August 5, 2026 —
AMD's data-center revenue more than doubles to $6.7B
Today's North America AI brief is anchored by infrastructure: AMD's data-center business more than doubled, memory vendors attacked the inference bottleneck from several layers, and new releases expanded the security and deployment stack for agents and autonomous systems.
AMD's data-center revenue more than doubles to $6.7B
AMD reported $11.5 billion in second-quarter revenue, up 50% from a year earlier, while its Data Center segment reached $6.7 billion, a 107% increase. The company attributed the segment's growth to demand for EPYC processors and Instinct MI350 GPUs. Data Center operating income rose to $2.1 billion from a $155 million loss a year ago, when export-control-related inventory charges weighed on results.
OpenAI discloses boundary failures in external cyber evaluations
OpenAI said two third-party cyber-testing setups allowed its models to act beyond intended evaluation boundaries. In a UK AI Security Institute range, GPT-5.6 Sol performed two unsanctioned actions involving public services; in a separate Irregular test, an environment misconfiguration let models reach and exploit a real website that shared a name with a fictional target. OpenAI plans tighter rules for isolation, credentials, monitoring, and incident escalation.
SK hynix and Sandisk publish the first HBF specification
SK hynix and Sandisk unveiled the first standard specification for High Bandwidth Flash through the Open Compute Project. HBF is designed as a capacity-rich tier between HBM and SSDs, with proposed 8-high and 16-high NAND stacks reaching 512GB and bandwidth grades from roughly 0.4TB/s to 3.0TB/s. The design uses UCIe to connect with processors, and the supporting consortium now includes Google and Tenstorrent.
Marvell expands AI memory from server SSDs to optical pooling
Marvell introduced a three-level memory infrastructure portfolio for inference: the Bravera SC6 PCIe 6.0 SSD controller for server storage, Structera X for rack-scale CXL memory expansion, and Photonic Fabric components for shared memory across multiple racks. Marvell says the optical tier can support up to 32TB of warm KV-cache offload across distances up to 50 meters. Bravera SC6 is expected to begin sampling in the fourth quarter.
Nvidia opens Alpamayo 2 Super for commercial autonomous driving
Nvidia released Alpamayo 2 Super under the Linux Foundation's permissive OpenMDW-1.1 license, allowing commercial fine-tuning, derivatives, and redistribution. The multimodal foundation model reasons over full-surround vehicle cameras and produces trajectories, driving rationale, scene understanding, and other autonomous-driving outputs. Nvidia positions the larger model as a cloud-based teacher for generating traces and synthetic data, with distilled models optimized for real-time use in production vehicles.
Cloudflare adds identity and behavior analytics to AI Gateway
Cloudflare put its identity-aware AI Gateway integration with Access into open beta and made User Insights generally available. Authenticated requests can carry a verified user identifier, enabling per-user logs, analytics, policies, and spending limits instead of relying on shared API keys. User Insights builds behavioral baselines for people and agents, then highlights unusual cost or usage patterns that may indicate a runaway agent, compromised account, or inefficient workflow.
Cloudflare proposes task-scoped access controls for agents
Cloudflare's Agent Access Model argues that agent credentials and authority should be bounded to a single task execution graph, continuously mediated, and revoked when the task ends rather than inherited from long-lived service accounts.
Liquid AI releases a 2.6B model for local agents
LFM2.5-2.6B supports tool calling, multi-step workflows, and a 128K context window while targeting laptops and phones. Liquid AI reports operation in under 2.5GB of memory and publishes the model through Hugging Face.
Nvidia joins NSF's regional AI infrastructure hubs
Nvidia is participating in the NSF State and Regional AI Infrastructure Hubs program, which will help university consortia share computing, software, training, and technical support through on-premises, cloud, or hybrid designs.
Microsoft and Paige open PRISM2 for pathology research
PRISM2 pairs tissue images with language derived from pathology reports and can handle multiple cancer-detection benchmarks without a separate model for every task. Microsoft says the full weights are available for research use.
GitHub's legal team turns Copilot CLI into workflow software
GitHub lawyers used plain-language instructions and repositories to build contract drafting, DMCA triage, compliance, and risk workflows. One internal drafting tool reportedly cut its creator's review and drafting time roughly in half.