Reflection AI is a U.S. artificial-intelligence company founded in March 2024 by Misha Laskin and Ioannis Antonoglou, both former Google DeepMind researchers. The company originally focused on autonomous coding agents and later repositioned itself as an open-weight frontier-model developer, pitching itself simultaneously as an open alternative to closed labs such as OpenAI and Anthropic and as a Western counterpart to Chinese open-model developers including DeepSeek and Qwen (Source: techcrunch.com).
| Field | Value |
|---|---|
| Type | AI research company (frontier-model developer) |
| Country | United States |
| Founded | March 2024 |
| Founders | Misha Laskin (CEO), Ioannis Antonoglou |
| Positioning | Open-weight ("open intelligence") frontier models; Western alternative to Chinese open models |
| Lead backers | Nvidia, Sequoia, Lightspeed, DST, GIC, B Capital, CRV, Eric Schmidt, Eric Yuan, Citi |
Founders and origins
Misha Laskin led reward modeling for Google DeepMind's Gemini project; Ioannis Antonoglou co-created AlphaGo, the system that defeated a human world champion at Go in 2016. The founders' pitch centers on the claim that frontier-scale models can be built outside the established large labs by recruiting the right research talent. The company launched in March 2024 and emerged from stealth in March 2025 with roughly $130 million in funding at an approximately $545 million valuation (Source: turingpost.com; Source: techcrunch.com).
Funding and valuation
In October 2025 Reflection AI raised $2 billion at an $8 billion valuation — about a 15x increase over its $545 million valuation seven months earlier — in a round backed by Nvidia, Sequoia, Lightspeed, DST, GIC, B Capital, CRV, Eric Schmidt, Zoom founder Eric Yuan, and Citi, among others. Subsequent reporting indicated the company was pursuing a further raise of roughly $2.5 billion at a valuation near $25 billion (Source: techcrunch.com; Source: turingpost.com).
Open-weight strategy and business model
Reflection AI describes its approach as "open intelligence." CEO Misha Laskin has said the company will release model weights for public use while keeping its training datasets and full training pipelines proprietary — an access-oriented definition of openness similar to Meta's Llama and Mistral rather than fully open-source development. The company has said it built a large-scale LLM and reinforcement-learning platform capable of training Mixture-of-Experts (MoE) models at frontier scale, first applied to autonomous coding and then extended toward general agentic reasoning (Source: techcrunch.com).
The stated commercial model is to let researchers use the models freely while drawing revenue from large enterprises building products on the models and from governments developing sovereign AI systems. Laskin has framed the open-weight push in national-competitiveness terms, arguing that if the United States does not field competitive open models, "the global standard of intelligence will be built by someone else." The launch drew public endorsements from White House AI and crypto adviser David Sacks and from Hugging Face CEO Clément Delangue as a development for American open-source AI (Source: techcrunch.com). As of mid-2026 the company had not yet released its first model, which it described as initially text-based with multimodal capabilities planned later.
Compute agreements
Nebius agreement
On July 14, 2026, Reflection AI signed an agreement worth more than $1 billion with Nebius for computing capacity through 2029, including access to Nvidia GB300 chips, following its earlier SpaceX deal (Source: reuters.com; bloomberg.com).
SpaceX compute agreement
On June 22, 2026, Reflection AI agreed to rent data-center capacity from SpaceX in a contract worth up to $6.3 billion, paying $150 million a month from July 1, 2026 through 2029 for access to Nvidia GB300 chips and supporting hardware at SpaceX's Colossus 2 data center near Memphis, Tennessee; either party may exit on 90 days' notice after the first three months. The deal — Reflection's first compute contract — is smaller than SpaceX's arrangements with Anthropic ($1.25 billion per month) and Google ($920 million per month). The Colossus facility was originally built by xAI, now part of SpaceX, which has leased out its chip capacity to outside labs, the Reflection deal following earlier Colossus arrangements with Anthropic, Google, and Cursor (Source: techcrunch.com; cnbc.com). Reflection presented the agreement as evidence of demand for open models following the U.S. government's restriction of Anthropic's closed Fable and Mythos models, part of the broader export-control dispute. The Wall Street Journal, which valued the contract at roughly $6.3 billion and reported the same $150 million monthly terms and 90-day exit, noted that it was among SpaceX's first capacity deals since the company's public debut earlier in June at a $1.77 trillion valuation; SpaceX shares fell about 15% on June 22 amid reports of a planned $20 billion bond sale (Source: wsj.com).
Pentagon classified-network cohort
Reflection AI was the only open-weight developer named in the May 1, 2026 Pentagon classified-network cohort, alongside xAI, OpenAI, Google, Nvidia, Microsoft, and AWS; on May 4, 2026, coverage reframed the cohort as "seven companies" after Oracle was dropped and Reflection was re-confirmed (see Dean Ball notes). The selection committed U.S. government procurement to an open-weight option: most leading U.S. frontier labs (OpenAI, Anthropic, Google) are closed-weight, while the leading open-weight competitors are Chinese (DeepSeek, Qwen, Z.ai), leaving Reflection in the U.S. open-weight frontier slot that would otherwise be empty. The same procurement decision excluded Anthropic, the subject of Anthropic v. United States (Pentagon ban challenge); the contrast between the two outcomes recurs in coverage of the cohort, which is documented on the CDAO and DOD pages. Inclusion in the IL5/IL6/IL7 cohort implies access to CDAO-vetted compute infrastructure; public coverage of the associated compute commitments is limited. The selection anchors Procurement-Driven AI Governance and figures in AI and National Security and US-China AI Competition: Different Races, Different Metrics.
Several forecasts have been attached to the cohort selection. Based on the May 1, 2026 selection, one forecast (issued May 2026) holds that by May 2027 Reflection will ship an open-weight frontier model competitive with the best Chinese open-weight model on at least one capability benchmark, with the resolution criterion that Reflection exceeds the best DeepSeek, Qwen, or Z.ai score on the same version of a public benchmark such as GPQA, MMLU-Pro, HumanEval, or METR HCAST. Reading the Pentagon-only inclusion as an opening, a second forecast (issued May 2026) holds that Reflection will be added to a civilian-agency procurement deployment by May 2027, resolved by a GSA, DHS, Treasury, or Commerce procurement that names Reflection. Treating the procurement endorsement as a commercial signal, a third forecast (issued May 2026) holds that Reflection will raise a Series B or later at a valuation above $10B by May 2027, resolved by a public funding announcement at a valuation of at least $10B. Reading the open-weight plus procurement-customer combination forward, a fourth forecast (issued May 2026) holds that by December 31, 2027 Reflection will become the first US open-weight lab to ship a successor model with capability gains comparable to Anthropic or OpenAI generation increments, resolved by a successor open-weight release with at least a 10% gain on a capability benchmark relative to the prior generation.
Relationships
- deploys-in: Chief Digital and Artificial Intelligence Office (CDAO), DOD — Department of Defense (AI Deployer) (Pentagon classified-network cohort, May 1, 2026)
- related: Open-Source AI / Open-Weight Models (open-weight model strategy)
- related: Procurement-Driven AI Governance — the structural concept the cohort selection anchors
- related: AI Sovereignty (sovereign-AI customer base)
- related: DeepSeek, Alibaba / Qwen Team (Chinese open models it positions against)
- related: Meta AI (comparable weights-released approach)
- related: Export Controls (AI) (open-weight framing amid the Anthropic restriction)
- related: AI and National Security, US-China AI Competition: Different Races, Different Metrics — the strategic frame
- related: Clawed, Anthropic v. United States (Pentagon ban challenge) — the same procurement decision excluded Anthropic
- related: Eric Schmidt, David Sacks (backer; public endorser)
- related: xAI (origin of the Colossus data center)
Sources
- TechCrunch / Reflection AI raises $2B to be America's open frontier AI lab, challenging DeepSeek (Oct 9, 2025): techcrunch.com
- TechCrunch / SpaceX inks compute deal with Reflection AI, an open source AI lab (June 22, 2026): techcrunch.com
- The Wall Street Journal / SpaceX Strikes $6 Billion Deal With AI Startup for Data-Center Space (June 22, 2026): wsj.com
- CNBC / SpaceX signs computing power deal with open-source AI startup Reflection worth up to $6.3 billion (June 22, 2026): cnbc.com
- Turing Post / Reflection AI Explained: $20B Valuation, No Model Yet: turingpost.com