The Omniscient Bulletin · 2026-05-07
The Omniscient Bulletin — May 7, 2026
Anthropic locks in xAI's Colossus 1 and eyes orbital compute, Trump weighs an executive order on pre-release AI vetting, and GPT-5.5 quietly catches the Mythos that Anthropic still won't ship.
Today's signal has a single undercurrent: the race for compute is reshaping everything downstream of it — which models get withheld, which regulations get written, and which enterprise tools finally start charging.
Frontier
Anthropic locks in all of SpaceX's Colossus 1 capacity, eyes orbital compute
Anthropic agreed Wednesday to take all 300+ MW of compute capacity at SpaceX's Colossus 1 in Memphis - turning Musk's operation into its largest single supplier overnight. Combined with existing Amazon (5GW), Google, Microsoft, and Nvidia deals, no frontier lab is more diversified across compute partners. The buried lead: Anthropic also expressed interest in working with SpaceX on multiple gigawatts of orbital - the first frontier lab to publicly cross the line into space-based AI infrastructure.
Policy
Trump pivots: pre-release AI vetting EO under consideration
Senior administration officials briefed Anthropic, Google, and OpenAI in late April on a draft executive order that would formalize federal pre-release vetting for frontier models - a sharp reversal of Trump's earlier laissez-faire stance. The catalyst: Anthropic's Mythos, withheld over cybersecurity concerns. CAISI's voluntary deals with five labs would become mandatory rather than optional. A White House official called the NYT report "speculation" - language the administration has used before reversing course on other policy fronts.
Research
GPT-5.5 edges Mythos on AISI expert cyber tasks
UK AISI's late-April evaluations show GPT-5.5 at 71.4% on expert offensive cyber tasks - slightly ahead of Anthropic's withheld Mythos Preview (68.6%) - though Mythos still leads on the full 32-step "TLO" attack chain, completing it end-to-end in 3 of 10 runs vs. GPT-5.5's 2 of 10. The capability convergence undermines Anthropic's central case for keeping Mythos private: if comparable performance ships on a public OpenAI endpoint, the restraint mostly serves Anthropic.
Agents
Glean stops searching, starts working: "AI coworker" launches
Glean shipped its biggest product update since founding today: Skills (codified workflows callable by name), Adaptive Reasoning (automatic model-tier selection per task), and multi-workstream delegation with approval gates - branded as a shift to "a proactive coworker that helps you execute work before you ask." If it sticks, Glean's positioning moves against Salesforce Agentforce and Microsoft Copilot, both still framed as in-app reactive assistants rather than persistent partners.
Industry
AMD beats Q1, raises Q2 guide; data center up 57%
AMD reported $10.25B in Q1 revenue Tuesday after the close - up 38% year-over-year and beating $9.89B consensus - with data center revenue up 57% to $5.8B. Q2 guide: $10.9–11.5B, midpoint above $10.52B consensus. Lisa Su flagged "strong and increasing confidence" in tens of billions of DC AI revenue next year. Shares ripped 16% Wednesday. Nvidia (May 20) still owns the AI chip narrative, but AMD is the only credible second-source story hyperscalers can finance against.
Agents
OpenAI Workspace Agents flips to credit billing
OpenAI's Workspace Agents - the cross-app successor to Custom GPTs that pipes into Slack, Salesforce, Notion, Microsoft, Google Drive, and Atlassian Rovo - exited free tier Tuesday into credit-based pricing. The transition is the first real monetization gauge for enterprise agent usage at scale. Whether revenue materializes will tell us more about agent ROI than any 2026 keynote has. Watch the next earnings cycle for the read.