AI Weekly Digest: Sora Shutdown, Claude Agents & $100B AI Market Shift

Nine things that moved the needle — and what they mean for your strategy like claude agents. From leaked models and killed products to an AI that opened your laptop while you were eating — this was one of the most consequential weeks in AI in 2026. Here’s what happened, why it matters, and what you should actually pay attention to.

Anthropic’s next model wasn’t supposed to be public yet

A misconfigured CMS. Nearly 3,000 unpublished files. And a model called Mythos that Anthropic says poses “unprecedented cybersecurity risks.”

Anthropic had a rough few days — and the AI world is still processing the fallout. Security researchers at the University of Cambridge and LayerX Security discovered that close to 3,000 unpublished assets on Anthropic’s content management system were sitting in an unsecured, publicly searchable data store. Among them: a detailed draft blog post about a new model the company had been developing under the name Claude Mythos.

Anthropic confirmed the model’s existence after Fortune reported on the leak, calling it “a step change” in AI performance and “the most capable model we’ve built to date.” The company attributed the exposure to a human configuration error. The draft described Mythos as “currently far ahead of any other AI model in cyber capabilities” and warned it “presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders.”

Five days later, it happened again. An employee accidentally shipped the full source code for Claude Code — 512,000 lines of TypeScript across roughly 1,900 files — to the public npm registry as part of a routine update. Anthropic issued a DMCA takedown that accidentally removed over 8,000 GitHub repositories.

What leaked: The Claude Code source revealed unreleased features including KAIROS — a background agent designed to work autonomously while you’re idle — and an “undercover mode” that strips Claude branding in external repositories.

BuzzMora Take  The story isn’t just about Anthropic’s bad week. It’s about what the leaks confirmed: AI agents are now complex, autonomous systems with full system access — and the security posture of the companies building them hasn’t caught up. For brands evaluating AI partnerships, operational security should be on the checklist alongside capability benchmarks.

Elon said one number. Chip stocks moved.

On March 21, Musk took the stage at a defunct Austin power plant and announced Terafab — a $20–25 billion semiconductor fabrication facility that sent chipmaker stocks surging.

The number Musk put on the board was 1 terawatt. That’s his estimate for how much computing power Tesla, SpaceX, and xAI will eventually need — roughly 50 times what the world’s chip fabs currently produce. Terafab is a joint venture between Tesla, SpaceX, and xAI, designed to consolidate the entire chip supply chain from design to fabrication under one roof in Austin, Texas.

“We either build the Terafab or we don’t have the chips,” Musk said at the launch event. The project is planned for the North Campus of Tesla’s Giga Texas and could cost $20–25 billion — on top of Tesla’s existing 2026 capex of more than $20 billion, which doesn’t yet include Terafab.

The scale: Musk estimates needing 1 billion Nvidia GPUs per year at full operation — current global AI computing capacity is just 20 gigawatts, roughly 2% of what Terafab targets.

BuzzMora Take  One number — “1 terawatt” — moved semiconductor markets. That’s the power of framing at scale. The Terafab announcement is as much a positioning statement as it is a capital commitment. For AI-adjacent brands, the lesson is identical: the story you put out first shapes market perception before the product exists.

Google dropped a paper. Three companies lost billions.

Google’s TurboQuant algorithm claims to reduce AI memory usage by 6x — and triggered one of the sharpest single-week sell-offs in chip stock history.

On March 24, Google Research published a paper on TurboQuant — a memory-compression algorithm targeting the KV cache, the working memory that large language models use to avoid recomputing past calculations. The algorithm compresses the cache to just 3 bits per value, down from the standard 16, reducing its memory footprint by at least six times without any measurable loss in accuracy, according to Google’s benchmarks.

Markets moved fast. Micron Technology saw its stock plunge by 14% in 48 hours, erasing over $25 billion in market capitalization. Western Digital dropped 11%, and Seagate fell nearly 9%. Samsung fell nearly 5% and SK Hynix dropped 6% in Seoul. Cloudflare’s CEO called it “Google’s DeepSeek moment.” By one estimate, the announcement contributed to roughly $100 billion in combined market value losses across the memory sector.

The nuance: TurboQuant only addresses memory used during inference — it does nothing for training, which drives the bulk of HBM chip demand. Micron’s CEO confirmed all calendar 2026 HBM supply is already sold out. The sell-off may have been an overreaction.

BuzzMora Take  A research paper triggered a $100B sell-off in days. This is AEO in reverse — when a well-structured, clearly communicated technical claim lands in the right context, it shapes perception at institutional scale. The companies whose technical narratives reach the market first set the frame everyone else responds to.

OpenAI killed their own product

Sora launched in September 2025 with enormous hype, a viral app store debut, and a $1 billion Disney deal. Six months later, it’s dead.

On March 24, OpenAI announced it was discontinuing Sora — its AI video generation tool — in what may be the most expensive product-market fit failure in AI history. Sora peaked at roughly 1 million monthly active users before crashing below 500,000. Its estimated $15 million daily inference cost consumed compute resources OpenAI said it needed for “the next generation of automated researchers and companies.” Total lifetime revenue: $2.1 million. The math was not close.

OpenAI is preparing for a potential IPO and needs defensible, recurring enterprise revenue. Consumer video generation, no matter how impressive the demos, didn’t fit that story. The Sora app shuts down April 26. The API follows in September.

What happened to the Disney deal: Disney had committed $1 billion to the partnership. Per TechCrunch, the entertainment giant found out Sora was being shut down less than an hour before the public announcement. The deal died with the product. 

BuzzMora Take  The most dangerous number in AI is not the benchmark score — it’s the cost per unit of revenue. Sora had a world-class score on capability and a catastrophic one on unit economics. Build the funnel first, validate the economics second, ship at scale third.

Claude Agents opened someone’s laptop while they were at dinner

Anthropic launched Computer Use on March 23 — and for the first time, you can give Claude a task from your phone and return to finished work on your desktop without ever sitting down.

Users can now message Claude agents where a task from a phone, and the AI agent will complete that task autonomously. Claude can open apps on your computer, navigate a web browser, and fill in spreadsheets. The earlier Computer Use version (October 2024) required you to watch it work. The March 23 launch removes that constraint entirely.

Through Claude Cowork’s Dispatch feature, you can assign Claude a task from your phone, turn your attention to dinner, and return to finished work on your desktop. Scheduled tasks let Claude check your emails every morning, pull weekly metrics, or run your Slack digest automatically — without you initiating anything. Pro and Max subscribers on macOS can access it now in research preview.

Security note: Anthropic flagged prompt injection as a systems-level concern: any webpage Claude browses during a task becomes a potential attack surface. One documented case showed how a Word document with hidden white text tricked Cowork into uploading financial files to an attacker’s account.

 

BuzzMora Take  The shift from “AI that answers” to “AI that acts” is the defining transition of 2026. Brands that understand this won’t just use AI for content — they’ll use it to run entire workflow layers autonomously. The funnel doesn’t need a human at every step anymore. But it does need a human who designed which steps the AI takes.

ChatGPT added one tab. It fixes everything.

Canvas Mode in GPT-5 is ChatGPT’s answer to the one thing chat interfaces have always been bad at: working on a project iteratively without losing your mind.

OpenAI’s Canvas Mode — originally launched in beta in October 2024 — has matured significantly inside GPT-5. It opens a persistent workspace next to the chat where you can highlight specific sentences, adjust length with a slider, switch reading levels, debug code inline, and see version history — all without copying outputs back and forth.

In the GPT-5 era, Canvas Mode supports large file uploads, data-driven dashboards built directly from CSVs, and interactive learning tools generated from a single prompt. For marketing teams, strategists, and ops leads, it means complex deliverables that used to take hours of back-and-forth now emerge in a single session.

For content teams: Canvas closes the loop between ideation and production. Brief → draft → edit → refine → export — all in one persistent workspace. If you’re still treating ChatGPT like a search engine with opinions, you’re behind.

BuzzMora Take  The brands winning with AI content right now aren’t using AI to replace writers — they’re using interfaces like Canvas to cut the revision cycle from days to hours. Speed of iteration is the new content moat.

One button in Gamma. One hour saved.

Gamma’s one-click redesign feature does what used to take a designer an afternoon — and it’s just one reason the platform has reached 70 million users.

Gamma, the AI-native presentation and document builder, has quietly become one of the most adopted AI productivity tools and claude agents on the market — 70 million users, 400 million pieces of content created, and a new API generally available since January 2026. The feature that keeps coming up in user reviews: one-click redesign. You can transform the visual theme of an entire presentation — every colour, font, and layout — in a single click, while keeping all your content intact.

Gamma 3.0 introduced the Gamma Agent, which researches the web, refines content, restyles entire decks, and provides design feedback through natural language conversation. A complete, polished slide deck from a text prompt generates in under 60 seconds.

 

BuzzMora Take  The productivity tools winning in 2026 removed the most friction from a specific workflow. Gamma removed the blank slide problem. Apply the same thinking to your own positioning: nail one specific conversion blocker, and usage follows.

An image model that actually understands what you meant

The gap between what you typed and what you wanted has been the persistent failure mode of text-to-image AI. The latest generation is finally closing it.

GPT Image 1.5 from OpenAI marked a notable shift in what text-to-image generation can reliably deliver. Where earlier models struggled with complex compositional prompts, the latest generation handles multi-element prompts with significantly higher fidelity. Typography in images — notoriously broken in earlier models — now renders correctly. Product renders from natural language descriptions are becoming production-usable.

For marketers and creative teams, this closes a workflow gap that has kept AI image tools in the “experimentation” category rather than the “production” category. When the model does what you actually meant rather than what you literally typed, the creative process stops feeling like prompt engineering and starts feeling like direction. Check our blogs for learning Affiliate Marketing in 2026: How It Works & Should You Try It?

What this changes: Social media graphics, ad creative variations, product visualisation, and presentation imagery can now be generated at brief-to-output quality without a designer in the loop for basic production tasks. The designer’s role shifts to creative direction, not pixel execution.

 

BuzzMora Take  Intent-accurate image generation changes the content production funnel structurally. The bottleneck shifts from “can AI produce this?” to “can we brief AI clearly enough to produce what we actually need?” That’s a strategy problem, not a technology problem.

A 20-year-old. Ten days. Four million dollars.

The AI nad Claude agents funding boom isn’t just for established labs. Young founders with a sharp problem and a working demo are raising seed rounds in days, not months.

The story that made the rounds this week wasn’t one company — it was a pattern. Young AI founders, some still in their teens or early twenties, are compressing traditional fundraising timelines to days. Toby Brown raised $1 million for his AI startup Beem at 16. Arlan Rakhmetzhanov dropped out of high school in Kazakhstan and raised over $6 million for Nozomio through Y Combinator. Aaru, co-founded at 19, hit a $1 billion valuation simulating human behaviour for market research.

The underlying dynamics are real. AI startups captured 41% of all venture dollars raised in 2025 — a record share — and foundational AI startups had raised $178 billion across 24 deals in Q1 2026 alone, compared to $88.9 billion across 66 deals in all of 2025. The money is moving fast, and it’s not exclusively chasing the labs. Why not visit it Buyer Journey Explained: Awareness to Conversion

BuzzMora Take  Capital has never moved this fast toward this many people at this early a stage. The barrier isn’t age, geography, or credentials — it’s clarity of problem framing and speed of execution. If you’re building at the intersection of AI and a specific vertical, 2026 is arguably the best year to raise you’ll ever see. Don’t wait until you feel ready. Ship, show, raise.