AI News: May 21, 2026

1. SpaceX S-1 reveals Anthropic will pay xAI $1.25B per month for 300 MW of Colossus 1 compute

Compute. SpaceX’s S-1 filing disclosed that Anthropic has signed a roughly 36-month, $40 billion-plus compute contract with xAI worth $1.25 billion per month, with the first two months at discounted rates during ramp-up and either side able to terminate on 90 days notice. The deal allocates 300 megawatts from xAI’s Colossus 1 data center near Memphis, Tennessee, with SpaceX framing it as a way to “monetize unused compute capacity,” a notable concession given that xAI is Anthropic’s direct competitor on consumer chat. Read together with Anthropic’s separate AWS Trainium expansion and its projected first profitable quarter, the contract makes clear that compute supply, not model design, is the binding constraint Anthropic is now solving against, and pushes xAI further into a neocloud business model for capacity Grok itself isn’t filling. Source

2. OpenAI lines up a September IPO with Goldman and Morgan Stanley after Musk lawsuit dismissed

OpenAI. Sam Altman is targeting a September IPO for OpenAI, with confidential regulatory filings possibly within weeks and Goldman Sachs and Morgan Stanley leading the offering. The decision follows the dismissal of Elon Musk’s lawsuit, which had directly threatened OpenAI’s structure, leadership, and finances; with that overhang cleared, the company is moving on a timetable that would make it the largest tech IPO of the year. The same window has SpaceX preparing its own IPO filing alongside xAI’s $6.4 billion operating loss disclosure, setting up parallel megacap listings that are likely to dominate AI-finance headlines through Q3. Source

3. xAI’s $6.4B 2025 loss and Q1 2026 $7.7B capex put it on a $31B annualized AI-spend pace

xAI. SpaceX’s S-1 filing also exposed xAI’s full-year 2025 financials: a $6.4 billion operating loss on $3.2 billion in revenue, more than quadrupling the prior year’s $1.56 billion loss on $2.62 billion in revenue, with Q1 2026 AI-segment capex of $7.7 billion alone, an annualized $31 billion run-rate. Grok features reached 117 million monthly active users out of 550 million combined across X and xAI as of March, and revenue grew via AI solutions ($465M), Grok and X subscriptions ($365M), data licensing ($88M), and ads ($116M). The filing also confirmed xAI’s plan to scale Grok to “multiple trillions of parameters” and an aspiration to deploy orbital AI compute satellites by 2028 as a long-run cost play, alongside a separate filing revealing $2.8 billion in new natural-gas turbine purchases over three years against an ongoing lawsuit over the Memphis facility’s existing generators. Source

4. Meta cuts 8,000 jobs while reallocating 7,000 staff to new AI teams in a flatter org

Meta. Mark Zuckerberg told staff in an internal memo on May 20 that Meta does not expect any further widespread layoffs this year, as the company finishes a restructuring that cuts roughly 10% of headcount, or about 8,000 positions, while reallocating roughly 7,000 employees to newly formed AI teams. The change moves Meta to a flatter organizational model intended to compress decision-making cycles around AI shipping and lifts the 2026 capital expenditure cap to a maximum of $145 billion, almost entirely for AI infrastructure. The cuts coincide with the Llama 5 push and the company’s quieter image-and-video model effort, but the headline takeaway is that AI-team headcount is now growing inside Meta even as net headcount falls. Source

5. Exa raises $250M Series C at a $2.2B valuation to scale search infrastructure built for AI

Funding. AI-native search startup Exa closed a $250 million Series C at a $2.2 billion valuation led by Andreessen Horowitz, intending to fund the next generation of search models and scale infrastructure for “hundreds of thousands of searches per second” against agent workloads. Co-founder Will Bryk pitched the round on the thesis that “as trillions of agents come online over the coming years, search needs will grow thousands of times beyond the total search volume of Google,” with agents requiring comprehensiveness, freshness, and precision well beyond human-grade search. Exa says it has more than 5,000 customers including Cursor, Cognition, HubSpot, and Monday.com, and that it overtook Google on code search inside six months after underperforming there earlier. Source

6. Parallel Web Systems and Tavily ride an AI-search funding wave as Exa hits $2.2B

Funding. Alongside Exa’s round, Parag Agrawal’s Parallel Web Systems closed a $100 million Series A at a $2 billion valuation led by Sequoia, and Tavily and TinyFish are circulating new rounds in the same space, suggesting the agent-grade search layer is settling into a multi-startup market rather than a winner-take-all. The thesis across the cohort is that agent workloads need a search API with a different shape from the consumer Google product — higher precision, freshness, and structured outputs — and that LLM labs are unlikely to own this layer themselves. Amazon, LinkedIn, and Reddit are simultaneously adding their own AI search front-ends, which gives the cohort both potential acquirers and competitors. Source

7. LinkedIn rolls out AI-slop detection, claims 94% accuracy on generic content

Platform. LinkedIn published an AI-slop policy that downranks AI-generated posts judged to lack authentic perspective and limits their reach to mostly the author’s first-degree network, plus throttles on bulk-generated comments and replies. The company claims its new detection systems correctly tag generic AI-generated content “94 percent of the time” in internal tests, though false-positive rates have not been disclosed, and is leaning on its over-100-million-verified-member base to differentiate authentic accounts from automation. The move is openly in tension with Microsoft’s parallel push to embed Copilot writing assistance, which days earlier shipped a browser-side Copilot feature that demoed on LinkedIn itself, and reads as much as an admission that the feed has been overrun as it does a corrective policy. Source

8. Figma ships a multi-agent AI assistant on the collaborative canvas

Design. Figma launched an AI assistant directly on its multiplayer design canvas, with the ability to generate new designs from prompts, edit existing files via natural-language commands, automate repetitive iteration work, and run multiple agents concurrently against the same file. The models are fine-tuned for design context, and the company plans to expand the agent surface from Figma Design into other Figma products once usage patterns settle. The launch sits alongside Figma’s existing Claude Code and Codex partnerships with Anthropic and OpenAI on the coding side and arrives against Q1 2026 revenue of $333.4 million, up 46% year-over-year, undercutting the narrative that AI tooling is collapsing the design platform category. Source

9. IrisGo emerges from Andrew Ng’s AI Fund with a $2.8M seed for a learn-by-watching desktop agent

Startup. IrisGo, backed by a $2.8 million seed led by Andrew Ng’s AI Fund with NVIDIA and Google also on the cap table, launched macOS and Windows betas of a desktop companion that learns user workflows by observing and then automates them without explicit instructions, demoed on a multi-step coffee order including item selection, payment entry, and checkout. The product ships with a built-in skills library for email drafting, invoice processing, report building, and document summarization, plus a Codex-style coding assistant, and pitches on-device processing for privacy with cloud offload only on explicit user authorization and end-to-end encryption. Co-founder Jeffrey Lai, a Carnegie Mellon alumnus like Ng, connected with the AI Fund through a shared contact, and IrisGo is one of the earliest Ng-backed bets on the post-Copilot desktop-agent category. Source

10. DeepSeek stands up a Beijing “Harness” team to build a Claude Code/Codex/Cursor rival

Open source. DeepSeek is assembling a dedicated “Harness” team in Beijing to build “DeepSeek Code,” a coding agent positioned against Claude Code, OpenAI Codex, and Cursor, with engineer Deli Chen’s job listing framing the strategy as “model plus harness equals AI agent.” The openings target a product manager and developer with hands-on experience using competing platforms, and require background in agent loops, multi-agent systems, context engineering, and MCP, signalling that the harness will be built around an MCP-native toolchain rather than a proprietary protocol. The move slots DeepSeek into the increasingly crowded harness-engineering race and is the company’s first explicit step into product-tier coding agents after years of focusing on raw model releases. Source