6 min read · 12 sources

Accenture is weighing an exit from Greater China. The IT consulting giant has retained JPMorgan to explore offloading its operations across mainland China, Hong Kong, and Taiwan in a transaction that could reach $1.5 billion.

The move comes after years of geopolitical friction culminated in Beijing’s 2023 restrictions barring foreign consultancies from advising state-owned enterprises. Paired with a fiscal Q3 revenue miss and a guidance trim to 3 - 4% annual growth that shaved 16% off Accenture’s stock, leadership is looking to shed geographically constrained units and free up balance-sheet capacity for global infrastructure and model deployments.

The rest of the industry faces infrastructure bottlenecks of a more physical kind: utilities cannot deliver power, while enterprises are deploying automated agents faster than their security teams can write policies.

GPT‑6.1 Sol’s 86.6 % accuracy on COBOL migration means legacy mainframes can be modernized with barely any manual testing.

Accenture reviews $1.5B Greater China divestment

Accenture’s potential carve-out, managed by JPMorgan, would offload its consulting footprints in mainland China, Hong Kong, and Taiwan. The firm may retain a minority stake, but the core business has faced severe friction since Beijing cracked down on international consulting firms accessing data from state-owned enterprises.

For enterprise operators, this retreat reflects the operational reality of regional IT fragmentation. Running unified IT consulting or shared services across China and Western markets has become nearly impossible under strict cross-border data controls and state procurement blacklists. If a deal completes, Accenture sheds compliance overhead and redeploys capital into core platform engineering, while multi-national clients operating in the region will likely need to re-contract through local domestic vendors.

OpenAI benchmarks jump on COBOL modernization

Source: openai.com ↗

OpenAI published new evaluation results targeting low-resource codebases, highlighted by GPT-6.1 Sol’s performance on COBOL. On benchmarks measuring behavior-preserving migrations from COBOL to modern target languages, GPT-6.1 Sol achieved an 86.6% accuracy rate under low reasoning effort settings. Its predecessor, GPT-5.6 Sol, scored just 5.4% on the identical workload.

Mainframe logic often resists static analysis and traditional transpilation because it leans heavily on implicit data layouts, file-record structures, and decades of side-effect accumulation. LLMs historically choked here by hallucinating edge-case math or dropping decimal alignment in packed fields. An accuracy leap of this scale suggests that reasoning models are finally handling language-level state preservation across legacy enterprise stacks, offering a credible alternative to multi-year manual re-platforming projects.

Power constraints sink Microsoft’s 23,040-GPU deal with Nscale

Source: cloudcomputing-news.net ↗

Microsoft cancelled its lease agreement for capacity at Nscale’s planned data centre in Loughton, Essex, after power delivery timelines collapsed. The site had been slated to host 23,040 liquid-cooled Nvidia GB300 GPUs across an initial 50 MW footprint, targeting a Q1 2027 delivery. The deal fell apart after UK Power Networks indicated that the transmission capacity needed to energize the facility would not arrive until the early-to-mid 2030s.

The cancellation isolates the single biggest failure point in hyperscale planning: power grid interconnection queues. Sourcing silicon is no longer the sole hard bottleneck; procuring high-voltage substation connections within a five-year deployment window is. While Nscale claims development will continue alongside its other Microsoft contracts, losing an anchor tenant for an entire 50 MW phase leaves the facility stranded without an immediate operator.

Changing workflows and the governance of accidental agents

Source: techradar.com ↗

Enterprises rolling out agents are discovering that tooling is not the bottleneck; organizational inertia is. Gallup workplace data shows that while 65% of employees report productivity gains from baseline AI tools, only 12% strongly agree that it has fundamentally altered organizational operations. The friction point lies in moving teams away from conversational, prompt-based assistants toward autonomous, asynchronous agents.

Compounding this is the issue of unintentional agent sprawl. Employees using sanctioned enterprise SaaS environments are assembling autonomous multi-step loops - such as continuous mailbox scanners that approve low-tier requests or draft downstream tickets - without security vetting. Gartner projects that 40% of enterprise applications will embed task-specific agents by year-end, while simultaneously forecasting that 40% of organizations will have to forcefully decommission or demote active agents by 2027 due to systemic governance voids.

Solving this requires shifting governance from static prompt scanning to runtime context orchestration. That means tying dynamic access control, persistent agent memory, retrieval mechanisms, and deterministic execution logs directly into production runtimes before these autonomous loops escape local desktop contexts.

Lenovo cuts Blackwell delivery times to 15 days

Source: computerweekly.com ↗

Lenovo launched AI Express, a hardware supply partnership with Nvidia designed to bypass typical lead times by shipping pre-validated clusters within 15, 20, or 25 business days. The small tier deploys a ThinkSystem SR650a V4 containing dual Nvidia RTX Pro 6000 Blackwell Server Edition GPUs aimed at 7B - 70B model inference. Mid-tier setups scale to an eight-GPU ThinkSystem SR675 V3, while the large configuration packs Nvidia’s HGX B300 platform inside a liquid-cooled ThinkSystem SR680a V4.

For teams deploying private clusters, lead times for modern Blackwell silicon regularly stretch past six months. Lenovo’s SKU-standardized pipeline trades customization for delivery speed, offering hardware allocations aimed at Asia-Pacific enterprises facing strict deployment windows and infrastructure scarcity.

Cloudflare wraps workspace governance around AI automations

Source: cloudflare.com ↗

Cloudflare expanded integrations for its Cloudflare OS platform, introducing native enterprise ties for Notion, GitHub, ClickHouse, Gong, Asana, Box, and Canva. The service shifts automation pipelines off local user workstations and runs them directly inside Cloudflare’s network boundary.

The platform relies on Cloudflare’s AI Gateway to handle policy enforcement, API routing, and spend auditing. By forcing agentic workflows to execute centrally, it prevents ungoverned data ingestion into public models, limits authorization scopes to target APIs, and monitors aggregate token consumption across fractured SaaS environments.

Active root zero-day hits Cisco Secure Email Gateway

Source: diesec.com ↗

Cisco issued an urgent advisory for CVE-2026-76519, an actively exploited relative path-traversal vulnerability in its Secure Email Gateway carrying a CVSS score of 9.8. Unauthenticated attackers can leverage crafted email attachments to overwrite arbitrary files and execute code with root privileges. The bug resides within the Advanced Malware Protection attachment parsing subsystem.

Systems are vulnerable if File Reputation is enabled and the appliance runs a Content Scanner Tools engine older than version 25.3.0.5525. Cisco has issued patches in AsyncOS 15.5.5-014 and 16.5.0-780. There is no fix for the 16.0 tree - the earlier 16.0.4-302 release only resolved a separate SQL injection flaw (CVE-2026-76461). CISA added the path-traversal bug to its Known Exploited Vulnerabilities catalog with an immediate compliance mandate.

The realities of GPU cluster observability

Source: hackernoon.com ↗

Monitoring distributed AI clusters continues to expose deep flaws in conventional observability stacks. As detailed in an analysis of GPU cloud performance debugging, traditional dashboards monitor subsystems in isolation - metrics show nominal health across local compute, InfiniBand links, and block storage, yet tenant distributed training workloads suffer massive step-time inflation.

The root cause usually stems from dynamic issues that aggregate metrics mask: individual GPU thermal throttling, transient fabric hash collisions on dense rail-optimized topologies, or tail-latency spikes in storage nodes feeding workers. Debugging these systems requires a real-time, topology-aware control plane that maps individual tensor parallel ranks and NCCL communication rings directly to the physical switches, PCIe trees, and power phases supporting them.

IBM targets commoditized models; OpenCode credential breach

Source: horsesforsources.com ↗

IBM is centering its enterprise go-to-market strategy on operational integration harnesses rather than standalone foundational models. CEO Arvind Krishna projects that open-weight models will serve roughly 80% of all enterprise inference by 2030. IBM claims its internal transformation reduced operating costs across its 280,000-person footprint by 30%, and is now pitching that orchestration framework as a sellable deployment harness to enterprise clients seeking structured runtimes for commoditized architectures.

Finally, developer toolmaker OpenCode announced a security incident affecting 1.4% of its console users. Threat actors breached an older administrative console component and extracted stored vendor API keys, subsequently generating unauthorized inference traffic on third-party accounts ranging from hundreds to over $1,000 per tenant. OpenCode clarified that its core open-source command-line tool was untouched, while Google and Anthropic revoked the compromised keys and OpenAI deployed usage credits to offset downstream billing impacts.