6 min read · 13 sources

The hyperscaler hardware moat just collapsed. Google and Oracle are partnering to deploy Google Tensor Processing Units directly inside Oracle Cloud Infrastructure data centers starting in 2027. It makes OCI the first third-party cloud provider permitted to run Google’s custom silicon, entirely changing the procurement calculus for large-scale AI engineering teams.

At the same time, OpenAI is reportedly preparing to buy Google TPUs directly to supplement its existing rentals, diversify away from pure Nvidia GPU clusters, and hedge its custom silicon development with Broadcom. The AI compute supply chain is fracturing across traditional platform rivalries, driven by power constraints, wafer limits, and unrelenting demand.

Eighth-generation TPUs will offer up to 12.6 PFLOPS of FP4 compute per chip with over 97% goodput, giving OCI customers non-Nvidia hardware without rewriting PyTorch pipelines.

Google TPUs break out of GCP onto Oracle Cloud

Source: cloud.google.com ↗

Google Cloud and Oracle announced a joint partnership to install Google TPUs natively within OCI. Starting in calendar year 2027, OCI instances will offer eighth-generation Google TPUs pushing up to 12.6 PFLOPS of FP4 compute per chip. That represents a 100x jump over the 92 TFLOPS delivered by Google’s first-generation hardware, targeting sustained training goodput above 97%.

For operators, this fundamentally alters multi-cloud capacity planning. Teams locked out of GCP allocations will be able to provision bare-metal TPU pods inside Oracle’s high-density data centers. Because Google’s open-source XLA and PyTorch/XLA software stacks have matured, running distributed training on TPUs no longer requires rewriting standard model definitions into proprietary frameworks.

Azure cloud revenue hits $33.17 billion as AI monetization hits

Source: channelnewsasia.com ↗

Microsoft reported fiscal Q1 Azure revenue of $33.17 billion, climbing 48% year-over-year and blowing past analyst expectations. Microsoft generated $18.6 billion in free cash flow - roughly $10 billion above market estimates - and guided for 49% to 50% Azure growth in Q2.

The report marks the first time Microsoft has explicitly broken out Azure revenue in raw dollar figures rather than isolated percentage growth rates. For infrastructure leads budgeting multi-year platform commitments, these numbers put an end to the narrative that hyperscaler capital expenditure on accelerators is an unmonetized balance sheet sinkhole.

OpenAI moves to procure Google silicon

Source: reuters.com ↗

OpenAI is reportedly planning to buy Google TPUs outright, according to reporting from The Information. The move expands OpenAI’s footprint past temporary cloud rental agreements and plants Google silicon directly into its long-term hardware baseline alongside Nvidia GPUs and custom chips in development with Broadcom.

Securing dedicated TPU clusters gives OpenAI structural redundancy. If TSMC packaging allocations or Nvidia rack lead times slip, workloads can shift to TPU clusters without halting deployment schedules. It also turns Google into a tier-one component vendor for its primary generative AI competitor.

Anthropic's IPO prospectus flags severe vendor concentration

Source: itpro.com ↗

Anthropic’s newly filed IPO prospectus provides a transparent look at the fragile economics of foundation model vendors. According to reporting on the filing, 47% of Anthropic’s projected 2025 revenue will flow entirely through its primary cloud distributors, Amazon and Google. In 2024, its two largest individual enterprise customers accounted for 12% of total revenue each.

While customer concentration among the top five accounts declined to approximately 25% over the trailing nine months, the distribution infrastructure remains completely tethered to two landlords. For infrastructure engineers building on foundation models, this underscores an operational hazard: downstream API availability, pricing stability, and compliance boundaries are directly tied to commercial agreements between rival hyperscalers. Portability wrappers across model APIs are no longer optional enterprise hygiene; they are survival requirements.

Claude synthesizes 62-CVE exploit chains in Chrome and Safari

Source: anthropic.com ↗

Anthropic published research demonstrating that an internal variant of Claude autonomously discovered and constructed five critical, zero-day vulnerability chains across production software. The findings included a one-click remote code execution chain inside Google Chrome that tied together a V8 heap corruption, a sandbox escape, a CacheStorage race condition, and a FedCM use-after-free bug.

The disclosure resulted in 62 separate CVEs patched in Chrome 156, alongside verified exploit chains against Safari and core OS networking stacks. This moves automated vulnerability tooling past the era of static AST analysis and targeted fuzzing. SREs and security operators must prepare for an environment where attackers use reasoning models to synthesize non-linear, multi-subsystem exploit paths faster than human response teams can trace patch diffs.

Oracle deepens OCI stack: Vera Rubin NVL72, native Snowflake, and Acceleron

Source: nvidianews.nvidia.com ↗

Oracle spent the week positioning OCI as the universal clearinghouse for high-performance enterprise workloads. First, NVIDIA and Oracle announced multi-gigawatt OCI Superclusters built on NVIDIA Vera Rubin NVL72 systems, alongside Vera CPUs. For mid-tier topologies, Oracle is rolling out HGX Rubin NVL8 servers pairing eight Rubin GPUs with dual 64-core Intel Xeon 6 processors, 4 TB of DDR5 memory, and 6.4 Tbps of RDMA fabric.

Second, Snowflake is landing as a native first-party OCI service. Customers can run Snowflake’s AI Data Cloud directly against OCI-hosted storage, connecting data lakes with Oracle Fusion, NetSuite, MySQL HeatWave, and native Oracle DBs without egress fees or detached IAM setups.

Third, Oracle rolled out updates to its Acceleron networking fabric. The revision adds front-end RDMA connectivity for distributed inference nodes, expanded NVMe-over-Fabrics block storage attachments, and a dynamic collective communication library designed to automatically steer cluster traffic around failed links during training runs without aborting checkpoint state.

GitHub Universe 2026: Multi-model Copilot Fusion and autonomous pipelines

Source: github.blog ↗

At GitHub Universe 2026, GitHub shifted its developer tooling toward an agentic delivery architecture. The headline launch is Copilot Fusion, an orchestration engine that automatically routes engineering tasks across different underlying models depending on context length and task complexity.

GitHub also launched a standalone Copilot workspace app designed to manage feature generation end-to-end - from initial issue description to pull request validation and merge. Enterprise accounts can now consume Copilot features without standalone seat licenses, billing usage against shared AI credit pools.

Infrastructure quick hits: Tesco dumps VMware, Entra kills SMS MFA, and Google ships ER 2

Source: itpro.com ↗

• Tesco begins migrating retail edge nodes to Proxmox: Tesco has opened recruitment for a Head of Software Engineering to lead a full migration from VMware to Proxmox VE. The initiative covers edge virtualization across thousands of supermarket locations, reinforcing the post-Broadcom shift of enterprise hypervisors toward open-source KVM orchestration.
• Microsoft sets sunset date for legacy Entra MFA: System administrators have until February 1, 2027, to migrate off SMS and voice MFA in Microsoft Entra ID workforce tenants. Passkeys will become the default authentication mechanism this September, forcing security teams to transition users toward FIDO2 hardware tokens or platform authenticators.
• Google launches Gemini Robotics ER 2: Google Cloud released Gemini Robotics Embodied Reasoning 2 into public preview. Built as an orchestration layer for physical robotics, the model processes multi-camera spatial video feeds, handles long-horizon task planning, and executes structured tool calls for factory and warehouse automation.
• Enterprise data fragmentation lowers output: A Forrester study run for Staffbase showed 55% of enterprise employees track company data across six or more distinct platforms. Thirty-five percent deal with conflicting internal technical documentation weekly, causing 48% of surveyed teams to report measurable delivery drag.