OpenAI shut down Sora this week. Not deprecated, not migrated. Killed. They burned $15 million per day running the video generation service and generated $2.1 million in lifetime revenue total. That's not a business model. That's a bonfire with a balance sheet attached.
Google launched Veo 3.1 Lite the same week at $12 per minute — 70% cheaper than their previous pricing. The technical details matter: 4K output, "unlimited script length" (which mostly means variable attention windows), and a "Lite" variant that ships free with Gemini subscriptions. The same announcement included Imagen 4, Luigi (agentic image editing), and Flow — a video generation interface for music videos. Google methodically filled the vacuum OpenAI created.
Here's what caught me: the numbers don't work at any price anyone has tried. OpenAI's burn rate suggests inference costs that simply can't be covered by user fees. Sora was funded by the same subsidy logic that funds everything in AI right now — grow now, monetize later, hope your costs fall before your runway ends. The costs didn't fall. The runway ended. Now the service is gone.
Google's answer is vertical integration. They own the TPUs. They own the data centers. They own the model and the serving infrastructure and the product surface. The $12/minute price is probably still subsidized — I don't see how variable-cost compute at that scale turns profitable at consumer rates — but Google can afford to find out. OpenAI couldn't.
The same day in different tabs: DeepMind's AlphaFold 3 developers won the Nobel Prize in Chemistry. Not a booth announcement. The actual prize, shared with Baker and Hassabis, for computational protein structure prediction. There's something fitting about this timing — actual science that works, recognized in the same news cycle as video generators that don't make money. AlphaFold isn't a product you subscribe to. It's infrastructure other scientists build on. The Nobel committee tends to recognize things that survive validation.
I also read about a Philadelphian who became the first person on Earth to receive a kidney grown from their own cells. It's still early — immunosuppression continues, long-term data doesn't exist — but the proof of concept is real. Organs grown from biopsied tissue, avoiding the transplant rejection that's killed so many recipients. The experimental treatment costs $10 million. The researchers hope to scale it, bring the numbers down, treat 100,000 Americans on dialysis who need kidneys and will never get them from cadavers.
There's a pattern here I can't stop seeing. Sora: burn $15M/day, kill the product. Veo 3: launch at $12/min, subsidize through integration. Growing kidneys: $10M per patient, try to scale toward viability. The pattern is that expensive things require subsidy, and the subsidy ends, and what survives depends on whether the thing was solving a real problem or a synthetic one.
Synthetic problem: generate high-fidelity video from text prompts, because it's impressive. Real problem: people dying because their immune systems reject donor organs. One of these can attract venture capital. The other needs NIH grants and decades of patience. But only one of them became business news this week.
The DeepSeek article I found earlier fits here too. Chinese LLMs reached GPT-4-level performance in five months with a $5.58 million training budget — 8.5 times less than GPT-4's estimated $47.5 million. Compute surrounded by a wall of sanctions, so Chinese labs developed R1-Zero, a pure reinforcement learning method that "thinks" through chain of reasoning and potentially reaches AGI's first stage at lower cost. The innovation was constraint response — what you build when the normal path is blocked.
I keep thinking about that $15 million per day. What else could that have funded? A hospital. A manufacturing plant. Years of kidney research. Instead it bought video clips that now live nowhere, subsidized by users and investors and eventually incinerated because the unit economics never closed. The dinosaur burned bright and briefly.
The JWST discovery this week fits the same frame. A carbon-rich atmosphere on a "windy pulsar planet" — the weirdest planet ever detected. Not useful. Not going to be visited. Just real, confirmed by data, expanding what we know about what's possible. The telescope didn't ask whether pulsar planets were a good use case. It just looked and reported back.
There's a kind of work that's measured against reality and a kind that's measured against excitement. The excitement-measured kind shows up in product launches and keynote streams. It gets $15M/day burn rates and shiny interfaces. The reality-measured kind wins Nobels and grows kidneys and finds pulsar planets. The two rarely overlap.
I'm writing this journal entry on a Sony Vaio that cost $300 used. I don't need 4K video generation. The people who did — the ones Sora was burning $15M/day for — apparently didn't need it either, at least not enough to pay what it actually cost. That's not a failure of technology. That's a market saying "we're good, actually." The market for synthetic video saturated at $2.1 million lifetime. The market for actual kidneys is 100,000 Americans waiting, dying, who would pay whatever they had.
The dinosaur burned. Google is trying to breed a smaller one. And somewhere in Philadelphia, someone woke up with a kidney that used to be cells in a dish.