Nvidia Wants to Buy the 'DeepSeek of the West': Inside the Talks for Reflection AI

On October 10, the Financial Times reported that Nvidia is in talks to deepen its investment in Reflection AI, the open-weight AI startup founded by former DeepMind researchers, or to buy it outright. The talks are at an early stage, an agreement could arrive in the coming weeks or the discussions could still fall apart, and neither company would comment. But the shape of the deal the FT describes matters more than whether it closes this month: one option on the table is a so-called acqui-hire, where Nvidia hires the startup's staff and licenses its technology rather than pursuing a full acquisition, potentially avoiding a lengthy regulatory review.

Nvidia is already a major financial backer and strategic investor in the company, having put roughly $800 million into Reflection, according to the FT. In other words, the chip giant would not be discovering a stranger. It would be buying more of something it already owns a large piece of, five days after that something finally showed the world its first model.

The model that started the clock

Reflection unveiled Beam on October 5, 2026, its first frontier open-weight AI model, and the timing of the acquisition talks five days later is almost certainly not a coincidence. Beam is a text-only sparse mixture-of-experts model with 501 billion total parameters but only 23 billion active per token, giving it what Reflection describes as the serving cost of a much smaller model while retaining frontier-scale capacity. The company says it pretrained the model on 23.8 trillion tokens, with a context window of one million tokens.

The company claims Beam matches Z.ai's GLM-5.2 on advanced reasoning and coding benchmarks while using roughly a quarter to a third of the inference compute, and reports 80.9 on SWE-Bench Verified and 80.1 on Terminal Bench v2.1. In its own benchmark tables, Moonshot AI's Kimi K3 still scores ahead on most tests, which is the comparison Reflection is inviting as it positions Beam as a Western alternative to the powerful Chinese open models. Those are company-reported numbers that await independent verification.

Why investors call it the DeepSeek of the West

Reflection was founded in 2024 by Misha Laskin and Ioannis Antonoglou, both formerly of Google DeepMind, to build powerful AI models that anyone can download and use rather than keeping them behind an API. Its money story has moved at the speed of the AI boom. In October 2025, Reflection raised about $2 billion in a round led by Nvidia, which put in roughly $800 million at an $8 billion valuation, according to reports. By March 2026, the company was in talks to raise $2.5 billion at a pre-money valuation of $25 billion, more than triple in about six months. Chief executive Laskin told CNBC in April that the startup was raising fresh capital at that $25 billion pre-money figure.

The valuation rests on a bet that open-weight models are about to become strategic infrastructure. Some investors have nicknamed Reflection the DeepSeek of the West, and the nickname is doing analytical work: just as DeepSeek proved that open, efficient models can reset the frontier cost structure, Reflection is selling the idea that Western labs can compete with lower-cost Chinese models such as DeepSeek, Z.ai, and Moonshot AI on their own terms. The Trump administration has expressed interest in fostering US-based open-weight models that can compete with Chinese alternatives, according to the FT, and the White House's AI Action Plan, released in July 2025, explicitly praised open-weight models and called for the federal government to encourage their development.

Nvidia sells the picks and shovels, funds the prospectors, and now may simply buy the claim.

The acqui-hire as a deal structure

The most interesting part of the FT report is not that Nvidia wants Reflection. It is the form the deal might take. An acqui-hire, hiring the team and licensing the technology instead of buying the company, has become a standard template in AI dealmaking precisely because it reduces antitrust exposure. Regulators have been circling the AI ecosystem's tendency toward consolidation, and a straight acquisition of a startup carrying a $25 billion price tag would invite a long, uncertain review. A talent-and-license transfer does the substantive work of an acquisition while wearing different legal clothing.

Nvidia has used a similar playbook before, in a large transaction involving Groq, the AI chip startup, which observers noted followed the hire-and-license structure. The pattern is now familiar across the industry: the biggest labs and chipmakers compete for AI talent and model technology, and the acqui-hire lets them move fast without the merger-review clock. If the Reflection talks end in this form, it will confirm the structure as the default deal technology of the AI boom, not a one-off maneuver.

What Nvidia is really buying

From a business perspective, there are three layers to Nvidia's interest. The first is straightforward demand economics. Reflection trains on Nvidia hardware, the company says its models run on Nvidia GPUs, and every open-weight model that succeeds in the wild expands the market for the chips it runs on. Nvidia has a long habit of investing in its own customers, and Reflection fits the template: vendor capital that comes back as purchase orders.

The second layer is strategic. Nvidia maintains its own open-weight model family, Nemotron, but observers have noted that it lacks a frontier-scale open model of its own. Owning or absorbing a lab that just released a 501-billion-parameter open model would fill that gap instantly, giving Nvidia a flagship open model to pair with its hardware and software stack. It would also give Nvidia a Western contender in the open-model race just as the category is getting a political tailwind.

The third layer is the one that matters most to everyone else. Open-weight models are the escape valve from platform lock-in: if the weights are public, customers are not tied to any single vendor's API. For Nvidia, that makes a frontier open model a perfect strategic asset. It keeps the value in the hardware layer, which Nvidia dominates, rather than the API layer, where it is merely one competitor among many. An open ecosystem runs on Nvidia chips either way. A closed one might not.

What to watch next

The FT's sources were clear that the talks could still collapse, and that is worth taking seriously. Deals at this altitude die over valuation: Reflection was last valued at $25 billion, and a buyer or strategic investor paying a premium on top of that number is betting that open-weight models become the standard substrate for enterprise AI rather than a niche for hobbyists and researchers.

If Nvidia closes this deal, in whatever form, it will be the biggest confirmation yet of a thesis that has been building all year: in the AI economy, the scarcest asset is not compute or data. It is the small number of teams that know how to turn compute into frontier models, and the licenses to what they build. Nvidia already sells them the chips. It may now decide to own the teams too.