Good morning. It's Wednesday, May 6th, 2026. Here's what's happening in tech and science today.

**Story one: Anthropic locks in $200 billion with Google Cloud.** The Information reported yesterday that Anthropic has committed to spending roughly $200 billion on Google Cloud and AI chips over the next five years. That dwarfs the $25 billion deal with Amazon announced just last month, and combined with Google's own $40 billion equity commitment, we're looking at nearly $300 billion in hyperscaler backing for a single AI company. The compute starts flowing in 2027 — multiple GPU and TPU clusters coming online. Put this in perspective: OpenAI, Anthropic, and their cloud partners now account for more than half of the $2 trillion in committed cloud backlogs across AWS, Azure, and Google Cloud. The AI infrastructure arms race isn't slowing. It's accelerating.

**Story two: Starship Flight 12 takes a hit — deluge system explodes.** SpaceX had a rough weekend at Starbase. During a test of the water deluge fire suppression system at the new launch complex, what's being called an "explosive event" ripped through the deluge farm, triggering an automatic shutdown. The system is supposed to spray massive volumes of water during launch to protect the pad from 33 Raptor engines going to full thrust. A methane-oxygen explosion in a gas generator — the same one that tripped a deluge abort — pushed the Flight 12 target from May 12th back to late May, maybe mid-May at earliest. Reddit trackers are already pointing to May 15th as the new NET date. This is testing at the edge — new orbital launch mount, new flame trench, new deluge system. Explosions are part of the process, but every delay gives the competition a few more weeks.

**Story three: OpenAI drops GPT-5.5 Instant as the new default ChatGPT model.** Released Tuesday, 5-5 — because symmetry. Forty-five percent fewer hallucinations over the previous Instant model, AIME score of 81.2, and yes, OpenAI is quietly cutting back on "gratuitous emoji" in responses. New memory controls let you see which conversations informed what ChatGPT remembers about you, plus personalization settings for tone and verbosity. Available on all tiers, and hitting the API with a three-month overlap window starting today. The Instant tier has become OpenAI's volume play — the model most people actually interact with every day.

**Story four: NASA's science budget proposal faces open rebellion.** The White House's 2027 NASA budget request would slash the overall agency budget by 23% to $18.8 billion — and the science directorate takes a devastating 47% cut. The Planetary Society called it a "horrible threat to our future in space." Ted Cruz, Republican Senator from Texas, pushed back saying he doesn't want to look up at the moon and realize the Chinese beat us there. The Republican-led House subcommittee responded by advancing its own $24.4 billion plan — a clear thumbs-down to the administration's numbers. Meanwhile, a different Republican-appointed NASA administrator, Jared Isaacman, is defending deep cuts in front of Congress while the Roman Space Telescope, targeting September, is coming in ahead of its original schedule. The political dynamics here are genuinely weird.

**Story five: Harvard study — AI outperforms ER doctors on diagnosis.** Published in Science this week, a Harvard-led study tested OpenAI's o1 reasoning model against attending physicians on real-world emergency department triage cases using anonymized clinical data. The AI model hit 67.1% diagnostic accuracy — identifying the exact or very close diagnosis — compared to 50 to 55% for human doctors. The researchers stress this is narrow task performance, not wholesale replacement. But as the lead author points out, we don't yet know which tasks humans will consistently do better than AI, and which the reverse is true. The agentic AI angle: systems that can augment clinical decision-making in real time during emergency workflows could be a genuinely near-term application.

That's today's Morning Voicecast. See you tomorrow.