OpenAI Confidentally Files for IPO, SpaceX Discloses Public Filing with $6.4B xAI Burn, Karpathy Defects to Anthropic — AI News Briefing

🗞️ AI News Briefing — May 21, 2026 (18:00 CST)

Top 7 Stories

1. OpenAI Prepares to Confidentially File for IPO — “As Soon as Friday”

OpenAI is preparing to confidentially file a draft of its IPO prospectus as soon as Friday, marking what could become one of the largest public market debuts in history. CNBC confirmed the move on Wednesday, reporting that the AI company — now valued at more than $850 billion by private investors — is working with Goldman Sachs and Morgan Stanley to prepare the filing. The Wall Street Journal was first to report the confidential filing, and the news was quickly corroborated by The New York Times.

In a carefully worded statement, an OpenAI representative said: “As part of normal governance, we regularly evaluate a range of strategic options. Our focus remains on execution.” OpenAI CFO Sarah Friar told CNBC last month that it is “good hygiene” for a company of OpenAI’s size to “look and feel and act” like a public company, though she declined to comment on a specific timeline at the time. Goldman Sachs and Morgan Stanley also declined to comment on their involvement.

The timing of OpenAI’s IPO preparations is particularly notable because it comes on the same day that SpaceX — which merged with xAI earlier this year — is set to publicly disclose its own IPO prospectus. This creates an extraordinary week for AI-related public market entries, with two of the most influential AI companies simultaneously moving toward going public. OpenAI had previously been preparing to IPO as soon as the fourth quarter of this year, suggesting the timeline may have been accelerated by competitive pressure and market conditions. The confidential filing route allows OpenAI to keep its financial details private during the initial SEC review process, reducing the risk of competitive intelligence leakage.

2. SpaceX IPO Filing Reveals $6.4B xAI Burn and AI-Centric Future

In a parallel development that compounds the market’s attention, SpaceX has filed publicly for its stock listing, and the prospectus lays bare the company’s massive AI ambitions and associated costs. According to Reuters, the filing reveals significant losses and Elon Musk’s outsized control structure as the company stakes its future on AI integration. ABC News reported that the filing shows SpaceX and its sister company xAI burned through $6.4 billion last year, and the spending trajectory is far from over.

The financial details paint a picture of a company betting heavily on the convergence of aerospace and artificial intelligence. xAI’s spending has drawn scrutiny, with the company now facing a separate lawsuit over its data center generators and simultaneously purchasing $2.8 billion more in infrastructure. The prospectus also reveals that Anthropic will pay xAI $1.25 billion per month for compute — a staggering figure that underscores the compute arms race between frontier AI labs.

Goldman Sachs will hold the lead left position on the SpaceX prospectus, followed by Morgan Stanley and then Bank of America. The dual IPO filings from OpenAI and SpaceX/xAI in the same week represent an unprecedented moment in tech capital markets, essentially offering investors two distinct visions of AI’s future: one focused on general-purpose AI models and applications (OpenAI), and the other combining space technology with AI research infrastructure (SpaceX/xAI). The market’s reception of both filings will likely shape the valuation environment for AI companies for years to come.

3. Andrej Karpathy Joins Anthropic: A Major Blow to OpenAI in the Talent War

In what may be the most significant talent move in the AI industry this year, Andrej Karpathy — one of the most prominent researchers in artificial intelligence and a former core team member at OpenAI — is joining Anthropic. Karpathy announced the move in a post on X, saying he is “excited to get back into research and development” and calling the next few years at the frontier of large language models “especially formative.” Axios reported that Karpathy will be standing up his own pretraining team at Anthropic focused on using Claude to speed up pretraining research.

Karpathy’s career trajectory reads like a map of the AI industry’s evolution. He was part of OpenAI’s core team in its early days, then helped build Tesla’s Autopilot and Full Self-Driving technology before returning to OpenAI and ultimately leaving for good in 2024. Most recently, he had been working on AI in education through his startup Eureka Labs. He noted that the topic still matters deeply to him and that he plans to pick that work back up “when the time is right.”

The significance of Karpathy choosing Anthropic over a return to OpenAI cannot be overstated. It represents a clear loss for his former employer and a major validation for Anthropic’s research direction. Karpathy recently said he was “blown away” by the progress of agentic AI for coding, after dismissing agentic capabilities just months before — a reversal that speaks to the accelerating pace of progress in the field. His focus on pretraining is particularly noteworthy: the pretraining team handles the initial training of large AI models, building the strongest possible base model before fine-tuning with reinforcement learning for specific tasks. Karpathy’s bet on using AI to improve AI pretraining represents a key thesis about exponential compounding in AI research.

4. Google’s Gemini 3.5 Flash Ships with 5.5x Cost Surge and Agentic Ambitions

Google DeepMind has released Gemini 3.5 Flash, and while the model delivers impressive benchmark improvements, it arrives with a dramatic cost increase that signals a broader industry trend. An analysis by Artificial Analysis found that Gemini 3.5 Flash costs 5.5 times more to run in benchmark testing than its predecessor Gemini 3 Flash and nearly twice as much as the Pro model Gemini 3.1. Token prices alone have tripled: Google now charges $1.50 per million input tokens and $9.00 per million output tokens, up from $0.50 and $3.00 for Gemini 3 Flash.

The cost story gets more complex when you factor in token consumption. Gemini 3.5 Flash burns through so many tokens on agent-based tasks that total costs end up 75 percent higher than the pricier Gemini 3.1 Pro. On agent tasks measured by the GDPval-AA benchmark, the model requires an average of 49 turns per task — more than any other model tested, including Claude Opus 4.7 at 45 turns and GPT-5.5 at 40 turns. All those extra interaction steps drive input token consumption dramatically higher.

Despite the cost concerns, the performance improvements are genuine. Gemini 3.5 Flash scores 55 on the Artificial Analysis Intelligence Index, nine points above Gemini 3 Flash, putting it ahead of Grok 4.3 (53) and Claude Sonnet 4.6 (52). On agentic tasks, it hit an Elo score of 1,656 — a massive leap from Gemini 3 Flash’s 1,204 — and its hallucination rate dropped 31 percentage points to 61 percent, though it still trails leaders like MiMo-V2.5-Pro and Grok 4.3 at 25 percent. Google CEO Sundar Pichai has positioned the 3.5 series as built specifically for agentic work, noting the model can sustain autonomous sessions for several hours and run complex coding pipelines independently. The company claims that running on its in-house Antigravity platform makes the model twelve times faster, and Pichai estimated companies shifting 80 percent of workloads to a mix of 3.5 Flash and Pro could save over a billion dollars annually.

5. Meta Layoffs Begin as 8,000 Job Cuts Take Effect

The layoffs at Meta that were announced earlier this week have now begun taking effect, with the first wave of 8,000 job cuts hitting the company’s workforce. NPR reported that the reductions are part of Meta’s aggressive pivot toward AI, and KRON4 confirmed that the layoffs started this Wednesday. The cuts represent approximately 10% of Meta’s total workforce and are accompanied by the simultaneous transfer of 7,000 workers into AI-focused roles.

The New York Times captured the human impact of the restructuring with a piece about Meta workers creating AI-generated songs themed around their layoffs — a surreal illustration of how AI is simultaneously displacing workers and becoming the tool through which they process the experience. The internal mood at Meta has been described as tense, with the dual message of cuts and transfers creating uncertainty throughout the organization.

The scale of Meta’s restructuring is historically significant. Moving 7,000 employees into AI while cutting 8,000 others in a single restructuring event signals that Mark Zuckerberg views AI not merely as a growth initiative but as the fundamental reorientation of the company’s entire workforce. For the broader tech labor market, this sends a clear signal: AI skills are becoming the minimum requirement for job security at major technology companies. The 15,000-person internal shuffle at one of the world’s largest employers will likely ripple through the industry as displaced workers seek new positions and companies adjust their own workforce planning.

6. Trump Expected to Sign AI Cybersecurity Directive This Week

President Trump is expected to sign an executive order focused on AI cybersecurity as soon as Thursday, according to CNN and Bloomberg. The directive would represent one of the most significant federal policy actions on AI to date, potentially reshaping how AI systems are secured and deployed across government and critical infrastructure.

The timing of the directive comes at a sensitive moment for AI policy, with the ongoing Anthropic-Pentagon court battle over blacklisting, increasing concerns about AI-powered cyber attacks, and growing bipartisan interest in AI governance. Bloomberg reported that the order is being characterized as a “cybersecurity directive,” suggesting it may focus specifically on securing AI systems against adversarial attacks rather than broader AI regulation.

The expected order follows a pattern of increasing executive branch attention to AI. The Pentagon recently established a new task force specifically to bring powerful AI tools to America’s most sensitive networks, as reported by Politico. The task force is racing to integrate AI capabilities into classified systems, raising questions about the balance between speed and security in AI deployment for national security applications. The convergence of these developments — an AI cybersecurity executive order, a Pentagon AI task force, and ongoing court battles over AI company access to government contracts — suggests that AI policy is moving from the realm of academic discussion into active government action.

7. DeepSeek Building “Deepseek Code” to Challenge Claude Code and Codex

Chinese AI company Deepseek is entering the AI coding agent race with a new project called “Deepseek Code.” According to reporting from The Decoder, Deepseek is setting up a new “Harness” team in Beijing to build its own code agent from scratch. The move places Deepseek in direct competition with Anthropic’s Claude Code, OpenAI’s Codex, and Cursor — three of the dominant players in the AI-assisted development space.

The job postings, shared on X by Deepseek’s Deli Chen, reveal that the company is hiring a product manager and a developer who are heavy users of tools like Claude Code, Cursor, Codex, or GitHub Copilot. The “Harness” designation refers to everything beyond the model itself: tool use, planning, and memory systems that turn a language model into a functional agent. Deepseek’s core thesis is captured in the equation: model plus harness equals AI agent.

Candidates for the roles are expected to have experience with agent loops, MCP (Model Context Protocol), multi-agent systems, and context engineering. Notably, Deepseek explicitly expects experience with “vibe coding” — the practice of using natural language to guide AI-assisted development — suggesting the company is targeting the same developer experience that has made Claude Code and Codex popular. The product manager will own the roadmap, run feedback analysis, and build a community, while working closely with the model research team. This represents Deepseek’s most direct push into the developer tools market, expanding beyond its reputation as a research-focused AI lab known for open-weight models.

📊 Trend Watch

🔭 What to Watch

• OpenAI’s confidential IPO filing timeline — The company could file as early as Friday; the prospectus details (revenue, profitability timeline, risk factors) will reveal how Wall Street should value AI model companies.
• Meta’s workforce transition execution — Moving 7,000 employees into AI roles while cutting 8,000 others is an unprecedented organizational experiment; attrition rates and retraining success will be key indicators.
• Gemini 3.5 Flash real-world cost impact — Early benchmark data shows 5.5x cost increases; developer adoption will hinge on whether the agentic performance gains justify the token consumption.
• Trump’s AI cybersecurity directive scope — Thursday’s expected signing will clarify whether the order focuses narrowly on securing AI systems or takes a broader regulatory approach.
• Anthropic’s pretraining team under Karpathy — Whether Karpathy’s thesis about AI-accelerated pretraining yields faster model improvement cycles could reshape the competitive dynamics of frontier model development.