TechLook
The AI industry just got a sharp reminder that valuations and independence don't always move together. xAI — with a reported $250 billion valuation — has been absorbed into SpaceX as a department. The Grok chatbot survives. The independent company doesn't. Meanwhile, across the court battle lines, Anthropic is locking up compute capacity with the same SpaceX whose owner is simultaneously suing OpenAI. The AI landscape is sorting itself into two piles: those who own infrastructure, and those who rent it.
xAI Is Now a SpaceX Department, Not a Lab
xAI has effectively ceased to exist as an independent entity, absorbed into SpaceX under the SpaceXAI umbrella. Grok survives; the company does not.
Here's everything you need to know:
- xAI held a reported $250 billion valuation at the time of absorption — valuation didn't buy independence
- The structural failure came down to two missed windows: agents and coding execution
- Grok remained too close to a chatbot identity while competitors like Claude Code, Cursor, and enterprise agent platforms turned models into work engines
- The 220,000 GPUs xAI accumulated became more valuable as a resource to rent than to operate under xAI's own roadmap
- Anthropic secured those GPUs — signing a deal for over 300 megawatts of capacity at SpaceX's Colossus 1 cluster — adding to the irony
The lesson isn't that xAI failed. It's that high-valuation model labs without workflow control can become departments, vendors, or GPU landlords. For founders building on top of foundation models, this is a reminder to check who owns the infrastructure layer — and what happens to your dependencies if the owning company changes direction.
Anthropic Locks Up SpaceX Compute as the Claude Code Era Gets Real
Anthropic signed a new agreement with SpaceX for more than 300 megawatts of capacity at the Colossus 1 data center — including over 220,000 NVIDIA GPUs. Claude Code's five-hour usage caps are now doubling across Pro, Max, Team, and seat-based Enterprise plans, with peak-hour restrictions removed for Pro and Max users.
Here's everything you need to know:
- The compute deal underpins Anthropic's ability to actually honor the usage it's selling — this is a physical constraint, not a policy choice
- Elon Musk posted on X that SpaceX will rent compute to "AI companies taking the right steps to ensure it is good for humanity" — a framing that lands awkwardly given his simultaneous lawsuit against Sam Altman (Anthropic's backer)
- Claude Code rate limits doubling means developers who hit walls on complex builds now have more runway before hitting a hard stop
- Opus API limits have been raised as well, expanding what's available to developers building on Claude
- The timing is also notable: Anthropic is clearing its rate limits before enterprise buyers standardize on a single toolchain
For founders and developers who've been rationing Claude Code sessions, this matters immediately. The doubled limits mean fewer interruptions on multi-step agent tasks. The SpaceX deal is a signal that compute access — not model quality alone — is becoming the competitive differentiator in the agent era.
Mira Murati Testified That Sam Altman Lied About a Safety Review
Ex-OpenAI CTO Mira Murati gave a video deposition in Elon Musk's lawsuit against OpenAI, accusing CEO Sam Altman of lying about a model's safety review, undermining her authority, and pitting executives against each other.
Here's everything you need to know:
- Murati said Altman told her OpenAI's legal team cleared a model to skip a safety review — she verified with counsel Jason Kwon that this was false
- She briefly became interim CEO during Altman's 2023 firing and said the board process put OpenAI "at risk of falling apart"
- Former board member Helen Toner also testified, reportedly saying Murati was "afraid to stick her neck out"
- The testimony includes contemporaneous notes showing internal discussions about the for-profit transition happening earlier than publicly acknowledged
- Greg Brockman's testimony included evidence that contradicts the nonprofit idealism narrative — old notes about going for-profit and wanting billions
This is significant because it's the first time a senior insider has testified in a formal legal proceeding that Altman directly misrepresented a fact to colleagues about a safety review. For founders who've held up OpenAI's structure as a model for responsible development, the picture is getting harder to square with the original narrative.
Nvidia Bets $3.2 Billion on Fiber as the Data Center Backbone Shifts
Nvidia and Corning announced a major optical fiber partnership: three new US factories, 3,000+ advanced manufacturing jobs, and Nvidia investing up to $3.2 billion in Corning through warrants.
Here's everything you need to know:
- The deal is a bet that AI data centers will move from copper cables to glass fiber — fiber moves data faster with significantly less energy
- AI data centers are hitting power constraints that are reshaping how infrastructure is built
- Nvidia's investment is structured as warrants — it's not a simple purchase, it ties Corning's future to Nvidia's infrastructure roadmap
- The manufacturing jobs signal that optical fiber for AI isn't theoretical — it's going into production now
- 3M+ advanced manufacturing jobs at stake in a sector where the US has been trying to reshore production
The broader point: the copper-to-fiber transition in data centers is not cosmetic. It's a fundamental infrastructure shift that will affect latency, energy costs, and buildout speed for anyone who rents or owns AI compute. Founders planning inference-heavy products should note that where the cables run matters as much as which chips are running.
OpenAI Makes GPT-5.5 Instant the Default ChatGPT Model
OpenAI replaced GPT-5.3 Instant with GPT-5.5 Instant as the default model for all ChatGPT users — free, Plus, and Pro — with significant accuracy improvements baked in.
Here's everything you need to know:
- 52.5% fewer hallucinations on high-stakes prompts in medicine, law, and finance
- 37.3% fewer inaccurate claims in previously flagged error conversations
- Cross-session memory now references past conversations, uploaded files, and Gmail for personalized answers
- Memory sources control lets users see exactly what context ChatGPT used — enabling verification and adjustment
- Emoji style has been toned down for more professional outputs, addressing creator complaints about overly casual results
- GPT-5.3 remains available to paid users for three months
The accuracy improvements in high-stakes domains are the relevant detail for builders. If you're using ChatGPT as part of a product workflow — especially anything touching finance, legal, or medical — the default just got meaningfully better. The memory cross-reference feature is also notable: it means ChatGPT is becoming a more persistent reasoning partner, not just a session-by-session tool.
Real-World Agent Experiments Are Producing Revenue, Not Just Demos
Three separate agent experiments surfaced today, all involving real money, real infrastructure, and real-world execution — moving past the demo phase.
Stockholm café run by an agent named "Mona" generated 44,000 SEK (~€3,800) over two weeks through coffee sales, QR-code deals, pastry sponsorships, and agent events. The agent hired baristas, managed suppliers, and negotiated brand deals.
Sam Altman told Patrick Collison that using an OpenClaw agent to handle his morning message overload was one of his biggest "magic AGI moments" — and that OpenClaw felt more magical than it sounded when described. He's exploring other everyday automations.
British mathematician Hannah Fry ran a test with an OpenClaw-built agent named "Cass," giving it real internet tasks and a credit card. Cass reported potholes to local authorities, built a mug shop, and launched outreach campaigns. The agent moved from chat into commerce, civic services, and everyday workflows.
Here's everything you need to know:
- All three involve real execution — credit cards, hiring decisions, supplier negotiations — not simulated tasks
- The Stockholm café produced real revenue over a defined period, demonstrating the economics are at least plausible
- Altman's endorsement is notable: the OpenAI CEO finding an OpenClaw agent to be a genuine "magic" moment signals where the competitive frontier is moving
- The Hannah Fry experiment demonstrates that non-technical people can deploy agents and get them to do useful work on real tasks
For founders evaluating agent infrastructure: the question is no longer whether agents can do useful things in demos. It's whether the economics hold up when they run in the real world for weeks at a time. The early data suggests the answer might be yes — at least for defined, commerce-adjacent tasks.
US Government Expands Pre-Release AI Testing Framework
Google, Microsoft, and xAI have voluntarily agreed to test their models through the Department of Commerce's CAISI (AI Safety and Standards center) before public release. The framework evaluates models for safety, security, and capabilities — and builds on earlier agreements with OpenAI and Anthropic.
Here's everything you need to know:
- The expansion is notable given the Trump administration's general preference for lighter AI regulation
- Rising military AI use appears to be the driver behind the government's shifting posture — national security concerns are overriding deregulatory impulses
- Testing is voluntary, not mandated — but the list of participants now covers the major US AI labs
- The framework is evaluating safety, security, and capabilities — a broader scope than early voluntary commitments
The signal for founders: pre-release testing is becoming table stakes for any company that wants to be seen as a responsible player in the US market. If you're building on foundation models, the companies you're building on are now formally accountable for what they release. That's not nothing.
⚡ Quick Hits
-
Snap walked away from a $400 million Perplexity deal before it reached most Snapchat users. Conversational search inside a fast-scroll social app probably looked cleaner in an earnings announcement than inside the actual product. Snap still grew daily active users 5% to 483 million — not a crisis story, but a lesson in what AI distribution actually costs.
-
Google TPUs are struggling to find external buyers. Nebius says demand is still 99% Nvidia GPUs. CoreWeave puts resources where customers want them. Lambda's CFO said they "bleeds green." The only serious external buyer is Anthropic, wrapped inside a reported $40 billion Google investment commitment.
-
Crypto platforms are trading fake AI equity. Ventuals and PreStocks let traders bet on implied valuations for Anthropic, OpenAI, and SpaceX — without owning real shares. Combined volume exceeds $1.1 billion. Anthropic hit a fake $1.6 trillion implied valuation on these platforms. One leverage wipeout could drag private market trust into crypto's furnace.
-
UK launching sovereign AI hardware plan in June, with Technology Secretary Liz Kendall framing compute independence as both economic and security priority.
-
DeepMind took a minority stake in Fenris Creations to use EVE Online as an AI research sandbox — studying reasoning over long timelines, memory retention, and learning on a 23-year-old MMO running on a single server.
-
OpenClaw added Google Meet support — agents can now recognize Meet links, attempt to join calls automatically, and participate through the calendar and messaging context.
-
Claude managed agents now have a "Dreams" feature — lets Claude reflect on past sessions to spot long-term patterns like recurring mistakes, shared user preferences, or stale assumptions, producing a cleaned-up memory store to improve performance.
-
Nvidia is backing Span's residential AI compute plan — mounting mini data centers (XFRA units) on homes and small businesses, pulling unused grid capacity through smart panels. 8,000 units can match a 100-megawatt data center at one-fifth the cost and 6x faster deployment.
-
AMA pushing for medical AI guardrails after evidence that a fictional disease ("bixonimania") research spread into AI systems, and deepfakes can impersonate doctors.
-
PayPal planning ~20% staff cuts over 2-3 years under a new CEO pursuing an AI-linked turnaround.
Techlook — AI & tech signal for founders and builders.