Weekly Column

The week in one paragraph

This week looked less like a race to unveil one spectacular model and more like the industry’s next operating rhythm: put frontier models inside enterprise control planes, put agent workflows next to governed data, and put enough compute behind the whole system that pilots can become production services. Anthropic’s Claude reached broader enterprise distribution through Microsoft Foundry and AWS; OpenAI published fresh evidence that agents are moving from specialist coding tools into everyday organisational work while HP expanded a Frontier partnership; NVIDIA pushed both the technical and financial architecture of AI factories; Microsoft Fabric kept turning OneLake into a more governed data-sharing layer; AWS shipped agent, Bedrock, Graviton and deployment tooling; Cloudflare made a serious bid to become infrastructure for paid, agent-mediated web access; and developer ecosystems kept converging around evaluation, security, and model choice. The through-line is clear: the winners are trying to own the boring but decisive middle layer between impressive demos and accountable systems.

The big AI/platform moves

The most commercially important AI news was Claude’s continued push into mainstream enterprise channels. Microsoft said Claude in Microsoft Foundry is now generally available, hosted on Azure and running on NVIDIA GB300 Blackwell Ultra infrastructure. The point is not merely that Azure customers can call another model. Microsoft is packaging procurement, billing, data zones, identity, networking and agent orchestration around Claude so enterprises can use it without building a parallel governance stack. AWS moved in the same direction by announcing Claude Sonnet 5 on Amazon Bedrock and Claude Platform on AWS. For buyers, the practical implication is model optionality without abandoning the cloud operating model they already have.

OpenAI’s week was more about deployment evidence than one-off spectacle. Its HP Frontier partnership update described HP moving from pilots into a broader portfolio of AI workflows across customer support, partner operations, device telemetry, security, ChatGPT and Codex. The most interesting detail was not the anecdote that one HP engineer used OpenAI models across 122 pull requests in 43 projects, or that a security team compressed remediation work from an estimated month into a day. It was the emphasis on Frontier as a connective layer for access, context, deployment and evaluation. That is where enterprise AI strategy is heading: not “which chatbot?” but “what system decides what context an agent can see, what tools it may use, how actions are reviewed, and how outcomes are measured?”

OpenAI’s research on agents reinforces that shift. The company said Codex has become the primary AI tool across every OpenAI department, including non-technical groups such as Legal and Recruiting, and that non-developer usage grew faster than developer usage. Taken cautiously, this is still a company reporting its own internal behaviour; taken seriously, it is a leading indicator that agents are escaping the IDE. If finance, legal, support and operations teams can safely use agentic tools for structured analysis, automation and data transformation, the enterprise software market will have to redesign workflows around supervised execution rather than static screens.

There was also a compute strategy story. NVIDIA announced a model for helping AI clouds procure and monetise large-scale accelerated infrastructure through revenue-sharing and credit support. Sharon AI is deploying up to 40,000 Grace Blackwell GB300 GPUs, while Firmus is building a DSX AI factory campus in Batam, Indonesia, expected to scale to 360 megawatts and up to 170,000 NVIDIA GPUs. This looks like hardware, but it is also finance and go-to-market. The bottleneck is no longer just chip supply; it is the ability to assemble capital, power, sites, customers and utilisation into an investable AI factory.

Data stack and enterprise software

The data platform news was quieter than the model news, but arguably more revealing. Microsoft Fabric shipped a cluster of governance and data-sharing updates: the Fabric data agent API became public, Delegated OneLake Shortcuts entered preview for secure cross-team and cross-tenant reuse, and Workspace Outbound Access Protection reached preview for Real-Time Intelligence. Fabric is leaning into a world where data does not move into a single perfect repository; it is reused through shortcuts, policies and APIs while agents and analytics systems sit on top.

That matters because agent adoption increases the blast radius of weak data controls. A dashboard user can only click what the product exposes; an agent can chain tools, retrieve context and generate outputs at scale. Fabric’s focus on OneLake security, data protection, recovery, private network support and event-flow controls is therefore not administrative plumbing. It is the condition for letting copilots and agents touch enterprise data without creating a shadow data lake of copied extracts.

AWS’s week had a similar production theme. Bedrock posts covered managed entitlements for multi-account model access, an AgentCore memory filtering pattern using metadata, a serverless agent-to-agent gateway, resilience patterns with LLM gateways, and model selection through an open source Bedrock Model Profiler. AWS also launched EC2 C9g and C9gd instances on Graviton5, claiming up to 25% better compute performance than Graviton4-based instances, and CloudFormation Express mode, advertised as accelerating infrastructure deployments by up to 4x. None of these individually changes the market narrative; together they show AWS making AI applications look like normal cloud workloads: governed, routable, resilient, observable and deployable by automation.

MongoDB marked 10 years of Atlas, noting availability across more than 125 AWS, Google Cloud and Microsoft Azure regions. Elastic argued that AI performance problems often come from poor data foundations rather than model choice, and cut serverless metrics pricing for time-series data. Databricks and Snowflake did not have a clearly material fresh leader signal in the primary sources reviewed this week, but the competitive pressure on both is obvious: the lakehouse/warehouse vendors are being judged not only by SQL performance, but by how well they provide governed context to agents, semantic layers, vector search, evals and real-time applications.

What the leaders are saying

• Justin Boitano, Vice President and GM of Enterprise Computing, NVIDIA — On Microsoft’s Claude-in-Foundry launch, Boitano said: “At NVIDIA, we use autonomous AI agents every day to help our teams move faster and think bigger. Anthropic’s Claude models bring strong reasoning, coding and enterprise capabilities that are valuable for complex technical work.” He added that Claude running in Microsoft Foundry on GB300 GPUs gives organisations the “performance, scale and security needed for production.” Why it matters: NVIDIA is positioning inference infrastructure as the foundation for enterprise agents, not just training clusters.

• Steve Sweetman, Azure Product Lead for Foundry Models, Microsoft — Microsoft’s launch post framed Claude in Foundry as “the production path enterprises have been asking for: true frontier model choice, Azure-native controls, simplified procurement, and faster time to value.” Why it matters: The model marketplace is becoming a governed cloud procurement and operations layer, which favours incumbents that already own enterprise identity, billing and compliance.

• James Manning, cofounder and CEO, Sharon AI — In NVIDIA’s AI compute announcement, Manning said: “This strategic collaboration with NVIDIA marks a pivotal moment in Sharon AI’s mission to deliver sovereign, large-scale AI compute infrastructure.” Why it matters: Sovereign and regional AI clouds are becoming a major channel for GPU demand, especially where governments and enterprises want local control.

• Tim Rosenfield, co-CEO, Firmus Technologies — Rosenfield said: “AI-native companies need access to scalable, energy- and cost-efficient compute infrastructure to compete globally.” Why it matters: Power efficiency and site economics are now strategic product features for AI infrastructure providers, not facilities footnotes.

• OpenAI Economic Research team — OpenAI wrote that “Codex became the primary AI tool for every department at OpenAI,” with Legal, Finance and Recruiting crossing into majority Codex usage around April 2026. Why it matters: The strongest agent adoption signal may be non-developer usage, because it expands the addressable market beyond software engineering.

• OpenAI Frontier/HP deployment team — In the HP partnership update, OpenAI described Frontier as connecting “access, context, deployment, and evaluation” as work moves from pilots to production, and quoted an HP engineer saying of the tool: “I am using it daily.” Why it matters: Enterprise AI buying is moving from seat-based experimentation toward governed portfolios of repeatable agent workflows.

• Cloudflare product team — Cloudflare’s Monetization Gateway post argued that “the agent becomes the primary buyer on the Internet, and the request becomes the transaction.” Why it matters: If agent traffic starts paying for APIs, data and MCP tools at request time, web infrastructure companies could become payment, identity and policy layers for autonomous software.

• Mark Zuckerberg, CEO, Meta — TechCrunch reported that Zuckerberg told staff that Meta’s AI agent efforts had not progressed as quickly as he had hoped. This is a reported paraphrase rather than a verified direct quote. Why it matters: Even the best-capitalised consumer AI companies are discovering that reliable agents are harder than demos, especially when product expectations involve autonomy, safety and consumer polish.

Products and repos worth watching

For developers, the week’s useful signal was around evaluation and operational safety. GitHub published an evaluation of the Copilot agentic harness across models and tasks, emphasising that harness quality can improve multiple GitHub and Microsoft surfaces at once. The important idea is that agent performance is not only a property of the base model. Tool search, MCP servers, task decomposition, retry policy, sandboxing and UI design all affect whether work gets done reliably and cheaply.

GitHub also described how it used secret scanning to drive more than 20,000 alerts across 15,000 repositories to “inbox zero” in nine months. That is worth watching because AI-assisted coding increases code volume and speed, which raises the value of automated security remediation workflows. Security tooling that can prioritise, route and close findings will become more important as agents produce more changes.

On the open-source and model-serving side, Hugging Face highlighted ScarfBench, an IBM Research benchmark for enterprise Java framework migration, and a post on running a vLLM server on Hugging Face Jobs in one command. AWS introduced an open source Bedrock Model Profiler for comparing model metadata, while its GovCloud update brought OpenAI GPT OSS and NVIDIA Nemotron models into a government-oriented Bedrock environment. NVIDIA’s BioNeMo Agent Toolkit, available through GitHub and integrated with Anthropic’s Claude Science public beta, is also notable: it packages scientific models and workflows as callable agent skills, including Evo 2, Boltz-2 and OpenFold3-related capabilities.

Cloudflare’s Monetization Gateway may be the week’s most provocative infrastructure product. It proposes charging for web pages, datasets, APIs or MCP tools behind Cloudflare using x402, with payment proof embedded in the request path and support for dashboard, API and Terraform configuration. Whether stablecoin settlement becomes mainstream is uncertain, but the architecture points to a real need: agents will need identity, permissioning, metering and payment when they access scarce resources.

Regulation, risk and market context

The policy backdrop is becoming inseparable from product strategy. OpenAI’s EU jobs transition report used European labour-market data to examine where AI may automate, augment or reshape occupations. The useful point for executives is not to treat this as a prediction machine; it is a planning framework for reskilling, workflow redesign and regulatory engagement. AI adoption will be judged by labour-market outcomes, not just productivity charts.

Security and sovereignty were the other big themes. AWS made OpenAI GPT OSS and NVIDIA Nemotron available on Bedrock in GovCloud, and NVIDIA highlighted Palantir’s use of Nemotron open models for air-gapped government and critical-infrastructure environments. This is a strong signal that “open model” does not simply mean hobbyist deployment. For regulated customers, open weights plus secure environments can mean inspectability, customisation and control over where sensitive data and model artefacts live.

There are also hype risks. TechCrunch’s report about Zuckerberg’s internal caution on agents is a useful counterweight to the week’s confident product launches. Agents remain brittle when tasks are long, tools are messy, incentives are unclear, or permissions are over-broad. Enterprises should ask vendors for evaluation data, incident handling, audit logs, cost controls and rollback paths. The more autonomous the workflow, the more procurement should resemble security architecture review.

Chip and supply-chain strategy remains just as important. Reports that Anthropic is discussing custom-chip work with Samsung, alongside OpenAI’s earlier custom-chip direction with Broadcom, fit a broader pattern: frontier labs want leverage over inference cost, capacity and availability. At the same time, NVIDIA is expanding beyond selling GPUs into financing and ecosystem structures that help keep its platform central. For AMD, Broadcom, TSMC, ASML, Arm, Micron and Marvell, the strategic question is how much AI infrastructure value shifts into accelerators, networking, memory, packaging, foundry capacity and custom silicon over the next cycle.

What to watch next week

Watch whether Claude’s broader availability through Azure and AWS translates into reference architectures, customer case studies and pricing clarity. Watch whether Microsoft Fabric’s data-agent API gets early ecosystem tooling from consultants, ISVs or internal platform teams. Watch Cloudflare’s Monetization Gateway waitlist and x402 adoption, because paid agent access could become a new business model for data and API owners. Watch GitHub, OpenAI, Anthropic and AWS for more public evals of agent reliability, not just capability. And watch the infrastructure market for signs that AI factories are being sold as financial products as much as technical systems: utilisation commitments, revenue-share deals, sovereign AI clouds and power-secured campuses may tell us more about the next year of AI than any single benchmark score.