Reflection AI, the $25 Billion 'DeepSeek of the West,' Explained

A two-year-old startup that has not shipped a single public frontier model is now being valued at $25 billion. Reflection AI, founded in 2024 by two former Google DeepMind researchers, is raising roughly $2.5 billion at that figure — up from a $545 million valuation less than a year ago — on a wager that is as much geopolitical as it is technical: that America needs an open-weight answer to China's DeepSeek, and that Reflection will build it. Days before word of the raise firmed up, the company underscored just how serious the bet is, signing a $6.3 billion compute lease with Elon Musk's SpaceX. For a firm with a waitlisted research tool and no flagship release, the numbers are staggering. Here is what is actually going on.

The company and the thesis

Reflection was started by Misha Laskin, who led reward modeling for DeepMind's Gemini, and Ioannis Antonoglou, a co-creator of AlphaGo — the system that beat the world's top Go player in 2016. Its backers read like a who's-who of AI money: Nvidia, Sequoia, and Lightspeed. After raising $2 billion in October 2025 at an $8 billion valuation, the company is now in talks for a round that could include JPMorgan Chase through its Security and Resiliency Initiative, a vehicle aimed at firms deemed critical to national security.

The pitch is the so-called "DeepSeek of the West." US governments, banks, defense contractors, and large enterprises increasingly want frontier-capable AI they can run themselves — with open weights they can inspect and host. Many of those buyers will not use closed American labs because of sovereignty concerns, and they will not touch Chinese open-weight models such as DeepSeek or Qwen because of security and legal risk. Reflection positions itself as the third option: American, open-weight, and frontier-scale.

Laskin frames the stakes in nationalist terms. "DeepSeek and Qwen and all these models are our wake-up call because if we don't do anything about it, then effectively, the global standard of intelligence will be built by someone else," he said. "It won't be built by America." The company has tied itself to the Department of Energy's Genesis Mission and Pentagon AI programs, and its business model leans on enterprises building on its models and governments developing "sovereign AI" — systems each nation controls.

So far, the product story is thin. Reflection's code research agent, Asimov, has sat behind a waitlist since early 2026, and no public frontier model has shipped.

The SpaceX deal and the Nvidia loop

On June 22, 2026, SpaceX signed a compute lease with Reflection worth up to $6.3 billion. Reflection will pay $150 million a month starting July 1, 2026, running through the end of 2029, for access to Nvidia's top-of-the-line GB300 chips at Colossus 2, SpaceX's data center near Memphis, Tennessee. The contract carries a flexible exit clause — either side can walk with 90 days' notice after an initial three-month period.

The deal is a milestone for SpaceX as much as for Reflection. Musk's company has turned Colossus into one of the world's largest third-party compute platforms, signing recent deals with Anthropic, Google, and Cursor; outside committed revenue now reportedly exceeds $80 billion through 2029. For Reflection, the lease answers the most basic question hanging over any frontier lab — where will the compute come from — without forcing it to build its own data centers.

It also exposes the financing geometry that critics keep circling. Nvidia is both an investor in Reflection and the supplier of the GB300 chips Reflection is now renting at $150 million a month. As one outlet put it, Nvidia "sells the picks and owns part of the mine." The same circular dynamic — chipmaker invests in customer, customer spends the money on the chipmaker's hardware — has drawn scrutiny across the AI buildout, and Reflection is now a textbook case.

What to watch

The harder question is whether Reflection's "open" really means open. The company plans to release model weights publicly while keeping its datasets and full training pipelines proprietary — give the weights away, build the ecosystem, monetize the infrastructure and services around them. DeepMind engineer Susan Zhang has publicly questioned whether that approach is genuinely open source, since the training data and process stay locked up. It is a meaningful distinction: open weights let buyers run and fine-tune a model, but they do not let anyone reproduce it, and "sovereignty" built on a black box is a narrower promise than the marketing implies.

That tension sits at the heart of the $25 billion bet. The valuation prices Reflection as if the open-weight frontier lab already exists; in reality, the company has a compute contract, a roster of marquee investors, and a thesis. The watch list from here is short and concrete. Does a frontier model actually ship, and how does it benchmark against DeepSeek, Qwen, and Meta's Llama line? Do the Genesis Mission and Pentagon ties convert into signed government revenue? And can a startup burning $150 million a month on rented chips justify a valuation built almost entirely on a story about American AI sovereignty — before that story has shipped a single product?

---

Sources:

- [Reflection eyes $2.5B raise at $25B valuation — TechFundingNews](https://techfundingnews.com/reflection-ai-25-billion-valuation-nvidia-deepseek/) - [SpaceX signs computing power deal with Reflection worth up to $6.3 billion — CNBC](https://www.cnbc.com/2026/06/22/spacex-ai-colossus-data-center-reflection.html) - [Reflection raises $2B to be America's open frontier AI lab — TechCrunch](https://techcrunch.com/2025/10/09/reflection-raises-2b-to-be-americas-open-frontier-ai-lab-challenging-deepseek/) - [SpaceX lands $6.3B compute deal and Nvidia is on both sides of the trade — TechFundingNews](https://techfundingnews.com/spacex-lands-6-3b-compute-deal-with-open-source-ai-startup-reflection-and-nvidia-is-on-both-sides-of-the-trade/)