DEV Community

TechSparkLive
TechSparkLive

Posted on

OpenAI Hits $500B While China Cuts AI Costs in Half

This week showed AI splitting into two camps: those spending billions on computing power, and those finding ways to do more with less. OpenAI became the world's most valuable private company while reversing course on copyright, China's DeepSeek cut costs in half with smarter technology, and everyone else is scrambling to build enough servers to keep up.


1. OpenAI Hits $500B Valuation Through Employee Stock Sale

OpenAI sold $6.6 billion in shares to employees, valuing the company at $500 billion, the highest ever for a private company. Buyers included SoftBank, Dragoneer, Thrive Capital, MGX, and T. Rowe Price.

This wasn't fundraising. The cash went directly to employees, not OpenAI, functioning as a retention tool to stop talent from walking out the door.

Why now? Meta has been aggressively poaching talent, nabbing at least 7 top engineers this summer with multi-million dollar bonuses. This stock sale lets employees cash out without leaving.

The risk: OpenAI's for-profit conversion isn't court-approved yet, and this valuation could complicate things if the conversion fails.

OpenAI can raise billions despite burning billions, but the $500B valuation only works if they successfully convert to for-profit and justify the massive spending.


2. DeepSeek Cuts AI Running Costs in Half

DeepSeek

DeepSeek's V3.2 model uses "Sparse Attention" to cut operational costs by ~50%. It's free and open-source on Hugging Face.

How it works: Instead of processing everything (expensive), it scans for the most relevant sections with a "lightning indexer," then focuses only on key parts through "token selection." Gets the same answer using half the computing power.

This tackles running costs, how much it costs to operate AI after it's built. While most companies focus on training costs, DeepSeek is making daily operation cheaper.

DeepSeek's January model already showed cheaper training, and now they're cutting running costs too, demonstrating a consistent pattern: smarter engineering over brute spending.

Bottom line: If this scales, billions in U.S. infrastructure spending could look wasteful. China's winning through efficiency while the U.S. bets on capital.


3. Claude Codes Autonomously for 30 Hours

Claude Sonnet

Claude Sonnet 4.5 launched with state-of-the-art coding performance at the same pricing: $3 input / $15 output per million tokens.

The breakthrough: It codes autonomously for 30 hours in enterprise tests, sets up databases, buys domains, runs security audits, and creates production-ready apps, not prototypes.

Validation:

  • Cursor CEO: "Best coding performance for long projects"
  • Windsurf CEO: "New generation of coding models"
  • Apple and Meta use Claude internally

Previous AI wrote code snippets. This manages multi-day software projects with minimal human input. The jump from coding assistant to coding employee.

Anthropic is defending its lead while redefining what coding AI means, but the question remains:

Will companies actually trust 30-hour autonomous sessions?


4. OpenAI Reverses Sora Copyright: Opt-Out to Opt-In

OpenAI switched Sora from "opt-out" to "opt-in" for copyright, meaning studios must now give explicit permission before their characters can be used.

What changed:

  • Copyright holders control exactly how characters appear
  • Can ban use entirely or set specific rules
  • Similar to face/likeness controls, but for fictional characters

Altman hinted at revenue sharing. OpenAI takes a cut, studios get paid. He admits "some edge cases will get through," acknowledging the system won't be perfect.

OpenAI is learning that AI video can't follow the same "move fast and break things" playbook as text. Hollywood has lawyers, lobbyists, and political power, and the opt-in model is OpenAI accepting that reality.

Sora


The infrastructure crunch is real. OpenAI, Meta, and Anthropic are spending hundreds of billions on capacity, except DeepSeek, which is finding ways to need less.

The industry is splitting between unlimited computing power and doing more with less. This week showed both strategies in action. The question is which one actually works when the bills come due.

Top comments (0)