Artificial Intelligence

Data Center Delays, Power Bills, and Open Models: AI's Infrastructure Bill Comes Due

By 老金带你玩AI · Nov 7, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Every AI data center built affects electricity bills, water use, and equipment supply chains halfway around the world. When a Texas governor pauses grid connections, Chinese suppliers' stocks drop the next day; when U.S. ratepayers get stuck with stranded infrastructure costs, that risk is now a real liability AI companies have to price in.

Summary

A global tech selloff on November 5-6 followed a simple worry: AI companies keep spending, but new U.S. data centers are stuck on grid connections, permitting, and local opposition. Texas' incoming governor wants to pause projects, the current governor has ordered ERCOT reviews and paused environmental permits, and a $1.5 billion campus in San Marcos was rejected by a 5-2 city council vote. The market reaction hit Chinese optical module vendors like Zhongji Innolight and Eoptolink, which issued reassurances that orders already cover 2027 and some of 2028.

The article contrasts U.S. grid bottlenecks with Chinese projects in Ulanqab, where wind, solar, storage, and data center load are dispatched together, and a dedicated line connects a wind farm 16 km away. China's green power rules require the source, facility, and line to be built and commissioned as one unit. It then looks at who pays for grid upgrades and backup capacity: Dallas Fed estimates data centers have raised U.S. wholesale power prices by about 6.9% nationally and 15% in Northern Virginia, and OpenAI has pledged to cover its own grid costs. On the model side, DeepSeek's MIT-licensed V4.1-Flash and open DeepInfer tools let teams self-host or pay per token, with idle-time Flash API pricing at 1 yuan per million input tokens and 4 yuan per million output tokens.

Takeaways

— On November 5, U.S. Eastern time, the Nasdaq fell 5.52%, the Philadelphia Semiconductor Index fell 12.25%, Nvidia fell over 9%, and AMD fell over 17%; the next day the Nikkei fell 5.96%, the KOSPI 6.02%, the Shanghai Composite 1.10%, and the ChiNext 3.96%.

— Texas Governor Abbott ordered a review of data centers queued in the ERCOT interconnection process on August 3, and in September asked to pause environmental permits until the review is complete.

— Incoming Texas governor Gina Hinojosa has proposed pausing data center projects until resident protections and community input are in place.

— In San Marcos, Texas, the city council voted 5-2 in February to reject a $1.5 billion data center zoning change, and in June excluded data centers from permitted land uses.

— Chinese optical module suppliers Zhongji Innolight, Eoptolink, and T&S Communications issued announcements on November 6 saying orders are normal; Zhongji Innolight said major customer orders cover most of 2027 and some customers gave clear 2028 demand guidance.

— U.S. Energy Information Administration data for the first nine months of 2025 shows wind and solar met 36% of electricity demand in Texas' ERCOT region, with storage shifting daytime generation to evening.

— China's Envision Xinghe base in Ulanqab dispatches wind, solar, storage, and data-center load together and supplies 80% of its campus electricity from its own wind farm.

— Zhongjin Data's Ulanqab base built a dedicated 16 km line to a wind farm and has maintained stable power since July 2025, supporting training of multiple large models.

— Ulanqab's average annual temperature is 4.3°C, giving data centers nearly 10 months of natural cooling; Century Internet uses reclaimed water for summer cooling and consumes 0.099 liters of water per kilowatt-hour in its 2025 operating projects.

— A Dallas Fed model revised in October 2026 estimated data center demand raised U.S. wholesale electricity prices by about 6.9% nationally and about 15% in Northern Virginia through 2025.

— OpenAI said on November 5 that for campuses it owns and operates, it will pay all electricity and related infrastructure costs itself and provide financial assurance before large expenditures.

— China's new-energy nearby-consumption charging rules, in effect since October 2025, let grids charge transmission and distribution fees for reserved capacity, with an option of capacity fee plus energy fee for high-reliability projects.

— DeepSeek-V4.1-Flash-1017 is released with MIT-licensed code and model weights, and DeepSeek has also opened DeepInfer inference tools.

— As of November 6, DeepSeek's official Flash API charges 1 yuan per million tokens for cache-missed input and 4 yuan per million tokens for output during idle periods; peak doubles both, and cache-hit input is cheaper.

— A hypothetical task of 1 million input tokens and 100,000 output tokens, all missing cache, costs 1.4 yuan at idle and 2.8 yuan at peak with DeepSeek Flash.

— China's Water Conservation Regulation blocks new water withdrawal permits for water-scarce projects that could reasonably use unconventional water but don't.

Conclusions

OpenAI's pledge to pay its own grid costs is a response to political risk, not charity; it turns regulatory uncertainty into a line item on the balance sheet, and it undercuts the argument that data centers only bring free investment.

The market's reaction shows how tightly coupled U.S. data center construction is to Chinese supply chains: a permitting delay in Texas moves optical module vendors' share prices in Shenzhen the next morning.

China's model of building the power source, the data center, and the transmission line as one synchronized project sidesteps the weakest link in U.S. expansion, which is the gap between generation capacity and the actual grid path that gets electricity into a building.

DeepSeek's MIT licensing and low per-token pricing put competitive pressure on proprietary API vendors, but the cost math still depends on a team's own operations budget; self-hosting shifts the infrastructure burden from the vendor to the developer.

The article repeatedly connects data center expansion to household water and power bills, a framing that reframes AI's infrastructure boom as a local public-utility dispute rather than a purely tech-sector story.

Concepts & terms

ERCOT

The Electric Reliability Council of Texas, the nonprofit grid operator that manages power flow for about 90% of Texas' electric load. It handles interconnection requests and grid dispatch.

Wholesale electricity price

The price paid for electricity between generators and retail suppliers, before retail markups and local distribution charges. Data center demand can raise this price for everyone buying from the same market.

Token

A small unit used by language models to meter text; roughly one token corresponds to a piece of a word or subword. Input tokens are the prompt material a user gives, output tokens are the text the model generates, and APIs often bill them separately.

Capacity fee plus energy fee

A two-part grid pricing structure: the capacity fee pays for reserved power supply and backup readiness, while the energy fee pays for the actual electricity consumed. It lets high-reliability data centers pay for the grid's standby capacity rather than only for usage.

Green power direct connection

A Chinese regulatory arrangement for new-energy projects serving data centers directly, requiring the power source, the data center, and the dedicated transmission line to be built and commissioned as a synchronized unit.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗