Wednesday, May 6 on the Edge of AI: Mira Murati testifies under oath, Anthropic taps SpaceX for compute, and DeepSeek eyes a forty-five billion dollar valuation. Let's get into it.
Mira Murati, OpenAI's former chief technology officer, testified under oath that Sam Altman lied to her about model safety standards [2]. The video deposition played in the Musk v. Altman trial on Wednesday. Murati said Altman falsely claimed that OpenAI's legal department had determined a new AI model did not need to go through the company's deployment safety board. When asked directly whether Altman was telling the truth, she said no. Murati also noted that during her tenure, Altman made her work more difficult. The trial has been running for weeks now, with testimony from engineers, investors, and board members. But this is the first time a former executive has gone on record saying the CEO actively misled the safety team. The signal: the case has moved past corporate governance disputes and into direct allegations that the CEO bypassed his own safety review process. That's a harder claim to walk back. Altman has denied the broader claims throughout the proceedings. My read: if the jury buys this, the governance narrative shifts entirely.
Different beat. Anthropic, the lab behind Claude, signed a computing agreement with SpaceX to meet surging demand for its artificial intelligence software as the company scales its Claude platform and raises usage limits for paying customers [3]. The deal gives Anthropic access to additional compute resources. Bloomberg reported the agreement on Wednesday. The partnership also ties into xAI's infrastructure, since SpaceX houses Musk's AI venture. Anthropic announced higher usage limits for Claude alongside the deal. Which matters because: frontier labs are no longer building their own data centers. They're renting capacity from whoever has the chips. My read: this is a pragmatic move. Anthropic needs compute now, not in eighteen months. SpaceX has the hardware. Anthropic has the models. It's a straightforward trade. The irony of a Musk company powering a rival lab's growth is not lost on anyone watching. And it shows how quickly the compute landscape is shifting.
Now, on the open-source front. DeepSeek, the Chinese open-source upstart, could hit a forty-five billion dollar valuation from its first external investment round [1]. The lab came to prominence in early 2025 after launching a model that trained on a fraction of the compute and cost of U.S. frontier models. TechCrunch reported the valuation target on Wednesday. What this changes: the efficiency narrative that made DeepSeek famous is now being priced into actual capital. A forty-five billion dollar tag puts it in the same conversation as mid-tier U.S. labs. The round would mark the first time outside investors have put money into the lab at scale. My read: if the valuation holds, it validates the thesis that you don't need a hundred billion in compute to compete. You just need better architecture. And better timing.
Last beat. Google has released multi-token prediction drafters for its Gemma 4 open model family that speed up text generation by up to three times [4]. Three times faster. A small auxiliary model suggests several tokens at once while the main model checks them in a single pass. The angle: open models are catching up on inference speed, which is where most of the actual cost lives.
Four stories, four different angles on the same week. The trial, the deal, the valuation, and the speedup.
That is the edge for today.