SpaceX Locks a $6.3B Open-Source Customer Into Colossus
SpaceX signs a computing deal worth up to $6.3 billion with Reflection AI for access to Nvidia GB300 chips at Colossus 2, while Five Eyes agencies warn that frontier AI models capable of devastating cyber attacks are months away, and Nobel laureate John Jumper leaves DeepMind for Anthropic.
June 23, 2026 | Reading time: 7 minutes | Issue #192
Lead
SpaceX has signed Reflection AI to a compute contract worth as much as $6.3 billion, according to materials viewed by CNBC. The open-source AI startup will pay $150 million per month starting July 1, 2026, for access to Nvidia GB300 chips at Elon Musk's Colossus 2 data center. The deal runs through 2029, though either side can terminate with 90 days' notice after the first three months.
The agreement is the largest commercial compute transaction disclosed at the facility and signals that Colossus is becoming a multi-tenant AI infrastructure platform rather than a private training ground for xAI's Grok. Reflection, last valued at $25 billion, is building "American open intelligence" and has pitched its open-weight models to governments and enterprises wary of closed-model dependency. The contract gives it a supply of top-tier GPUs at a moment when chip access is the binding constraint for frontier labs.
For SpaceX, the deal arrives days after a record $75 billion IPO and the disclosure of more than $100 billion in cash. It also follows last week's Anthropic suspension of Fable 5 and Mythos 5 for foreign nationals, which has intensified demand for open-source alternatives. By selling capacity to Reflection, SpaceX is turning a strategic asset into a recurring revenue stream while positioning itself alongside cloud providers and sovereign compute projects.
Five Eyes Warns Frontier Cyber AI Is Months Away
Signals intelligence agencies from Australia, Canada, New Zealand, the UK, and the US issued a rare joint statement warning that frontier AI models capable of devastating cyber attacks on governments and businesses are "mere months away." The Guardian reported the statement late on Monday, framing it as a public intervention following the US Commerce Department's June 13 directive that forced Anthropic to disable Fable 5 and Mythos 5 for foreign nationals.
The agencies said AI will improve cyber defense over time but is already accelerating "the speed, scale, and sophistication of cyber threats." They called for a "whole-of-organisation and whole-of-society response" and argued cyber risk can no longer be treated as purely technical. The warning comes as Anthropic's own models are cited by national security authorities as justification for export controls — a circularity that has drawn both support and skepticism from the security community.
Open-Source Pulse
Reflection's $6.3 billion Colossus contract lands just as closed-model access becomes a policy variable. The startup has not yet released a public frontier open-weight model, but it has been building relationships with the Department of Energy's Genesis Mission and broader Pentagon AI efforts. A Reflection spokesperson told CNBC that "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."
The timing is the point. Anthropic's export ban on Fable 5 and Mythos 5, announced on June 12, created immediate demand for models that governments and companies can inspect, modify, and run locally. Reflection's contract gives it the GPU supply to train such models at scale. SpaceX, in turn, gains a high-profile customer that validates Colossus as a neutral compute platform rather than a Musk-family asset. The deal also complicates any narrative that open-source AI is limited to small models or research releases.
Policy & Power
Trump signed two executive orders aimed at accelerating quantum computing while mitigating its security risks. The Wall Street Journal reported the orders on June 22, noting the administration set a 2028 target for a "commercially relevant" quantum computer and a 2031 deadline for federal agencies to migrate to post-quantum cryptography. The first order directs a whole-of-government strategy to promote quantum computing, sensing, and networking. The second requires agencies to inventory cryptographic systems and begin shifting to NIST-approved post-quantum standards.
The quantum push is orthogonal to the immediate AI export-control fight but shares a common mechanism: Washington is treating advanced computing as a national-security domain where private-sector leadership and federal resilience must advance together. Alphabet, IBM, and IonQ publicly supported the orders, while crypto markets sold off on concerns that post-quantum migration will be slower and more expensive than the White House assumes.
Compute Watch
Upscale AI raised a $190 million Series A-1 at a $2 billion valuation, bringing total funding to $500 million in under 18 months, Fortune reported. Premji Invest led the round, with new investors including Nvidia, Salesforce Ventures, Temasek, and Seligman Ventures. Upscale is building an open-standard networking fabric for AI data centers that lets GPUs from different manufacturers communicate at full speed, challenging Nvidia's NVLink and InfiniBand lock-in.
The round follows Groq's $650 million raise, also announced June 22, led by Disruptive and Infinitum. Groq is scaling inference capacity after Nvidia hired away founder Jonathan Ross and other executives in a $20 billion "not-acqui-hire" deal last December. Together, the two financings show that investors are still betting on alternative AI silicon and networking even as Nvidia dominates the GPU layer.
India Lens
Reliance Jio is embedding AI across India's largest consumer infrastructure layer. At its annual shareholder meeting on June 19, the conglomerate announced Jio Call Agent, an AI assistant that can join phone calls, transcribe conversations, and perform tasks such as booking cabs and ordering food. The service, activated by saying "Hey Jio," is expected to launch later this year for Jio's more than 500 million users, TechCrunch reported.
Reliance also unveiled an AI-powered MyJio app and TeleFrame, a home display that uses AI agents to surface proactive recommendations. The approach mirrors Ambani's earlier framing that "India should not be a mere consumer of AI created elsewhere." Amazon is simultaneously testing Alexa+ in India with Hindi support, signaling that global platforms see voice as the critical interface for the country's non-English majority.
Europe
Google DeepMind has lost two top AI executives to rivals in a single week. Nobel laureate John Jumper, who led the AlphaFold team, announced on June 19 that he is leaving for Anthropic. His departure followed the earlier exit of Gemini co-lead Noam Shazeer, who joined OpenAI. Alphabet shares fell 5% on Monday, June 22, as investors digested the implications. DeepMind responded by announcing a $75 million research partnership with A24 to develop AI tools for filmmaking, a move that gives it a consumer-facing collaboration even as its research leadership bench thins.
The talent migration is a European problem too. DeepMind's London headquarters has been the continent's strongest concentration of frontier AI research. Losing Jumper to a US rival, and Shazeer before him, weakens the UK's claim as a counterweight to Silicon Valley in model development. A24 may help DeepMind retain cultural relevance, but it does not replace a Nobel-class scientist.
The View
Three signals this week point to the same restructuring of the AI stack. SpaceX selling Colossus capacity to Reflection shows that compute is becoming a traded commodity with geopolitical pricing power. The Five Eyes warning shows that governments now view frontier models as imminent strategic weapons, not future risks. And the migration of DeepMind stars to Anthropic and OpenAI shows that talent, not capital, is the scarcest input. The common thread is control: of chips, of models, of people. The labs that can secure all three are consolidating power; the ones that cannot are becoming suppliers to the winners.
The Miss
The arXiv paper "Randomized YaRN Improves Length Generalization for Long-Context Reasoning" (2606.23687) proposes a simple tweak to positional interpolation — randomizing the scaling factor during training — that yields better generalization to contexts longer than those seen during training. The result is intuitive but underexplored: instead of teaching a model a fixed recipe for stretching attention, randomized scaling forces it to learn a more robust mapping. For any application that processes long documents, legal contracts, or multi-turn agent traces, this kind of cheap architectural intervention may matter more than another scaling law headline.
Pull Quotes
"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." — Reflection AI, CNBC
"Frontier AI models are anticipated to exceed current industry expectations, fundamentally transforming both offensive and defensive cyber capabilities. The timeline is not years, it is months." — Five Eyes agencies, The Guardian
"India should not be a mere consumer of AI created elsewhere. It must become a creator, adopter, and a global leader in AI." — Mukesh Ambani, Reliance shareholder meeting
Reads & Links
SpaceX signs $6.3B compute deal with Reflection AI — CNBC on the Colossus 2 contract and what it says about SpaceX's AI infrastructure strategy. https://www.cnbc.com/2026/06/22/spacex-ai-colossus-data-center-reflection.html
Five Eyes warns frontier AI cyber attacks are months away — The Guardian on the rare joint statement following the Anthropic export ban. https://www.theguardian.com/technology/2026/jun/22/anthropic-claude-fable-ai-model-artificial-intelligence-national-security
Nobel laureate John Jumper leaves DeepMind for Anthropic — TechCrunch on the latest high-profile AI talent migration. https://techcrunch.com/2026/06/20/nobel-laurate-john-jumper-is-leaving-deepmind-for-rival-anthropic/
Trump signs quantum computing executive orders — The Wall Street Journal on the administration's 2028 quantum-computer target and 2031 post-quantum migration deadline. https://www.wsj.com/politics/policy/trump-seeks-to-boost-quantum-computing-with-new-executive-orders-b2200634
Upscale AI raises $190M to build open networking for AI data centers — Fortune on the Series A-1 led by Premji Invest with Nvidia and Salesforce Ventures. https://fortune.com/2026/06/22/nvidia-upscale-ai-next-ciscoand-seligman-ventures-premji/
Reliance Jio announces AI Call Agent for 500M users — TechCrunch on Ambani's plan to embed AI in Indian phone calls and homes. https://techcrunch.com/2026/06/19/billionaire-ambani-wants-ai-in-every-call-app-and-home/
Randomized YaRN paper — arXiv on a low-cost intervention for long-context generalization. https://arxiv.org/abs/2606.23687
Compute is becoming a traded commodity with geopolitical pricing power.