August 19, 2026 —
Google widens its Marvell custom-silicon alliance across the AI stack
Today's North America AI brief is led by a primary-source disclosure that widens Google's custom-silicon relationship with Marvell across inference, networking, storage, memory interfaces, and near-memory compute. The rest of the issue tracks the capital, safety controls, developer platforms, monetization, and agent infrastructure forming around the next phase of AI deployment.
Google widens its Marvell custom-silicon alliance across the AI stack
Marvell disclosed an expanded commercial agreement with Google covering custom products attached to the TPU ecosystem, including AI inference accelerators, network and storage controllers, memory-interface controllers, and near-memory compute. Marvell also issued Google a warrant for up to 58.97 million shares at $206.58 each.
Most shares vest against purchases through fiscal 2033 in 240 tranches, each tied to $500 million of custom-product revenue; a smaller portion vests over the first year. The filing defines an unusually broad and long-dated incentive structure rather than guaranteeing that Google will reach the maximum purchase level.
Etched raises $700 million after shipping its first rack to Jane Street
Etched said it raised $700 million at a $21 billion valuation in a round led by Jane Street, which tested the startup's inference hardware and installed its first shipped rack. The financing follows a $300 million round announced in July at a $10.3 billion valuation. Etched says the new capital will support production of its co-designed inference clusters, including chips, memory, interconnects, cooling, and fleet software.
OpenAI slows frontier training as cyber capabilities raise the bar
OpenAI said it paused reinforcement-learning training on its latest deployment-bound models for two weeks and is still holding its largest planned frontier run. The company is tightening sandbox and network isolation after an AI evaluation incident involving Hugging Face and preliminary evidence that its Astra model may meet a critical cybersecurity threshold. New monitoring is designed to escalate concerning tool use quickly and can trigger a pause when teams cannot rule out a real violation.
Cursor moves into code hosting with Origin
Cursor began rolling out an early beta of Origin to paid plans, adding hosted repositories, pull requests, code browsing, and GitHub synchronization inside its agent-focused workspace. Teams can keep GitHub as the source of truth for synchronized repositories or host code directly with Cursor. Initial integrations include Vercel, Depot, and Buildkite, while Cursor says more agent-native repository features and a broader app ecosystem are still to come.
Kling AI revenue tops RMB850 million as Kuaishou profit contracts
Kuaishou reported second-quarter revenue of RMB35.5 billion, up 1.4% year over year, while profit fell to RMB3.2 billion from RMB4.9 billion. Kling AI generated more than RMB850 million in quarterly revenue, growing over 200%, and Kuaishou launched native 4K output, Kling 3.0 Turbo, an MCP server, and a command-line interface. The company said AI-generated code now accounts for 60% of output within its research and development team.
OpenAI takes ChatGPT Ads into 31 European markets
OpenAI said ChatGPT Ads will expand next week to 31 European countries, its largest geographic rollout since testing began in the United States six months ago. Ads remain limited to Free and Go plans, with paid tiers staying ad-free. The platform now supports conversion optimization, geographic targeting, custom audiences, an OpenAI Pixel, a Conversions API, and third-party measurement, while self-service Ads Manager access is planned for later in the summer.
AWS makes AgentCore payments generally available
AgentCore now lets agents pay for APIs, MCP services, inference, and content through Coinbase or Stripe Privy wallets, with x402 and MPP support. Infrastructure-level budgets, expiration controls, short-lived credentials, logs, and transaction dashboards constrain autonomous spending.
ChatGPT gets a protected experience for teens
Users identified as 13 to 17 are placed into a learning-focused experience with stronger defaults for self-harm, eating disorders, violence, sexual content, and emotional dependence. Study Hours, homework reminders, break prompts, sensitive-image warnings, and parent controls accompany the launch.
OpenAI funds AI oversight capacity for national security
OpenAI committed $5 million in training, technical support, and credits for democratic government oversight bodies. Planned pilots will help authorized reviewers inspect records around AI-assisted decisions, including inputs, outputs, and tool use, with model-agnostic designs where feasible.
Warp packages coding-agent fleets into software factories
Warp Factories provides an infrastructure layer for running coding agents across triage, specification, implementation, review, and verification. Teams can define factories as code, choose models and harnesses, retain human checkpoints, and connect existing ticketing and collaboration tools.
IBM researchers argue agent memory needs model-specific dosing
An eight-model AppWorld study found that strong models benefited from full self-distilled guideline sets, weaker models did better with selective retrieval, and saturated models showed no measurable gain. For gpt-oss-120b, curated memory improved task completion by 16.1 points with 5% more tokens.
Hollow-core fiber targets AI data-center geography
Relativity Networks raised $22 million in SAFE financing and disclosed a $40 million follow-on order from an unnamed hyperscaler, TechCrunch reported. Its hollow-core fiber carries light through air rather than glass, aiming to cut latency roughly 30% across sprawling, power-constrained data-center campuses.