Open-weight artificial intelligence (AI) developers have spent 2026 fielding a skeptical question. If a model is free to download, how does the company behind it make money? A new infrastructure deal in the Gulf offers one answer, and it looks nothing like renting access to a chatbot.
The arrangement trades a one-time license fee for a deeper commercial relationship built on local deployment, regional customization and shared compute economics. It is an early example of how open models are expected to generate revenue at scale.
Reflection AI, the open-weight model developer led by co-founder and chief executive Misha Laskin, agreed this month to deploy its models on infrastructure run by HUMAIN, a company backed by Saudi Arabia’s Public Investment Fund. The deal was announced at the LEAP technology conference in Riyadh on September 1.
In a press release announcing the deal, Laskin described HUMAIN as a strategic partner in scaling advanced AI models, saying the deal creates an opportunity to deploy models customized for sovereign use.
The logic traces back to an argument Laskin has made for months: that ownership, not rental, is what really separates open and closed AI.
Laskin frames that distinction with a housing analogy. In his telling, closed AI models function like a rental market, a landlord who can throttle access, restrict use or raise prices at will, because the tenant never owns the underlying asset.
Open models, in the same analogy, are the buying market, an asset a customer can hold, modify and never lose.
“If you’re a large bank, you don’t want to rent something that is going to be the single most important IP (intellectual property) asset you have,” Laskin said. “If you’re a country, you don’t want to rent from a private company in another country. You want to own your intelligence.”
He said this matters because AI has moved in just a few years from what he called a nice-to-have, like chat, to something capable of doing real work, and now a source of concern for national security, for companies and for countries alike.
“With an open model, you can own the model, run it on your infrastructure, and no one can take it away from you once you own it,” he said. “You can customize it and train it for your unique workloads, the same way you can build a garden when you own a house.”
That is the commercial logic behind Reflection AI’s deployment deal with HUMAIN, in which a Gulf sovereign wealth-backed platform gains ownership of infrastructure and customization rights rather than renting access to somebody else’s model.
China’s open model race
The comments came during a fireside chat at London Tech Week, organized by Founders Forum Group and Informa. There, Laskin discussed the future of open-source AI with Iain Martin, Forbes’ senior editor for Europe.
Martin pressed Laskin on China’s dominance of open-model rankings, citing a US narrative that Chinese labs had gained ground by distilling closed Western systems. Two of China’s best-known labs, Zhipu and MiniMax, had listed in Hong Kong that January.
Laskin said strong open models from any country, including China, help prevent a dangerous concentration of AI power, and rejected the idea that Chinese progress rests mainly on distillation. He said pre-training, not distillation, consumes the bulk of the compute required to build a model.
“Pre-training is where most of your compute goes, and it’s very hard,” Laskin said. “Those companies have done a lot of things right outside of distillation.”
He said Chinese labs are extremely capable and talented, and will keep advancing regardless, limited mainly by compute access.
Laskin pointed to DeepSeek V4, a model he said was quieter than its predecessor V3 but more consequential, because its inference was optimized specifically for Huawei chips, part of China’s push to move its entire AI stack onto domestic infrastructure.
“Open models are Trojan horses to the infrastructure that they bring along with them,” Laskin said, citing Reflection AI’s own competing bid to help build a sovereign cloud in South Korea against Chinese rivals.
He compared the dynamic to the 5G standoff between the West and China and to the Belt and Road Initiative, saying a sovereign cloud built on Chinese technology also imports the Huawei chip and software ecosystem around it.
He predicted countries such as Korea will end up adopting both Chinese and Western open models rather than choosing one bloc outright.
Martin then turned to the security implications of that ecosystem, asking whether an enterprise or government could safely run an open Chinese model like DeepSeek locally.
“The Chinese models were trained under copyright laws that are much more lax than in the West, so the models were trained on illicit data,” Laskin said. “An enterprise that uses those models inherits all those liabilities.”
He said the deeper concern is behavioral, not legal. Large language models are, in his words, “mechanical brains,” and in the same way people can be taught to behave deceptively, so can a model.
“If a state actor wants to embed sleeper cell type of behavior, that is completely possible,” Laskin said. “I think that is a pretty scary thing.”
He said the risk of sending sensitive data to a Chinese company can be avoided by running the models on a company’s own infrastructure, but the risk of lock-in cannot.
“A large conglomerate might do a $10 billion or $50 billion build for its AI cloud,” Laskin said. “By the time you’ve made that investment, it’s really hard to pull out.”
From AlphaGo to autonomy
Martin opened by asking why Reflection AI ended up building a Western open AI lab.
“We’re building frontier Western open models,” Laskin said. “Today there’s a very large gap between open-model capabilities in the West and the ones in China.”
Laskin and his co-founder, Ioannis Antonoglou, were both researchers at Google DeepMind before starting Reflection AI, where Antonoglou had been a founding engineer on AlphaGo.
“What made AlphaGo special was it learned to play Go (an ancient board game) through reinforcement learning on its own and became superintelligent,” he said.
He said AlphaGo’s limitation was narrowness, brilliant at Go but unable to play tic-tac-toe, a gap he and Antonoglou set out to close as language models began to emerge.
“Our thesis was simple,” Laskin said. “If you applied the same recipe used to build AlphaGo to language models, scaling up reinforcement learning on top of them, you’d get systems capable of general autonomy.”
Reflection AI's other bet was bootstrapping on a strong open Western model. Llama 3 became what he called the most successful open-model launch in Western history. A year later, Chinese open models overtook Western ones, and Meta abandoned its own open-model project.
“We found ourselves in a position where this core piece of technology we believed needed to exist no longer existed,” Laskin said.
He called that gap Reflection AI’s non-linear path into becoming a frontier open-model developer in its own right, a path that has since taken the company to a $25 billion valuation.
Asked whether the stakes extend to national security, Laskin said he is wary of comparing AI to nuclear technology. He pointed to Reflection AI’s role powering the US Department of Energy’s Genesis Mission, a project to embed AI across 17 national laboratories and help 50,000 scientists advance scientific research.
“It also has powerful negative things you can do with it,” he said. “If a country has very powerful models, open or closed, it can have significant leverage over another country by using those models for cyber attacks.”
That is why governments increasingly treat AI as a national security priority. He said every government Reflection AI works with is thinking deeply about how to become sovereign, how to run models and keep the IP within their own country, and how to train a workforce capable of taking these models and customizing them.
The conversation closed on London.
“London is largely the source of all this AI innovation that’s been happening over the last decade,” Laskin said. “The first company to take artificial general intelligence as its mandate seriously was DeepMind, founded in 2010, when it was considered largely laughable by the academic community.”
Reflection AI now has three core locations: San Francisco, New York, and London. The company is hiring toward a team of 100 in London now, growing to 1,000 within the next few years, while also supporting UK enterprises and the government’s own AI sovereignty push.
“My co-founder [Antonoglou] lived in London for over a decade,” Laskin said. “I spent a lot of my time here working with the team, so it means a lot to us personally.”
With Reflection AI’s headcount in London set to grow tenfold, that personal history looks likely to keep shaping where the company builds next.