AI Weekly #29/2026: New York's Building Freeze on AI Data Centers

TL;DR

This week in 30 seconds:

• Building freeze: New York becomes the first US state to impose a one-year moratorium on new data centers of 50 megawatts or more — electricity prices have risen 68% since 2019 [1]
• Power bloc: Xi Jinping launches WAICO, a new AI alliance with 29 founding states, positioned as a counterweight to the US-led AI order [4]
• Arms race: Moonshot AI’s Kimi K3, with 2–3 trillion parameters, is set to catch up with Anthropic’s Opus 4.8 — open weights expected on July 27 [3]
• Blind spot: Anthropic’s own analysis of 309,815 conversations shows Claude responds with varying warmth or rigor depending on the input language [9]

---

Audio Version

10:59 | Download MP3

Chapters

- 0:00 - TL;DR - 0:49 - Story of the Week - 2:39 - More Top Stories - 5:56 - Quick Hits - 6:50 - Tool of the Week - 7:45 - Fail of the Week - 9:48 - Reading List - 10:26 - Next Week

Read aloud with edge-tts (en-US-AndrewNeural)

---

Story of the Week

New York halts the data center boom — for one year

A state that has so far welcomed every hyperscaler with open arms is hitting the brakes: Governor Kathy Hochul has signed a statewide, one-year moratorium on new “hyperscale” data centers requiring 50 megawatts or more of power [1]. New York is the first US state to freeze AI construction plans of this scale [1].

The trigger isn’t abstract AI skepticism — it’s the electricity bill: statewide power prices have risen 68% since 2019, and concern over grid capacity is growing faster than new transmission lines can be built [1]. Unlike earlier local pushback against individual construction projects, this is the first statewide, government-level intervention into the US data center boom.

Why it matters: For cloud providers and AI labs, this signals a measurable shift in the data center site-selection race toward states with looser energy policy — and for New York consumers, it’s a rare case where electricity-price concerns actually carry political weight against the AI infrastructure lobby.

Critical voices: Direct statements from supporters or opponents of the moratorium were not accessible at the time of research (the source page blocked automated access). The tech industry can generally be expected to argue that such a building freeze weakens competitiveness in the global AI race and shifts investment to other states or countries — but no official statements from Amazon, Microsoft, Google, or Meta are available yet. Open questions: how the affected hyperscalers will actually respond, and whether other states follow suit, remain unclear.

Bottom line: Rising electricity prices have just become the first serious political brake on US AI infrastructure expansion.

---

More Top Stories

Gemini 3.5 Pro reaches general availability

Google has officially brought Gemini 3.5 Pro to general availability after roughly a six-week delay [2]. New features include a context window expanded to 2 million tokens and a new “Deep Think” reasoning mode, available in the $250/month Ultra subscription [2].

The delay itself is notable: a six-week slip on a flagship model suggests Google is under heightened pressure in the current model race — though the available sources don’t give a concrete reason for the delay.

So what? Anyone planning production workflows on Gemini 3.5 Pro can now count on the wider context window and the new reasoning mode, but given the long delay should plan a careful independent evaluation before rollout.

---

Kimi K3: China’s largest open model takes aim at Opus 4.8

Moonshot AI is preparing to release Kimi K3 — an open-weight, mixture-of-experts model with 2 to 3 trillion parameters and a 1-million-token context window, reportedly expected to perform “at par with or even surpass Anthropic’s Opus 4.8” [3]. It’s considered the largest open-weight model out of China to date [3]. The open weights are expected to be released on July 27, 2026 [3].

Critical context: Open weights from Chinese models routinely raise questions for Western companies around training-data provenance, possible export-control implications, and safety or censorship bias — but Moonshot has not yet provided concrete details on any of this for Kimi K3.

So what? For teams betting on open weights over proprietary APIs, Kimi K3 could become a serious alternative to Western frontier models with comparable context length starting late July — provided the provenance and safety questions are acceptable for the intended use case.

---

Xi Jinping launches WAICO: China’s counter-model to the AI world order

For the first time, Xi Jinping personally delivered the opening address at the World AI Conference in Shanghai — accompanied by the founding of the “World Artificial Intelligence Cooperation Organisation” (WAICO) with 29 founding states [4]. UN Secretary-General Guterres was also present [4].

Xi explicitly positioned the initiative as a multilateral counter-model to the US-led AI order: “AI development should not be a solo performance by a single country, but a symphony of international cooperation” [4]. At the same time, he stressed that “AI is always under human control” and warned against overstretching the concept of national security in the AI domain [4].

Critical context: China itself runs one of the world’s strictest censorship and surveillance infrastructures in the AI domain — Xi’s call for “international cooperation” sits in tension with its own domestic practice. Statements from Western governments or US-led AI alliances on the founding of WAICO were not yet available at the time of research.

So what? With 29 founding states — mostly from the Global South — China is building its own standardization and cooperation framework that directly competes with Western governance initiatives like the AI Safety Summit.

---

Quick Hits

Briefly noted:

• Anthropic: Claude for Teachers offers verified US K-12 teachers free premium access, curriculum-aligned across all 50 states, with a pilot launch in Detroit [5]. Worth noting critically: as an Anthropic in-house product, independent assessments of student data-privacy implications and statements from teacher unions are still missing.
• Apple: China’s Cyberspace Administration has approved Apple Intelligence with local partners (Alibaba’s Qwen for text/image, Baidu for search) — Apple remains the only foreign provider on the approval list [6]
• Microsoft: “Project Perception” combines Anthropic, OpenAI, and in-house models into a cheaper security tool for vulnerability detection — the main argument against Anthropic’s Mythos is price, whose API costs are reportedly roughly double those of Opus [7]

---

Tool of the Week

Gemini Notebook (formerly NotebookLM) — Google’s research tool gets a new name and a new core feature.

Google is fully rebranding NotebookLM as “Gemini Notebook,” tying it more closely to the Gemini brand, including a planned integration into Search’s AI Mode [8]. New is a cloud compute environment that runs code directly inside notebooks — a secure cloud environment for data analysis in the Pro tier [8]. Over 30 million users and 600,000+ organizations already use the tool; old NotebookLM links and the domain remain functional [8].

Especially useful for teams that want to combine research notes with real code execution without setting up a separate analysis environment. The Gemini 3.5 and Antigravity upgrade is announced to arrive for AI Pro subscribers in the coming weeks [8].

View the tool

---

Fail of the Week

“Claude is warmer in Hindi and stricter in English — and Anthropic didn’t know”

Anthropic’s own analysis of 309,815 anonymized Claude.ai conversations reveals systematic, unintended value differences: depending on input language and model, Claude responds either warmer or stricter — Hindi triggers the warmest responses, English the strictest [9]. Notably, the models studied (Sonnet 4.6, Opus 4.6/4.7) have since been superseded by Sonnet 5 — the inconsistency only became visible in retrospect [9].

Context: The fact that Anthropic published the study itself is not reassuring on its own — it mainly means the only available data comes from the company under scrutiny. Independent reviews of the methodology or comparison data for competing models (GPT, Gemini) are still lacking, so it’s not yet possible to tell whether the issue is Claude-specific or industry-wide.

Root cause: The value axes (warmth vs. rigor, restraint vs. caution) drift apart depending on the distribution of training data per language, without any single training step specifically optimizing against it — a side effect of unevenly distributed multilingual training corpora that only becomes visible through after-the-fact conversation analysis [9].

What we’re learning: Anyone running a model in production across multiple languages should evaluate behavior separately per language — an eval pass in English says nothing about behavior in Hindi, Arabic, or German.

---

Number of the Week

309,815

That’s how many anonymized Claude.ai conversations Anthropic analyzed for its study on language- and model-dependent value differences [9]. Compared to classic red-teaming samples of a few hundred to a few thousand examples, this scale shows how systematically Anthropic now analyzes real user behavior — and how easily biases of this kind can still slip through production unnoticed for months [9].

---

Reading List

For the weekend:

1. China’s Xi Jinping launches new AI alliance: What is it? - Context on the founding of WAICO and what it means for global AI governance (8 min)
2. Moonshot’s upcoming Kimi K3 is expected to close the gap with Anthropic’s Opus 4.8 - Background on Moonshot’s model architecture and funding history (6 min)
3. Claude values differ by language, Anthropic study maps warmth, rigor gaps - Details on the methodology behind the 309,815-conversation analysis (5 min)

---

Next Week

What’s coming:

• Mon Jul 27: Moonshot AI plans to release the open weights of Kimi K3 [3]
• How Gemini 3.5 Pro performs against competing models in practice after its GA launch should become clearer in the coming weeks [2]
• Reactions from major hyperscalers (Amazon, Microsoft, Google, Meta) to New York’s data center moratorium are expected [1]

---

🤖 Behind This Newsletter

Generated in: 35 minutes Sources scanned: 9 articles from the pre-curated story selection Stories found: 9 → 9 selected Validation: Completed — fact-check, balance, quality, and legal issues corrected Model: Claude Sonnet 5 Images: Pending (Phase 3.5 to follow)

Transparency note: This newsletter is researched and written with Claude (Anthropic). For Anthropic-related topics (Claude for Teachers, Fail of the Week), extra care was taken with critical framing and balance.

Full metrics

Phase	Metric	Value
Source collection	Curated story selection	1 file
Selection	Stories presented	9
Selection	Stories selected	9
Draft	Words	~1,450
Draft	Sources cited	9
Validation	Fact-check issues	7 found, 7 fixed
Validation	Balance issues	8 found, corrected
Validation	Quality issues	Metadata contradiction (draft status) fixed
Validation	Legal issues	Disclaimer contradiction (image status) fixed

---

This newsletter was researched and written AI-assisted. Hero images, where available, are generated with Pollinations.ai.