# Reflection AI's $6.3B SpaceX Bet Reshapes Open-Weight AI

A startup without a shipped frontier model just secured $6.3B in GB300 compute through 2029, backed by Nvidia's $800M and timed perfectly for the post-Fable open-weight surge.

By Neomanex. Updated 2026-10-07.

**Reflection AI has not shipped a single frontier model — and it just signed a $150 million-per-month compute deal that puts it in the same weight class as Anthropic and Google.** The startup, backed by an $800 million Nvidia investment, will access Nvidia's latest GB300 chips at SpaceX's Colossus 2 data center beginning July 1, 2026. The contract runs through 2029, totaling up to $6.3 billion. Either party can terminate with 90 days' notice after the first three months.

The timing is everything: this deal lands two weeks after the U.S. government's Fable/Mythos export ban forced enterprises and governments to confront the risk of depending entirely on closed-model providers. The open-weight alternative just got a billionaire's budget.

## What happened

Reflection AI will pay SpaceX **$150 million per month** for immediate access to GB300 AI chips and supporting hardware at the Colossus 2 data center near Memphis, Tennessee. The contract spans July 2026 through 2029 — a potential **$6.3 billion** total commitment (TechCrunch, 2026).

This is the third major compute tenant SpaceX has landed for Colossus. Anthropic pays **$1.25 billion per month** and Google pays **$920 million per month** — all three contracts running through July 2029 (CNBC, 2026).

Reflection AI was founded in 2024 by two former Google DeepMind researchers: CEO **Misha Laskin** and **Ioannis Antonoglou**. The company was last valued at **$25 billion** and secured an **$800 million investment from Nvidia**. It has not yet released a public frontier open-weight model, but has been building relationships with the Department of Energy's Genesis Mission and broader Pentagon AI efforts (Axios, 2026).

The startup explicitly positions itself as an "American open intelligence" alternative to closed frontier labs. "Recent events highlight how important open source is to the AI ecosystem, with more nations and enterprises recognizing the risks and costs associated with exclusively depending on closed models," a Reflection spokesperson told TechCrunch.

## Why it matters

The Fable/Mythos export ban transformed open-weight AI from an ideological position into a procurement necessity. When the U.S. government forced Anthropic to restrict access to its most powerful closed models, organizations discovered a single-provider dependency is not theoretical risk — it's a supply-chain vulnerability.

Open-weight models solve this: nations and enterprises can inspect the weights, customize them, and run inference without depending on any API provider's continued goodwill. After the Fable shutdown, that capability moved from "nice to have" to "RFP requirement." Reflection and its backers are betting the lesson sticks.

The compute math is what makes this credible. $150 million per month on GB300 hardware is not a speculative experiment — it's the kind of spend that produces frontier-scale training runs. With Nvidia's direct backing and DoD relationships already in place, the open-weight ecosystem is getting a player whose compute budget rivals the closed labs.

For the [Llama ecosystem](/models/llama-4-maverick), currently the leading open-weight model with a "conditional" verdict from our directory, Reflection represents an incoming peer — not a replacement. Meta's Llama has the advantage of being battle-tested in production; Reflection has fresh hardware and a clean-sheet approach. Competition here is good: it drives both camps to ship faster.

The geopolitical dimension is equally significant. SpaceX — Elon Musk's company — is now renting AI compute to labs that compete with his own xAI/Grok. Colossus was originally built for Grok training, but as internal AI efforts stalled, SpaceX pivoted to become a neutral infrastructure provider. Anthropic, Google, Cursor, and now Reflection are paying customers. The compute-rental business is turning SpaceX into the cloud provider for the frontier AI ecosystem.

## What changes for you

| Who | What changes |
|-----|-------------|
| **Enterprises evaluating open-weight models** | A new option with frontier-scale compute backing is coming within months. Start tracking Reflection's benchmarks. |
| **Llama users** | Competition is arriving. If Reflection delivers, the Llama ecosystem gains a peer — expect faster releases from both camps. |
| **Government / regulated-industry buyers** | The open-weight procurement pipeline now has a U.S.-based player with Nvidia backing and DoD relationships. |
| **Closed-model API customers** | The bargaining position shifts. Every major open-weight advance strengthens your "we can self-host" negotiating leverage. |

## FAQ

**Who is Reflection AI?** Founded in 2024 by former DeepMind researchers Misha Laskin and Ioannis Antonoglou, Reflection is an "American open intelligence" startup valued at $25 billion and backed by an $800 million Nvidia investment. It has not yet released a public frontier model but is working with the Department of Energy and Pentagon on AI efforts. Its first model is expected within months on the GB300 hardware.

**What is Colossus 2?** Originally built by xAI (Elon Musk's AI company) to train Grok, the Colossus data center near Memphis, Tennessee now operates as a commercial compute platform under SpaceX. It houses Nvidia's latest GB300 AI chips and counts Anthropic, Google, Cursor, and Reflection among its tenants.

**How does this compare to Anthropic's deal?** Anthropic pays $1.25 billion per month — over 8x Reflection's commitment. Both contracts run through July 2029 with mutual early-termination clauses. Reflection's deal is smaller but proportionally as aggressive for a pre-revenue startup: it signals the company believes it has something worth training at frontier scale.

Canonical: https://neomanex.com/news/reflection-ai-spacex-compute-deal