended June 30, 2026.…

21. HUMAIN Node: benchmark table and access tiers Inline ↗

    5 passages checked · read September 12, 2026

    …for HUMAIN M3 2 Request limited preview Once approved, you can use HUMAIN M3 with the Saudi alignment guardrail -- made for everyday use and testing.…

    …The guardrail adds a little latency, with thinking and streaming off.…

    …Scores are HUMAIN's evaluation of the previewed humain-m3 checkpoint; seven public Arabic benchmarks, equally weighted.…

    …Against the reference model it builds on, HUMAIN's Arabic post-training adds nine points on average -- the difference between speaking Arabic and being built for it.…

    …Global, in-Kingdom, or sovereign hosting Choose where each workload runs -- without changing your code.…

22. Hugging Face: Open Arabic LLM Leaderboard 2 Inline ↗

    3 passages checked · read September 12, 2026

    …Furthermore, we identified a silent bug in one of the main tasks, AlGhafa, which inadvertently impacted model rankings.…

    …AraTrust [8]: a dataset comprising of 522 human-written multiple-choice questions covering different aspects related to…

    …Finally, we introduce the ALRAGE benchmark: Arabic Language Retrieval Augmented Generation Evaluation.…

23. OALL: raw per-category evaluation results for ALLaM-7B-Instruct-preview, February 19, 2025 Inline ↗

    3 passages checked · read September 12, 2026

    …4945.569198387, "total_evaluation_time_secondes": "4227.222765372", "model_name": "ALLaM-AI/ALLaM-7B-Instruct-preview", "model_sha": "661e616893aaa90a934e02de5b5030a8c0899e01", "model_dtype": "torch.bfloat16",…

    { "config_general": { "lighteval_sha": "?", "num_fewshot_seeds": 1, "override_batch_size": 1, "max_samples": null, "job_id": 0, "start_time": 718.346433015, "end_time": 4945.569198387, "total_evaluation_time_secondes": "4227.222765372",…

    …{ "acc_norm": 0.6239110287303058, "acc_norm_stderr": 0.006595552160200563 }, "community|alghafa:meta_ar_msa|0": { "acc_norm": 0.7240223463687151, "acc_norm_stderr": 0.01495010300247536 },…

24. Omniscient: Fable 5.1 review Inline ↗

    1 passage checked · read September 12, 2026

    …Inside Claude Fable 5.1: One Model, Five Products, and the Question a Model Name No Longer Answers | Omniscient Media Home › AI Research ›…

25. eCFR: License Exception AIA Inline ↗

26. BIS: AI Diffusion Rule press release Inline ↗

    3 passages checked · read September 12, 2026

    …Controls May 12, 2025 WASHINGTON, D.C.-- Today, the Department of Commerce (DOC) initiated a rescission of the Biden Administration's AI Diffusion Rule, while announcing additional steps to strengthen…

    …BIS plans to publish a Federal Register notice formalizing the rescission and will issue a replacement rule in the future.…

    …Under Secretary of Commerce for Industry and Security Jeffery Kessler has instructed BIS enforcement officials not to enforce the Biden Administration's AI Diffusion Rule, stating: "The Trump Administration will…

27. eCFR: advanced-computing export controls Inline ↗

28. Federal Register: UAE favorable treatment rule Inline ↗

29. Commerce: UAE and Saudi chip exports Inline ↗

30. CNAS: Saudi Arabia and advanced chips Inline ↗

    1 passage checked · read September 12, 2026

    …Under its initial terms, the Saudi AI champion Humain could purchase hundreds of thousands of advanced NVIDIA and AMD chips to build a 1 gigawatt (GW) data center.…

31. eCFR: Data Center VEU guidelines Inline ↗

32. Congress: GAIN AI Act Inline ↗

    1 passage checked · read September 12, 2026

    …``(4) Country of concern.--The term `country of concern' means-- ``(A) a country listed in Country Group D:5 or E in Supplement No. 1 to part 740 of the Export Administration Regulations; ``(B) the Macau Special…

33. Congress: Chip Security Act Inline ↗

    1 passage checked · read September 12, 2026

    …integrated circuit product to be outfitted with chip security mechanisms that implement location verification, using techniques that are feasible and appropriate on such date of enactment, before it is exported,…

34. SDAIA: AI Ethics Principles Inline ↗

    2 passages checked · read September 12, 2026

    …Optional Registration Optional Registration aims to motivate target entities to consider AI ethics when building and developing…

    …Motivational Badges To motivate the implementing agencies to register and work with the principles of AI ethics, the Authority…

35. Saudi Arabia: Personal Data Protection Law Inline ↗

    1 passage checked · read September 12, 2026

    Personal Data Protection Law Article 1 For the purpose of implementing this Law, the following terms shall have the meanings assigned thereto, unless the context requires otherwise: 1-Law: The Personal Data Protection Law.…

36. HUMAIN: Ministry of Interior partnership announcement (unreachable this session; see following footnote for verified substance) Inline ↗

    4 passages checked · read September 12, 2026

    …Prince Abdulaziz bin Saud bin Naif bin Abdulaziz, Minister of Interior, the Ministry of Interior today announced the launch of a strategic partnership with HUMAIN, a PIF company delivering full-stack artificial intelligence capabilities…

    …capabilities in artificial intelligence and advanced computational infrastructure to support its operational and administrative activities.…

    …It represents a key milestone in the Ministry’s efforts to modernize its technological environment and enhance digital capabilities across its internal functions.…

    …The launch is designed to empower Ministry employees with access to advanced AI tools that support faster information retrieval, the performance of administrative and…

37. Asharq Al-Awsat, via Saudi Press Agency: Interior Ministry-HUMAIN agreement Inline ↗

    3 passages checked · read September 12, 2026

    …(SPA) Assistant Minister of Interior for Technology Affairs Prince Dr. Bandar bin Abdullah bin Mishari affirmed that the strategic agreement signed by…

    …This contributes to the modernization of the technological environment and the enhancement of the efficiency of operational and administrative functions.…

    …The assistant minister further clarified that the agreement aims to strengthen the ministry’s digital capabilities across its internal functions and to adopt advanced domestic technologies in large language models,…

38. Freedom House: Saudi Arabia, 2026 Inline ↗

    2 passages checked · read September 12, 2026

    …Financials Careers Contact Us Sign Up for Updates Freedom in the World 2026 Saudi Arabia Not Free 9 100 PR Political Rights 1 40 CL Civil Liberties 8 60 Last Year's Score & Status 9 100 Not Free A country or…

    …The regime relies on pervasive surveillance, the criminalization of dissent, appeals to sectarianism and ethnicity, and public spending supported by oil revenues to…

39. Human Rights Watch: Saudi Arabia, 2026 Inline ↗

    1 passage checked · read September 12, 2026

    …Saudi authorities continue to use vague counterterrorism and cybercrime laws to silence dissent.…

40. Justice Department: Ahmad Abouammo sentencing Inline ↗

41. ODNI: assessment of Saudi government's role in Khashoggi killing Inline ↗

42. Semafor: HUMAIN on customer compute prices Inline ↗

    1 passage checked · read September 12, 2026

    …any slowdown in demand from AI companies to use its data centers, which are about 30% cheaper for compute customers than elsewhere in the world, he said.…

43. Arab News: HUMAIN IPO preparations Inline ↗

    2 passages checked · read September 12, 2026

    …Just 16 months after its launch, Saudi Arabia’s Public Investment Fund-owned Humain has begun preparations for a potential public offering, with its CEO publishing an announcement seeking a specialized team for the task.…

    …The company had previously set a timeframe extending to 2029 for listings on the Saudi and New York stock exchanges.…