The $200B Compute Pact, Five Eyes Move on Agents, Defender Window Collapses: Thursday Briefing, May 7, 2026

On May 5, Anthropic confirmed a $200B, five-year commitment to Google Cloud for compute — reportedly more than 40% of the revenue backlog Google disclosed last week. The same week, Google announced plans to invest up to $40B in Anthropic in cash and compute, deepening a relationship that already pairs with Anthropic’s expanded Google-plus-Broadcom footprint. The capital flows in both directions, and the dependency is increasingly two-sided.

That two-sided dependency turns out to be load-bearing on the income statement, too. Fortune reports that of Alphabet’s record $62.6B Q1 profit, roughly $28.7B — about half — came from a mark-to-market revaluation of its ~14% Anthropic stake, not from operating earnings. Analysts have begun raising the question explicitly: how much of headline Big Tech AI profit is operating performance versus equity-revaluation accounting on private holdings? That question is now part of how diligence reads quarterly results.

Microsoft’s 2026 Work Trend Index reports that 78% of knowledge workers now use AI agents at least weekly, up from 12% in 2024. Microsoft paired the number with an “agent-first” Copilot strategy — positioning agents as the next operating layer for work across Microsoft 365, Agent 365, and Copilot Chat. The implication for product planning is concrete: agentic surfaces stop being a roadmap item and become the default expectation for productivity software in 2026 procurement.

Anthropic answered with vertical specialization: ten preconfigured financial-sector agentstargeted at investment banks, asset managers, and insurers, automating typical workflows out of the box. Independently, Meta’s unified AI agent platform reportedly recovered hundreds of megawatts of internal infrastructure power, with engineers describing 30-minute investigations replacing 10-hour ones, the platform generating review-ready code. And on the eval side, the new ClawBench benchmark covers 153 tasks across 144 live production sites; Claude Sonnet 4.6 scores 33.3% — a more honest read on real-world agent reliability than sandboxed suites, and a useful counterweight to Stanford’s 12% → 66% jump on real computer tasks.

The cyber agencies of the US, UK, Canada, Australia, and New Zealand released their first joint guidance on agentic AI security, identifying five risk categoriesand best practices spanning the AI lifecycle. The document is the cleanest cross-governmentbaseline to date and is likely to anchor the next round of compliance frameworks — expect procurement language to start citing it within the quarter.

That baseline lands the same week researchers reported a structural collapse in the disclosure-to-exploit window: from five months in 2023 to roughly 10 hours in 2026, with frontier LLMs doing much of the offensive heavy lifting from bug discovery to working exploit. In a meaningful share of cases, the working exploit lands before a patch exists. Pair that with last week’s Mandiant 2026 M-Trends finding that 28.3% of CVEs are exploited within 24 hours of disclosure, and the operating assumption for defenders has shifted: a near-zero patch window for high-severity vulnerabilities is the new baseline.

The Pentagon moved earlier this year to sever ties with Anthropic after the company declined contract terms allowing Claude for “all lawful purposes”, including autonomous weapons and mass surveillance. Last month, a California federal judge blocked the administration’s ban — the first high-profile court intervention testing the limits of government procurement leverage against a frontier lab’s use-policy red lines.

The case sets up a precedent that will shape vendor posture for the rest of the year: how far can a buyer — even the federal government — push a model provider to drop safety carve-outs as a condition of contract? Other labs are watching, and so are general counsels writing the next round of enterprise MSAs with AI providers.

On the model side, multiple sources — including details surfaced from the Claude Code source — point to Claude Opus 4.7 for a May 2026 release at the same $5/$25 per million token pricing as Opus 4.5, holding Anthropic on a roughly quarterly cadence and on price parity. Anthropic has also moved 1M-token context windows for Opus and Sonnet 4.6 to standard pricing with no surcharge — full-book and large-codebase analysis now lands inside a normal API plan and resets pricing pressure on competitors. Prediction markets and Sam Altman commentary continue to point to a May–Julywindow for GPT-6, with long-term memory, expanded agentic capabilities, and personalization as the disclosed focus areas. Kimi K2.6 from Zhipu reportedly topped Claude, GPT-5.5, and Gemini in a recent programming challenge — further evidence that Chinese frontier models are at parity-or-better on code-specific evals.

On the research side, Google’s TurboQuant (ICLR 2026) combines PolarQuant vector rotation with Quantized Johnson–Lindenstrausscompression to cut KV-cache memory overhead — the dominant bottleneck for long-context inference. Cloudflare shipped a new LLM-serving architecture that separates input processing from output generation on different optimized systems with a custom inference engine; disaggregated prefill/decode is becoming table stakes. UPenn’s Mollifier Layers integrate classical mathematical smoothing into neural networks for inverse PDEs (TMLR / NeurIPS 2026), and UH Mānoa shipped a physics-informed ML algorithm that better preserves physical-law constraints in fluid dynamics and climate modeling. Meta capex is tracking $115–135B for 2026 — nearly double 2025 — confirming that compute supply remains the binding constraint.

Four threads to track. First, the actual Opus 4.7 drop and whether 1M-token standard pricing pushes competitors to follow. Second, Pentagon v. Anthropic procedural updates — the next ruling will materially shape AI-procurement clauses. Third, how quickly the Five Eyes agentic-AI guidance shows up in concrete procurement language and internal IAM/observability roadmaps. Fourth, independent ClawBench-style numbers for GPT-5.5, Gemini 3.1 Ultra, DeepSeek v4, and Kimi K2.6 — live-site benchmarks will increasingly drive procurement, not closed-set evals.