This week's lens: agents became infrastructure. Microsoft and AWS shipped the governance and toolkit layers that turn agent platforms into managed enterprise products. NIST/CAISI's first independent eval of DeepSeek V4 Pro shows the open-weight gap is real on held-out benchmarks (roughly 8 months behind US frontier) while cost efficiency holds up. The Pentagon's IL6/IL7 vendor list and NVIDIA's equity-backed IREN deal complete the picture: frontier AI is now a national-security supply chain and an AI-factory build-out priced in gigawatts. The main report's structure and the new Healthcare AI section are untouched.
The report now includes a standalone AI in Healthcare section that treats healthcare as value chains, reimbursement paths, incentives, liability, regulation, and data flows, not just "models for clinicians." The section itself was not edited in this run; mentioned here so readers know where to find it.
CAISI's independent evaluation used non-public ("uncontaminated") benchmarks across cyber, software engineering, sciences, abstract reasoning, and math. Public-style evals look close to the US frontier; CAISI's held-out evals do not. V4 Pro lands roughly comparable to GPT-5 (released about 8 months earlier), with an IRT Elo near 800 versus GPT-5.5 around 1260. The cost-efficiency story holds: V4 Pro was cheaper than GPT-5.4 mini on 5 of 7 benchmarks. The cleanest framing: public evals look close, held-out evals show a gap, cost efficiency is real.
Microsoft Agent 365 went GA at $15/user/month as the enterprise governance and control plane for agents (discovery of shadow agents, identity, observability, policy). AWS made the Agent Toolkit and AWS MCP Server generally available, treating agent tooling as a first-class AWS interface alongside the SDKs. Anthropic shipped Dreaming (cross-session memory consolidation in research preview), Outcomes (rubric-graded loops), and Multi-agent Orchestration on its managed-agent platform. A2A protocol crossed 150+ organizations with Google, Microsoft, and AWS integration. Net: there is now a real "agent control plane" layer above models.
Coatue's "services as software" framing reads AI as a re-pricing of the software stack from per-seat to per-output, with the cited 3.8% consumption share of traditional SaaS as the gap between narrative and commercial reality. Orb's State of AI Agent Pricing reinforces the same point from a billing-system vantage: most production agent companies are on hybrid pricing today (subscription floor plus usage and outcome components). Outcome pricing is the strategic direction, hybrid/usage pricing is current commercial reality, and subscriptions are increasingly used as a revenue floor inside hybrid models rather than as the value-capture engine.
The Department of War's May 1 release named eight vendors cleared for IL6 (Secret) and IL7 (most sensitive compartmented) deployment on GenAI.mil: OpenAI, Google, NVIDIA, Microsoft, AWS, SpaceX, Oracle, and Reflection AI. The Pentagon framed this as a diversity-of-supply move including open-source weights. Brookings put DoD's potential AI contract value at $90.7B in 2026 and 98.9% of all federal AI spending. The exclusion narrative around Anthropic is covered separately in earlier reporting; we are not re-attributing causality here.
NVIDIA and IREN announced a multi-part deal: a $3.4B five-year managed GPU cloud contract at IREN's Childress, Texas campus; a five-year warrant for NVIDIA to buy up to 30M IREN shares at $70 (about $2.1B notional); and an intent to build up to 5 GW of DSX-aligned AI factory capacity across IREN's pipeline, with the 2 GW Sweetwater campus on the Texas grid as the flagship. The structural read: NVIDIA is now using equity to anchor preferred infrastructure partners, extending DSX as the reference architecture for AI factory construction.
The Healthcare AI section in the main report was added in a previous refresh and is intentionally untouched this week. Mentioned here only so readers know the section is current as of its original publish and was reviewed, not rewritten, on May 13.
Anthropic adds SpaceX compute (more than 300MW and over 220,000 NVIDIA GPUs, per Anthropic's own framing) on top of Amazon for primary training and Bedrock distribution, Google/Broadcom for TPU capacity, and Microsoft/NVIDIA for Azure capacity. SpaceX/xAI is simultaneously circling Cursor, one of the coding distribution points where Anthropic's models are economically powerful (Cursor and GitHub Copilot together represented roughly 30% of an earlier Anthropic revenue milestone, per reporting). The result is a checks-and-balances structure: everyone gets stronger, everyone gets more exposed.
Under the Apr 27 partnership reset, Microsoft is no longer OpenAI's exclusive cloud and the IP license through 2032 is non-exclusive; OpenAI continues a (capped) revenue share to Microsoft through 2030 while Microsoft no longer pays revenue share to OpenAI. The strategic cost of cutting the tether is real: OpenAI now has to assemble compute and distribution from multiple patrons (Microsoft, Oracle, AWS, plus its own buildout) without a single dominant protector. The Incumbents section now frames OpenAI/Microsoft/Oracle/AWS as one of three coalitions under stress, alongside the Anthropic/Amazon/Google/SpaceX/Cursor diversified-but-interdependent stack and Meta's separate vertically controlled stack.
Aligned model labels and tooltips across index.html, slides.html, slides.js, and slides-data.js so the report and deck no longer contradict each other on the OpenAI / Claude / Gemini lineup (GPT-5.4 + GPT-5.5 default Instant, Claude Opus 4.7, Gemini 3.1 Pro). The inference-cost collapse window is reconciled to 36 months end-to-end (2023–2026). The Manus → Meta acquisition card was softened to a reported deal with no in-repo primary source. The OpenClaw narrative is now an explicit rise-and-fall arc (Jensen-era hype → systematic absorption by Anthropic). CapEx is anchored to ~$695B midpoint / up to $725B top end with per-company numbers consistent between report and deck. Rolling phrases like "this week", "active talks", and "doubled in 3 months" were replaced with dated snapshots. "Microsoft restarted Three Mile Island" softened to the 20-year PPA / targeted-restart framing.
Anthropic signed for the full compute capacity of Colossus 1 (300+ MW), with the deal reportedly extending to interest in space-based compute. Rival labs are now willing to rent each other's factories — compute scarcity beats ideology.
OpenAI raised $4B+ for The Deployment Company at a $10B valuation (TPG, Brookfield, Bain) with access to 2,000+ portfolio clients; Anthropic was reported as forming a parallel adoption entity with Blackstone, Hellman & Friedman, Goldman, Apollo, GIC, and Sequoia. Frontier labs are copying the Palantir FDE model and becoming deployment companies, not just model APIs.
Replaces GPT-5.3 Instant: AIME 2025 jumps 65.4 → 81.2, MMMU-Pro 69.2 → 76, with lower hallucination in law, medicine, and finance and preserved low latency. API exposed as chat-latest. The default model is compounding quietly at consumer scale.
Microsoft remains primary cloud partner but OpenAI can now serve products on any cloud; the Microsoft IP license through 2032 is non-exclusive; OpenAI revenue share to Microsoft continues through 2030 (capped) while Microsoft no longer pays revenue share to OpenAI. AI cloud exclusivity has cracked — distribution and compute are becoming multi-cloud.
Big 4 hyperscalers plus Oracle now guide to roughly ~$695B midpoint, up to $725B at the top end — about $100B above prior projections. Per-company anchors used across the report and deck (CapEx explorer in sec-7 and slide 7): Amazon ~$200B, Alphabet ~$180B, Microsoft ~$145B, Meta ~$125B, Oracle ~$50B. The infrastructure chapter got bigger, not smaller; the bottleneck is memory, power, and data-center execution.
DoD signed formal agreements with SpaceX, OpenAI, Google, NVIDIA, Reflection, Microsoft, AWS, and Oracle to deploy frontier AI on classified networks for lawful operational use. Anthropic was excluded amid an ongoing supply-chain-risk dispute. Frontier models are now national-security infrastructure — safety posture has become a go-to-market constraint.
Ten templated workflows (pitchbooks, KYC, month-end close, etc.) deployed via Claude Cowork, Claude Code plugins, and Managed Agents — with Excel, PowerPoint, Word, and Outlook add-ins inbound and a Moody's MCP app covering 600M+ public and private companies. Agents are moving from generic chat into packaged vertical workflows.
Recurring patch-notes log. Reported items are sourced from primary releases or first-party reporting; dollar figures and benchmarks reflect what was published in-window and may be revised in the next refresh. Each item is also reflected, in concise form, in the relevant section of the main report — this page is the running changelog with full sources.