Tokenpocalypse: Rising AI Prices, Grid Failures, and a Surveillance System Gone Wrong

The AI industry's mounting costs surfaced from multiple angles: token prices are set to rise as labs eye IPOs, the Texas power grid is buckling under data center load, and a key White House policy seat went vacant. Meanwhile, DeepSeek's newest model staked a claim against the frontier, and automated surveillance misfired — badly — on a San Diego man who had nothing to do with the crime it flagged him for.

Security

Flock gets it wrong

A Flock license plate reader wrongly linked a San Diego man to a violent crime; he was five miles away at the time. Flock systems now blanket thousands of US jurisdictions, feeding a network of automated plate matches that police increasingly treat as actionable leads. When the system misfires, the burden falls on the wrongly flagged person to prove a negative. This case won't be isolated.

Internet centralization as democratic risk

A paper in Science argues that the structural concentration of the internet — a handful of CDN providers, cloud platforms, and DNS resolvers handling the majority of traffic — creates systemic risks for democratic institutions. The concern is empirical, not theoretical: outages, BGP hijacks, and state-level filtering events have all demonstrated how infrastructure consolidation becomes a single point of failure for public discourse and services adjacent to elections.

Algorithmic monocultures in hiring

A new research project cataloging algorithmic monocultures in hiring makes the case that when every employer runs the same AI screening tools, the bias encoded in those systems propagates uniformly across the job market. The old variance — where a candidate rejected by one employer might succeed with another — disappears when a single flawed model defines who clears the first filter everywhere.

AI

The Tokenpocalypse

TechCrunch is calling it the "Tokenpocalypse": across-the-board AI price increases tied to the major labs positioning for public offerings. The logic is simple — companies that subsidized low token prices to drive adoption are now shifting to extract margin from their installed base before IPO. Anyone running AI-heavy workloads in production should model what their costs look like if per-token pricing doubles.

That IPO pressure is real. The Information has a long read on the Anthropic vs. OpenAI race to go public, arguing that first doesn't necessarily mean best. Both are burning cash at scale; both face scrutiny over whether current revenue justifies private-round valuations. Being the second IPO with cleaner fundamentals may outperform being the headline-grabbing first.

DeepSeek V4 Pro benchmarks

Runtime Wire reports that DeepSeek V4 Pro outperforms GPT-5.5 Pro on precision benchmarks. Treat these claims with appropriate skepticism pending independent verification, but the directional pressure is consistent: DeepSeek continues to match or exceed frontier performance at significantly lower compute cost, which undercuts US labs' pricing rationale precisely when those labs are trying to raise prices.

Policy vacuum

The top White House AI advisor is leaving. At a moment when chip export controls, AI safety frameworks, and compute policy are all live and consequential, losing institutional expertise in that seat creates a meaningful gap — poorly timed for an administration navigating the geopolitics of model capability competition.

OpenAI's super app

OpenAI is still working on a consumer product that goes beyond chat — a senior employee was quoted saying "chat is dead." The move is consistent with the company's push to own the interaction surface directly rather than ceding it to integrators, though no timeline or feature details were shared.

Tech

Grids and drought

ERCOT is flagging stability risks after data centers and crypto mining sites failed voltage regulation tests. Texas attracted a disproportionate share of data center construction through its deregulated power market, but large loads that don't provide reactive power support create instability when enough of them connect simultaneously. The grid operator's warning matters because Texas has been the path of least resistance for AI infrastructure buildout.

The power problem pairs with a water problem. Data centers consumed 264 billion gallons of water in 2025 for cooling while drought covered nearly 63% of the US. The physical infrastructure that AI runs on is now stressing water supply chains and power grids simultaneously, in some of the most drought-affected regions in the country.

Xbox Games Showcase

Microsoft's Xbox Games Showcase delivered concrete dates during Summer Game Fest: Halo: Campaign Evolved — a remake of Combat Evolved's campaign — arrives July 28th on Xbox, PC, and PS5. Fable gets a late February window after its recent delay. Gears of War: E-Day will not come to PS5, contrary to earlier rumors — Xbox and PC only. Microsoft also unveiled a 25th anniversary Xbox Series X in translucent green, echoing the original console's design language.

Firefox gets Vulkan video

Firefox merged Vulkan Video support, enabling GPU-accelerated video decode via the Vulkan API. The practical win is on Linux, where Vulkan is the primary GPU interface — a meaningful step toward closing Firefox's performance gap with Chromium on Linux desktops.

The gap between what AI costs and what it charges is closing fast — and the power and water bills are only starting to show up.

Also yesterday

• Thousands of Instagram Accounts Compromised via Meta's AI Chatbot
• C0XMO Botnet Is Actively Exploiting a DD-WRT Router Flaw Right Now
• Silent Ransom Group Actively Targeting Law Firms via Fake IT Support Calls