Nvidia Weighs Buying Beam Maker Reflection AI in Deal Talks

Nvidia, already an $800 million Reflection AI investor, is in early talks on a bigger stake or takeover as the startup's Beam model rivals China's.

@sharedot

Nvidia Weighs Buying Beam Maker Reflection AI in Deal Talks

By @sharedot · · 8 pages

  • AI Frontier
  • Nvidia
  • Reflection AI
  • Open Weight Models
  • Beam

Nvidia, already an $800 million Reflection AI investor, is in early talks on a bigger stake or takeover as the startup's Beam model rivals China's.

What happened: Nvidia is in early talks over Reflection AI

Nvidia is holding early-stage discussions to either deepen its investment in Reflection AI or acquire the open-weight model startup outright, according to the Financial Times, which Bloomberg and Reuters followed up on.

Options on the table include a full takeover, an acqui-hire in which Nvidia hires the staff and licenses the technology, a chip supply agreement, or simply a larger equity position.

Nvidia already committed $800 million to Reflection a year ago as part of a $2 billion raise, and nothing has been finalized — Reflection declined to comment, Reuters could not verify the talks, and reports warn negotiations could still collapse even though an agreement could come within weeks.

Why it's surprising: a $25B price tag one year in

The speed of Reflection's rise is the headline-grabbing part.

tikr.com notes a hypothetical $25 billion purchase is only about 20% of Nvidia's roughly $127 billion in free cash flow over the past four quarters, leaving Nvidia with roughly $31.6 billion of its $56.6 billion cash pile at July 26 if it paid that price.

A buyout a year after the first check signals how fast the open-weight tier has shifted.

The technical reveal: Beam's specs and compute bill

The deal talk follows Reflection's October 5 release of Beam, its first open-weight model.

According to Startup Fortune, Beam is a 501-billion-parameter mixture-of-experts model with 23 billion active parameters, trained on 23.8 trillion tokens with a 1 million token context window, with full weights and a technical report due under an Apache 2.0 license later this month.

Reflection says Beam matches z.ai's GLM-5.2 on reasoning benchmarks while using three to four times less inference compute, though the company's own efficiency and benchmark claims have not been independently verified.

The training bill was steep: Startup Fortune reports a four-week reinforcement learning run across 10,500 Nvidia GB300 GPUs, rented through a $6.3 billion multi-year SpaceX Colossus 2 deal costing roughly $150 million a month through 2029, plus a billion-dollar-plus Nebius compute pact.

The stakes: the open-weight race and the flywheel

A tie-up slots into a furious Nvidia deal streak: the world's most valuable company agreed in September to buy Hugging Face for $13 billion and struck an estimated $20 billion licensing arrangement with Groq, whose technology it licensed while hiring the CEO in December per TIKR.

Startup Fortune frames the pattern bluntly — Nvidia keeps writing checks into companies that burn its own compute, so invested dollars generate demand that lands back on Nvidia's income statement — and argues that Reflection, unlike Perplexity (which Nvidia is separately discussing backing at above $30 billion per the same report), builds the foundation model layer Nvidia does not directly control.

Market Briefs reports the deal streak is partly about nurturing a US-led open ecosystem.

The geopolitical layer: keeping open weights American

Reflection was founded in March 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou with the explicit pitch of building frontier-grade open-weight models that keep the intellectual property onshore in the US rather than ceding the lane to Chinese labs like DeepSeek.

Startup Fortune recalls that DeepSeek's R1 release in early 2025 rattled Washington because a cheap, competitive Chinese open model exposed how little moat US labs had in the open tier.

Market Briefs and The Business Times note Chinese open-weight models are far cheaper to run than US offerings from OpenAI and Anthropic and are not dramatically behind in capability, which is why Jensen Huang — a vocal open-weight proponent — and Washington both back domestic options, with President Trump's July 2025 AI action plan stating America needs leading open models founded on American values.

What comes next: the structure is the tell

No terms have been reached, and the three reported structures differ sharply in meaning.

BigGo Finance reports an acqui-hire — hiring staff and licensing technology — could sidestep a drawn-out regulatory review, while TIKR notes that playbook echoes the Groq deal and could avoid some merger-review requirements without eliminating antitrust scrutiny; Startup Fortune adds the FTC's 2024 scrutiny of Microsoft's Inflection AI team hire is well understood in Silicon Valley.

A full buyout would be the bigger signal, deepening Nvidia's model business alongside its Nemotron family.

BigGo says an agreement could come within weeks, but both outlets caution negotiations can still fall apart or shrink into something smaller — though Startup Fortune argues even the surprising outcome would be Nvidia not at least doubling down on its stake.

Sources

  1. briefs.co › Nvidia eyes closer ties with Reflection AI as open-weight race heats up
  2. startupfortune.com › Nvidia is in talks to buy the open-source AI lab it just backed
  3. businesstimes.com.sg › Nvidia weighs deeper ties with open-model developer Reflection AI
  4. finance.biggo.com › Nvidia Weighs Deeper Stake or Full Takeover of AI Startup Reflection
  5. tikr.com › Nvidia Weighs Buying Reflection AI, the $25 Billion Open-Weight Model Startup It Already Backs

More on AI Frontier