HUMAIN’s Switzerland Pitch Still Runs Through China

HUMAIN chief executive Tareq Amin is hiring an IPO team, 16 months after Saudi Arabia’s wealth fund launched the company as a global AI bet. On September 7 he posted on LinkedIn that he wants “a few exceptional individuals to join our IPO preparation team at HUMAIN,” to work on investor strategy and the pitch.

Four days earlier, in an interview, Amin described HUMAIN as a marketplace, “a Switzerland for open models.” That same day in Riyadh the company unveiled humain-m3, an Arabic system it commissioned from China’s MiniMax. The neutrality line is the product. The stack is not neutral.

Saudi’s Flagship Model Starts in Shenzhen

Amin, a Jordanian-American executive who previously led Saudi Aramco’s tech arm, has been clear that HUMAIN cannot train a frontier model the way OpenAI does. When his team built Allam, the company’s earlier Arabic model, it used about 4,000 GPUs. OpenAI, he said, has hundreds of thousands. Scientists with that depth are scarce. Compute at that scale is scarce. Short term, he said, open-weight models are “the strategy to guarantee sovereignty.”

humain-m3 is that strategy in public. HUMAIN commissioned it. MiniMax delivered it. The model is a 428-billion-parameter mixture-of-experts system built on the MiniMax-M3 line and further trained on more than one trillion tokens of Arabic-native text. HUMAIN put a research preview on HUMAIN Node on September 3 at the LEAP conference and said it expects to release weights under the MiniMax Community License after safety training, aimed at October.

https://x.com/HUMAIN/status/2095482874673631495

On seven public Arabic tests, HUMAIN’s own table puts humain-m3 at an equal-weighted average of 89.37%, ahead of GPT-5.6 SOL at 87.30% and Claude Opus 5 at 87.34%, and 9.03 points above the MiniMax M3 reference at 80.34%. It led five of the seven tests. On ALRAGE, a retrieval task, it trailed both GPT-5.6 SOL and Opus 5.

ARABIC BENCHMARK SCORES

Those scores are HUMAIN’s. They are also a map of dependence. A US-aligned Gulf flagship just put a Chinese open model under its national Arabic product, on campuses that will run American chips. That is a working hedge against closed US APIs. It is not a third pole.

Arabic is spoken by hundreds of millions of people, yet it remains significantly underrepresented at the frontier of artificial intelligence. With humain-m3, we are investing in changing that.

Tareq Amin, CEO of HUMAIN, September 3, 2026

Amin is right that Arabic is under-served. The fix he shipped still carries a Shenzhen serial number.

Blackstone Puts $3 Billion Into the Campus

The other half of the pitch is physical. Amin sells predictable power tariffs, empty land, and a build cost he puts at $11 to $12 million per megawatt, among the lowest total costs he sees anywhere. “It becomes a no-brainer to start looking seriously at investing here,” he said. He is not shy about the offtake either. “My entire capacity is sold before it’s deployed.”

On October 28, 2025, that pitch became a term sheet. HUMAIN and AirTrunk, the Asia-Pacific data-center platform backed by Blackstone and Canada Pension Plan Investment Board, agreed to an approximately US$3 billion investment for a campus in the kingdom. AirTrunk’s Middle East entry rides on HUMAIN’s national mandate. Blackstone chairman Stephen A. Schwarzman called AI one of the firm’s “highest conviction themes” and said Blackstone is “the largest provider of data centers globally.”

Land is the easy part. HUMAIN has said it holds 211 plots. Amin says he is not worried about grid capacity. The constraint he will talk about is commercial: lenders want named offtakers before they fund the next hall. Cheap Saudi power is the lure. American private capital is the valve.

A smaller site shows the same pattern on the Red Sea. On August 31, HUMAIN and DataVolt said they would develop about 100 megawatts as the first slice of a 360-megawatt phase inside a planned 1.5-gigawatt campus at Oxagon in NEOM, aimed at 2028. The kingdom’s own live load is still modest. Official figures put operating data-center capacity at 467 megawatts in the first quarter of 2026.

The Outside Money Amin Still Needs

HUMAIN is a Public Investment Fund company, chaired by Crown Prince Mohammed bin Salman, and PIF is still the largest shareholder. Amin’s target is different. “My goal is to be 30-70, 30% shareholder financing and 70% third party,” he said. “We’re not there yet, but I’m getting close to it.” He wants HUMAIN opened to global investors so the company is no longer a slide in a sovereign deck.

He will not give a single capital number. He will give a unit cost and a declared build. Industry builds, he said, run $11 to $12 million per megawatt. HUMAIN has declared a build toward 3 gigawatts. At his own range that is $33 billion to $36 billion. A longer AMD joint-venture target, stated with AMD chief executive Lisa Su, is 1.9 gigawatts by 2030 and 6 gigawatts by 2034.

HUMAIN BY THE NUMBERS

• Declared build: 3 gigawatts, which at $11 to $12 million per megawatt implies $33 billion to $36 billion.
• Financing mix: Amin wants 30% from shareholders and 70% from outside parties, and says the company is not there yet.
• Blackstone check: about $3 billion with AirTrunk for a Saudi campus, announced October 28, 2025.
• xAI stake: $3 billion into xAI’s Series E on February 18, 2026, later converted into SpaceX shares.

The xAI cheque is the one Amin can already point to as done. HUMAIN put $3 billion into the Series E just before SpaceX took xAI in, and the holding converted into SpaceX paper. That followed a November 19, 2025 pact, struck at the U.S.-Saudi Investment Forum, to jointly develop more than 500 megawatts of AI compute in the kingdom, including a flagship site of 500 megawatts or more, and to put Grok into Saudi use. The early win is real. It is also a Musk-line bet, not a Swiss one.

Amin is preparing a large venture fund as well, with offices planned in the United States, France, and the United Kingdom, and has said it should be bigger than a $10 billion vehicle floated in May 2025. The fund is another way to pull third-party money into the same story PIF started.

What HUMAIN Still Buys From Nvidia

Amin can promise low-cost power. He cannot stamp a GPU. On November 19, 2025, HUMAIN said it would deploy up to 600,000 of NVIDIA’s latest systems over three years in Saudi Arabia and, for the first time, in the United States, including NVIDIA GB300 platforms. The first cluster, announced at launch in May 2025, is about 18,000 GB300 GPUs. First chips arrived at the end of December 2025.

The same forum stacked two more American pipes. AWS will deploy and manage up to 150,000 NVIDIA GPUs in a dedicated “AI Zone” in Riyadh, under a partnership the firms dated to May 2025 at more than $5 billion. On September 3, HUMAIN said the next phase aims to make up to 50 megawatts of that AI Zone available by 2028, to put HUMAIN Fabric on AWS Marketplace, and to put Allam on Amazon Bedrock. AWS has also said its first Saudi cloud region remains on track for December 2026. HUMAIN named AWS its preferred AI partner. HUMAIN Chat, aimed at more than 400 million Arabic speakers, is to be trained with NVIDIA Nemotron open models.

HUMAIN COMPUTE DEALS

NVIDIA founder Jensen Huang said AI is essential infrastructure, like electricity, and that every country will build it. Saudi Arabia can host the halls. The accelerators still ship from the same companies that supply Virginia and Oregon. If Washington tightens export rules again, Amin’s offtake math moves with them.

Mistral, Reflection, and a French Door Into Europe

The Switzerland line is a shelf, not a lab. Amin wants HUMAIN to host open models so a customer in Riyadh can run them on kingdom soil if a US or Chinese API goes dark. The shelf filled up in ten days around LEAP.

THE OPEN-MODEL SHELF

• Mistral, August 24, 2026: A pact the firms put in the hundreds of millions of Euros, covering Arabic, cybersecurity and voice, with Mistral exploring HUMAIN data centers and a joint push into regulated Saudi industries.
• Reflection AI, September 1, 2026: The US lab’s first generally available model is slated to run on HUMAIN kit later in 2026, with both sides exploring training offtake. Reflection chief executive Misha Laskin called HUMAIN a partner “to deploy models customized for sovereign AI.”
• MiniMax, September 3, 2026: humain-m3, the Arabic flagship, commissioned by HUMAIN and delivered from the MiniMax-M3 line, with open weights planned after safety work.
• Allam: The earlier Arabic model Amin says was trained on about 4,000 GPUs, now headed toward Bedrock as a hosted option rather than a from-scratch rival to OpenAI.

Mistral’s own note defines the product Amin is buying: data stays inside customer lines, weights can be adapted and owned, and training can run in a jurisdiction the customer picks. That is a commercial answer to export controls and closed APIs. It still means the intelligence is designed in Paris, New York, or Shenzhen, then parked in Riyadh. Amin said he does not think anyone in the Middle East has yet built a state-of-the-art frontier model. Europe is trying, he said, “and I think eventually we will catch up, but we need a bit of time.”

Every Megawatt Is Sold Before It Exists

Amin will not raise the 3-gigawatt target on demand. He could, he says, if the crown prince asks. “If I get asked by His Royal Highness to accelerate, I will absolutely accelerate.” Until then he is matching build to offtake, because third-party debt and equity start with one question: who is the buyer, and on what terms.

“The world is underestimating what is going to come,” he said. “And I think the world does not have enough power capacity to go build for this demand.” That is why he wants contingencies. It is also why a company that sits on cheap gas and empty desert still moves at the speed of hyperscaler contracts. PIF can wait. Blackstone, bank syndicates, and GPU lessors will not.

The mix he describes is split on purpose. Some halls are build-to-suit, with hyperscalers buying the GPUs. Some revenue is meant to come from inference APIs that sell frontier models in a HUMAIN marketplace. The first path is how you fill a campus. The second is how you tell a public-market story that is more than renting power to Amazon and xAI.

Amin Is Staffing an IPO Team

The LinkedIn post on September 7 makes the 70% goal operational. Amin had already talked, in 2025, about a dual listing on the Saudi exchange and a US market in 2028 or 2029. A HUMAIN spokesman declined to give a timetable. Hiring a small IPO-prep group is not a listing. It is a signal that the PowerPoint year is over and the roadshow year has a date on the calendar.

THE 16-MONTH BUILD

1. May 2025: PIF launches HUMAIN as President Donald Trump visits the kingdom; Crown Prince Mohammed bin Salman chairs the company; a first NVIDIA cluster of about 18,000 GB300 GPUs is lined up.
2. October 28, 2025: AirTrunk and Blackstone agree an about $3 billion Saudi campus with HUMAIN.
3. November 19, 2025: HUMAIN expands with NVIDIA, xAI, and AWS in Washington, adding the 600,000-GPU plan, the 500-megawatt-plus xAI site, and the Riyadh AI Zone.
4. February 18, 2026: HUMAIN invests $3 billion in xAI’s Series E; the stake later converts into SpaceX shares.
5. August 24 to September 3, 2026: Mistral, Reflection, and MiniMax land on the open-model shelf; humain-m3 goes to preview at LEAP.
6. September 7, 2026: Amin posts that he is hiring an IPO preparation team.

From launch to that post is 16 months. In that span HUMAIN went from a fund vehicle to a company that wants outside shareholders, a marketplace of other people’s models, and a listing file.

WHAT WE KNOW

• The hire: Amin is recruiting an IPO preparation team for investor strategy and pitch materials, per his September 7 LinkedIn post.
• The window he has named before: a dual listing path he discussed for 2028 or 2029, on the Saudi market and a US exchange.
• The capital rule he set: 30% shareholder money and 70% third party, which he says he has not reached.

WHAT IS UNCONFIRMED

• A listing date: a HUMAIN spokesman declined to comment on timing.
• The size of the new venture fund: Amin said it should exceed the $10 billion idea from May 2025, without a final figure or legal structure.
• How much of the 3-gigawatt build is contracted versus announced: Amin says capacity is sold before it is deployed, without publishing the offtake book.

Amin said the story from a year ago was a PowerPoint. People can now see halls, chips, a chat app, and a model card. They can also see who built each layer. The IPO team’s job is to sell a Saudi AI company. The books will still read like a joint venture with NVIDIA, AWS, Blackstone, Mistral, and MiniMax, with PIF on the register until the 70% arrives.

Disclaimer: This article is news reporting and analysis of HUMAIN’s public statements, partnerships, and capital plans, and it is for information only. It is not investment advice, an offer of securities, or a recommendation to buy or sell HUMAIN, PIF-linked assets, SpaceX, MiniMax, NVIDIA, or any other share or fund. Readers who are weighing an investment should consult a licensed financial adviser who can review their own objectives and risk limits. Figures, deal terms, listing windows, and model scores reflect the company and partner statements cited here and can change as filings, offtake contracts, and product releases move.