Suprvisr TLDR - March 30 2026

Suprvisr AI just launched Cumulus — a Canadian-hosted, audit-ready AI platform built for teams that need to move fast without handing their data to a foreign cloud. Cumulus runs on Canadian infrastructure, meets federal and provincial compliance requirements out of the box, and gives your team access to frontier models with full logging, role-based access, and policy controls.

No procurement headaches. No shadow AI. No six-month rollout. Start your 30-day free trial today and see what responsible AI adoption actually looks like.

OpenAI will shut down Sora on April 26. The video-generation tool that was supposed to be the future of creative AI is being killed so the company can redirect GPUs toward a new reasoning model internally codenamed Spud. The pivot cost them a $1 billion Disney partnership that was in final-stage negotiations. Disney pulled out after the shutdown announcement, citing "platform instability."

Sora hit 3.3 million downloads at launch but dropped to under 1 million monthly actives within weeks. Lifetime revenue was approximately $2.1 million against an estimated $15 million per day in compute costs. The app became better known for deepfakes of deceased celebrities than for any legitimate creative use.

Sam Altman is now betting on a "super-app" pivot — one model that handles text, code, images, video, and voice in a single interface. Meanwhile, Anthropic has quietly captured 73% of enterprise AI spend, and OpenAI's consumer-first strategy is starting to look like a very expensive hobby.

That is Anthropic's annualized revenue run rate. The company now has over 500 customers paying more than $1 million per year. Claude Code alone is on a $2.5 billion run-rate, making it one of the fastest-growing developer tools ever shipped. The average Anthropic enterprise user generates $211 per month in revenue. The average OpenAI consumer user generates roughly $25 per week. The margin structures are not even in the same conversation.

GTC 2026 was NVIDIA's clearest signal yet that it intends to own the full AI stack, not just the chips. The Vera Rubin platform promises a 10x reduction in inference costs, which matters because inference — not training — is where enterprise money actually gets spent.

The bigger move was NemoClaw, NVIDIA's enterprise-grade wrapper for the open-source OpenClaw robotics framework. NemoClaw adds safety filtering, skill certification, and audit logging on top of OpenClaw's 900+ community-contributed manipulation skills. It is NVIDIA's play to become the middleware layer between open-source robotics research and corporate deployment — the same playbook Red Hat ran with Linux, but for robot arms.

Anthropic has been quietly tightening its 5-hour rolling session limits on Claude Pro and Max plans, particularly during peak North American business hours. Screenshots from Max subscribers showed rate-limit warnings appearing after as few as 15 messages in a session. Anthropic confirmed the adjustments, saying roughly 7% of users were affected and pointing to a GPU supply bottleneck that is being addressed with new capacity coming online in Q2.

The tone-deaf part was the suggestion that users "try again during off-peak hours" — a response that landed poorly with paying customers who chose the $200/month Max plan specifically to avoid limits. Claude Code users were hit hardest, with some reporting a rapid rate-limit drain bug that burned through session allowances in minutes rather than hours.

1. The White House gave AI copyright a green light. The new National Policy Framework for AI explicitly endorses fair use for training data, settling (for now) the legal ambiguity that has haunted every foundation model company. Anthropic's recent fair use ruling in Thomson Reuters v. Anthropic is now effectively federal policy. This doesn't end the lawsuits, but it gives every AI company a much stronger legal footing and signals that the US government views training-data access as a competitive necessity, not a copyright violation.

2. Google shipped Gemini 3.1 Flash Live. The model supports native audio-to-audio processing — no speech-to-text-to-LLM-to-text-to-speech pipeline. It scored 90.8% on ComplexFuncBench, the best result any model has posted on multi-step function calling. Anthropic still has no voice offering, which is starting to look like a gap rather than a deliberate choice.

3. Deloitte got caught using AI hallucinations in a government report. A $1 million healthcare consulting report submitted to the Newfoundland provincial government contained fabricated citations generated by an AI tool. Deloitte issued a partial refund and blamed a "quality assurance gap." The incident is now Exhibit A in every AI governance pitch deck on the planet.

Figure AI's Figure 03 launched with a Vision-to-Action architecture that skips the traditional perception-planning-execution pipeline entirely. The robot sees, decides, and moves in a single forward pass, hitting 99% task accuracy in warehouse trials. That matters for Ontario because Toyota's Cambridge and Woodstock plants are already in the Figure deployment pipeline.

Meanwhile, Canadian startup Mirsee unveiled the MH3, a made-in-Canada humanoid designed for variable tasks in manufacturing and logistics. Ontario now has two humanoid robotics programs converging on its automotive corridor, which is quietly becoming the densest cluster of humanoid deployment outside of Asia.

If your organization is thinking about what this means for workforce planning, automation strategy, or procurement — we should talk.

Stop outsourcing your tools. Start building with them. Claude Code just crossed a $2.5 billion revenue run-rate and now accounts for an estimated 4% of all GitHub commits. This is not a novelty — it is a production-grade development environment that is reshaping how software gets written.

We are running a Code with Claude workshop for teams that want to start using Claude Code seriously. The format is hands-on, the scope is practical, and the goal is to get your team shipping with AI-assisted development inside of a week. Three principles we teach: start with a pain point, not a project; treat Claude like a junior dev, not a magic wand; own what you build.

If that sounds useful, reach out. We will set something up.