7 min read · 11 sources

Alibaba is going back to shareholders for about $10 billion to keep building AI infrastructure, and the timing tells you where the rest of the hyperscalers are heading. The raise lands days after the company reported a 75% jump in capex tied largely to AI: chips, data centers, and model development. It is one of the clearest signals yet that the AI buildout is now a balance-sheet event, not an operating-expense line.

The constraint is no longer money or silicon. It is the grid. PJM, the largest US grid operator, is proposing rules that would let it cut power to some new large customers, with AI data centers explicitly named, before anyone else during supply crunches. The industry is being told, in effect, that bringing a multi-hundred-megawatt facility online is no longer enough; you also have to negotiate how often you are allowed to go dark.

The rest of the day’s news is a mix of new tooling, new bill shock, and the slow emergence of data foundations as the actual blocker on agentic AI.

95% of 1,500 enterprise architects surveyed by Cloudera delayed or cancelled an AI project in the past year, some six or more times.

PJM wants AI data centers on a kill switch

PJM’s proposal would make new “large-load” customers interruptible, meaning the grid operator can drop them first when supply gets tight. The mechanism is contractual: in exchange for faster interconnection, the customer accepts being shed.

For operators, the practical consequences are: planning around a new class of SLA, sizing on-site storage and ride-through to survive curtailment events, and modeling revenue as a function of expected interruption hours rather than a flat uptime percentage. Power availability is now a siting constraint as real as fiber and water. Expect land selection for new builds to start weighting PJM territory, ERCOT, and MISO interconnection queues very differently.

Alibaba raises $10B as the AI capex curve steepens

Source: links.tldrnewsletter.com ↗

The planned $10B share sale is a straight funding round for AI infrastructure: chips, data centers, and model development. The 75% capex jump is the more interesting number, because it implies Alibaba is willing to compress margins today for compute capacity that will not generate revenue for several quarters.

For engineers outside the hyperscalers, the takeaway is that the rest of the industry is now competing for the same limited pool of HBM, advanced packaging, and grid interconnect. Lead times for capacity are unlikely to shorten in 2026, and the discount that smaller buyers used to get is closing.

Ramp opens its model router as Router.com

Source: ramp.com ↗

Ramp has turned the model-routing infrastructure it built for its own 100-plus internal AI use cases into a public product. It exposes a single API across providers, with three routing strategies: a Flex tier that auto-picks discounted model variants when latency holds, Shadow models that sample real traffic against a candidate model without affecting production, and Benchmark Routing that picks per-workload based on scores you control. Ramp says it cut its own LLM costs by around 30% without sacrificing quality or uptime.

The pitch to engineers is simple: stop re-evaluating every new model release and every provider price change by hand. Router.com sits between your code and the model APIs and absorbs the churn. The risk to watch is lock-in to Ramp’s routing logic, and the operational cost of debugging when a request is silently steered to a different model than the one in your logs.

GitLab 19.3 puts agentic AI behind the same compliance wall as your code

Source: helpnetsecurity.com ↗

GitLab 19.3 is the release enterprises running regulated workloads have been waiting for. The Duo Agent Platform now runs inside GitLab Dedicated’s single-tenant environment, so prompts, responses, and tool calls never leave the customer’s boundary. You can also bring your own model for inference and route it through a new Dedicated AI Gateway. Secrets Manager is moving from “pipeline-only” to a general add-on, billed via GitLab Credits on GitLab.com, with first-class support for Kubernetes, Terraform, OpenTofu, and custom tools.

Two new agentic security features matter for AppSec teams: Bulk SAST False Positive Detection and Agentic SAST Vulnerability Resolution, both aimed at cutting the human loop on triage. For platform teams, the practical question is whether GitLab is now a viable single-vendor alternative to stitching together a coding assistant, a secrets backend, and a separate SAST pipeline. The 19.3 release makes that pitch a lot more honest.

Data silos are now the rate-limiter on agentic AI

Source: cio.com ↗

Two surveys, Google/MIT and Cloudera/Wakefield, land on the same conclusion from different angles. The Cloudera survey of 1,500 enterprise architects and cloud infrastructure leads found that 95% delayed or cancelled an AI project in the past year; some organizations did so six or more times. In the Google/MIT survey of 300 IT and product leaders, more than half paused an agent deployment to fix the data layer underneath, and most said legacy systems have a “significant negative impact” on AI ROI.

The implication for platform teams is that the next big internal program is not another model or another agent framework. It is the data plumbing: multimodal, context-aware, low-latency, governed. The data warehouse, the lakehouse, and the integration tier were built for humans running dashboards. Agents need something closer to a real-time, permissioned, semantically rich substrate, and the orgs that win the next phase of agent deployment are the ones that fund it.

The GPU bill is becoming the new AWS bill

Source: cio.com ↗

A GPU-cloud developer-relations lead argues that AI teams are replaying 2010s-era cloud cost mistakes, with an extra zero attached. GPUs cost roughly 10x per hour what typical cloud compute does, and utilization patterns are worse: reserved capacity runs 24/7 whether or not the workload is there. The piece cites Flexera’s State of the Cloud finding that organizations waste more than a quarter of cloud spend, and the author’s own experience cutting about $220,000 a year by consolidating three analytics platforms.

The FinOps lesson is the same one cloud teams learned the hard way: tag everything, separate bold-bet spend from operating cost, and put utilization dashboards in front of the people who can act on them. The new wrinkle is that GPUs depreciate fast, so idle time is not just wasted spend, it is a depreciating asset doing nothing.

Nvidia backs Cloverleaf for gigawatt-scale AI factories

Source: hpcwire.com ↗

Nvidia has taken a minority stake in Cloverleaf Infrastructure, and the two are co-developing AI factory sites across the US. Cloverleaf’s claimed pipeline is already at multiple gigawatts, which puts it in the same conversation as the big hyperscale campuses, just with Nvidia as a strategic rather than a pure vendor.

For operators, the question is what “AI factory” actually means as a build standard: standardized power and cooling blocks, prefabricated halls, and a delivery cadence that looks more like semiconductor fabs than traditional data centers. The Cloverleaf-Nvidia deal is a bet that the model is the spec and the building is the line.

Sam Altman on scaling compute as the most expensive project in history

Source: davidsenra.com ↗

In a long-form conversation with David Senra, OpenAI CEO Sam Altman frames the company’s bet as one about coordinating chips, fabs, data centers, power, finance, policy, and supply chains into a single platform, more than about any one model. He reiterates the “one consumer interface, one API” thesis and notes that OpenAI spent roughly four and a half years without shipping a product before GPT. The podcast is light on engineering specifics and heavy on platform philosophy, but the compute-coordination framing is useful context for anyone trying to forecast GPU supply, data-center power, or model pricing over the next 24 months.

Anthropic will let enterprises keep retained data in their own cloud

Source: qz.com ↗

Anthropic is planning to let enterprise customers keep required retained data inside their own cloud accounts rather than inside Anthropic’s systems. The details are thin in the announcement, but the direction matters: regulated buyers, especially in financial services and healthcare, have been blocked on retention rules they cannot satisfy with a vendor-managed store. Expect this to narrow the procurement gap between Anthropic and the hyperscaler-hosted model APIs for those buyers.

Portworx 3.6.2.2 tightens vSphere TLS and flags a KubeVirt DR gotcha

Source: docs.portworx.com ↗

Portworx Enterprise 3.6.2.2, released August 21, requires Operator 26.3.1+, Stork 26.4.0+, and a supported kernel. The headline change is real TLS certificate verification for vCenter API connections during cloud-drive operations on vSphere, so you can now supply a CA cert and have Portworx actually check the vCenter identity.

The release notes also document two Major-severity known issues, PWX-57471 and PWX-58026, that affect async-DR restore of KubeVirt VMs. Inherited volume-affinity placement from the source cluster can cause restore failures when root and data disks have mismatched replication factors or placement strategies. The workaround is to align replication factors up front with pxctl volume ha-update --repl before you run async-DR. If you run Portworx on vSphere with KubeVirt, plan the upgrade and audit replication-factor alignment on every VM you intend to fail over.