SoftBank's AI-datacenter arm files for a ~$50B IPO, with OpenAI as tenant and future shareholder

Good morning 👋 SoftBank's datacenter arm just filed for a $50B IPO. The backlog number is big, but the real story is buried in the warrants: OpenAI is set to become a shareholder in the company it rents its compute from.

In today's issue:

• 🔭 SoftBank's AI-datacenter arm files for a ~$50B IPO, with OpenAI as tenant and future shareholder
• 🧠 AWS Agent Registry catalogs every agent, tool, and skill you've built
• 🔬 Fingerprinting anonymous models, and sycophancy that survives fine-tuning
• 🏛️ OpenAI backs California's companion-chatbot safety bill
• 🛠️ Turn your agent's tools into one MCP server, not five
• 🔗 Anthropic's second $35B compute deal this month, OpenAI's tools land in Japanese city halls

🔭 THE ONE THING

🏗️ SoftBank's AI-datacenter arm files for a ~$50B IPO, with OpenAI as tenant and future shareholder

SB Energy filed its S-1 with the SEC today, applying to list on Nasdaq under ticker SBE, and is reportedly seeking $5-7B at a roughly $50B valuation. The pitch rests on a $439B contracted backlog of AI data-center and power deals, with JPMorgan, Goldman Sachs, Morgan Stanley, Citigroup, and Mizuho running the book. Look past that headline number to the $5.5B in warrants OpenAI holds against its 20-year Ohio lease: OpenAI is set to become a shareholder in the same company it pays rent to for compute capacity. Investors who flinched at Nvidia's circular stake in OpenAI should recognize the shape here. Treat the backlog as contracted capacity, not booked revenue, until pricing tells you what the market actually thinks it's worth.

🧠 MODELS & RELEASES

• 📦 AWS pushed Agent Registry to general availability: a single governed, searchable catalog for every agent, tool, and skill your org has built. If your platform team is the one fielding "wait, does an agent for this already exist" Slack messages, that's the itch it's meant to scratch. AWS blog

🔬 RESEARCH HIGHLIGHTS

• Stealth-release forensics A four-stage protocol fingerprints anonymous API models using archived launch configs, tokenizer diffs, and behavioral probes, no weight access required. Run against known releases it matched 7 of 10 exactly, and its prospective call on an unlabeled August drop landed on the right version line before the vendor confirmed it. Handy checklist if you're vetting an unlabeled endpoint before you sign a vendor contract. Paper

• Sycophancy transfers from teacher to student, and you can't filter it out Training OLMo 3 on preference pairs from different teacher models, the authors found student sycophancy tracks teacher sycophancy almost linearly, across DPO and six other preference-optimization objectives. No single bad example is the culprit either. The signal is smeared across the whole dataset, so removing a small chunk of it doesn't help. If a fine-tune inherited an agreeable base model, this is probably why. Paper

• Your coding agent's token budget is measuring the wrong thing Replaying 55 archived coding-agent trajectories, researchers found that instructions, tool outputs, and agent state compress and decay at different rates, so two agents with identical context windows can end up delivering very different context to the model. Compression tricks tuned on one task set didn't carry over to new ones. Worth measuring what actually survives in context before optimizing how much of it you allow. Paper

🏛️ POLICY & REGULATION

• OpenAI backed California SB 1119, Sen. Padilla's companion-chatbot bill: age assurance, annual child-safety risk assessments, third-party audits reported to the state AG, parental controls on memory and time limits, and a flat rule against self-harm content for minors. Binds any operator of a companion chatbot in California, not just OpenAI. Worth noting: none of it is a stretch for a company already running most of these controls voluntarily. Endorsing the bill costs little and buys goodwill ahead of whatever federal youth-safety rule comes next.

🛠️ TRY THIS

Turn your agent's tools into one reusable MCP server, not five one-off integrations

SoftBank's datacenter arm just filed to raise close to $50B on a bet that agent workloads keep multiplying. More agent surfaces means more places your tools get rebuilt from scratch, unless you stop building them per-surface now.

1. Write the tool logic as a stateless MCP server (FastMCP, stateless_http=True) instead of hardcoding it into one agent.

2. Containerize it and deploy once to a managed runtime. AWS's version is AgentCore Runtime.

3. Front it with an OAuth gateway. AWS does this with two Cognito pools, one for inbound auth, one for outbound.

4. Register the same server against every surface that needs it this week: Amazon Quick, a Claude Code plugin, your own chat UI. No rewrite required.

Prompt: Rewrite this tool function as a standalone MCP server: strip out framework-specific glue, add a stateless HTTP transport, and list the OAuth scopes a gateway would need to call it safely.

Worth a look

• MiMo-Code Xiaomi's open coding-agent CLI, 12.9k stars. Build/Plan/Compose agent modes, SQLite project memory that survives across sessions, subagents that fan out for parallel work. A real free option if you're pricing agentic coding tools. link
• PixelRAG 9.8k stars. Retrieves by rendering documents as screenshots and embedding them with a vision model, so tables and layout survive instead of getting flattened by an HTML parser. Ships with a prebuilt 8.28M-article Wikipedia index and a Claude Code plugin. link

🔗 QUICK LINKS

• OpenAI Japan's Polimill is running GPT models and Codex inside municipal offices there, turning buried admin paperwork into something staff can actually search. source
• Decathlon AWS's Chronos-2 pushed the retailer's demand forecasts up 11 to 15 accuracy points while trimming the weekly pipeline down to something one team can run. source
• Anthropic Signed a $35B cloud deal with Nvidia-backed Lambda, its second megadeal this month after $45B with Nscale. Compute scarcity, not model quality, is the Anthropic story right now. source

See you tomorrow.

Pradeep Perugu