Three threads pull through today’s news: the IPO calendar is colliding with the compute bill in real time, the open-weight stack is multiplying across audio, video, and unified multimodal at once, and the layer beneath agent workflows is consolidating around new runtime primitives.
The IPO Calendar Meets the Compute Bill
Two of this week’s biggest AI finance stories landed together: Anthropic locked in a $45 billion compute commitment with SpaceX while disclosing Q2 revenue tracking toward $10.9 billion, and OpenAI lined up its own September IPO window after Musk’s lawsuit was dismissed.
- Anthropic commits $45 billion to SpaceX for three years of compute — The $1.25 billion-per-month contract through May 2029 annualizes at $15 billion in compute spend, more than four times Anthropic’s most recent funding round, with a 90-day mutual exit clause that complicates the headline number.
- Anthropic’s Q2 revenue is projected to reach $10.9 billion — Disclosed to investors mid-fundraise, with growth outpacing Google and Facebook in the quarters before their IPOs. Anthropic also told investors it does not expect full-year profitability, because the SpaceX compute bill changes the math.
- OpenAI eyes September IPO after Musk lawsuit dismissed — Goldman Sachs and Morgan Stanley are retained, confidential SEC paperwork may file within weeks, and the structural conversion to a public benefit corporation can now proceed.
Two Stories About AI Pricing, Both True Today
The cheap-AI narrative and the rising-flagship-prices narrative both landed today, and procurement teams will have to decide which one to budget against.
- Cheap models are already eating OpenAI and Anthropic’s IPO story — Artificial Analysis priced ten identical evaluations: Claude at $4,811, GPT at $3,357, DeepSeek at $1,071. Enterprises are routing around frontier defaults using the “advisor model” pattern.
- Cheap AI was yesterday’s story. Prices are climbing again. — Tomasz Tunguz charts three years of LLM pricing and finds all three frontier vendors raising rates on their flagship tiers. Cuts came when cash was plentiful; increases come now that margins matter.
Open Models Multiply Across Modalities
Four open-weight releases today push the same trend forward: capable foundation models landing across audio, video, and unified multimodal at sizes deployable on a single GPU.
- Stability AI ships Stable Audio 3.0 with open weights and six-minute tracks — Three of four variants ship as open weights with commercial licensing and indemnification, the first time the Stable Audio family can credibly compete with Suno and ElevenLabs on length, ownership, and legal exposure.
- ByteDance open-sources Lance, a 3B unified model for image and video — Six tasks folded into one transformer trained from scratch on 128 A100 GPUs, deployable on a 40GB GPU.
- Meta AI’s WavFlow generates synchronized audio in raw waveform space — Flow matching applied directly to raw waveforms, skipping the latent audio codec that most generation systems rely on.
- OpenAI model disproves Erdos geometry conjecture from 1946 — Nine credentialed mathematicians including Noga Alon and W.T. Gowers verified the proof, though DeepMind’s FunSearch and AlphaGeometry got to AI-assisted mathematics first on different problems.
The Agent Plumbing Layer Consolidates
Google shipped an agent runtime designed for hundreds of millions of agents, Chrome’s Lighthouse started auditing whether your site is ready for them, and Spotify published the funnel that should sit between offline evals and online experiments.
- Google ships Agent Executor, an open-source runtime for long-running agents — Durable execution, secure isolation, session consistency, connection recovery, and trajectory branching, plus an Agent Substrate layer over Kubernetes designed for sub-second tool calls at scale.
- Google’s Lighthouse now audits for llms.txt under “Agentic Browsing” — Chrome’s Lighthouse added a check for Jeremy Howard’s machine-readable site summary file, placing it under a new Agentic Browsing category alongside accessibility tree integrity and WebMCP.
- Spotify treats LLM evals as a funnel, not a fork — Only 12% of Spotify’s A/B tests ship a positive result. The proposed framework: evals filter candidates before the experiment, and experiment outcomes calibrate the evals afterward.
Today’s Quick Hits
- No-filter pretraining outperforms curated data at scale — An arXiv preprint argues that in the high-compute, data-scarce regime, large models benefit from including low-quality data, inverting the orthodoxy behind FineWeb, Dolma, and DCLM.
- LiteFrame cuts Video LLM inference latency 35% by shrinking the vision encoder — Targets the ViT bottleneck that token-reduction approaches leave behind, enabling 8x more video frames per fixed compute budget.
- Alibaba unveils the Zhenwu M890 to chase AI agent workloads — A new domestic Chinese AI accelerator pitched at agent-style inference rather than training, entering a market where Huawei’s Ascend 910B is already deployed at Baidu, ByteDance, and state-backed institutes.
- The agent-training loop is just five steps, whatever the framework — A pure-Python toy demonstrates that TRL, Unsloth, and PRIME-RL all reduce to the same prompt, action, environment, reward, gradient cycle.