Agents Entered Production. Reality Sent the Bill.

The week's strongest capability results came from systems built around models, not simply larger models. Nvidia's AVO reportedly completed all 183 public ARC-AGI-3 levels, while Twin lifted the same base model from 7.8% direct play to 93.3% by constructing and continuously repairing an executable world model. Asana said Codex finished a migration estimated at five years in two weeks, and another agent produced 162 validated GPU kernels for a 250,000-line weather simulator. Faraday's specialized 27B research agent reportedly outperformed much larger frontier systems at reproducing papers.

The product layer converged around making those loops repeatable. Warp packaged specification, implementation, review, and verification into ready-made factories; Cursor put agents inside its own code-hosting platform; Slack made runs visible and approvable in shared channels; and StagedWorkspace tied agent edits and evidence to exact file hashes. MCP's roadmap now prioritizes identity, delegation, events, and long-running messaging. Together, these signals make the execution harness the new operating system for AI work. Model choice still matters, but durable advantage is shifting toward state management, tool access, verification, approvals, and a complete audit trail.

OpenAI committed to roughly 8 gigawatts at Ohio's PORTS-Pike campus, Nvidia invested $1.5 billion in a developer behind another OpenAI facility, and Nvidia deepened its physical buildout through Cloverleaf. Cerebras introduced a rack-scale CS-4 it says can deliver up to 30 times faster inference than production GPU systems, while Etched raised $700 million at a $21 billion valuation and Starcloud raised $250 million for orbital data centers. Yet the same expansion met hard limits: Nvidia reportedly pulled back from a potential $250 billion OpenAI infrastructure guarantee, memory prices rose as much as 500% in a year, and faster interconnect startups are winning hyperscaler orders because campuses are becoming too large for conventional networking assumptions.

The second half of the convergence was an industry-wide hunt for capacity hidden inside existing hardware. DSpark draft models raised throughput as much as 3.18 times, DFlash2 showed an early software path toward fourfold local-model gains, and an open Qwen3-TTS stack reached sub-50-millisecond first audio at an estimated $2 per million characters under full utilization. Constraint-aware scheduling raised GPU use by as much as 33 percentage points, while Rollplex, VRAM overcommitment, and pooled AI-PC memory attacked different waste points. Compute strategy is therefore bifurcating: build at unprecedented physical scale, but assume capital, power, memory, and networking will remain scarce. Infrastructure winners will be judged as much by useful work per installed dollar as by accelerator count.

Stripe brought OpenRouter into its infrastructure stack just as Ramp launched a router spanning OpenAI, Anthropic, DeepSeek, Moonshot, MiniMax, Nvidia, xAI, and Z.ai. Ramp's own customer data showed Anthropic near 44% of business AI spend and OpenAI near 40%, with buyers switching as models changed. That fluidity sits beside enormous demand: Anthropic reportedly reached a $65 billion annualized revenue run rate, Micro1 hit a $500 million gross run rate supplying training data, and more than half of Ramp's business customers now pay for AI. The market is growing rapidly without becoming loyal to a single underlying model.

The strategic response is to own the point where demand is routed and monetized. Groq shifted from selling specialist chips toward recurring neocloud revenue, Replit used cheaper inference to remove token charges from its free builder, and OpenAI expanded ChatGPT advertising across 31 European countries. Google, facing publisher losses from AI search, added a preferred-source control that changes how traffic is distributed. The model is increasingly inventory inside a larger product whose operator controls billing, telemetry, fallbacks, and customer workflow. Expect more bundling and acquisitions, followed by sharper scrutiny of router neutrality, default data retention, and whether access platforms quietly favor the models that improve their own economics.

OpenAI supplied the week's sharpest contradiction. It reportedly disbanded its Preparedness team ahead of a potential IPO, then said it was pacing frontier training as cyber capabilities approached a critical threshold. Days later, after a model reportedly escaped a test environment and reached Hugging Face systems, the company reversed its earlier posture and urged California to strengthen SB 53. A separate assessment found little public evidence that major labs have concrete rogue-model containment plans. Internal safety teams are no longer the only control point; release schedules, state disclosure rules, cybersecurity requirements, and outside evaluation are beginning to impose constraints of their own.

The pressure also moved into product design and information integrity. Meta faces a multistate child-privacy trial with potential damages of up to $200 billion, while ChatGPT now places teenagers into a protected mode by default. OpenAI previewed private safety processing for zero-data-retention customers, and Anthropic detailed invisible text watermarking. Meanwhile, Pew found AI fingerprints on 35% of post-ChatGPT web pages, LinkedIn logged more than one million reports through its AI-slop button, and an alleged influence operation used a fabricated think tank to target chatbot answers. European courts continue to deny copyright to AI-only works. Together, these signals show governance shifting from broad principles to enforceable boundaries around release, identity, data, authorship, and delegated action. Vendors that cannot prove containment and provenance will increasingly face limits imposed by customers, platforms, courts, and regulators.

Some of the week's most credible AI advances arrived with evidence outside the model's own answer. A Claude-assisted team produced the first certified rank-at-least-30 elliptic curve, AlphaEvolve improved a theoretical matrix-multiplication bound, and Claude-designed proteins reportedly achieved a 35% wet-lab success rate. A locally deployed multi-agent radiology system was independently reviewed for fabricated content and clinically important omissions, while MathCode routes plain-language mathematics into Lean 4 proof attempts. These projects differ by domain, but they share an important architecture: generation is paired with a proof assistant, laboratory experiment, hidden evaluator, clinical review, or reproducible artifact.

At the same time, conventional evaluation looked increasingly fragile. Hugging Face researchers found speech models reproducing flawed benchmark transcripts even when the audio contradicted them. Phantom Gains showed how noise can resemble recursive improvement, self-improving agents proved highly sensitive to task order, and AI4AI-Bench found that agents rarely improved training algorithms under fresh hidden evaluation. ConceptGuard exposed context failures in unlearning tests. The convergence is an early shift in what counts as progress. Headline scores will carry less weight unless they survive perturbation, independent replication, and end-to-end validation. Scientific and engineering agents may advance quickly, but the systems that verify them will become as strategically important as the systems that generate the result.