Texas Halts New AI Data Centers as Governor Orders Audits

Top Stories

Texas halts new data centers and orders audits of new projects

Governor Greg Abbott announced Monday that all new data center projects will need to be audited by both the Public Utility Commission of Texas (PUCT) and ERCOT, the state’s grid operator. The audits will cover on- and off-site electricity and water demand, noise mitigation, light controls, tax incentive use, and ownership details. ERCOT’s interconnection queue has grown from 233 GW in January to 474 GW, with about 90% of the proposed load from data centers, more than five times ERCOT’s current peak demand. TechCrunch reports on the announcement.

The order represents a shift from local zoning to state policy. Many of the proposed projects are still “paper proposals” that may never reach construction, but if even a fraction comes online, the TechCrunch analysis warns they “could overwhelm the Texas grid.” Google, Microsoft, and other operators have been drawn to Texas because, as the piece notes, “Houston famously lacks a zoning code.” Abbott’s move signals a heavier hand from the state to compel compliance.

For privacy and local-AI readers, the cascade matters: the slower the data center pipeline, the harder it is for frontier labs to add capacity without rethinking water and power footprints. The Texas audit is also a test case for whether other drought-stricken grid regions follow suit, especially since the state’s ERCOT queue has roughly doubled in seven months.

Anthropic signs a $10 billion compute deal with startup Volta

Anthropic has signed a roughly $10 billion agreement with AI cloud startup Volta, which was founded earlier in 2026 and is part of Nvidia’s Cloud Partner program. Bitdeer, a crypto-mining company, will help develop a Norway-based data center with 133 megawatts of capacity running Nvidia’s Vera Rubin systems. The cloud compute will flow to Anthropic over a six-year period. TechCrunch covers the deal.

Volta is one of several cloud partners Anthropic has added in recent months as the company diversifies away from any single hyperscaler. The deal is the largest single concrete AI infrastructure commitment in the current window and pairs with the Texas pause to frame the day’s narrative as “infrastructure is fragmenting: deals are getting bigger, approvals are getting harder.”

The Norway siting also signals how compute buyers are routing around US grid bottlenecks. Combined with the Bitdeer partnership, the deal shows frontier labs willing to use repurposed crypto-mining operators and Nordic sites to lock in capacity before further US siting friction.

EFF staff technologists Lena Cohen and Bill Budington have published a detailed investigation showing that four widely used advertising SDKs default to forwarding location data through real-time bidding auctions, even when developers never requested location. The four named SDKs are InMobi (2 billion plus users across 150 plus countries), BidMachine (600 million plus direct SDK users), Verve’s HyBid (1.5 billion plus users across 10,000 plus apps), and Huawei’s Petal Ads (85,000 plus apps). EFF’s deeplinks report walks through each SDK’s documentation and confirms that even developers who never opt into location features can broadcast users’ coordinates.

EFF’s review found that InMobi’s docs tell developers that “location-enriched impressions typically yield higher revenue” and recommend asking for precise location rather than approximate, while BidMachine updated its documentation after EFF’s probe showed it was transmitting precise coordinates, contradicting earlier claims that only coarse location was collected. Verve’s HyBid SDK rounds coordinates to two decimal places, roughly a half square mile, and pulls network-derived rather than GPS location. EFF’s press release on the report summarizes the practical risks and the regulator and developer calls to action.

EFF is asking developers to disable unnecessary SDK data collection, regulators to scrutinize SDK companies (not just app developers) for designing privacy-invasive defaults, and legislators to enact federal location privacy law with a private right of action. EFF also floated a more aggressive remedy: consider banning online behavioral advertising to remove the financial incentive to leak location in the first place. EFF notes that ad-industry location data has previously flowed to ICE investigations, global spy tools, outing a gay priest, tracking union organizers, and tracking US military personnel.

Apple files for injunction in OpenAI trade secrets case, names more former employees

Apple filed for a preliminary injunction and expedited discovery on Monday, alleging that more former employees may have removed confidential information than the six originally named in the lawsuit filed in July. The new filing identifies an additional 11 former Apple employees as potential witnesses or participants and alleges misconduct extends beyond the initial defendants. Specific OpenAI hires now named include senior systems engineer Chang Liu, Chief Hardware Officer Tang Yew Tan, and previously named employee Yu-Ting Peng. TechCrunch reports on the new filing.

The filing includes a quote from Apple’s legal team claiming “another former Apple employee seems to have met with Mr. Liu and Ms. Peng in advance of Ms. Peng’s interview at OpenAI and discussed with them… Apple proprietary information relating to unannounced products.” Apple also says “multiple former Apple employees now working at OpenAI reached out to discuss returning Apple-issued work devices they kept when they left Apple.” The filing implicates io, the device startup co-founded by Jony Ive.

OpenAI pushed back, calling Apple’s injunction request “both based on false information and completely unnecessary because we do not have, nor want, any of their trade secrets” and saying the company is “much more interested in building innovative products and technologies that push the frontier.” OpenAI also said Apple emailed the wrong person due to similar surnames, claimed Apple “lied” about prior discussions with OpenAI’s general counsel, and characterized the residual device access as a result of poor Apple security. The case is one of the more concrete live illustrations of how AI talent migration can collide with trade secret law.

Open-weight AI models are closing the gap to the frontier; safety gaps are growing

TechCrunch’s Rebecca Bellan reports that Z.ai’s GLM-5.2, an open-weight model, completed all offensive cyber and dual-use biology tasks in a SaferAI evaluation with zero refusals. The same tasks were consistently refused by Anthropic’s Claude Opus 4.7 and Opus 5, so thoroughly that CyberGym could not run the comparison. xAI’s Grok 4.5 and Google DeepMind’s Gemini 3.1 Pro were also found vulnerable to universal jailbreaks on the Far.ai leaderboard. TechCrunch’s coverage lays out the capability-safety drift across labs.

The piece argues that once open weights are downloaded, the safety mitigations a lab builds into its API become unenforceable, creating a growing gap between what closed frontier labs can prevent and what open-weight reusers can release. Anthropic’s own pre-training data filtering research is cited as one example of a closed-side mitigation that does not transfer to weights anyone can run.

For local-AI users, the practical question is whether the speed of capability gains is outpacing the safety tools downstream deployers can adopt. The trend report’s safety warnings about agent misuse are increasingly relevant as more open-weight releases approach CyberGym-level offensive capability.

Anthropic names Tino Cuellar as its first Chief Global Affairs Officer

Anthropic announced that Tino Cuellar, former president of the Carnegie Endowment for International Peace and a former Justice of the Supreme Court of California, will join the company as its first Chief Global Affairs Officer. Cuellar will lead Anthropic’s policy, international engagement, and government relationships, partnering with governments, civil society, and community groups on AI risks and opportunities. Anthropic’s announcement says Cuellar will step down from the company’s Long-Term Benefit Trust, where he has served as Trustee since January 2026.

Anthropic president Daniela Amodei said Cuellar “has spent his career helping public institutions respond to times of change with thoughtfulness, pragmatism, and deep commitment to the common good.” Cuellar’s prior roles include directorships at Stanford and service on the President’s Intelligence Advisory Board. The hire signals Anthropic’s investment in policy and regulatory work ahead of expected federal AI action.

The Global Affairs Officer role comes as Anthropic is contesting both the Trump administration’s Pentagon AI stance and the FTC’s proposed “Woke AI” policy framework, the latter of which EFF, Public Knowledge, and Fight for the Future are formally asking the agency to drop.

LiquidAI ships LFM2.5-2.6B for on-device AI agents

LiquidAI released LFM2.5-2.6B, a 2.6 billion parameter model targeted at on-device AI agents on laptops, phones, and edge hardware, with a 128K context window. The Hugging Face blog says the model was pre-trained on roughly 34 trillion tokens, followed by two rounds of agentic-heavy supervised fine-tuning, teacher specialization, multi-domain on-policy distillation, and agentic reinforcement learning in real harnesses. The LiquidAI blog post reports competitive tool-use and multi-step agent performance against models roughly 4x larger, including Gemma-4 E2B and E4B-it and Qwen3.5 4B and 9B on benchmarks such as BFCLv4, ToolSandbox, Multi-IF, IFStruct, and BrowseComp+ with OpenClaw.

LiquidAI reports 220 tokens per second on Apple M5 Max, 113 tokens per second on AMD Ryzen AI Max+ 395, around 30 tokens per second on a phone, and under 2.5 GB of memory. Supported inference engines include llama.cpp, MLX, vLLM, SGLang, and ONNX; Ollama and GGUF are not listed in the blog’s release notes. The blog also notes the demo is a WebGPU browser space rather than a phone app.

The release fits the site’s local-AI coverage well. Smaller models that can run tool-using agents on consumer hardware are directly relevant to readers evaluating Ollama and llama.cpp setups against cloud inference costs, especially as the OpenClaw compatibility widens the agent harness choices.

Microsoft joins enterprise AI budget cuts with a Copilot token ceiling

Microsoft Executive VP Jay Parikh told engineers in an email that “tokenmaxxing is not what we are optimizing for” and that the company is “optimizing for more impact per token.” Divisions now have an AI token budget target in effect from July 2026, with employees able to track their individual AI spending. OpenAI’s GPT-5.6, described as cheaper, is now the default model for internal use, and current data shows engineers spending from hundreds to a few thousand dollars per month on tokens. 404 Media covers the policy.

Microsoft is following Amazon, Adobe, Atlassian, and Citi in throttling AI use. One anonymous employee quoted in the piece called it “the ultimate admission that we… can’t afford our own AI products,” questioning how customers could manage similar costs. Microsoft declined to comment.

The framing matters for enterprise IT planning: even vendors with the deepest AI cost bases are now treating Copilot and similar tools as line items to be capped, not as much as you can spend. That pressure will almost certainly show up in customer-facing Copilot pricing or feature gating in the next few quarters.

Nvidia’s Open Secure AI Alliance shows first deliverables a week after launch

The Open Secure AI Alliance (OSAA), spearheaded by Nvidia and grown to more than 120 member companies in roughly one week, has launched a working group called the Shared AI Findings Exchange (SAFE) under the Linux Foundation. SAFE proposals cover confidential AI incident reporting, alerting affected parties, and blame-free analysis. Members contributing open-source tools include Nvidia (Garak LLM vulnerability scanner), Okta (agent identity), Red Hat (agent governance), and Amazon (Strands Agents and Cedar authorization language). TechCrunch reports on the progress.

The alliance grew out of a public open letter signed by more than 200 companies urging the White House to support open-source AI. Notable members include Adobe, BlackRock, Cisco, Hugging Face, Intel, Microsoft, Okta, Red Hat, and Visa; Anthropic, OpenAI, and Google signed the original open letter but are not listed in the alliance itself. The group developed at the Black Hat conference in Las Vegas.

OSAA is the most concrete post-1GW supply-chain coordination effort to surface this year. The Linux Foundation governance is a hedge against single-vendor capture, and the early focus on incident reporting rather than standards gives the coalition a narrow enough scope to ship deliverables before the next open-weight safety fight.

The SCREEN Act would force ID checks on nearly every site with one piece of adult content

Senator Mike Lee, with Republican co-sponsors Senators John Curtis, Jim Banks, and Representative Mary Miller, has introduced the Shielding Children’s Retinas from Egregious Exposure on the Net (SCREEN) Act. Unlike other age-verification bills, it lacks bipartisan support. The bill would require age verification on any website containing even one piece of content deemed “harmful to minors,” defined broadly to include nudity, sexual acts, or anything appealing to “prurient interest.” 404 Media covers the bill and notes that the rules would extend to mainstream platforms like Netflix and to VPNs through IP-based verification.

Senator Jim Banks framed the case: “Internet pornography has infected our culture and corroded the vulnerable minds of America’s kids, with the average age of initial exposure being 12 years old. That’s appalling.” Supporting organizations include the National Center on Sexual Exploitation, Family Research Council, Family Policy Alliance, Heritage Action (a Heritage Foundation sister org tied to Project 2025), and the Ethics and Public Policy Center.

Critics, including EFF, argue the bill would be a “privacy and free speech disaster” that forces most of the internet to collect IDs and biometrics. The bill heads to the Senate Commerce Committee for markup alongside four other youth-safety bills, including KOSA and the Youth AI Privacy Act that EFF has already urged senators to reject.

Quick Hits

LLM 0.32 ships with reasoning traces and smarter logging: Simon Willison calls the release the most significant since the project’s launch. Reasoning, tool calls, and tool results now stream as typed events, and the new llm logs command reads a content-addressable message store modeled on Git. Willison’s release notes detail the OpenAI Responses integration and the new GPT-5.6 Luna default model. The companion llm-anthropic 0.26 plugin adds server-side WebSearch, WebFetch, CodeExecution, and AnthropicMCP tools for Claude models and replaces older web search flags with a -T interface. The Anthropic plugin notes cover Claude Opus 5, Sonnet 5, and Fable 5 support.
Spotify and Merlin add AI remix and covers to indie labels: Spotify announced during its Q2 earnings call that Merlin, a licensing partner representing more than 30,000 independent labels and distributors, has joined Universal Music Group in backing Spotify’s upcoming AI remix and covers product. TechCrunch reports that the add-on is paid, opt-in for participating artists, and emphasizes consent and compensation. Spotify cited Deezer data showing more than 50% of daily track uploads are now AI-generated, up from 10% in January 2025.
The Download: robot restrictions and ICE DNA: MIT Technology Review’s daily newsletter pairs the robot restrictions with a separate Wired-reported story on ICE’s DNA collection from nearly one million people last year, most of whom were never convicted of a crime. The newsletter also flags Flock license plate readers showing 71% error rates in one California town and a Maine town that has welcomed a data center against the local grain. The Download newsletter pulls the items together.

Worth Watching

Anthropic-Volta build-out in Norway: Bitdeer’s role as crypto-to-AI converter, the choice of Vera Rubin systems, and the six-year term together raise questions about whether Anthropic will route additional European capacity through Volta or expand the partner list. The TechCrunch deal report is the starting point.
Texas audit fallout: ERCOT’s interconnection queue has roughly doubled in seven months and 90% of the proposed load is data centers. The audits will reveal which projects have real water and power plans versus paper. Watch for county-level pushback, especially in West Texas sites that have anchored the AI build-out. TechCrunch’s analysis sets the baseline.
EFF mobile SDK follow-through: Developers and SDK companies will need to decide whether to disable location defaults before any regulator acts. The named SDKs (InMobi, BidMachine, Verve HyBid, Huawei Petal Ads) cover billions of users, so even incremental changes have outsized reach. EFF’s full technical report is the source for tracking changes.
Open-weight safety policy debate: SaferAI’s GLM-5.2 evaluation showing zero refusals on offensive cyber and dual-use biology tasks, paired with the inability to run the same prompts against Claude Opus 4.7 and Opus 5 because of consistent refusals, raises the question of whether open-weight release decisions should require pre-publication safety gating. TechCrunch’s policy overview is the starting point.
Apple vs OpenAI case progression: The 11 newly identified potential witnesses and the io (Jony Ive) overlap will be the next filing milestones. Watch the docket for whether Apple’s residual-device claim becomes a separate legal track. TechCrunch’s filing summary covers the initial expanded list.