This Week in AI: Ads Enter Answers, Stripe Buys Model Distribution, and Compute Gets Physical

Omneky added self-service ChatGPT campaign launching alongside Meta, Google, LinkedIn, TikTok, and Reddit. Advertisers can connect an OpenAI ad account, generate creative, set contextual targeting, launch campaigns, and compare performance with other channels from the same platform. Similarweb's new AI Ads dataset adds a competitive view of creatives and landing pages appearing in ChatGPT, Google AI Mode, and AI Overviews.

These are company product announcements, not independent proof of advertising performance. The important change is that conversational advertising is gaining the buying and measurement infrastructure expected around an ordinary media channel.

The AI visibility question is expanding. Brands may now compete for both earned inclusion in an answer and paid placement around it. Media teams will need to judge these placements alongside search, social, and display rather than treating them as a separate experiment.

The latest comparison of the large agency groups showed Publicis and Omnicom growing while WPP continued its attempt to move from a classical holding-company structure into one integrated business. Across the groups, AI platforms, identity data, and common operating layers are becoming table stakes rather than distinctive claims.

The agency networks are responding to consolidation, margin pressure, changing client expectations, and the need to make shared data and AI investments usable across businesses that historically operated with more independence.

The strategic unit is shifting from the individual agency brand toward the shared system underneath it. If every group has an agentic platform, differentiation depends less on announcing the platform and more on whether talent, data, and client work can move across organizational boundaries.

OpenAI agreed to secure approximately 8 gigawatts of IT capacity at the PORTS-Pike campus in Ohio, working with SB Energy, NVIDIA, and the U.S. Department of Energy. The first 800 megawatts are expected in 2028, while later phases require new generation and transmission. OpenAI says the project could create 35,000 construction jobs and 2,500 long-term operating jobs through its buildout.

The employment, water, tax, and community-benefit figures are forward-looking company projections. Development still depends on infrastructure, permits, environmental reviews, financing, and phased delivery.

Frontier AI is now an industrial development question. Model strategy increasingly depends on land, power, transmission, cooling, financing, chip supply, and local consent. These constraints will shape who can scale, how much intelligence costs, and which regions capture the economic activity around it.

Stripe confirmed its acquisition of OpenRouter, which gives developers and companies one interface for accessing and switching among hundreds of AI models. Axios reported a value above $8 billion, although Stripe did not disclose the price.

OpenRouter already used Stripe for global payments, usage pricing, tax, and fraud protection. The acquisition brings the model access layer and the economic infrastructure around model consumption under one owner.

This is more than a payments acquisition. OpenRouter sits at the point where developers compare model capability, price, latency, and availability. Stripe can now connect model consumption, usage billing, fraud controls, and monetization.

Axios reported that executives are shifting from blunt claims about replacing workers toward language about role redesign and changing skill needs. The communication change comes as public anxiety rises and companies face pressure to show investors a return on AI spending without telling employees they are simply a cost to remove.

Companies including Etsy, Patreon, and Microsoft have recently separated specific workforce cuts from direct AI replacement while acknowledging that AI is changing their operating models and required capabilities.

The wording is changing because the evidence is more complicated than the early efficiency narrative. AI may remove tasks, change hiring, alter team design, or support a wider restructuring. Leaders lose trust when they use AI as a broad explanation without showing what work actually changed.

CentaurBench compared models that completed work directly with models that guided a lower-capacity worker model. Rankings in automation and assistance were only modestly related; the best automation model lost the augmentation comparison on five of seven tasks, and on three tasks no tested assistant improved on the unaided worker. A separate StagedWorkspace paper found that knowledge-work agents performed better when they could access both native files and parsed views, and when reviewers could see version-linked changes.

Both papers are fresh working papers with limited task sets. Their findings should be treated as early evidence rather than settled conclusions about every model or work setting.

Most AI evaluations ask whether a model can produce the final answer. Real work often asks a different question: can the system help a person or another agent improve the work without losing context, versions, or accountability? Assistance quality and automation quality should be evaluated separately.