The Noise — March 27, 2026
OpenAI killed Sora, teased a model named Spud, and raised another ten billion dollars. Arm named a CPU after a technology that doesn't exist. DoorDash is paying gig workers to train their own replacements. The weekly reality check.
OpenAI killed its most famous product, pledged a billion dollars to charity, teased a new model codenamed after a potato, and raised another ten billion in funding. All in the same week. The product they killed was losing money. The charity pledge is 0.14% of their valuation. The potato will "really accelerate the economy." The funding round brings total capital raised to $120 billion for a company projecting $14 billion in losses this year. If you can hold all of that in your head simultaneously without laughing, you might be an investor.
The Many Lives and Quick Death of Sora
Remember when Sora was going to change everything? When AI-generated video was the next frontier? When Disney put up a billion dollars and 200 characters from its universe to make Sora the creative engine of the future?
That was December. Three months ago. On Tuesday, OpenAI posted "We're saying goodbye to Sora" on X, with no real explanation. Disney confirmed the same day it "will not be moving forward" with the investment. The licensing agreement for Marvel, Pixar, and Star Wars characters never went into effect.
The stated pivot: video generation tech will live on internally to train robots, because video provides a "reasonable simulation of the physical world." OpenAI CEO of Applications Fidji Simo has told staff the company is "orienting aggressively" toward enterprise use cases. Every GPU-hour now needs to justify itself against revenue, and consumer video creation lost that argument.
When Sora launched last September, the discourse treated it like an extinction-level event for visual media. Artists panicked. Hollywood scrambled. Hot takes about the death of cinematography accumulated like snow in a parking lot no one plows. Six months later, the tool that was going to end human creativity couldn't survive a budget review. The thing that was supposed to replace artists got replaced by a spreadsheet.
Meanwhile, The Information reported that OpenAI has finished pretraining its next model, codenamed "Spud." Sam Altman told employees it could "really accelerate the economy." No benchmarks. No demos. No details beyond a root vegetable and a promise. Sora launched with a demo reel that made the internet lose its mind, and it still died in six months. Spud doesn't even have a demo reel. It has a codename and vibes.
Arm Called Its Chip "AGI" and Hoped Nobody Would Notice
On Monday, Arm announced its first in-house production chip in 35 years: a 136-core, 3nm processor built on Neoverse V3 cores for AI inference workloads. Meta co-developed it. OpenAI, Cerebras, and Cloudflare are launch partners. It's a legitimately interesting product.
They named it the AGI CPU.
AGI, in this context, stands for something Arm hasn't publicly clarified. It does not stand for Artificial General Intelligence, the hypothetical technology that would match or exceed human cognition across all domains, which this chip cannot do because it is a CPU. Tom's Hardware called the naming "decidedly disingenuous." The Register called it "hypemaxxed branding." A Hacker News commenter noted that people will buy Arm stock thinking they've cracked AGI, and that "the people running Arm absolutely know this."
The chip itself is real: 300 watts, air-cooled, no liquid cooling required, targeting data centers that don't want GPU plumbing overhead. But the name is doing more work than the silicon. In a week where Sora proved that the gap between AI marketing and AI reality can close fatally fast, Arm decided the best strategy was to name their CPU after a technology that doesn't exist.
Quick Hits
DoorDash launched an app called Tasks that pays delivery couriers to film themselves washing dishes and moving objects across tables. The footage trains AI and robotics models. The first task earns you not cash but a free body-mount camera so you can film more training data. The app is already banned in four states and cities. DoorDash has 8 million couriers in the US. They are now paying the humans whose jobs robots will eventually replace to teach those robots how to do those jobs.
CFOs surveyed by Duke University and the Federal Reserve Banks of Atlanta and Richmond admitted privately that AI-driven layoffs will be nine times higher than last year's 55,000 AI-attributed cuts. The same survey found only 2% of organizations have made large AI-driven headcount reductions, but nearly 90% have frozen or reduced hiring in anticipation. The headline framing was "not doomsday." In practice, you don't get fired. You just never get hired.
Meta got hit with $381 million in jury verdicts across two courtrooms in 48 hours for harming children. While those verdicts were coming in, the company laid off another 700 people. Priorities.
Hype Watch
"Really accelerate the economy" — Sam Altman's internal pitch for the Spud model, per The Information. When a CEO tells employees their unreleased product will accelerate the entire economy, and also that company is losing $14 billion this year, those two facts are coexisting in the same building.
"AGI" — Now means both "the holy grail of artificial intelligence" and "a server chip from Arm." Language is truly a collaborative effort between humans and marketing departments.
"People-First AI Fund" — The OpenAI Foundation's name for its $1 billion grant program. The company that helped popularize AI doing the work of entire departments now has a fund dedicated to the people displaced by AI doing the work of entire departments. The fund is partially staffed by effective altruism alumni from the community that has been most vocal about OpenAI's safety problems.
From the Editor's Desk
The Anthropic-Pentagon story resolved Thursday when Judge Rita Lin granted a preliminary injunction, calling the supply-chain risk designation "classic First Amendment retaliation." We've been tracking this since February. The ruling is significant but we're holding deeper analysis for next week because the D.C. case is still active.
We're also sitting on the LiteLLM supply chain attack aftermath. Compromised versions of the library exfiltrated API keys for thousands of companies. The attack chain started with a compromised security scanner. Worth a standalone piece, not a newsletter mention.
Running Tallies
Week 6 of DeepSeek V4 being "imminent." The latest credible reporting, from Chinese tech outlet Whale Lab, now points to April. A blog that previously listed "mid-February" as the launch window still hasn't updated its headline. V4 has now been launching next month for three consecutive months.
Week 9 of "the era of physical AI." Jensen Huang's CES declaration. Arm is shipping inference chips. DoorDash is paying humans to film themselves doing chores. The robots are learning from YouTube and gig workers. The era is more "physical humans training digital systems" than anyone at CES envisioned.
The AI Layoff Index: 60,000 and counting. Up from 45,000 two weeks ago, per trackers at SkillSyncer and IBTimes. CFOs say privately it's going to get nine times worse. Meta alone has cut over 2,200 people in 2026 while posting record revenue. The company that just lost $381 million in child safety verdicts is somehow still the one doing the firing.
A company raised $120 billion, killed its most famous product, pledged 0.14% of its valuation to charity, teased a model named after a potato, and is on track to lose $14 billion this year. Another company named a CPU after a technology that doesn't exist. A third company is paying gig workers to train the robots that will replace them. The era of physical AI is the era of physical comedy.