Welcome back, revolutionary…
OpenAI burned $15 million a day running Sora and made $2.1 million in total. The app closes April 26, and Disney's only got to know an hour before us.
In this week's edition:
What Sora's real numbers reveal about the gap between hype and paying customers
Anthropic's step change model leaked before they could announce it
Where $297 billion went in 90 days (and what it means for everyone else)
Half a million robotaxi rides a week, $300 million in political spending, and a sales agent that works for free
OpenAI Killed Sora. The Numbers Explain Why.
Sora launched in February 2025 as OpenAI's flagship bet on AI video. Fourteen months later, they pulled the plug. The app closes April 26, the API limps on until September, and users have been told to grab their content while they still can.
The economics tell the story better than any press release.
Need to know:
Each 10-second clip cost about $1.30 to generate: four parallel GPUs churning for 40 minutes at nearly $2/hour. At scale, that's roughly $15 million a day in inference costs against $2.1 million in total lifetime revenue. Read that ratio again.
Downloads fell 66% from their peak. Active users dropped from 1 million to under 500,000. Meanwhile, every deepfake and copyright complaint added moderation cost without adding a cent of revenue.
Disney had pledged $1 billion and a three-year character licensing deal covering 200+ properties (Marvel, Pixar, Star Wars). No contract was signed. No money moved. Disney reportedly found out about the shutdown less than an hour before everyone else.
The Sora research team pivots to "world simulation" for robotics. The successor, codenamed Spud, targets July 2026 as an API-only product. No consumer app this time.
A jaw-dropping demo and a viable product turned out to be completely different things. The people sharing Sora clips on Twitter weren't the same people willing to pay for them. If you're building anything on top of AI right now, this is the number to tattoo on your forearm: the constraint isn't what the model can do, it's what each output costs to produce. Capabilities without unit economics is just an expensive magic trick.
Breach
Anthropic's Most Powerful Model Leaked Through an Unsecured Data Cache
Someone at Anthropic misconfigured a content management system, and close to 3,000 unpublished assets spilled onto the open web. Among them: draft documentation for Claude Mythos (internal codename: Capybara), a new model tier sitting above Opus that Anthropic calls a "step change." Fortune found the publicly searchable data store on March 26 and broke the story.
Need to know:
Capybara scores "dramatically higher" than Opus 4.6 on coding, academic reasoning, and cybersecurity benchmarks. Anthropic frames it as an entirely new tier, not an upgrade.
The leaked docs show Anthropic privately warning government officials that Mythos makes large-scale cyberattacks meaningfully more likely in 2026. The model is expensive to run and nowhere near general release.
A handful of early access customers are testing it now. Anthropic locked down the data store after Fortune reached out, but by then the documents had already been downloaded and circulated widely.
The company that positions itself as the responsible AI lab just accidentally published its own most capable (and, by its own assessment, most dangerous) model's documentation through a misconfigured cache. If you're evaluating which AI providers to build on, the question isn't just about model quality. It's about operational competence.
Market Signal
AI Just Swallowed 81% of All Venture Capital on Earth
Investors put $297 billion into 6,000 startups globally in Q1 2026. That's up 150% year over year and the biggest single quarter in venture history. AI took $239 billion of it. Four of the five largest venture rounds ever recorded closed in the same 90-day window.
Need to know:
The mega-rounds: OpenAI ($122 billion), Anthropic ($30 billion), xAI ($20 billion), and Waymo ($16 billion) collectively raised $186 billion, or 64% of all global venture investment in the quarter. Late-stage funding reached $244 billion across 582 deals.
US-based companies captured 83% of global venture capital, up from 71% in Q1 2025. The concentration is accelerating.
Startup M&A hit $56.6 billion in Q1, the third-highest quarter since the 2022 downturn. OpenAI alone has completed six acquisitions in 2026 already, nearly matching its total for all of 2025.
When 81 cents of every venture dollar flows into one technology category, that's not a trend. It's gravity. Marketing, compliance, logistics, professional services: they're all getting pulled into the orbit whether they signed up or not. The question for smaller operators is simple: are you standing where the money is landing, or where it's leaving?
Round Up: Building the Machine
Waymo doubles to 500,000 paid rides per week: Tenfold growth in under two years, from 50,000 weekly trips in May 2024 to half a million now across 10 US cities. Seven Sun Belt cities (Austin, Atlanta, Miami, Dallas, Houston, San Antonio, Orlando) were all added in the past year alone. Fleet size: still over 3,000 robotaxis on the same 5th-gen system. Year-end target: 1 million rides per week.
OpenAI acquires Astral and Promptfoo: Six acquisitions in 2026 so far. Astral (March 19) brings the widely used Python tools uv and Ruff to OpenAI's Codex team, which has hit 2 million weekly active users with 5x usage growth since January. Promptfoo adds AI security testing to OpenAI Frontier, the enterprise platform for building AI coworkers.
Runway launches $10M fund and Builders program: Pre-seed and seed checks up to $500,000 for startups building across AI, media, and world simulation. The Builders program offers 500,000 API credits and access to Characters, Runway's real-time video agent API. Founding cohort includes Cartesia, MSCHF, and Oasys Health.
Meta launches Business AI for small businesses: A free AI sales agent trained on your catalog, website, and ad performance. Handles product recommendations, objection handling, and purchase guidance across Messenger, WhatsApp, and your website. Available now for eligible US advertisers.
Innovation Council Action preps $100M midterm spend: A new PAC led by former White House deputy chief of staff Taylor Budowich, endorsed by outgoing AI czar David Sacks. The group scores lawmakers on alignment with the administration's AI deregulation agenda. Combined with Leading the Future ($50M) and Meta's state-level super PAC ($65M), total AI election spending now exceeds $300 million.
Google brings Gemini to Maps with Ask Maps: Conversational AI built into Google Maps, rolling out in the US and India. Users can ask complex natural-language questions ("Where can I charge my phone without waiting in a coffee queue?") and get personalized recommendations. Paired with a redesigned Immersive Navigation using 3D maps and real-time road data.
Tool Shed
Jupid: AI tax filing that connects to your bank, maps transactions to IRS categories with 96% accuracy, and finds deductions averaging $1,249/year.
Pixero AI: Give it a URL and it builds, launches, and manages your Meta ad campaigns end-to-end in about 9 minutes.
Stamp: AI secretary that autonomously manages your email and calendar, only asking for approval before making changes.
Lightfield: AI-native CRM that reads your emails, meetings, and calls to build and update your customer records automatically.
Lexaclaw: Startup compliance assistant that tracks obligations, fills government forms, and generates legal documents.
Peopling: AI voice roleplay for practicing difficult workplace conversations, from performance reviews to client pushback.
Letterbook: Customer support platform with an AI agent that resolves tickets across email and web forms without human handoff.
Crossnode: Monetize your AI agents without coding or backend setup; handles billing, auth, and deployment.
Quick Bytes
Q1 2026 global venture funding hit $297 billion across 6,000 startups, with AI accounting for 81% of every dollar invested, up from 55% a year ago.
Zero-click searches have risen to 65-70% of all Google queries, and click-through rates drop from 15% to 8% when an AI Overview is present.
Organic click share is down 11 to 23 percentage points across multiple verticals between January 2025 and January 2026, even as AI-driven referral traffic grew 187% in the same period.
60% of small businesses now have at least partially adopted AI in customer service and marketing, nearly double the figure from 2023.
AI-native referral traffic is dominated by ChatGPT at 55-60% share, followed by Perplexity (18-22%), Gemini (10-14%), and Microsoft Copilot (6-9%).
