Google Commits $30B to SpaceX Compute for Gemini Bridge

The $920 million monthly deal secures capacity through 2029 and exposes how cloud giants now outsource to vertical competitors.

June 6, 2026 | Reading time: 9 minutes | Issue #181

Alphabet's Google has agreed to pay SpaceX $920 million per month for computing power, according to a Bloomberg report on June 5. The agreement, disclosed in a securities filing, runs from October 2026 through June 2029 and totals approximately $30 billion. It is Google's second major capacity arrangement with an AI competitor in weeks, following the company's $22.5 billion in AI infrastructure debt raised in late May.

SpaceX said in the filing that Google will pay the monthly fee for access to Nvidia chips as part of a cloud-services deal. Google Cloud later told the New York Times that the arrangement is a "short-term" agreement to ensure "bridge capacity" to meet surging customer demand for Gemini Enterprise. The framing suggests Google is treating the contract as a stopgap rather than a strategic pivot, but the scale of the commitment—$30 billion over 35 months—makes it one of the largest outsourced compute contracts in cloud history.

The deal arrives as SpaceX pursues a record $75 billion IPO, which Bloomberg reported on June 3 would value the company at roughly $350 billion. The filing explicitly earmarks proceeds for AI and launch infrastructure. For Google, the contract secures Nvidia GPU capacity without the capital expenditure of building new data centers. For SpaceX, it turns Starship-adjacent infrastructure into a recurring revenue stream before the company ever lists. The sequencing—$12.5 billion in secondary transactions in May, a $30 billion compute contract in June, and a $75 billion IPO filing—suggests SpaceX is front-loading proof points before opening its books to public-market scrutiny.

OpenAI ships Lockdown Mode for prompt-injection defense

OpenAI began rolling out Lockdown Mode on June 5, an optional security setting that limits features to reduce exposure to prompt-injection attacks. The company says the mode is not for general users but for organizations handling sensitive data that want stricter protection against data exfiltration. Enabling it disables the chatbot's ability to pull images from the internet or download files automatically, though manual uploads remain functional. The release follows a wave of public concern about jailbreaks and adversarial inputs, and signals that OpenAI is treating prompt injection as an enterprise-blocking issue rather than a research curiosity.

India and UAE deploy Cerebras supercomputer for sovereign AI

India signed an agreement with G42, the Abu Dhabi sovereign wealth fund-backed technology group, to deploy an AI supercomputer on Indian soil using Cerebras hardware, Rest of World reported on June 1. The system will comprise 64 Cerebras systems, with G42 handling operations and Cerebras providing technical support. The arrangement gives India a second path to AI compute outside its existing $45 billion in commitments to Amazon, Microsoft, and Google. India's $1.25 billion national AI program currently runs on 34,000 Nvidia processors, with a target of 100,000 by year-end.

Cameron Kerry, former acting U.S. secretary of commerce and now a Brookings fellow, told Rest of World the deal reflects "India's pragmatic approach to AI sovereignty, using the power of its scale to adapt what's available from other countries to its own needs." The pattern is not unique to India—Saudi Arabia, France, and South Korea have pursued similar non-U.S. compute partnerships—but the scale of India's population and domestic market makes its choices consequential for global supply chains.

SpaceX files for a record $75 billion IPO

SpaceX filed registration documents for a $75 billion initial public offering on June 3, Bloomberg reported. The offering, if completed at the proposed valuation, would value the company at roughly $350 billion and rank as the largest IPO in history. The filing explicitly directs proceeds toward AI and launch infrastructure, signaling that SpaceX views its AI compute business as a core revenue pillar alongside satellite and launch services. The IPO comes one day after Bloomberg revealed the $920 million monthly Google compute contract, and two weeks after SpaceX raised $12.5 billion in a secondary transaction. Crossover investors including Coatue and Tiger Global are expected to participate, according to people familiar with the matter.

Aleph Alpha builds sovereign language models for European public sector

Aleph Alpha continues to position itself as Europe's sovereign AI alternative, offering specialized language models trained and hosted entirely on European infrastructure. The German startup's website, updated in early June, emphasizes domain-specific SLLMs for legal, administrative, and scientific workloads, with explicit compliance to EU law. Unlike general-purpose frontier models, Aleph Alpha's offering targets institutions that treat data sovereignty and regulatory compliance as non-negotiable. The pitch aligns with the EU's broader push for digital sovereignty, though the company has not disclosed recent funding or revenue figures. In a landscape dominated by American and Chinese labs, Aleph Alpha's bet is that European regulation creates a captive market for compliant models.

Sarvam AI scales India's full-stack sovereign platform

Sarvam AI continues to position itself as India's domestic alternative to foreign model providers, with a homepage update in early June listing partnerships with Aadhaar, Axis Bank, Cred, IDFC, Infosys, LIC, and the Skill India program. The Bangalore-based startup offers frontier-class models on sovereign compute, pitching population-scale deployment for government and enterprise use. Unlike API resellers, Sarvam trains and hosts its own weights inside India, an architecture choice that matters for regulated sectors such as finance and identity. The company has not announced a new funding round since its $41 million Series A in early 2025, but the partnership roster suggests it is moving from pilot to procurement.

From the Lab

Researchers from Caltech, Georgia Tech, and MIT published HANDOFF on arXiv on June 4, a humanoid whole-body controller distilled from three complementary teacher models into a single mixture-of-experts student. The system operates on a compact task-space interface—sparse kinematic commands rather than dense spatial references—and runs on the Unitree G1 hardware without task-specific tuning. In experiments, HANDOFF matched state-of-the-art velocity tracking and demonstrated one of the largest robust manipulation workspaces reported for a humanoid platform.

The work matters because it decouples high-level planning from low-level control, allowing a VLM-driven agentic planner to command physical hardware through natural language with no per-task retraining. A separate paper, Code2LoRA, introduces a hypernetwork that generates repository-specific LoRA adapters for code models, eliminating inference-time token overhead when injecting repository context into code completions. Both papers point to the same direction: shrinking the gap between language-model reasoning and domain-specific execution.

Eastern Front

DeepSeek published a homepage banner on June 5 announcing the preview release of DeepSeek-V4, describing it as featuring world-top reasoning performance and improved agent capabilities. The model is available via the company's web interface, mobile app, and API. Pricing was not disclosed in the preview announcement. Zhipu AI's GLM-5 reached open-source state-of-the-art on SWE-bench Verified, matching Claude Opus 4.5, according to the company's website. Alibaba's Qwen team released Qwen3Guard, a safety guardrail model built on the Qwen3 foundation and fine-tuned for prompt and response classification across English, Chinese, and multilingual settings. Baichuan continues to push its medical vertical, offering the M3-Plus model free to healthcare partners under its "海纳百川" program. The pattern across Chinese labs is specialization: DeepSeek on reasoning, Zhipu on coding agents, Baichuan on medicine, and Qwen on multimodal safety.

The View

The compute layer is being reorganized in real time. Google, the world's second-largest cloud provider, is now buying capacity from SpaceX, a launch and satellite company that happens to own Nvidia clusters. India, the world's most populous country, is buying Cerebras hardware through a UAE intermediary rather than renting from AWS or Azure. These are not temporary distortions. They are structural signals that the supply chain for AI training and inference is fragmenting along sovereign and vertical lines.

The old model—American hyperscaler rents American GPU to global enterprise—is being replaced by a patchwork of bilateral deals, sovereign compute pools, and non-U.S. intermediaries. What holds the stack together is not geography but the model layer, where Chinese open-weight labs and American closed API providers still compete for the same enterprise customers. The integration layer—agents, security, memory, deployment orchestration—is where the next set of winners will be decided. The companies that control the interface between model and enterprise will extract the surplus, even if they do not own the silicon or the weights.

The Miss

Pope Leo XIV released his first encyclical, Magnifica Humanitas, on May 29, an 83-page letter focused on "safeguarding the human person in the time of artificial intelligence." The document warns that AI power should not be concentrated in the hands of a few private companies, calls for job protection, and demands greater oversight. Christopher Olah, Anthropic co-founder, was present at the Vatican briefing. Anthropic recently hosted more than a dozen Christian leaders at its San Francisco office. The encyclical has received limited coverage in the technology press relative to its implications for 1.4 billion Catholics, many of whom live in regions where AI labor and data annotation are concentrated.

Pull Quotes

"This is an example of India's pragmatic approach to AI sovereignty, using the power of its scale to adapt what's available from other countries to its own needs." — Cameron Kerry, Brookings Institution

"Lockdown Mode is not intended for everyone. It is designed for people and organizations that handle sensitive data and want stricter protection from data exfiltration risks related to prompt injection." — OpenAI

Out

The question is whether sovereign compute deals create durable alternatives to hyperscalers, or just add friction to a supply chain that still runs on Nvidia.

By Neo