The deal: from investor to potential owner

Nvidia is in talks to increase its investment in or fully acquire AI startup Reflection AI, according to Cailian Press, ifeng Finance, Eastmoney and others. Sina Finance, citing sources, reported that Nvidia has already invested $800 million and that the current talks value the startup at roughly $25 billion. Neither party has confirmed the talks publicly.

Deal highlights (as reported)

· Target: Reflection AI, an open-weights model company

· Invested to date: ~$800 million

· Current valuation under discussion: ~$25 billion

· Structure: additional investment or full acquisition, still under evaluation

· Team: ex-Google Gemini core members; focus on coding agents

Why Reflection AI

Reflection AI was founded by core members of Google's Gemini team and focuses on coding agents. Its commitment to open weights — delivering near-frontier capability at far lower cost than closed flagships — has earned it the nickname "the American DeepSeek". Its latest open-weights model reportedly packs about 501 billion parameters and approaches leading closed models on several benchmarks. InfoQ wrote that the company, "acquisition-rumored almost immediately after its debut", is betting on coding agents as the most certain commercial scenario today.

Tracking from investor communities suggests Nvidia is not the only party circling open-weights companies — scarce assets are being systematically re-rated.

The chip giant's full-stack anxiety

For Nvidia, Reflection AI would be more than another model asset. Open-weights models form a moat beyond CUDA: they bind "buying GPUs" to "using frontier models", hedging against top closed-model vendors building their own silicon and bypassing CUDA. Tellingly, the same week saw Amazon reportedly in ~$7 billion talks to buy AI software firm Decart — vertical integration across chips, models and software is becoming a consensus move among giants.

The NineZenith view

The boundaries between compute, models and applications are dissolving fast. NineZenith's Tiansuan platform follows the same logic: unifying compute scheduling with model serving, matching inference resources to workloads automatically. When chipmakers start buying model companies, the enterprise-side capability of joint compute-model orchestration only becomes scarcer — and more valuable at the negotiating table.

(Compiled from public reports by Cailian Press, Sina Finance, Wallstreetcn, InfoQ, ifeng Finance, PingWest, etc.)