The AI Brief

Gov. Kathy Hochul signed an executive order on July 14, 2026 imposing the nation's first statewide moratorium on new hyperscale data center construction in New York, pausing state permitting for up to one year. The order directs state regulators to develop new standards covering environmental impact, energy demand, water usage, and grid stability before new facilities can break ground.

Hochul opted for an executive order rather than the legislative moratorium bill the state legislature had passed earlier this year, her office described the bill as too complex and needing additional work, while the executive order takes effect immediately upon signing. The governor simultaneously announced plans to repeal sales-tax exemptions for large data centers in New York. Alongside the data center action, Hochul signed several tech-governance measures including an AI deceptive practices act and a ban on AI-generated child sexual abuse material.

Tech industry groups have argued that data center restrictions cede competitive ground to China and eliminate local jobs. The moratorium is expected to carry political weight for Hochul's reelection campaign, as rising energy bills and water-use concerns have become mobilizing issues in tight New York congressional districts this fall. Moratoriums have been proposed in at least a dozen states but have previously not advanced; New York's executive order provides a faster-moving model. The state's current footprint in hyperscale infrastructure is limited, Virginia and Texas remain dominant, but the precedent applies nationally to any state where a governor can act unilaterally on permitting.

The Financial Times, Bloomberg, and Reuters all reported on July 14 that DeepSeek, the Hangzhou-based AI lab, has opened preliminary discussions with new investors about a round targeting a pre-money valuation of roughly $71 billion, up from the approximately $52 billion post-money valuation it achieved when it closed its first-ever external fundraising of about $7 billion around the end of May 2026. TechCrunch reported the new round is targeting approximately $1.5 billion. Among reported prospective backers are Tencent and Beijing's national artificial intelligence industry investment fund. DeepSeek declined to comment.

Simultaneously, Bloomberg reported that DeepSeek has begun IPO preparations, engaging accounting firms and banking advisors, with a target of completing financial statements by December 2026 and filing for a mainland Chinese listing as early as late 2026 for a 2027 debut. No exchange has been formally identified. Founder Liang Wenfeng holds approximately 78% of equity and was the single largest investor in the May round, contributing roughly $3 billion of his own capital. DeepSeek's aggressive capital push is driven by three stated priorities: constructing a gigawatt-scale proprietary data center campus, accelerating development of in-house AI inference chips to reduce reliance on Nvidia and Huawei Ascend processors, and scaling AI agent product development.

DeepSeek's commercial footprint has grown substantially: as of June 2026, the company accounted for nearly 23% of tokens processed through Vercel's enterprise AI gateway, with Anthropic holding 32%. Its DeepSeek-V4-Flash API model ranked first globally by API call volume. The IPO trajectory tracks alongside confidential filings already made by Anthropic and OpenAI targeting valuations up to $1 trillion, and SpaceX's June 12 listing that raised $86 billion, the largest IPO in market history.

Google DeepMind CEO Demis Hassabis published a personal manifesto on July 14, 2026, titled "A Framework for Frontier AI and the Dawning of a New Age," calling for the US to establish a Frontier AI Standards Body. In an exclusive interview with Axios, Hassabis said he has held private talks with the Trump administration, European officials, and rival lab leaders over recent months, and described the White House's response as "very positive." He told Axios his timeline is "months," with the body ideally operational before year-end.

The proposed body would be modeled on FINRA, a federally overseen public-private partnership funded by industry, with a majority-independent board of Turing Award winners and credentialed technical experts alongside open-source representatives. It would develop assessment protocols covering cybersecurity, biological, and nuclear risk domains, with tests refreshed quarterly and benchmarks updated as capabilities evolve. Initially, frontier labs would voluntarily share models up to 30 days before release. Once the protocol is "shown to be effective and robust," Hassabis writes, formalization "could quickly follow", making the review a mandatory deployment gate for the US market. The rules would apply to all frontier-class models "no matter their country of origin or whether they are open or closed."

Hassabis cited the Trump administration's improvised June shutdown of Anthropic's Fable 5 and Mythos 5 as "a bit of a wake-up call", proof Washington needs a structured mechanism rather than ad hoc directives. He told Axios he believes AGI is "probably only a few short years away" and that the window to establish governance is "precious." Agentic-specific tests would look for deception attempts, safety-guardrail bypassing, and model reasoning transparency. The body would eventually build held-out evaluation capacity independent of what labs can optimize against, and would coordinate with US National Labs for national-security-relevant testing.

Anthropic launched Claude for Teachers on July 14, 2026, giving verified US K-12 educators free access to premium Claude capabilities for one year, including Claude Code, Claude Cowork, a Learning Commons connector carrying academic standards for all 50 states, and a library of teaching skills built with learning scientists. Sign-up is open through June 30, 2027. The product is teacher-facing only, consistent with Claude's 18-and-over policy; a dedicated district and school offering is described as "coming soon."

The platform ships with nine education connectors at launch: ASSISTments, Brisk Teaching, Canva Education, Coteach, Diffit, Eedi, MagicSchool, Snorkl, and TeachFX. Anthropic developed the offering in partnership with Learning Commons (a Chan Zuckerberg Initiative project), Teach for America, and the American Federation of Teachers, whose president Randi Weingarten called the product's privacy commitments a "Gold Standard" for K-12 AI. Data from Claude for Teachers is not used for model training, and student data is covered by a K-12 Data Processing Addendum written to align with FERPA. Anthropic will pilot an evaluation of the product in the Detroit Public Schools Community District to study impact on educator well-being and instructional practice.

Claude for Teachers enters a market where OpenAI, Google, Microsoft, and Khan Academy already have teacher-specific AI products. The edtech category is politically complex: the American Federation of Teachers, the Trump administration, and Bill Gates have all encouraged AI adoption in classrooms, while a growing parent-led backlash against technology use has prompted some major school districts to rethink ed-tech contracts. Anthropic's partnership with the AFT is a deliberate hedge against that headwind. The Gates Foundation partnership announced alongside the launch is aimed at co-developing tools for K-12 student outcomes over the longer term.

Disclosure: Claude, which generates this brief, is built by Anthropic.

Baichuan Intelligence, a Beijing-based AI lab, released Baichuan-M4 on June 22, 2026, developed in collaboration with a Tsinghua University research team. The model is designed as a clinical-grade agent system for continuous-care scenarios, multi-turn patient management, chronic disease follow-up, and multimodal medical image understanding, rather than single-turn medical question-answering.

On the HealthBench Hard subset, Baichuan-M4 scored 15.9 points above GPT-5.5, the second-best model, per the arXiv preprint (arXiv:2606.08982). On all three HealthBench dimensions, overall, hard, and professional, Baichuan-M4 leads the field. Its hallucination rate of 3.3% is the lowest publicly reported for any large-scale medical AI model. The model is open-source and available on Hugging Face.

The architecture is built around three components: Baichuan-Harness, a unified training-and-deployment runtime that enforces action constraints, long-term patient memory, and multi-agent coordination; a core reasoning model trained on a continuous-care reinforcement-learning framework integrating span-level reward modeling (SPAR++) and curriculum learning; and a clinical tool layer for multimodal perception across documents, X-rays, and dermatology images. Each generated medical conclusion is cited to a specific passage in an authoritative paper or clinical guideline. The model's predecessor, Baichuan-M3, had already surpassed GPT-5.2 on HealthBench overall. OpenAI's GPT-5.5 had previously held the strongest general-purpose model position on this benchmark. The M4 result is a vendor-published benchmark on the lab's own arXiv preprint; independent replication has not yet been published.