May 7 Roundup: Anthropic buys compute at hyperscale, Google rewires search around human context, OpenAI sharpens the agent race, and Washington keeps building the AI rulebook

1. Anthropic locks in SpaceX compute and turns capacity into a product feature

Anthropic announced a partnership with SpaceX that gives it access to all of the compute capacity at SpaceX’s Colossus 1 data center in Memphis. According to Anthropic, the agreement adds “more than 300 megawatts of new capacity (over 220,000 NVIDIA GPUs) within the month,” and the company says the new supply is large enough to immediately raise Claude Code and Claude API limits.

“We’ve signed an agreement with SpaceX to use all of the compute capacity at their Colossus 1 data center. This gives us access to more than 300 megawatts of new capacity (over 220,000 NVIDIA GPUs) within the month.” — Anthropic

Anthropic paired the infrastructure announcement with three customer-facing changes: doubled five-hour Claude Code rate limits for Pro, Max, Team, and seat-based Enterprise plans; removal of peak-hours limit reduction for Pro and Max; and materially higher Claude Opus API limits. CNBC added useful context here: Anthropic says the deal will directly improve capacity for Claude Pro and Claude Max subscribers, and it comes after the company publicly admitted that demand had strained reliability during peak periods.

There is also a stranger, more forward-looking piece to this deal. Anthropic said it has “expressed interest” in partnering with SpaceX on “multiple gigawatts of orbital AI compute capacity.” That is not an operational roadmap yet, but it tells you how these labs are thinking: not just renting cloud capacity, but imagining compute as strategic infrastructure on the scale of telecom, logistics, or energy.

Sources: Anthropic announcement, CNBC

2. Google pushes multimodal retrieval deeper into the enterprise stack

Google’s latest Gemini API update is less flashy than a new model launch, but it may be more useful for actual systems builders. The company expanded File Search so retrieval-augmented generation workflows can now process text and images together, use custom metadata filters, and cite specific pages from source documents.

“You can now build retrieval-augmented generation (RAG) systems with multimodal data and custom metadata. We’re also introducing page citations to improve grounding and transparency.” — Google

The three additions matter for different reasons. Multimodal indexing means an enterprise knowledge base can search visual and textual artifacts together rather than treating screenshots, diagrams, scanned PDFs, and design references as second-class data. Metadata filtering gives teams a more practical way to fence searches by department, document status, geography, or workflow stage. And page-level citations attack one of the biggest trust problems in enterprise AI: users need to verify where an answer came from, not just receive a plausible summary.

What I like about this update is that it is not trying to pretend hallucinations are solved. Instead, Google is narrowing the surface area where hallucinations become expensive by improving evidence retrieval and traceability. That is exactly where enterprise AI needs to go if it wants to move from “helpful assistant” to operationally trusted system.

Practice areas: Systems Architecture, Agentic AI

Source: Google blog

3. Google Search keeps borrowing from humans because AI summaries still need help

Google is also changing the front-end experience of AI Search and AI Overviews by adding “preview of perspectives” from sources like Reddit, forums, and social media, plus clearer context around who said what. The company is also surfacing subscription-linked news sources more prominently inside AI responses.

“AI responses will now include a preview of perspectives from public online discussions, social media, and other firsthand sources.” — Google, via its search update

This is a fascinating admission wrapped as a product upgrade. Users increasingly append “Reddit” to queries because they want lived experience, not optimized content farms. Google is effectively productizing that behavior inside AI search. The Verge framed this as Google trying to make it easier to find information from sources users “know and trust,” while TechCrunch noted the underlying tension: if AI Overviews are supposed to answer questions, but also increasingly act as source directories, they are drifting back toward something closer to classic search.

That tension is healthy. AI summaries are useful, but they are brittle when they flatten contested, experiential, or niche subjects into a single answer. Pulling in human discussion threads may not eliminate errors, but it creates a more honest interface for ambiguity.

Practice areas: Digital Marketing, Systems Architecture

Sources: TechCrunch, The Verge

4. OpenAI is making the default AI experience feel cleaner — while old AGI questions come back

OpenAI’s newest cycle is really two stories. First, GPT-5.5 and GPT-5.5 Instant continue the company’s push toward more capable, tool-using agents that can plan, research, code, and operate software with less hand-holding. OpenAI describes GPT-5.5 as a model that “understands what you’re trying to do faster and can carry more of the work itself,” highlighting gains in coding, computer use, and knowledge work.

“Instead of carefully managing every step, you can give GPT‑5.5 a messy, multi-part task and trust it to plan, use tools, check its work, navigate through ambiguity, and keep going.” — OpenAI

Second, The Verge surfaced language from OpenAI’s 2019 contract with Microsoft, revealed in the Musk v. Altman trial exhibits, defining AGI as “a highly autonomous system that outperforms humans at most economically valuable work.” That definition matters because it ties one of AI’s biggest philosophical questions directly to contract structure, ownership, and control. It is a reminder that AGI is not only a scientific milestone; it is a governance and commercial trigger.

Meanwhile, OpenAI’s May 5 update on GPT-5.5 Instant shows the company also understands something simpler: users value tone, clarity, and reliability. The default ChatGPT model is being positioned as smarter, more concise, and more accurate for mainstream use. That sounds incremental, but it is probably the right move. In enterprise adoption, reducing noise is often more valuable than adding yet another headline capability.

Practice areas: Agentic AI, Systems Architecture

Sources: OpenAI GPT-5.5, OpenAI GPT-5.5 Instant, The Verge

5. Anthropic’s “dreaming” preview shows where agent design is heading

Anthropic also rolled out a research-preview feature called “dreaming” for Claude managed agents. The idea is simple and important: give agents a structured way to review prior sessions, identify repeated mistakes, notice patterns, and improve over time.

“Dreaming” will allow Claude to review previous sessions to “find patterns and help agents self-improve.” — Anthropic, via The Verge

This kind of reflective loop is one of the missing pieces in most production AI systems. Today, many agents can execute tasks, but they do not build durable process memory unless teams bolt on custom infrastructure. If Anthropic can make session review, team preference learning, and error pattern recognition manageable at the platform layer, that could make long-horizon agents more useful and less brittle.

There is also a bigger design lesson here. The most capable agents will not just act; they will observe themselves acting. That raises obvious governance questions — what gets remembered, how preferences are stored, and how systems avoid reinforcing bad habits — but it also points to a more realistic future for enterprise copilots: not one-shot answer engines, but software coworkers with bounded memory and iterative learning.

Practice areas: Agentic AI, Security

Source: The Verge

6. The U.S. AI Action Plan shows policy is now an operating constraint, not background noise

On the policy side, the AI Action Plan at AI.gov lays out a very direct federal posture. Its three pillars are to accelerate AI innovation, build American AI infrastructure, and lead in international AI diplomacy and security. The subpoints are even more revealing: streamlined permitting for semiconductor manufacturing and energy infrastructure, secure-by-design AI, an AI evaluations ecosystem, export controls, stronger incident response, and explicit efforts to counter Chinese influence in governance bodies.

“The United States is in a race to achieve global dominance in artificial intelligence.” — AI.gov

This is not the language of soft guidance. It is industrial policy, national security policy, infrastructure policy, and regulatory design collapsing into one frame. And when you line it up with the week’s commercial news — giant compute deals, enterprise agent rollouts, model access controls — the public and private strategies look increasingly aligned.

Jason Calacanis’s This Week in AI positioning and Peter Diamandis’s latest “Solve Everything” framing both reflect the same zeitgeist from a different angle: mainstream tech voices are now treating AI as a civilization-scale replatforming story, not just another software category. Diamandis and collaborators describe an “accelerating path toward a singularity,” while Washington is building the procurement, energy, and export layers around that belief.

Practice areas: AI Regulation, Security, Systems Architecture

Sources: AI.gov Action Plan, This Week in AI, Peter Diamandis podcast