OpenAI Killed Sora 30 Minutes After a Disney Meeting. The Kill List Is the Strategy Now.

$15M/day to run, $2.1M lifetime revenue. The pivot to Codex puts them behind Claude Code — in a market China is about to commoditize from below.

THE NUMBER: $15 million / $2.1 million — the daily operating cost of Sora vs. its lifetime revenue. When a product costs 2,600x more to run per day than it has ever earned, killing it isn’t a choice. It’s arithmetic. The question is what that arithmetic tells you about everything else OpenAI is doing.

---

OpenAI killed Sora this week. Not quietly — 30 minutes after a working session with Disney, whose $1 billion investment was predicated on Sora existing. Downloads had collapsed from 3.3 million in November to 1.1 million by February. The app consumed computing power at a rate that made ChatGPT look like a calculator. And in the same breath, OpenAI’s CEO of applications told staff: “We cannot miss this moment because we are distracted by side quests.”

The kill list is the strategy now. Sora gone. In-app checkout gone. Consumer experiments trimmed to the bone. What’s left: ChatGPT as a funnel, Codex as the revenue engine, Pentagon contracts for credibility, PE deals for distribution, and ads for the freeloaders who won’t convert. Sarah Friar told CNBC that OpenAI needs to be “ready to be a public company.” Translation: the research lab is becoming a business, and businesses don’t burn $15 million a day on products with no revenue model.

But the business they’re becoming may already have a winner. Anthropic’s Claude Code commands 42-54% of the coding market. Ramp’s corporate spend data shows enterprise customers choosing Anthropic over OpenAI at a 3:1 ratio. OpenAI’s response? Double headcount to 8,000 by year-end while simultaneously telling current staff their projects are being deprioritized. In a market where elite AI talent can walk across the street tomorrow, that’s not a reorganization. It’s a loyalty test.

Meanwhile, the market OpenAI is pivoting toward is being repriced from below. Chinese open-source models surpassed U.S. models in Hugging Face downloads for the first time. Cursor built its IDE on Kimi at one-eighth the cost of frontier models. Pinterest, Airbnb, and Notion are incorporating Chinese open-source into production. And Jensen Huang told the world that token budgets should be part of compensation packages — up to half base salary.

Three stories. One throughline. OpenAI spent two years trying to be everything. Now it’s racing to be enough. The coding market it’s pivoting to has a leader. The model layer underneath is commoditizing. And the unit of value itself — the token — is becoming something nobody expected: a wage.

OpenAI’s Focus Era: Strategy Is What You Kill

💲 On Monday afternoon, teams from Disney and OpenAI sat down for a working session on Sora. Thirty minutes after that meeting ended, Disney was informed that OpenAI was killing the app. The $1 billion investment that Bob Iger had called “breathtaking” — his signature final deal as CEO — evaporated between meetings.

Sora was supposed to be the consumer AI product that transcended chat. Instead it became a case study in what happens when a research lab launches products without business models. Downloads peaked at 3.3 million in November 2025 and fell to 1.1 million by February. The app consumed computing resources that dwarfed ChatGPT’s. And the content it generated kept going sideways — users created copyrighted material from across pop culture, and AI-generated voices of James Earl Jones as Darth Vader dropped slurs in Fortnite, prompting a SAG-AFTRA unfair labor practice charge. The Writers Guild said Disney’s deal with OpenAI “appears to sanction its theft of our work.”

But Sora’s death isn’t the story. The kill list is.

Fidji Simo, OpenAI’s CEO of applications, told staff at an all-hands: “We cannot miss this moment because we are distracted by side quests.” Sarah Friar told CNBC they need to be “ready to be a public company.” Sam Altman’s “code red” from late last year has become company doctrine. The remaining focus areas: a “super app” combining ChatGPT, Codex, and Atlas; an enterprise sales buildout including a “technical ambassadorship” team; and ads rolling out to free-tier and Go-tier users at $60 CPM.

Here’s what the reorganization reveals. For two years, OpenAI ran like Y Combinator — Altman’s former institution — placing small bets across a wide surface area: video, hardware, browsers, robots, consumer apps. That’s how you find product-market fit when you’re a startup. It’s how you hemorrhage capital when you’re spending $100 billion over four years on data centers and preparing for an IPO. The transition from research lab to operating company requires killing your darlings, and Altman is finally swinging the axe.

The question is whether the focus came too late. While OpenAI was trying everything, Anthropic was doing one thing: winning enterprise. Claude Code didn’t just ship features. It shipped trust. Cisco’s security leaderboard. Ramp’s spend data. Seventy percent win rate in direct comparisons. Anthropic’s run rate reportedly jumped from $14 billion to $19 billion in a single quarter. That’s not marketing. That’s invoices.

OpenAI’s Codex crossed $1 billion in annualized revenue in January — impressive until you consider it’s chasing a competitor that’s been pulling away for a year. And here’s the talent problem nobody at OpenAI wants to talk about: a large corps of engineers just spent months or years busting their asses on projects — Sora, checkout, consumer experiments — that got stripped away overnight. The Sora research team was told to pivot to robotics. In a tight market for AI engineering talent, “your project no longer exists, now go work on something else” is not a retention strategy. It’s a resignation trigger. Jerry Tworek, OpenAI’s VP of Research, already left after struggling to get resources. He won’t be the last.

None of this means OpenAI is desperate. The free tier is being monetized intelligently — every ChatGPT user is either a future paying customer (the $20-to-$200 escalator is real, and anyone doing serious work hits the token ceiling fast) or an ad impression at $60 CPM. That’s a rational funnel. But rational isn’t the same as winning. When you’ve spent two years ceding the enterprise market to a focused competitor, “we’re focused now too” is catching up, not leading. And catching up with a corps of demoralized engineers who just watched their work get killed is catching up on hard mode.

The signal: The kill list tells you more than any press release. OpenAI is doing what every pre-IPO company eventually does. But the wonder is whether they squandered an early lead that no amount of focus can recover — not because the technology isn’t good enough, but because the humans who build it have memories, and options.

China’s Open Source Playbook and the Global Repricing

🇨🇳 While OpenAI pivots to coding, the coding market itself is being repriced from below.

Chinese open-source models surpassed U.S. models in Hugging Face downloads for the first time in 2025. The top five models on OpenRouter’s leaderboard — Mimo V2 Pro, Step 3.5 Flash, Deepseek V3.2, MiniMax M2.5, and GLM 5 Turbo — are all Chinese. Qwen has over 113,000 variations on Hugging Face. And the adoption isn’t theoretical. Airbnb uses Qwen to power its AI customer service. Pinterest’s CTO Matt Madrigal confirmed Qwen outperforms off-the-shelf proprietary models by 30% on shopping relevancy. Notion, Cursor, and a growing list of U.S. companies are incorporating Chinese open-source into production stacks.

Clément Delangue, co-founder of Hugging Face, named the structural advantage: “For Chinese AI labs right now, the default is open source. The standard is open source. Versus in the U.S., the default and standard is closed source.” China isn’t winning on model quality alone. It’s winning on volume, on cost, and on the strategic decision to make openness the default while U.S. labs debate the topic.

The cost differential is staggering. Hosting your own open-source model or using mixture-of-experts architectures reduces inference costs by 5-10x compared to frontier APIs. Cursor proved the point by building Composer 2 on Kimi — one-eighth the cost of Claude. Some U.S. developers may bristle at a Chinese model at the base of their coding tools. The rest of the world — the developers in Ho Chi Minh City, Bangalore, and Kyiv — will shrug and keep coding.

And this is where the story stops being about models and starts being about economics. Tomasz Tunguz published his “AI’s Bundling Moment” thesis this week, arguing that SaaS rewarded specialization — Salesforce owned CRM, Slack owned messaging, Dropbox owned storage — but AI rewards breadth. Harvey went from legal AI to legal plus professional services. ElevenLabs went from text-to-speech to full voice agents. The playbook is: bundle first, dominate later. When models change every 42 days, buyers want stable platforms, not best-of-breed stacks.

But Tunguz’s thesis undersells the darker implication. When open-source models are effectively free, communication costs are zero thanks to the internet, and the talent pool in Vietnam, Ukraine, and India is deep, skilled, and living at 10-20% of U.S. costs — any software business built by a U.S. company that needs a certain margin to placate investors will get competed away by a lower-cost version built offshore. It’s the Upwork thesis for the AI era, except the tools are better, the models are free, and the talent pool is global.

Which raises the question every coding-focused AI company needs to answer: do you need frontier LLMs to compete, or will specialized models do the trick? If it’s the former, token costs are coming down anyway as sufficiently capable open-source models get run locally with zero inference cost beyond electricity. If it’s the latter, the frontier labs may need to pivot yet again — because specialized models can be fine-tuned from open-source bases at a fraction of the price. Either way, the cost of intelligence is falling. And when the cost of the commodity falls, the companies built on commodity margins fall with it.

Why this matters: If you’re building a software business that depends on frontier AI pricing holding steady, stress-test your model against a world where inference costs drop 80% in 18 months. That’s not a prediction. It’s a trendline. And if your competitive moat is “we built this in the U.S.,” remember that the developer in Bangalore has the same open-source models, the same Stack Overflow, and a cost of living that makes your margins look like a luxury tax.

The Token Becomes the Wage

💲 Jensen Huang dropped a line at GTC that most people treated as a throwaway. He suggested companies should offer token budgets as part of compensation packages — up to half of base salary. For a senior engineer, that’s $250,000 a year in AI compute credits.

Dismiss it if you want. But follow the logic.

Engineers at frontier AI companies are already comparing token usage on internal leaderboards. Agent workloads consume 15x more tokens than chat interactions. As coding tools move from copilot to autopilot — where agents run multi-hour sessions, debugging, testing, and shipping code autonomously — token consumption per engineer is growing exponentially. The cost of an AI-augmented engineer isn’t salary plus benefits. It’s salary plus benefits plus compute. And the compute line item is growing faster than the other two combined.

If tokens are part of your comp package, they’re a commodity. And commodities get priced by willing buyers and willing sellers.

Here’s how the market works today. Anthropic sells Claude Opus output tokens at $25 per million. OpenAI prices GPT-5.4 output at $15. Kimi K2.5 output? $2.20. These are primary market prices, set by the labs. But the moment tokens become compensation — the moment an engineer receives $125,000 in Claude tokens as part of their package — a secondary dynamic emerges naturally. That engineer might need GPT-5.4 tokens for a specific task. Or they might realize their work can be done on Kimi at one-eleventh the output cost, and would rather convert the difference to cash.

Think of it like airline seats. The airline sets ticket prices in the primary market. But empty seats at departure represent wasted capacity. So discount markets, last-minute fares, and aggregators emerge to fill them. The same dynamic applies to compute. Models sitting at 40% utilization have excess capacity. Price discovery — willing buyer, willing seller, settled in dollars — drives utilization up. But it also forces the labs to acknowledge something they’d rather not: that their model has slack in demand. No lab wants to admit excess capacity. But the market will find the clearing price whether the labs participate or not.

We’re not talking about a crypto-style exchange. We’re talking about something simpler: a marketplace where compute is bought and sold at transparent prices, tied to API keys. A token wallet. The infrastructure already exists in fragments across every cloud provider’s spot instance market. Somebody just needs to stitch it together.

The deeper question is what happens when this market matures. If tokens are compensation, and compensation gets arbitraged, and the arbitrage reveals that 90% of enterprise workloads can be served by open-source models at near-zero cost — the frontier labs’ pricing power evaporates. Not because their models aren’t better. Because “better” doesn’t matter when “good enough” is free. The gap between frontier and open-source is narrowing every quarter. At some point — maybe this year, maybe next — the gap shrinks below the threshold where most buyers can tell the difference.

Jensen doesn’t care which way this plays out. He’s selling the GPUs to all of them.

The action item: If you’re negotiating compensation at an AI company, pay attention to the token component. If you’re building on frontier APIs, model your costs against open-source alternatives at 12-month intervals. And if you’re a lab CEO, ask yourself: would you rather build the exchange, or let someone else build it and discover that your pricing was a fiction?

On Our Radar

We’re tracking a few stories that didn’t make the main briefing but deserve your attention:

Halter raises $220M at $2 billion valuation. Virtual fencing for cattle has gone global. Halter’s GPS collars train livestock to stay within invisible boundaries using audio cues and mild vibration. It sounds like niche agtech until you realize livestock management is a $130 billion global industry and the same spatial intelligence that keeps a cow in a paddock is the same capability every robotics company is chasing. When AI fences cows better than barbed wire, the technology is real. Everything else is a matter of application.

LiteLLM supply chain attack. Version 1.82.8 of the popular AI proxy library — 97 million monthly downloads — was compromised with base64-encoded malware exfiltrating SSH keys, cloud credentials, and API keys. NASA, Netflix, Stripe, and Nvidia among those exposed. Andrej Karpathy called it “software horror.” As we hand agents the keys to our systems, one compromised dependency can empty the vault.

Arm makes its first chip in 35 years. After decades licensing IP, Arm built physical silicon: the AGI CPU with 64 CPUs, 8,700 cores, 2x performance-per-watt vs. x86. Meta co-developed it. OpenAI, Cloudflare, and SAP signed on. Targets a $1 trillion market. The architecture wars are back.

GPT-5.4 ships. Available in ChatGPT and via API. Strong benchmarks. Incremental, not paradigm-shifting. The model treadmill keeps spinning.

What This Means For You

OpenAI just demonstrated what every company eventually learns: strategy is subtraction. The kill list tells you more than any product roadmap. But they’re subtracting toward a market that already has a leader and a cost structure that’s collapsing from below.

Stop confusing focus with advantage. OpenAI’s pivot to coding is rational. It’s where the revenue is. But arriving focused doesn’t mean arriving first. If your strategy depends on matching a competitor who’s been building trust and market share for three years, you’re playing their game on their field.

Stress-test your margins against open-source at 12-month intervals. Chinese models aren’t a curiosity — they’re in production at Airbnb, Pinterest, and Notion. If you’re building on frontier APIs without an exit ramp, you’re one pricing disruption from a margin crisis. The performance gap between proprietary and open-source is narrowing faster than your procurement cycle can adapt.

Model the full cost of an AI-augmented workforce. Salary plus benefits plus compute. When tokens become compensation, every efficiency gap between frontier and open-source shows up in your P&L. The company that figures out token economics first builds a structural advantage that compounds every quarter.

The kill list is the strategy. But the market doesn’t care what you’ve killed. It cares what you can build with what’s left — and what it costs.

Three Questions We Think You Should Be Asking Yourself

If Chinese open-source models are good enough for Pinterest and Airbnb, why are we paying 8x for frontier APIs?

Run the comparison. Not on benchmarks — on your actual workloads. The performance gap between proprietary and open-source has narrowed to the point where the question isn’t quality. It’s whether your procurement process has caught up to reality.

What happens to our talent strategy when our best engineers’ projects get killed overnight?

OpenAI just reassigned the Sora team to robotics. In a market where top AI engineers can leave tomorrow, “pivot and be grateful” isn’t a retention plan. If your company is reorganizing around AI, have the honest conversation about continuity before your best people have it with a recruiter.

If tokens become part of compensation, what does that do to our cost structure?

Jensen’s suggestion isn’t theoretical anymore. Agent workloads are growing exponentially. The cost of an AI-augmented workforce isn’t salary plus benefits — it’s salary plus benefits plus compute. If you haven’t modeled that third line item for 2027, you’re budgeting for a world that no longer exists.

“The essence of strategy is choosing what not to do.”

— Michael Porter

— Harry and Anthony

Sources:

• OpenAI Enters Its Focus Era by Killing Sora (WIRED)
• Why OpenAI and Disney Ended Their Deal (Variety)
• OpenAI Looks to Double Headcount to Catch Up With Anthropic (Futurism / Financial Times)
• Why China Is Winning The Open Source AI Race (Forbes)
• Tomasz Tunguz: AI’s Bundling Moment
• Halter raises $220M at $2B valuation
• LiteLLM Supply Chain Attack — Andrej Karpathy
• Arm AGI CPU Announcement
• Ramp corporate spend data: Anthropic vs OpenAI
• Cisco LLM Security Leaderboard
• Clément Delangue on Chinese open-source dominance (Hugging Face)