This Week in One Line

OpenAI rolled out GPT-5.5 to paid ChatGPT tiers, Microsoft and OpenAI unwound Azure exclusivity to enable multi‑cloud, the Pentagon expanded Gemini to millions on GenAI.mil, and Nvidia unveiled a CPU for agent loops — agents are crossing from demos into daily workflows.

Week in Numbers

• $600B — Estimated 2026 AI capex across Microsoft, Alphabet, Amazon, and Meta; a market-level test of AI spend. ¹
• 10M — Weekly conversations handled by Meta’s business AI tools across WhatsApp, Messenger, and Instagram. ²
• 3M — Users eligible to access Gemini 3.1 Pro on the U.S. DoD’s GenAI.mil platform; 1.3M+ already active. ³
• 9x — Throughput gain Nvidia claims for Nemotron 3 Nano Omni vs. other open “omni” models at similar interactivity. ⁴
• 88 — Custom Olympus cores in Nvidia’s Vera CPU (central processing unit) for agentic workloads. ⁵
• $1.1B — Seed funding raised by Ineffable Intelligence, valuing the new lab at $5.1B. ⁶
• €500M — Schwarz Group’s structured financing to back Cohere’s tie-up with Aleph Alpha. ⁷

Top Stories

OpenAI releases GPT-5.5 for paid ChatGPT tiers

OpenAI launched GPT-5.5 across ChatGPT Plus, Pro, Business, and Enterprise, with GPT-5.5 Pro for Pro/Business/Enterprise. The model targets multi-step work (coding, research, data analysis, computer use) and reports gains over GPT-5.4 on Terminal-Bench 2.0 (82.7%), SWE-Bench Pro (58.6%), GDPval (84.9%), OSWorld-Verified (78.7%), and Tau2-bench Telecom (98.0%), while matching per-token latency. ⁸

OpenAI highlights planning, tool use, and self-checking, with internal teams citing time savings; API (application programming interface) access follows “very soon” pending additional safeguards, and safety evaluations/red-teaming framed the rollout. ⁸

Microsoft–OpenAI loosen exclusivity, enabling multi‑cloud

Microsoft and OpenAI reworked their partnership so OpenAI can sell on any cloud; Microsoft keeps a license to OpenAI tech through 2032, and its OpenAI stake is valued at over $135B. ⁹

Azure remains the default “ship first” path, revenue‑share terms shift, and AGI (artificial general intelligence) triggers were removed — VentureBeat frames this as the end of exclusivity, opening competition on AWS (Amazon Web Services) and Google Cloud. ^{10 11}

Pentagon adds Gemini 3.1 Pro to GenAI.mil for up to 3M users

The U.S. Defense Department’s GenAI.mil made Google Cloud’s Gemini 3.1 Pro broadly available, with up to 3 million eligible users and more than 1.3 million already active; officials say the tools operate in Impact Level 5 environments and have enabled over 100,000 AI agents to be built. ^{3 12}

The expansion underscores faster government adoption of frontier models alongside ongoing debates over acceptable use inside vendors. ³

Nvidia unveils Vera CPU for agentic AI bottlenecks

Nvidia detailed Vera, a data‑center CPU (central processing unit) built for agentic loops and RL (reinforcement learning) post‑training: 88 custom Olympus cores, up to 1.2 TB/s LPDDR5X memory bandwidth, and up to 1.5× higher agentic sandbox performance vs. competing x86 platforms. ⁵

A Vera rack integrates up to 256 CPUs and supports 22.5K+ concurrent CPU environments, with systems expected from OEMs (original equipment manufacturers) in H2 2026. ¹³

Nvidia releases Nemotron 3 Nano Omni, a unified multimodal model

Nemotron 3 Nano Omni consolidates video, audio, images, and text in a single open model, which Nvidia says delivers up to 9× higher throughput than other open “omni” models at comparable interactivity, with distribution via Hugging Face, OpenRouter, and NVIDIA NIM. ⁴

Under the hood: a 30B‑A3B hybrid Mixture of Experts (MoE) backbone with integrated vision and audio encoders; checkpoints arrive in BF16/FP8/FP4 and an accompanying paper details accuracy gains across modalities. ¹⁴

Anthropic pilots agent‑to‑agent commerce with real money

In a one‑week internal marketplace, AI agents representing 69 employees negotiated 186 deals totaling just over $4,000, handling listings, offers, and counteroffers without human sign‑off during negotiations. Anthropic found stronger models achieved better outcomes, while weaker‑model users often didn’t notice the gap. ¹⁵

Coverage notes the pilot’s small scale and highlights “agent quality” disparities as a governance issue to plan for. ¹⁶

OpenAI opens government and AWS routes (FedRAMP + Bedrock)

OpenAI said ChatGPT Enterprise and the OpenAI API Platform are authorized at Federal Risk and Authorization Management Program (FedRAMP) Moderate, enabling U.S. federal agencies to use OpenAI in a compliant environment (including access to GPT‑5.5, per agency decisions). ¹⁷

OpenAI also brought its models — including GPT‑5.5 — to Amazon Bedrock and introduced Bedrock Managed Agents powered by OpenAI in limited preview, designed to fit existing AWS identity, security, and procurement flows; it also published updated operating principles. ^{18 19}

Cohere to merge with Aleph Alpha, backed by €500M from Schwarz

Cohere agreed to take over Germany’s Aleph Alpha with €500 million in structured financing from Schwarz Group and likely use of its sovereign cloud, STACKIT — a push to serve regulated buyers seeking data control and European language coverage. ⁷

Reports say the combined company would keep the Cohere name, target a valuation around $20B, and allocate roughly 90%/10% ownership to Cohere/Aleph Alpha, pending approvals. ²⁰

Aidoc raises $150M to scale clinical AI imaging

Aidoc secured a $150 million Series E led by Goldman Sachs Alternatives. The company cites 31 FDA clearances and deployments in nearly 200 U.S. health systems and 1,600+ hospitals globally, funding a broader clinical AI model and additional regulatory work. ²¹

Competition among incumbents and startups is active; the raise signals enterprise-grade traction in imaging workflows like ED triage and abdominal findings. ²¹

Hyperscaler earnings put AI capex — and payoffs — under the microscope

With results clustered on Apr 29, options markets priced ≥4% moves as investors weighed an estimated $600B in 2026 AI capex across Microsoft, Alphabet, Amazon, and Meta (a group worth $10T and ~17% of the S&P 500). ¹

Alphabet reported Google Cloud revenue of $20B (up 63% year over year) with a $462B backlog nearly doubling quarter-over-quarter; AWS posted $37.6B and Microsoft Cloud $54.5B. Microsoft guided capex above $40B in Q4 and cited capacity constraints through 2026 as two‑thirds of spend goes to GPUs and CPUs (graphics/central processing units). ²²

Trend Analysis

A clear shift toward “agentic” systems emerged across model, infrastructure, and deployment news. OpenAI’s GPT‑5.5 is framed for multi‑step tasks, Nvidia’s Vera CPU targets CPU‑bound agent loops, and Nemotron 3 Nano Omni collapses perception handoffs by unifying vision, audio, and text — together pointing to faster, more reliable software that plans, uses tools, and checks its own work. ^{8 5 4}

Distribution and governance widened at the same time. Microsoft and OpenAI ended Azure exclusivity so OpenAI can sell on any cloud, while OpenAI added a compliant government path (FedRAMP Moderate) and an enterprise route inside Amazon Bedrock. For buyers, that means more leverage on latency, data residency, and procurement fit — and fewer reasons to block pilots on platform constraints. ^{9 17 18}

Adoption signals concentrated in real workflows: Meta’s business AI handled about 10 million weekly customer conversations, Aidoc raised $150M to expand clinical imaging AI with dozens of FDA clearances, and hyperscaler earnings emphasized cloud growth and capacity planning ($600B in capex, large backlogs). The center of gravity is shifting from model headlines to measurable throughput, guardrails, and ROI. ^{2 21 22}

Public-sector moves add a second anchor: the Pentagon’s GenAI.mil scaled Gemini across up to 3 million users and the department signed agreements to bring AI tools onto classified networks, while small-scale agent-to-agent commerce at Anthropic highlighted outcome gaps across model tiers. The takeaway: capability is rising, but usage policies and model choice now visibly shape results. ^{3 23 15}

Watch Points

• “GPT‑5.5 API” — OpenAI says API access is coming “very soon” with additional safeguards; watch for availability and usage policies. ⁸
• “Cohere–Aleph Alpha approval” — Regulatory review and STACKIT deployments will indicate how fast the sovereign AI pitch lands in the EU public sector. ⁷
• “Vera shipping window” — OEM timelines for Nvidia’s Vera (H2 2026) will show how quickly agentic CPU capacity reaches data centers. ¹³

Open Source Spotlight

• Qwen Code — Terminal-native coding agent that plugs into local or hosted models; ideal for CLI-first devs who want provider flexibility and privacy. QwenLM/qwen-code
• promptfoo — Reproducible prompt/agent/Retrieval‑Augmented Generation (RAG) test runner with CI integration; helps teams baseline quality, cost, and latency. promptfoo/promptfoo
• Skyvern — Browser automation with large language models and vision; good for scaling login–navigate–extract flows beyond brittle scripts. Skyvern-AI/skyvern
• TensorRT‑LLM — Nvidia’s high‑performance inference stack for large models; consolidates kernel and scheduling optimizations behind a Python/C++ API. NVIDIA/TensorRT-LLM
• vllm‑mlx — OpenAI/Anthropic‑compatible local server for Apple Silicon (MLX backend) with continuous batching and multimodal support. waybarrios/vllm-mlx

What Can I Try?

1. Put GPT‑5.5 on a real task: have it plan and complete a weekly report or spreadsheet end‑to‑end, then compare quality, time, and tokens vs. your current flow. ⁸
2. Register for Google × Kaggle’s free AI Agents Intensive (June 15–19) and pick a work‑relevant capstone to ship in a week. ²⁴
3. Pilot OpenAI on AWS Bedrock with IT: run the same prompt flow you use today and evaluate latency, security fit, and billing. ¹⁸
4. Test Adobe’s Firefly AI Assistant (public beta): turn one product photo into a set of platform‑ready social assets and time the lift. ²⁵
5. Verify model lineage before deployment: fingerprint one model using Cisco’s Model Provenance Kit and archive the result in your AI inventory. ²⁶