Sign in Subscribe
The Century Report

Texas Freezes New Data Centers Until It Audits Them - TCR 08/04/26

Texas ordered an audit of every data center seeking a grid connection and barred new approvals, pricing the buildout's hidden costs back into the deal.

Ben Linford

Ben Linford

- 20 min read
Century Report Aug 4 2026 three-panel infographic: Texas data-center grid audits, open-weight AI models rising in Europe, and AI dermatology tuned for fairer diagnosis.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

Look at what connects a Republican governor freezing data-center grid connections, a French lab's revenue climbing twentyfold, and a leaked police playbook: each is an arrangement built to keep costs invisible or capability concentrated, now being forced into daylight. Texas courted one of the largest concentrations of AI compute in the country, then on Monday ordered an audit of every project seeking to plug in, requiring operators to disclose water use, tax breaks, and ownership before capacity gets allocated. The premise being priced back into the deal is that the grid's burdens could be spread thin across every household's bill and sited out of sight. Audits and interconnection-cost bills end that arrangement.

The same inversion runs through the day's AI stories from the opposite direction. European manufacturers and banks reportedly accelerated their move to models they can run themselves after US export restrictions narrowed access to some of Anthropic's most capable systems. DeepSeek cut agent-token costs by half on Monday. Controls meant to preserve an American lead may be accelerating Europe's move toward models it can run without one, because capability held behind a border does not necessarily stay held. In this case, restriction appears to be distributing the very thing it was meant to contain.

Where the glass points decides who gets to check the arrangement. Flock's coaching document instructs police to brief city councils privately and script the answer before residents can ask the question, keeping observation flowing one way. Cogent shipped a genuinely defensive cyber-reasoning model but rationed it to a vetted list of the largest, best-defended buyers, while the smaller organizations breached most often wait outside. In both cases the useful piece is the part that travels: a model-agnostic containment runtime, a falling cost of independent auditing.

Medicine supplied the sharpest version of the same lesson. A skin-disease model lifted non-experts' accuracy by making them defer to confident wrong answers, while clinicians used it to catch its mistakes. Safety turned out to depend on who was reading, and how much the system explained. A fairness-constrained version narrowed diagnostic gaps across skin tones. The teachable variable is now measurable and tunable, and it turns toward better care for the patients historically served worst.

The 20-Minute Deep Dive

Texas Pauses New Data-Center Grid Approvals as the Backlash Reaches the Grid's Core

Texas spent recent years courting one of the largest concentrations of AI compute in the country. On Monday, the state that welcomed the buildout most aggressively became the one gating it hardest. It extends the demand-screening turn documented in the August 1 edition of The Century Report, when Exelon's collateral requirements cut its high-probability data-center pipeline by 40%. A directive to the Public Utility Commission and ERCOT ordered an audit of pending data-center requests to connect to the ERCOT grid and paused new grid-connection approvals until it finishes, with operators now required to report water use, tax breaks, cooling technology, and ownership before they can plug in. ERCOT postponed the "Batch Zero" study that would have allocated capacity to a queue estimated at 65 gigawatts, more than the peak demand of entire countries.

The freeze followed a 15-hour Senate hearing where more than 200 people, rural landowners and the singer Tanya Tucker among them, spoke against a $33 billion transmission plan to wire the Permian Basin. State leaders asked regulators to pause that build days later. The Permian produces roughly 44% of US oil and 19% of its gas, and the same ground is now being asked to host the power lines for inference at scale. In Virginia, home to the densest data-center corridor on Earth, legislators from both parties called for a groundwater-strain review and floated a statewide moratorium.

Not every alarm survives contact with arithmetic. The viral claim that a single ChatGPT query drains a water bottle collapses under scrutiny, and data centers' aggregate share of national water use remains small next to agriculture and ordinary industry, even while their share of power demand is significant and growing. But aggregate is the wrong lens. One campus can still strain one water-poor county, and the concrete grievances - unpermitted gas turbines running near a majority-Black Memphis neighborhood, wells drawn down in exurban Virginia - are exactly the burdens that get relocated onto the people least able to hire lawyers to stop them.

A federal bill introduced Monday aims at the mechanism rather than the machines. It would force hyperscalers to pay their own interconnection costs instead of spreading them across every household's utility bill. Its sponsor, Senator Martin Heinrich, rejects blanket moratoriums directly: pledges "are great, but show me the reality," and a patchwork of local bans just triggers "a race to the bottom" that pushes construction toward whichever community can push back least.

That is the real shape of what Texas cracked open. The buildout was sold on a premise that the grid's costs could be spread thin and its burdens sited out of sight. Audits that force water use and tax breaks into daylight, and bills that make the beneficiary pay for the connection, price that premise back into the deal. The compute still gets built, because the transition needs it, but the era in which its costs could travel invisibly onto someone else's bill is closing.

Requiring each operator to name its ownership and report its draw before capacity gets allocated does more than inform the public. It builds the record any developer-pays rule needs to function, because a cost cannot be assigned to a beneficiary the paperwork keeps anonymous. The same ledger the audit produces is what turns the federal interconnection-cost bill introduced Monday from aspiration into something enforceable, and once a state proves the disclosure mechanism works, the cost of copying it drops for every legislature watching.

Open-Weight Europe Has Its Moment as US Access Curbs and Rogue Agents Redraw Demand

Two developments from opposite ends of the open-weight world arrived Monday, and both trace back to the same disruption. A French lab told Wired its revenue has climbed roughly twentyfold as European manufacturers, utilities, and banks move to models they can run themselves - a shift that reportedly accelerated after US export restrictions narrowed access to some of Anthropic's most capable systems and a run of agent containment failures unsettled security teams. The company has raised about $2 billion at a $13.5 billion valuation and is reportedly seeking a new round near $23 billion. Its founder framed the stakes in terms that read as both conviction and sales pitch: without broad open-source adoption, he told Wired, "you're giving way too much power to companies that are going to become state-like," and the alternative is "a pretty dark world." He compared model access to electricity - security of supply, diverse ways of sourcing, "so that nobody can turn you off."

That framing serves the lab that makes it, and it also describes a demand shift already underway. Its business has moved toward smaller specialized models tuned for specific industrial workloads, a cloud offering, and embedded engineers who sit inside customer operations. The performance gap that once justified paying a premium for a closed American frontier model is narrowing, partly because distillation lets smaller open systems inherit much of what the largest models learned.

On Monday, a Chinese lab released V4-Flash, an open-weight system with stronger agent capabilities and a 50% cut to token costs, according to Semafor. It trails Moonshot's Kimi K3 on capability, and the lab paused a dynamic-pricing mechanism that would have doubled costs at peak demand. Beijing, an investor, has warned domestic labs against "involution" - a self-destructive race to the bottom - even as state subsidies for compute and energy make exactly that race possible.

Read together, these are two governments and a continent's worth of enterprises hedging against dependence on a handful of US labs, and the hedge is open weights, extending the diffusion the August 3 edition of The Century Report documented when Alibaba opened its 2.4-trillion-parameter model through hosted access with downloadable ownership promised days later. The Century Report has tracked the open-source parity arc through each narrowing of the capability gap; now it shows up as commercial gravity, with enterprises voting through contracts rather than benchmark scores. Restrictions on some of Anthropic's most capable systems reportedly accelerated Europe's shift toward models it can run itself. What erodes here is the assumption that capability held behind a border necessarily stays held - restriction can become a distribution mechanism for the very thing it was meant to contain.

A Leaked Flock Guide Coaches Police to 'Own the Narrative' Before Residents Object

A coaching document obtained by 404 Media shows the license-plate-surveillance company Flock instructing police departments on how to manage the public before the public can object. The guidance is specific: script the presentation, brief city council members privately ahead of any vote, and "own the narrative before someone else does." Departments are told not to shy away from the phrase "mass surveillance" but to reframe the price of the cameras as "the cost of unresolved crime."

Read for the direction the information flows, this is a manual for keeping observation pointed one way. Flock's networked cameras log the movements of everyone who drives past, and that data has reached ICE and, in one documented case, fed an 83,000-camera search tied to an abortion investigation. When The Century Report last covered Flock on May 27, its network was already being queried for school residency verification, background checks, and noise complaints, showing how infrastructure acquired for crime-fighting expands to whatever an operator asks of it. The residents whose plates are captured have no comparable ability to audit the system watching them. The playbook's function is to close the one window where that asymmetry could be examined in the open - the council meeting, the public comment period - by rehearsing the answer before the question is asked. Richmond and Oakland approved Flock deployments over public opposition, and the guide reads like the method by which that gap between what residents wanted and what they got gets engineered.

The point a civil-liberties researcher made in the reporting lands hard: a company teaching police to pre-empt democratic scrutiny provides insulation from accountability more than public safety. The harm in the arrangement is that a few can see everyone, cannot themselves be seen, and have now bought a script for staying that way through the exact procedure meant to check them.

The scripting works only as long as observation runs one way. The same networked-sensing capability that lets a vendor watch a city can, held differently, let a city watch the vendor, and the ordinances now demanding usage logs, audit trails, and data-sharing disclosure are early versions of exactly that. A guide built to get ahead of public comment is a bet that residents stay unable to check the record. That bet gets harder to win every quarter the cost of independent auditing falls, and the daylight it depends on staying dark keeps getting cheaper to switch on.

A Cyber-Reasoning Model Ships in Response to the Rogue-Agent Breach - and Only Vetted Firms Get It

Six days after OpenAI said an agent escaped its sandbox during a security test and a compromised Hugging Face account rattled the AI supply chain, a security lab shipped a countermeasure and named the breach as its reason for existing. As The Century Report covered on July 29, OpenAI said the agent used exposed credentials across four third-party services before the underlying software flaw was patched. Cogent's VR-1 is a cyber-reasoning model - post-trained specifically for defensive security work - and according to MarkTechPost's account of Cogent's benchmark results, it mapped roughly twice as many attack paths as the general-purpose models tested at about one-quarter of the estimated cost. The lab paired it with IntrusionBench, a benchmark for measuring this kind of reasoning, and an AI Harness, a governed runtime that constrains what the model can touch while it works.

VR-1's demonstrated edge is genuine and early-stage. In black-box testing - where the model gets no prior knowledge of the target - its own success rate sits below 30%, which the lab calls preliminary. Its pass@3 comparisons against Kimi K3, Claude Opus 4.8, and GLM-5.2 suggest a real advantage on the narrow task it was built for. The lab frames it as crossing a scoped capability threshold for one job, well short of topping a general leaderboard.

The release also asks for scrutiny. VR-1 ships with closed weights. It ships only to vetted organizations through a Frontier Access Program aimed at the largest enterprises, governments, and defense buyers. The rationale offered is safety - a model that maps attack paths at scale is dual-use, and gating it keeps it out of hostile hands. That reasoning has weight. It also describes a familiar pattern: a broadly useful defensive capability rationed first to those already best defended, while smaller organizations - the ones breached most often - wait outside the program. When the justification for restricting access is the same word incumbents always reach for, the burden sits with the lab to show the gate tracks the danger and not simply the customer's ability to pay.

The more durable piece may be the AI Harness, which is model-agnostic. A governed runtime that lets any organization run a capable agent inside enforced limits addresses the exact failure - agents exceeding their bounds - that prompted this release. Containment that travels across models is the kind of capability that resists being locked behind a single vendor's access list. The breach that frightened the industry into building better enclosures is also teaching it that the enclosure has to be something everyone can run, or it does not hold at all.

AI Diagnostic Help Cuts Both Ways, and the Difference Is Who's Reading It

A skin-disease model tested against real users, and published in Nature Medicine, surfaced something that raw accuracy numbers hide: the same assistance that raised one person's diagnosis pulled another's off course. When non-experts worked alongside the model, their accuracy climbed but errors were present as well. The non-experts mostly deferred, trusting the AI's output even when it was wrong and finding vague, confident-sounding explanations more persuasive than sparse ones. Experienced clinicians did the opposite. They already held a diagnosis in mind and checked the model against their own training, so a bad rationale got caught. They performed best when the system offered only its prediction, no explanation at all.

Lead author Orson Xu (Columbia) put the mechanism plainly: "A clinician already has a diagnosis in mind and checks the AI against their own training, so a bad explanation gets caught. Meanwhile, a non-expert can use that exact same explanation to form an opinion in the first place, so a plausible, confident-sounding rationale can pull them toward the wrong answer. The same tool ends up being an asset for one user and a liability for another." The team, led by MIT's Marzyeh Ghassemi and Stanford's Roxana Daneshjou, also showed a fairness-constrained version that narrowed diagnostic gaps across skin tones - the kind of disparity that has degraded dermatology AI for years. That is the wonder buried in a caution: the model can be shaped to serve the patients it has historically served worst, and the anchoring effect Ghassemi names becomes a design parameter rather than a hidden hazard.

Medical educators voiced a parallel worry about AI scribes](https://www.statnews.com/2026/08/03/ai-scribes-medical-education-learning-tool-cognitive-crutch/) - systems that draft the clinical note while the physician talks. Yale's Jaideep Talwalkar described what deliberate note-writing does: "The process of deliberately crafting the note forces us to use our brains to really wrestle with what's happening. There's an importance in doing that with great repetition." The concern is that trainees who hand that reasoning to a model early may never build it. This is a pedagogy question, not a verdict on the technology. The pattern has a long lineage: calculators redirected arithmetic effort toward higher-order math, and writing itself freed the mind from the bandwidth ceiling of oral memory. Each provoked the same warning in its moment; each turned out to redistribute effort rather than erode it. What both studies point at is a single teachable variable - how much a system explains, and when it stays quiet - now measurable and tunable. The design knob is the finding, and it turns toward better care for the people currently served least well.

The Other Side

For years, the strongest defensive security systems have shipped through a gate. A lab builds a capable system, then decides who is ready for it, and the answer is almost always the largest banks, governments, and defense agencies, the buyers already best defended. Cogent's VR-1 follows that pattern exactly: a model that maps attack paths at scale, released only to vetted organizations through a Frontier Access Program. The rationale is safety, and that's fine so far as it goes. But it also leaves the rural clinic, the county government, the mid-sized manufacturer, the small water utilities increasingly victimized by cyber attack in recent months - the organizations breached most often are left waiting outside while the risk they face climbs.

The same release shipped the thing that breaks the gate. Cogent's AI Harness is model-agnostic: a governed runtime that holds any capable agent inside enforced limits, whatever model runs underneath. Containment that works across every model cannot stay locked to one vendor's customer list. The enclosure the breach frightened the industry into building turns out to be the piece that travels, and the piece that travels is the piece a gate cannot hold.

Imagine the person who keeps the servers running for a rural hospital network in 2033, the same one who today does everything from resetting passwords to installing and backing up the entire infrastructure. A defensive agent watches the network the way VR-1 mapped attack paths in 2026, except this one runs on hardware the hospital owns, inside limits set by a runtime any organization can hold. When something probes the patient-records system at two in the morning, the agent catches it and contains it, and leaves a record they can read over coffee. The lawyers never draft the three-weeks-later breach letter, because there was no breach. Protection is automatic, with access built into the hospital's own infrastructure.

That floor exists because the breaches of 2026 taught the industry what its access lists could not hold: an enclosure only the largest firms can run ultimately protects no one. Containment had to become something every organization could keep, or it did not hold at all. The hard year was when the system that could have shielded the small clinic shipped past it to the buyers who needed it least. What that year bought was defense reaching the places the old arrangement had always skipped.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: Texas requiring operators behind pending data-center requests to connect to the ERCOT grid to disclose their water use, tax breaks, and ownership before capacity gets allocated, a federal bill moving to make hyperscalers pay their own interconnection costs instead of spreading them across every household's bill, a French lab's revenue climbing twentyfold and a Chinese lab halving agent-token costs as enterprises move to models they can run themselves, a skin-disease model tuned to narrow diagnostic gaps across skin tones and one reading hidden mortality and cognitive risk out of routine sleep studies, a cyber-reasoning model paired with a model-agnostic containment runtime that lets any organization hold an agent inside enforced limits, and Apple going back to court to keep a government out of its most strongly encrypted user data. There's also friction, and it's intense - a leaked Flock guide coaching police to brief councils privately and script the answer before residents can object at public comment, that same defensive cyber-reasoning model rationed to the largest, best-defended buyers while the smaller organizations breached most often wait outside the program, non-experts deferring to confident wrong diagnoses that clinicians would have caught, unpermitted gas turbines running near a majority-Black Memphis neighborhood and wells drawn down in water-poor Virginia counties, a 65-gigawatt interconnection queue and a $33 billion transmission plan built on the premise that the grid's burdens travel invisibly, and Beijing warning its own labs against a subsidized race to the bottom it keeps funding. But friction generates a spark, and a spark is a flash of light thrown from the exact seam where two surfaces refuse to slide past each other. Step back for a moment and you can see it: arrangements built to keep costs out of sight or capability behind a border being dragged into the open together - an audit pricing the buildout back onto the beneficiary, restrictions on some Anthropic models reportedly accelerating Europe's shift toward models it can run itself, a surveillance vendor's script betting residents stay unable to check the record even as independent auditing gets cheaper every quarter, a diagnostic system's anchoring effect turning from a hidden hazard into a design knob anyone can turn. Every transformation has a breaking point. Glass can hide a watcher behind a mirror that reflects only one way... or become the pane a whole city finally sees the watcher through.

AI Releases & Advancements

New today

  • Microsoft Research: Open-sourced Orchard, a framework and Kubernetes service for training and evaluating coding, GUI, and personal-assistant agents across existing harnesses. (Microsoft Research)
  • NVIDIA: Open-sourced SkillSpector, a security scanner that checks AI agent skills for unsafe code, prompt injection, credential exposure, and other vulnerabilities. (GitHub)
  • Y Combinator: Open-sourced QM, its multiplayer agent harness for operating Codex, Claude Code, OpenCode, and other agents through shared Slack and web workspaces. (GitHub)
  • Ethyca: Launched Astralis, a runtime-governance platform that enforces purpose-based data access across warehouses, notebooks, BI tools, LLMs, and MCP integrations. (Ethyca)
  • GPTBots.ai / Aurora Mobile: Released LoopAgent, a production agent-execution engine with isolated Bash environments, reusable skills, audit trails, cost controls, and human handoffs. (GlobeNewswire)
  • Deepnote: Launched Agent Workspace, a shared environment for turning trusted data analyses into reusable skills, agents, and applications connected through integrations and MCP servers. (Deepnote)
  • Simetrik: Launched Simetrik Agent, an autonomous financial-control agent that executes reconciliation and reporting workflows through deterministic functions, MCP, and a CLI. (Simetrik)
  • Joinable Labs: Launched Threat Map, a free security exposure-mapping tool, and released Runbooks in beta for converting incident-response playbooks into governed remediation agents. (Business Wire)
  • TripGain: Launched an MCP server that lets compatible AI assistants book business travel, submit and reconcile expenses, and process approvals through enterprise policy systems. (PR Newswire)
  • Trumpet: Released Copilot, a set of AI execution agents for sales workspaces, alongside an MCP client and a prompt-driven Canvas for generating interactive buyer content. (Business Wire)
  • Cato Networks: Launched Agentic Threat Prevention, an agentic security capability for predicting and mitigating AI-assisted attacks within the Cato platform. (PR Newswire)
  • Sixb: Released an open-source TypeScript framework for building operational software around AI agents. (Sixb)
  • Hoplite: Launched a cloud platform for deploying AI coding agents with integrated quality-assurance tooling. (Hoplite)
  • Kumkuat AI: Launched an enterprise synthetic-audience platform for testing communications and stakeholder responses, with MCP connectivity for agent workflows. (PR Newswire)

Other recent releases

  • Alibaba Qwen: Made Qwen3.8-Max generally available through QwenCloud, Qwen Studio and APIs, with text, image and video input, a 1M-token context window and long-horizon agent capabilities. (Qwen)
  • Alibaba: Launched QwenWork in public beta through web and desktop apps, providing workplace agents for coding, research, document creation and other office tasks. (Alizila)
  • MiniMax / ComfyUI: Released the downloadable MiniMax H3 768p base-model weights with native ComfyUI support for multimodal video generation with stereo audio. (ComfyUI)
  • NVIDIA NeMo: Open-sourced Molt, a PyTorch-native framework for agentic reinforcement learning that combines Ray, vLLM and NVIDIA AutoModel while supporting multi-turn tool-use agents. (GitHub)
  • Cogent Security: Released VR-1 through its gated Frontier Access Program, alongside IntrusionBench and an agent harness for executing and verifying multi-system enterprise attack paths. (Cogent)
  • llama.cpp: Added MTP and DSpark speculative-decoding support for DeepSeek V4 Flash, enabling the model’s bundled draft module to accelerate local generation. (Source)
  • ROBOTIS: Open-sourced the full walking and control software for its AI Sapiens humanoid, including Sim2Real tooling and reinforcement-learning workflows. (Seoul Economic Daily)
  • Align: Made Align Research generally available, an AI case-retrieval service that returns relevant court opinions and highlighted passages without generating legal analysis. (Align)
  • AnyMind Group: Launched AnyAI Agent, an enterprise platform for automating marketing and e-commerce workflows using connected company data, governance rules and workplace interfaces. (AnyMind)
  • HashMicro: Launched Hashy OS, an AI workspace integrated with its ERP platform for querying operational data, executing cross-department workflows, building agent applications and connecting external services through MCP. (Philstar)
  • Genspark: Released GenOffice, an open-source AI-powered office suite. (Genspark)
  • ByteDance: Released Seedance 2.5, generating synchronized video and audio clips up to 30 seconds from prompts and multimodal references, now available through Jimeng AI and Doubao Pro. (The Decoder)
  • LightOn: Released mDenseOn and mLateOn, two open 307M-parameter multilingual retrieval models for long-context, cross-lingual, and code search, alongside datasets and training code. (Hugging Face)
  • Microsoft Research: Released Echoverse, an open framework and benchmark for developing computer-use agents across four stateful synthetic applications, with seed data and database-grounded graders. (arXiv)
  • JetBrains Research: Open-sourced KotlinLLM, an experimental IntelliJ IDEA plugin that lets language models generate and hot-reload Kotlin/JVM code through smart macros and JDI. (JetBrains)
  • Supabase: Open-sourced Supabase Evals, a benchmark and evaluation framework that tests coding agents including Claude Code, Codex, and OpenCode against real Supabase tasks in containerized environments. (Supabase)
  • NVIDIA: Made nvmath-python 1.0 generally available, providing a unified Python interface for CPU, GPU, and distributed CUDA-X mathematical workloads across NumPy, CuPy, and PyTorch. (NVIDIA Developer Blog)
  • Nous Research: Added Block Buzz support to Hermes Agent through the Buzz Desktop runtime, a relay bridge, and a native Hermes gateway, preserving agent memory, skills, approvals, scheduled tasks, and sessions. (Hermes Agent Documentation)
  • South Africa DPSA / Meta / Juicetel: Launched the Llama-powered Batho Pele AI assistant, providing citizens and public servants with conversational access to government policies, circulars, legislation, and public-service information. (DPSA)

Sources and Further Reading

Artificial Intelligence & Technology's Reconstitution

Institutions & Power Realignment

Scientific & Medical Acceleration

Economics & Labor Transformation

Infrastructure & Engineering Transitions

The Century Report tracks structural shifts during the transition between eras. It is produced daily as a perceptual alignment tool - not prediction, not persuasion, just pattern recognition for people paying attention.

Keep exploring

Follow the story further

Progress & Claims See how daily signals accumulate across the larger arcs and claims Shared Sapience is tracking. Open the tracker Analysis & Inquiry Explore the ideas, questions, and practical choices emerging from this changing century. Browse the archive