U.S. AI Unease Reaches 52% as Data Center Resistance Raises Costs

01U.S. public unease over AI rises to 52% as data center resistance begins translating into corporate costs

Americans increasingly encounter artificial intelligence through chatbots, AI search and features added to products such as email and televisions. But many consumers do not see those tools delivering benefits large enough to offset concerns about job losses, intellectual property disputes and the local burden of the data centers needed to run them.

That dissatisfaction is now affecting both election calculations and project costs. On Wednesday, Axios reported that the National Republican Senatorial Committee sent leading AI companies a memo warning that U.S. data centers were hurting the Republican Party’s chances in a key Ohio election. At the same time, Pew Research reported that 52% of Americans were “more concerned than excited” about the increased use of AI in daily life, up from 37% in 2021.

Other polling points in the same direction. More than 70% of Americans said AI was advancing too quickly in a May Economist/YouGov survey. A CNBC poll of people aged 18 to 34 found that, when shown the names of nine leading AI executives, most respondents did not trust them to act responsibly on AI. The sources did not provide the surveys’ complete sampling methods or margins of error.

The resistance is particularly tangible around data centers, where communities are being asked to absorb local costs for infrastructure supporting products they may not value. The Wall Street Journal reported that opposition to planned AI data centers across the United States has pushed technology companies to improve their offers with job guarantees, clean-water investments and other community benefits. One Louisiana parish project included $50,000 bonuses for teachers. The total added cost of such concessions has not been quantified.

Daily AI products provide another focus for frustration. Consumers see AI inserted into familiar services, while hearing about its use in school cheating and systems trained on other people’s intellectual property. The benefits, by contrast, may amount to summarized webpages or conversational televisions rather than higher pay, shorter working weeks or other clearly felt gains.

Some industry leaders now describe the problem as more than poor communication. Airbnb CEO Brian Chesky said part of the backlash stems from companies failing to produce AI products that ordinary people genuinely like. Anthropic CEO Dario Amodei called the negative perception a “crisis of trust,” acknowledging that AI companies, including Anthropic, have not yet fulfilled their major promises to benefit the world.

02GPT-5.6 Sol price drops 50% on OpenRouter to $2.50 per million input tokens and $15 per million output tokens

OpenRouter has announced a 50% price cut for GPT-5.6 Sol, the flagship model in OpenAI’s GPT-5.6 series. The model is designed for complex reasoning, coding, agentic workflows and long-horizon problem solving, including command-line and multi-step coding tasks.

The new OpenRouter listing prices GPT-5.6 Sol at $2.50 per million input tokens and $15 per million output tokens. Cache reads cost $0.25 per million tokens, while cache writes cost $3.125 per million. Web search is priced separately at $10 per 1,000 calls.

Those figures are OpenRouter’s prices and service terms, not evidence of a matching reduction through OpenAI’s direct channel or other platforms. The source does not provide the previous prices, the date when the reduction took effect or whether any other distributor has made the same change. OpenRouter also says caching and discounts can push customers’ actual average costs below providers’ posted rates.

The service includes a 1.05 million-token context window and supports responses of up to 128,000 tokens. GPT-5.6 Sol accepts PDF, image and text files as inputs and returns text. Developers can use function-calling tools and tool_choice, as well as structured outputs defined through a JSON schema.

OpenRouter offers the model through three providers: OpenAI, Azure in Europe and Amazon Bedrock in the United States. Its platform routes each request according to the selected mode: Balanced weighs price and speed, Nitro prioritizes the fastest option, and Exacto targets the highest tool-calling accuracy. If an upstream provider returns an error, OpenRouter says it can send the request to another healthy provider, provided the customer’s filters allow it. Users can also pin a specific provider or exclude providers through routing controls.

The listing gives developers clearer inputs for recalculating the cost of reasoning-heavy applications and agent workflows, but a complete comparison still requires information it does not supply. That includes the former price, the effective date of the cut, comparable rates on other channels and the workload-specific effect of caching, output length, web searches and provider selection. The model was released on July 9, 2026, and has a February 16, 2026 knowledge cutoff.

03Cerebras Launches CS-4 Wafer-Scale Inference System, Claims Up to 30 Times the Speed of Production GPU Systems

Cerebras has introduced CS-4, a rack-scale AI inference system for hyperscale data centers and the first version of its new Nexus platform architecture. The system uses wafer-scale processors—chips built around an entire silicon wafer—to address the difficulty of generating responses quickly from increasingly large AI models.

Compared with the previous generation, CS-4 combines three WSE-3 Turbo wafers in each system, with each wafer delivering up to twice the earlier speed, according to Cerebras. The company claims the complete system can generate tokens—the units of text processed by a model—up to 30 times faster than production GPU systems while providing up to 10 times the throughput per watt of CS-3.

Cerebras has not supplied independent benchmark results, identified the GPU systems used for the 30-times comparison, or disclosed the full test conditions. It also has not provided pricing, customer names, delivery timing or availability details.

The computing upgrade is paired with a redesigned connection system. CS-4 introduces programmable wafer I/O that Cerebras says doubles bandwidth and allows wafers to connect directly within or across racks without a network switch. Wafer-to-wafer latency can fall as low as two microseconds.

That low-latency link is intended to preserve interactive generation when a model must be distributed across several wafers. Cerebras says CS-4 can exceed 1,000 tokens per second on models containing more than 10 trillion parameters, though that result is also a company claim rather than an independently verified measurement.

The physical system has been reorganized around modular compute, power and I/O components. Its Wafer-Scale Backpack folds a wafer, power conversion, direct liquid cooling, high-speed I/O and control electronics into a compact three-dimensional assembly. Cerebras says this design uses 50% fewer components and can reduce data-center deployment time from days to hours.

Power delivery sits 0.5 millimeters from the processor, compared with roughly 50 millimeters on conventional GPU boards, according to the company. Cerebras says this reduces board-level losses and enables twice as much power to reach the WSE-3 Turbo, supporting higher operating frequencies and faster token generation.