SpaceX Files for Record $1.75 Trillion IPO Following xAI Merger
In what could become the largest initial public offering in history, SpaceX confidentially filed its draft registration statement with the SEC on April 1, targeting a staggering $1.75 trillion valuation. The company is reportedly looking to raise up to $75 billion — more than three times the size of the biggest U.S. IPO to date.
The filing comes on the heels of SpaceX's February merger with Elon Musk's xAI, which valued the combined entity at $1.25 trillion. Musk has framed the merger as a step toward building "the most ambitious, vertically-integrated innovation engine on (and off) Earth," with a particular focus on space-based data centers to meet the soaring electricity demands of AI infrastructure.
A public S-1 filing is expected in late April or May, with a June Nasdaq listing as the most likely scenario. At a $1.75 trillion valuation, SpaceX would rank above every S&P 500 company except Nvidia, Apple, Alphabet, Microsoft, and Amazon.
Anthropic's Next-Gen Claude Mythos Exposed in Security Breach
Anthropic is dealing with the fallout from a significant security breach that exposed details of its unreleased Claude Mythos model, internally codenamed Capybara. A security researcher discovered that a misconfigured data store had exposed nearly 3,000 internal files, including a detailed draft blog post describing what Anthropic calls "by far the most powerful AI model we have ever developed."
Claude Mythos is positioned above the existing Opus tier and is reportedly already in early access trials with select cybersecurity partners. The leak has drawn particular attention from the cybersecurity community, given the model's purported capabilities. According to prediction markets, Polymarket assigns roughly a 25% probability of a public announcement by April 30.
The timing is especially sensitive for Anthropic, which has strong commercial incentive to launch Mythos before its anticipated IPO. The leaked source code was reportedly cloned over 8,000 times on GitHub before being taken down.
Google Open-Sources Gemma 4 Under Apache 2.0
Google released Gemma 4 on April 2, marking a significant milestone for open-source AI. For the first time, Google's Gemma models ship under the OSI-approved Apache 2.0 license, giving developers full freedom to modify, distribute, and commercialize the models without restriction.
Gemma 4 comes in four sizes — 2B, 4B, 26B, and 31B parameters — and is purpose-built for advanced reasoning and agentic workflows. The models feature context windows up to 256K tokens, native vision and audio processing, and fluency in over 140 languages. Google describes them as delivering "unprecedented intelligence-per-parameter."
Since launching the first Gemma generation, developers have downloaded Gemma models over 400 million times, spawning more than 100,000 community-built variants. The move to Apache 2.0 puts Gemma in direct competition with Meta's Llama 4 Maverick, which offers a 10 million token context window with 400B parameters.
Google's TurboQuant Slashes LLM Memory Use by 6x
In a separate but equally impactful development, Google's TurboQuant algorithm is generating waves across the AI infrastructure world. Set to be formally presented at ICLR 2026 in Rio de Janeiro later this month, TurboQuant compresses the key-value (KV) cache of large language models to just 3 bits per value — down from the standard 16 — reducing memory footprint by at least six times with no measurable loss in accuracy.
The algorithm uses a two-stage approach: first, PolarQuant converts data vectors from Cartesian to polar coordinates, separating magnitude from direction. Then, the Quantized Johnson-Lindenstrauss method handles compression. On H100 GPUs, 4-bit TurboQuant delivers up to 8x performance gains over unquantized 32-bit keys.
The implications are enormous. As AI models balloon in size — Meta's Llama 4 Maverick runs 400B parameters, Anthropic's Mythos reportedly reaches into the trillions — memory efficiency becomes a critical bottleneck. TurboQuant could significantly reduce the cost and hardware requirements for deploying frontier models, already rattling chip stocks as investors reassess GPU demand projections.
California Draws a Line on AI Regulation with First-of-Its-Kind Executive Order
Governor Gavin Newsom signed a first-of-its-kind executive order establishing new AI procurement standards for California, directly challenging the Trump administration's push for weaker federal oversight. The order requires AI companies seeking state contracts to disclose their policies on content moderation, model bias, civil rights protections, and user safety.
Key provisions include:
Independent review authority: California's Chief Information Security Officer can overrule federal designations of AI companies as supply chain risks — effectively allowing the state to continue purchasing from vendors the White House may try to block.
AI watermarking: The California Department of Technology must develop recommendations for watermarking AI-generated images and manipulated video, a first at the state level.
The executive order positions California as a de facto national standard-setter, particularly as the federal government advances the TRUMP AMERICA AI Act and a broader National Policy Framework for Artificial Intelligence. With most major AI companies headquartered in California, the state's procurement rules are likely to influence industry practices nationwide.