Benjamin Fanjoy

US AI start-up Reflection AI is reportedly preparing to launch its first open-weight model. Before the model has even been released, the company, founded around two and a half years ago, has already secured more than US$7 billion (about HK$54.6 billion) in computing capacity.

Axios reported on October 4, citing sources, that the model is mainly targeting China’s DeepSeek and Qwen. Its initial performance is expected to lag behind the most advanced closed-source systems in the US. Reflection has not yet announced a firm launch date or benchmark results.

Where the money and computing capacity are coming from

Nvidia’s US$800 million investment formed part of Reflection’s US$2 billion funding round last October. Nvidia led the round, which valued the company at US$8 billion (about HK$62.4 billion). Bloomberg reported that Reflection is now seeking fresh funding at a valuation of US$25 billion (about HK$195 billion).

There are three arrangements relating to computing capacity:

  • SpaceX: Under a deal signed in June, after an initial ramp-up period, Reflection will pay SpaceX’s AI division US$150 million (about HK$1.17 billion) a month from July 1 this year through the end of 2029 for access to Nvidia GB300 chips at the Colossus 2 data centre in Memphis. The total value could reach about US$6.3 billion (about HK$49.1 billion), equivalent to an annual fee of about US$1.8 billion (about HK$14 billion). The contract allows either party to terminate it by giving 90 days’ notice after the first three months.
  • Nebius: Under a July agreement, Nebius will sell Reflection more than US$1 billion (about HK$7.8 billion) worth of computing capacity through 2029. Nebius is listed on Nasdaq (NBIS.US).
  • Shinsegae: South Korean retail group Shinsegae signed a memorandum of understanding with Reflection in March to explore a joint venture for a 250-megawatt “sovereign AI factory”. Shinsegae will be responsible for land, power, permits and financing, while Reflection will handle equipment design and operations. The memorandum expresses an intention to co-operate and is not a confirmed customer contract.

Nvidia is betting on both sides

Nvidia is Reflection’s largest investor. At the same time, the GB300 chips Reflection is leasing from SpaceX are chips that SpaceX purchased from Nvidia. Nvidia is therefore both a shareholder in its customer and an indirect supplier of that customer’s computing capacity.

Analysts say the arrangement is linked to Nvidia’s growth considerations. Leading customers are developing their own chips, while major cloud companies are diversifying their procurement, meaning Nvidia’s position as the sole supplier may not last. The open-weight route could spread demand for computing capacity from a small number of major technology companies to a larger pool of medium and large enterprises and government bodies. Under this logic, Nvidia is seeking not only to sell GPUs, but also to ensure that the industry’s technology roadmap continues to develop around its CUDA software ecosystem. According to a 2025 report by Wallstreetcn, Nvidia also committed during the same period to invest up to US$100 billion in OpenAI, placing bets on both closed-source and open-source models.

What is the “AI factory” selling?

Reflection’s proposition is to combine customers’ private data, open-weight models and dedicated computing capacity into an integrated on-premises system. Compared with closed-source APIs that charge by the token, it offers three main selling points:

  1. Data does not need to be uploaded to third-party servers, making it easier for financial institutions, healthcare organisations and government bodies to meet compliance requirements.
  2. Models can be fine-tuned in depth, rather than being limited to prompt engineering and product settings.
  3. Customers can avoid the risk of price increases if API providers become more concentrated in future.

The trade-off is that customers must provide well-organised, high-quality proprietary data and computing infrastructure, while Reflection itself operates a capital-intensive business. Its annual fee of about US$1.8 billion will only be sustainable if the number of customers and their ability to pay grow in tandem.

In reality, leading closed-source models still have an edge in complex reasoning, multimodal capabilities and very long context windows. Many enterprises simply need a model that is “good enough” and may not require a fully controlled system. Reflection’s opportunity lies in markets where data security and customisation needs prevent customers from using public APIs. It remains unclear how much of total AI spending this market represents.

Three indicators to watch

  1. The technical baseline of its first model: This will determine whether it remains a peripheral option or can make it on to the procurement lists of mainstream enterprises.
  2. The number of unaffiliated customers: Shinsegae, SpaceX and Nebius are closer to strategic endorsements and resource exchanges. Over the next six months or so, a key question will be whether Reflection can build a portfolio of reference customers without relying on the halo effect of its investors.
  3. The direction of computing prices: If the cost of leasing GB300 and subsequent chips continues to rise, the unit economics of the “AI factory” will be difficult to sustain. If increased supply brings rental prices down, the model will become significantly more attractive to medium-sized enterprises.