Artificial intelligence has become an increasingly costly business expense, causing a growing number of companies to turn to cheaper, open-weight models that can be customized and run on their own infrastructure, giving them more control over their data. The vast majority of these open-weight models are made in China, though, triggering data privacy concerns and broader anxieties about the United States falling behind in a strategically important area of AI development.
What Is Reflection AI?
Reflection AI is an American AI lab founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou. The company develops open-weight foundation models and autonomous agents, positioning itself as an alternative to closed frontier models and open models from China.
Reflection AI wants to change that. Founded by former Google DeepMind researchers, the startup has been called the “DeepSeek of the West,” referencing the Chinese AI company DeepSeek that shocked the AI industry with its first powerfully efficient open-weight model in 2025. While other frontier labs have dabbled in open-weight models, Reflection AI is the first American company to devote themselves to it completely.
Reflection has already shown strong momentum, attracting billions of dollars from some of the biggest names in the AI industry. And the company says its debut model, Beam, rivals the top Chinese systems while using a fraction of the computing power. But what exactly is Reflection AI building, who is backing its ambitions and how will it compete with the world’s top AI labs?
What is Reflection AI?
Reflection AI is an AI company founded in 2024 by Misha Laskin and Ioannis Antonoglou, two former Google DeepMind researchers. Laskin worked on reward modeling for Google’s Gemini project, while Antonoglou was a founding engineer at DeepMind who helped develop AlphaGo, the AI system that defeated a world champion Go player in 2016.
Unlike closed, proprietary models developed by frontier labs like OpenAI, Anthropic and Google, Reflection AI is building open-weight models that can be downloaded, customized and deployed on an organization’s own infrastructure. The weights — the numerical parameters that determine the strength of connections in a neural network — are made public, so other researchers can analyze and customize them. Silicon Valley was built on a culture of open-sourcing technology for others to build upon, but Reflection AI has warned that frontier labs and their closed models are “creating a runaway dynamic that locks everyone else out” of AI development.
Open-weight models have risen in popularity recently as companies seek cheaper alternatives to the hefty token costs frontier labs charge to access their models via API. And almost all of the powerful open-weight models being developed come from China, sparking concerns about it overtaking America’s lead in the AI arms race. The Chinese company Deepseek, for example, rocked the AI industry and the U.S. stock market when it developed a powerful open-weight model that could operate more efficiently and at a lower cost than American frontier models. While Meta, France-based Mistral and other Western companies have developed some powerful open-weight models, Laskin has said “there’s a DeepSeek-shaped hole in the U.S.,” and that American companies are seeking a homegrown alternative that offers the control, affordability and customization that comes with an open-weight model.
Some have argued that open-weight models are dangerous, as they could be used for cyberattacks or other nefarious activities. These concerns have motivated the U.S. government to pressure Anthropic and OpenAI to restrict their most powerful models, but those efforts aren’t foolproof. After all, it was an unreleased OpenAI agent that hacked into Hugging Face, which then used a Chinese open-source model to counteract the attack. In the wake of the incident, Hugging Face CEO Clément Delangue, Nvidia CEO Jensen Huang and dozens of other tech leaders advocated for government policies that would facilitate the development of more open-weight models in the U.S. Reflection AI has positioned itself as the vanguard of that movement.
But how does Reflection AI plan to make money with an open-weight model that is free to download and customize? The company believes large enterprises and governments want the ability to control where their data is processed, how much they are paying in computing costs and how the model can be adapted to their specific needs. By making its models widely available, Reflection believes it will generate a large user base that would be willing to pay for other services, like deployment support, software tools, model customization and AI infrastructure build-outs.
What Products Has Reflection AI Built?
During its brief tenure, Reflection AI has evolved from building software development agents to large open-weight models that rival the world’s top open models.
Asimov
Reflection AI’s first product was Asimov, a code-comprehension agent introduced in July 2025. Instead of focusing on code generation, Asimov was designed to help engineering teams understand existing software. In addition to learning the codebase, it also ingests architecture documentation, chat histories and GitHub threads, and it stores tribal knowledge as engineers interact with it.
Asimov can help developers trace how different components interact, understand the reasoning behind architectural decisions and answer questions about unfamiliar systems before making changes. The product reflects the company’s belief that stronger software comprehension is an important step toward more capable autonomous coding agents.
Beam
In October 2026, Reflection introduced its first open-weight model, Beam, which is focused on coding and agentic performance. The company claims that Beam’s reasoning performance matches that of GLM-5.2, a powerful open-weight model developed by Chinese company Z.ai, but requires three to four times less computing power for inference.
Beam has 501 billion total parameters, but it activates only 23 billion per token. Its sparse mixture-of-experts architecture routes each token through a subset of specialized neural network components, or “experts,” rather than activating every parameter, greatly reducing the computational requirements for inference.
As an open-weight model, Beam can be downloaded and deployed on an organization’s own infrastructure, allowing organizations to customize the model for their own purposes. It also gives them greater control over their data, as they no longer need to send sensitive information to a third-party AI lab.
Reflection AI says it pretrained Beam on 23.8 trillion tokens drawn from the web, public sources and proprietary licensed datasets. It then used large-scale reinforcement learning to improve the model's reasoning and agentic capabilities.
Enterprise and Sovereign Solutions
Reflection AI is also developing solutions for enterprise organizations and governments, working in collaboration with these organizations to develop open-source software, turnkey AI factories, and custom AI solutions tailored to their specific needs. The company recently partnered with Shinsegae Group to build a sovereign AI cloud for South Korea, for example.
Reflection AI’s open models are also used by the U.S. Defense Department in classified environments, and the Energy Department is using the company’s models to accelerate scientific research through its Genesis Mission. Laskin told Axios that open-weight models were appealing for this work because scientists need to be able to analyze, understand and customize the model they are working with.
How Does Reflection AI Get Its Computing Power?
Building large AI models requires tremendous amounts of computing power. Reflection used more than 6,000 Nvidia GB300 NVL72 GPUs to pretrain Beam, for example, and the subsequent four weeks of reinforcement training took 10,500 GPUs.
Reflection AI has secured its computing resources through partnerships with several top tech companies. In June 2026, it agreed to pay SpaceX $150 million per month for three years to access Nvidia GB300 GPUs and supporting hardware at SpaceX’s Colossus 2 data center near Memphis. The deal was reported to be worth up to $6.3 billion. The following month, Reflection signed a $1 billion deal to access Nvidia’s latest chips from AI infrastructure company Nebius.
How Much Is Reflection AI Worth, and Who Backs It?
Reflection AI has seen a jaw-dropping level of investment in just a couple years, and investor interest shows no signs of fading anytime soon.
Founded in 2024, the startup emerged from stealth in March 2025 with approximately $130 million in funding and a reported valuation of $545 million. The funding announcement included a $25 million seed round led by Sequoia Capital and a $105 million Series A round led by CRV and Lightspeed Venture Partners.
In October 2025, the company raised $2 billion in funding at an $8 billion valuation. That round was led by Nvidia with participation from Sequoia, Lightspeed, Disruptive, DST, B Capital, Citigroup, former Google CEO Eric Schmidt and 1789 Capital.
Just five months later, Reflection AI reportedly sought to raise $2.5 billion at a $25 billion valuation — this time with participation from JP Morgan. When a CNBC journalist later asked Laskin to confirm if his company raised “as much as $2 billion” at a $25 billion valuation, Laskin confirmed that was accurate.
As of October 2026, Nvidia, one of Reflection’s top investors, is reportedly considering either growing its equity stake, acquiring the company outright or pursuing an “acqui-hire,” in which Nvidia would license the company’s technology and hire away its top employees instead of acquiring the company.
Frequently Asked Questions
Who founded Reflection AI?
Reflection AI was founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou. Laskin previously led Gemini reward modeling, while Antonoglou was an early DeepMind researcher and co-creator of AlphaGo.
Is Reflection AI open-source?
Reflection AI positions itself as an open-weights laboratory rather than strictly open-source. While it publicly releases weights for models like Beam, it relies on custom commercial and revenue-sharing licenses rather than traditional open-source terms.
Who are the main investors in Reflection AI?
The main investors in Reflection AI include Nvidia, Sequoia Capital, Lightspeed Venture Partners, and CRV, which led its primary funding rounds. Other notable backers include Citigroup, DST Global, B Capital and former Google CEO Eric Schmidt.
