Plumbing the
information age

Helios connects 72 MI455X GPUs into a single scale-up domain with 31 TB of unified HBM4 memory, 2.9 exaflops of peak FP4 compute and 260 TB/s of scale-up bandwidth, alongside AMD's 6th Gen EPYC Venice CPUs and Pensando networking silicon. AMD claims the rack delivers up to 30% more tokens per dollar than Nvidia's Rubin NVL72 system and 50% more memory capacity. A new Fabric Manager layer adds zero-touch provisioning, fabric-wide observability and virtual pods that carve the 72-GPU domain into isolated failure zones. For network operators evaluating AI cluster architectures, Helios represents a significant open-standard alternative to Nvidia's vertically integrated approach. The use of open Ethernet and Pensando DPUs rather than Nvidia BlueField preserves vendor flexibility for switching, NIC, and fabric control—critical for teams already operating Arista, Cisco, or Juniper fabrics. The system is in production now, with volume deployments ramping through the second half of 2026.

Presence connects AI agents to internal company systems with shared context, policies, permissions, guardrails, and evaluations across voice and chat. The launch signals that enterprise-grade agent deployments now require centralized access control, audit trails, and policy enforcement—not just model capability. For SREs and platform teams, this reflects the market moving away from ad-hoc agent prototypes toward managed agent fleets. On the same day, OpenAI also announced a data center campus that could exceed $30 billion. The infrastructure play underscores that frontier labs are building vertically: models, orchestration, and compute capacity all controlled in-house to manage latency and cost at scale.

NSF awarded $83 million through IDSS to expand access to data infrastructure resources researchers can use alongside computing and AI resources to accelerate discovery and innovation. IDSS supports national priorities to strengthen U.S. leadership in AI, build an AI-ready workforce and ensure researchers have the tools and infrastructure needed to compete globally. For MLOps practitioners, this investment signals growing recognition that data engineering and infrastructure are core constraints in production AI systems. The program focuses on connecting scientific data with compute, instruments, software and AI resources—a systems-level view that treats data pipelines and observability as foundational.

In a 12-hour window on July 23, the 10 most important disclosed startup rounds added up to $623 million, with more than 72% going to Etched and Humanoid, and nearly 81% landing in AI hardware, sensing, robotics, or physical-AI data infrastructure. The market is telling founders that the next margin pool in AI may sit below the application layer, in compute, perception, and deployment systems. For infrastructure teams, this signals a shift in where optimization returns are highest: not in training efficiency alone, but in inference economics, hardware-software co-design, and task-specific accelerators. The concentration of capital in hardware and hard-to-replicate infrastructure suggests the commoditization of general-purpose GPU training is complete.

Anthropic released a data connector for Claude that allows instances to safely connect to enterprise databases and APIs without requiring custom integration code. For MLOps and platform teams, this represents a shift in how LLM vendors approach production integration—treating data connectivity as a first-class concern rather than leaving it to customers. The $200 million Economic Futures Research Fund signals long-term commitment to studying labor economics impacts of AI deployment at scale. Together, these announcements position Anthropic as focused on both operational integration (data connectors) and responsible scaling (economic research).

ChatGPT Health rolled out to all U.S. users 18 and older on web and iOS on July 23, six months after a limited January 2026 pilot that drew tepid feedback and prompted a months-long rebuilding effort. Users can integrate data from services like Apple Health, Function, and MyFitnessPal, plus medical records from hospital systems like Epic and Oracle Health. Paid subscribers on Plus and Pro plans access GPT-5.6 Sol, which scores 88.0% on completeness in HealthBench Professional versus 53.2% for the free tier's GPT-5.5 Instant. The timing is significant: the rollout was announced a day after a Florida-based pastor sued the company for giving a near-fatal suggestion not to consult a doctor, highlighting the reputational and liability stakes. OpenAI incorporated user feedback and improvements made to its frontier models since January. For practitioners, this signals how aggressively major vendors are pushing personalization at the edge of regulated domains—expect scrutiny around model behavior, data retention, and liability frameworks.

Claude Code's /code-review command now runs as a background subagent, keeping review work out of conversations and maintaining stacked slash commands as its review target. Published artifacts can now call MCP connectors each time someone views them, pulling live data through the viewing account's own connections with viewer approval before each connector call. The Windows path fix addresses a practical operational concern: paths with \u-prefixed segments (like C:\Users\unicorn) were being corrupted into CJK characters in tool inputs, making those files inaccessible. Screen reader mode enables accessible sessions, expanding utility for practitioners requiring ADA compliance or assistive tech support. The update also surfaces HTTP status and error text for failed MCP servers in 'claude mcp list' and /mcp commands, improving debuggability of agent-tool integration failures. For MLOps and SRE teams, the MCP connector refresh is significant: live-data artifacts reduce snapshot staleness and enable real-time dashboard updates without rebuilding.

Lumen Technologies completed its acquisition of Alkira for approximately $475 million in July 2026, strengthening its east-west data center interconnect (DCI) proposition for AI workload connectivity. The Gartner report highlighted how Lumen's advantage lies in integrating three typically fragmented elements: extensive fiber backbone, dense metro reach, and edge computing capabilities paired with Alkira's programmable control plane. This unified enterprise service model delivers WAN-to-edge-to-cloud connectivity with orchestration, automation, and SLA-driven performance—critical for AI infrastructure where constraints exist in data center interconnect, hyperscaler connectivity, and high-capacity routes moving large data volumes between enterprise, cloud, and edge environments. The company's shift toward higher-margin enterprise and edge computing services represents a strategic pivot from consumer fiber following its February 2026 sale of mass-market fiber assets to AT&T. For network operations, Lumen is deploying Blue Planet's AI agents across its OSS infrastructure to automate complex workflows in fiber backbone and edge systems, continuing the industry-wide trend of AI-based operations support systems to reduce operating costs and accelerate provisioning.