🎯 Top 3 Things to Know
1. Anthropic signed a $1.8 billion, seven-year compute deal with Akamai, the largest contract in Akamai's history. The number matters less than what it signals about scarcity. Dario Amodei said Anthropic is "working as quickly as possible" to secure compute after 80x growth in annualized usage in Q1, and Akamai is the third capacity backstop after Google Cloud and SpaceX in six months. A CDN company is now an AI cloud, by demand more than design. The market is shifting from a handful of hyperscalers to anyone with stranded GPU capacity willing to sign a seven-year take-or-pay. Akamai's stock closed up 27 percent, its largest single-day move in 22 years. Worth watching whether other edge operators follow into the AI cloud category, and whether Anthropic's forward commitments outpace its enterprise revenue ramp. Bloomberg · TNW
2. The White House is studying an AI security executive order modeled on FDA drug approval, the first serious break from the administration's hands-off posture. NEC director Kevin Hassett told reporters the draft would create a road map for proving frontier models safe before public release, comparing the structure to how the FDA evaluates drugs. The catalyst is Anthropic's Mythos disclosure last month, which surfaced thousands of zero-day vulnerabilities across major operating systems and browsers in weeks of internal testing. The current draft stops short of mandatory pre-release approval, but the framing matters. After eighteen months of "remove barriers" rhetoric, the executive branch is now openly discussing pre-clearance for cyber-capable models. Worth watching what triggers the gating threshold and how it interacts with the EU AI Act's August 2 obligations. Bloomberg · Federal News Network
3. Cloudflare cut 1,100 employees, roughly 20 percent of staff, while reporting record revenue and reframing the move as a shift to an "AI-first operating model." The contradiction is the story. Q1 revenue grew 34 percent year-over-year to $639.8 million, the largest quarter in the company's history, alongside an explanation for why a fifth of headcount is no longer needed. CEO Matthew Prince attributed the cuts to a 600 percent jump in internal AI usage over three months, concentrated in support and operational roles. Severance is unusually generous: full base pay and US healthcare through year-end. This is the first major US tech company to frame layoffs not as a downturn response but as a permanent capacity reset from internal AI deployment, with strong financials underneath. Worth watching whether peers with similar internal AI curves announce comparable resets in the next two earnings cycles. TechCrunch · CNBC
🚀 Frontier Models & Features
GPT-5.5 Instant became the default ChatGPT model for all users on May 9, replacing GPT-5.3 Instant in the consumer product and API (chat-latest). Positioned on faster, more concise responses with stronger personalization across past chats, files, and connected Gmail. No new headline benchmarks; a maturation pass on the April 23 base model.
OpenAI
🔬 Research Worth Reading
LongSeeker: Elastic Context Orchestration for Long-Horizon Search Agents (Tongyi Lab, Alibaba). arXiv 2605.05191
- TL;DR: Treats the ReAct context window as a resource the agent actively curates, not a passive scrollback. The agent compresses and evicts observations between rounds based on a learned relevance signal, instead of letting traces and tool calls accumulate until the window thrashes.
- Stat: Outperforms ReAct baselines across BrowseComp, BrowseComp-ZH, xbench, and GAIA, with the largest gains on tasks past 50 reasoning steps where standard agents have already paged out the original goal.
- Apply it: If a long-horizon agent loses the plot after extended tool use, the fix is rarely a bigger context window. Add an explicit eviction policy, log what gets evicted, and validate that the original task spec survives to the final step.
SciResearcher: Scaling Deep Research Agents for Frontier Scientific Reasoning (Zheng, Wang et al.). arXiv 2605.01489
- TL;DR: A fully automated pipeline that synthesizes scientific tasks from arXiv evidence, then trains an agent to do information acquisition, tool-integrated reasoning, and multi-step computation against them. Frames frontier science as a data problem, not a prompting problem.
- Stat: Generates training tasks across 50+ scientific subfields without per-domain expert annotation.
- Apply it: When building research-style agents in narrow technical domains, stop hand-curating eval sets. Borrow SciResearcher's loop: harvest primary sources, generate verifiable tasks from them, and use the same loop to grade outputs.
🏢 Enterprise in the Wild
Akamai pivots into AI cloud capacity. The Anthropic deal contributes an expected $20-25 million in Q4 2026 revenue, ramping through 2033. The May 8 analyst briefing repositions edge inference as the primary growth driver over CDN. The Boston Globe
Healthcare ambient documentation crosses from pilot to default in large Epic deployments. Multi-system aggregate data this week shows a 68 percent reduction in documentation errors and 33 percent reduction in perceived workload across deployed clinical teams. Adoption follows the EHR trajectory: slow for two years, near-mandatory inside eighteen months. ACL Digital
🛠️ Tooling & Ecosystem
OpenAI Agents SDK gained native sandbox execution and a model-native harness. Long-running file and tool agents can now execute in a managed sandbox without bolt-on infrastructure, and the SDK includes a default harness rather than leaving teams to roll one. Closes a gap that pushed many shops to LangGraph or custom orchestration. OpenAI
Microsoft Agent 365 May update added DLP controls for grounding data and registry sync. Identity, security, and compliance for enterprise agents inside the existing Microsoft management plane. Aimed at organizations whose blocker is governance, not capability. Microsoft
⚖️ Policy & Regulation
EU AI Act August 2 deadline now under three months out. General-purpose obligations have been in force since August 2025; the full Act applies on August 2, 2026, including high-risk system requirements. Enforcement guidance from the European AI Office is still being finalized, with the latest tranche expected this month. European Commission
📌 Watch List
- Compute scarcity: third major Anthropic capacity backstop in six months. Forward commitments are the dominant signal of build-cycle size.
- AI-driven workforce resets: Cloudflare (record revenue + 20 percent cut) is the first clean version of a story peers have been hinting at.
- FDA-style model gating: a US AI security EO with pre-clearance language would be a step-change from the Dec 2025 framework.
- Long-horizon agent context: LongSeeker joins a growing line treating context as managed resource, not fixed window.
- Edge as AI cloud: CDN and edge operators have stranded GPU capacity. The boundary between "edge" and "AI cloud" is collapsing.