AI News Roundup: SpaceX Files for $1.75T IPO, OpenAI Kills Sora, Mercor Breach Rocks AI Industry

SpaceX Files for Historic $1.75 Trillion IPO After xAI Merger

SpaceX filed confidentially with the SEC on April 1 for what would be the largest IPO in history, targeting a $1.75 trillion valuation and aiming to raise up to $80 billion. The filing comes two months after SpaceX officially acquired Elon Musk’s AI company xAI in a deal valued at $1.25 trillion — already the biggest merger of all time.

The combined entity brings together SpaceX’s Starlink network (now at 9.2 million subscribers), xAI’s Grok models, and an ambitious plan to build orbital data centers. Musk argued that “global electricity demand for AI simply cannot be met with terrestrial solutions” and that space-based compute is the path forward. The IPO is targeting a June listing.

OpenAI Shuts Down Sora — A $1M-Per-Day Money Pit

OpenAI confirmed that Sora, its AI video generation tool, will be discontinued on April 26, with API access following on September 24. The decision came after Sora’s user base collapsed from a peak of roughly one million to fewer than 500,000, while the product was burning through approximately $1 million per day in compute costs.

The fallout extends beyond OpenAI. Disney had committed $1 billion to a Sora partnership but reportedly learned about the shutdown less than an hour before the public announcement, killing the deal. CEO Sam Altman is redirecting compute toward coding tools and enterprise customers — a strategic pivot that mirrors rival Anthropic’s focus areas. Users are being urged to download their content before the cutoff dates.

Mercor Breach Exposes AI Training Secrets Across the Industry

Mercor, a $10 billion AI data startup that supplies training data to Anthropic, OpenAI, and Meta, confirmed it was the victim of a major cybersecurity breach linked to a supply-chain attack on LiteLLM, a widely used open-source library. Attackers published malicious versions of the LiteLLM PyPI package (versions 1.82.7 and 1.82.8) designed to harvest credentials.

The Lapsus$ extortion group claims to possess 4TB of stolen data, allegedly including candidate profiles, employer data, source code, API keys, and — most critically — details about how major AI labs train their models. Meta has indefinitely paused all work with Mercor while the investigation continues. The incident underscores just how vulnerable the AI supply chain remains, even at the highest levels of the industry.

Meta Unveils Four Generations of Custom MTIA AI Chips

Meta revealed an aggressive custom silicon roadmap: four new generations of its MTIA chips (300, 400, 450, and 500) to be deployed across its data centers by end of 2027, with a new chip dropping roughly every six months. The MTIA 300 is already in production handling ranking and recommendation workloads.

All four chips are built on the open-source RISC-V instruction set architecture, manufactured by TSMC, and co-developed with Broadcom. Meta chose RISC-V over Arm to avoid licensing dependencies. The MTIA 500 will deliver a 50% bandwidth boost over its predecessor, with each generation adding more HBM memory (up to 288 GB on the MTIA 400). The move is a clear signal that Big Tech is accelerating its push to reduce reliance on NVIDIA.

Noah Labs Gets FDA Breakthrough for Voice-Based Heart Failure Detection

Noah Labs, a Berlin- and Boston-based medtech company, secured the FDA’s breakthrough device designation for Vox, an AI that can detect signs of heart failure from a five-second voice recording. The algorithm was trained on more than three million voice samples and has been validated in five multicenter clinical trials with partners including Mayo Clinic and UCSF.

Vox is designed for remote patient monitoring, letting heart failure patients record brief daily voice clips that the AI analyzes for subtle vocal changes associated with worsening cardiac function. The company expects EU MDR approval by mid-2026 and says the FDA designation will expedite its US commercial timeline.

Google’s TurboQuant Slashes AI Memory Costs at ICLR 2026

Google DeepMind unveiled TurboQuant at ICLR 2026, an algorithm that dramatically reduces the memory overhead caused by the KV cache — one of the biggest bottlenecks in running large language models with long context windows. The technique combines PolarQuant vector rotation with Quantized Johnson-Lindenstrauss compression to cut memory needs by up to six times.

The breakthrough could accelerate the industry’s shift from raw parameter scaling to efficiency-first development, making million-token context windows far more practical to serve at scale.