Reflection, an AI startup that received an $800 million investment from NVIDIA (NVDA), is preparing to release its first open-weight model.
Sources said the initial version of Reflection's model is expected to trail the most advanced frontier models in the United States, but it could compete directly with leading Chinese open-weight models, according to an Axios report on October 4.
People familiar with the matter said the model is intelligent enough to help enterprises build proprietary AI systems at low cost. In some cases, combining a somewhat less capable open model with a company's own high-quality data can produce results comparable to the most expensive frontier AI systems.
Open-weight models: China leads, the West races to catch up
The key difference between open-weight models and closed-source models such as ChatGPT and Claude is that developers can freely download and modify the models themselves within the terms of the license. This feature has driven rapid adoption among individual developers and small and medium-sized enterprises.
However, the world's most powerful open-weight models almost all come from China. On platforms that offer a large number of AI models simultaneously, open models sometimes account for the majority of total usage by combined market share. That said, open-weight models still hold a relatively limited share of the enterprise AI market, where spending is much larger. Multiple AI company executives and analysts noted that enterprises typically pay for AI services through APIs and enterprise accounts, and the penetration rate of open models in this market remains low.
Reflection's entry, along with open-weight models expected from other Western vendors this month, will introduce new variables into this landscape. Sources said that after Reflection releases its model, other Western vendors are also expected to follow with open-weight models this month.
Reflection CEO Misha Laskin previously worked at Google. In an interview with CNBC, he acknowledged that it will still take time for the company's AI model to reach the most advanced level, comparing it to a rocket: "They're a bit like rockets. To build a large rocket, it takes time."
Ahead of the model launch, Reflection has held discussions with relevant parties in Washington and other regions to brief them on the upcoming model and the operating logic of its "AI factory." Reflection has not yet disclosed the specific release date or performance data of the model.
Analysts believe Reflection's entry marks a new front in AI industry competition. In the past, market attention focused on the battle among frontier closed-source models from OpenAI, Anthropic and Google. Now, another question is becoming increasingly important: who can build the most powerful open-weight model?
"AI factory": Reflection's core ambition
Reflection's true goal goes far beyond releasing a single model. The company's core vision is to build what it calls an "AI Factory" — allowing enterprises or government agencies to combine their proprietary data, Reflection's models and their own computing power to build highly customized, lower-cost localized AI systems.
Institutions with highly confidential data, such as hedge funds and trading firms, have shown strong interest in this model. Reflection has already begun putting this concept into practice: in March, the company signed a memorandum of understanding with South Korea's Shinsegae Group to build a 250-megawatt AI factory in South Korea, running Reflection models and equipped with NVIDIA (NVDA) chips.
It is worth noting that the "AI factory" is not an original concept from Reflection — it is precisely the core vision that NVIDIA (NVDA) CEO Jensen Huang has emphasized repeatedly for years. Huang has described a future in which AI computing hardware is combined with open models, allowing enterprises to build their own AI systems while retaining data ownership and control. Huang has long championed the open AI ecosystem and has given strong support to Reflection's transition toward a top-tier open AI system.
NVIDIA bets $800 million as its computing footprint accelerates
NVIDIA (NVDA) plays a key role in Reflection's strategic layout. NVIDIA (NVDA) has invested $800 million in Reflection, making it one of the company's most important financial backers.
To support model training and future expansion, Reflection has recently locked in computing resources intensively:
SpaceX agreement: In June, Reflection signed an agreement with SpaceX (SPCX) to pay SpaceX $150 million per month from July 1, 2026 through 2029, after the initial ramp-up period, in exchange for access to NVIDIA (NVDA) GB300 chips in the Colossus 2 data center. The agreement allows either party to terminate the contract with 90 days' notice after the first three months.
Nebius agreement: In July, cloud computing company Nebius (NBIS) agreed to sell more than $1 billion in computing capacity to Reflection, with the contract extending through 2029. On the day the agreement was announced, Nebius shares fell.
Through the SpaceX agreement, Reflection also secured a rental channel for NVIDIA (NVDA) hardware, further deepening its ties with NVIDIA (NVDA).
Comments