AI’s New Bottleneck Is the Power Grid

The Grid Becomes the Bottleneck

America’s largest power grid is preparing to cut electricity to major data centres when supply runs short. AI workloads are about to become interruptible demand.

That makes this the development worth sitting with, because the constraint here isn’t a better accelerator or a smarter model. It’s whether enough generation exists when the grid is under stress.

PJM Interconnection said on July 28 that its latest capacity auction failed to secure enough power generation. Its response is a framework allowing electricity curtailment for data centres and other large customers rated at 50 megawatts or more. The cuts won’t begin before June 2027, and affected customers will be compensated.

PJM says curtailments should be infrequent and limited in duration, but their actual frequency and length are unknown. Comparable demand-response programs can give anything from roughly 30 minutes to several days of notice. That range is operationally significant: some computing can pause or move, but tightly coupled training runs and customer-facing services are much less forgiving.

For organisations building in PJM territory, reliable AI capacity can no longer be budgeted as racks plus electricity at an assumed availability. It may require flexible scheduling, backup capacity or more expensive on-site generation. My read is that diesel generators and similar measures could preserve uptime while shifting pollution and financial costs away from the grid rather than solving the underlying shortage. Compute expansion has become an electricity-reliability problem.

AMD Secures Power with the Hardware

That closes the grid story, but it also makes AMD’s latest deal much easier to understand.

AMD and Core Scientific announced an infrastructure partnership on July 28 that gives AMD access to more than 500 megawatts of US data-centre capacity beginning in 2027. Core Scientific separately described about 530 megawatts under 15-year agreements across five sites, with the relationship potentially expanding to 2.5 gigawatts.

The companies plan to develop physical infrastructure around AMD Instinct GPUs, EPYC processors and the ROCm software stack. The capacity is intended for customers in AMD’s ecosystem, rather than one named hyperscaler. AMD will also receive market-priced warrants to buy Core Scientific shares, subject to commercial conditions.

None of this capacity is available immediately, and the larger expansion, deployment schedule and customer demand remain forward-looking company claims. Still, AMD can now approach prospective customers with a path to powered space alongside its chips and software.

For large AI buyers, that changes the competitive question. Accelerator specifications still matter, but a system that can actually be installed and energised may beat a faster one waiting for a substation. My inference is that chip companies now have to secure land, buildings and electricity nearly as aggressively as they design silicon.

GitHub Adds a Human Checkpoint

The next constraint is less physical, but just as capable of stopping a deployment: whether the software feeding it can be trusted.

GitHub introduced two supply-chain protections on July 28. Potentially malicious Actions workflows in public repositories can now be held before execution. A collaborator with write access must approve a held workflow through an authenticated web session, and GitHub says repositories don’t need extra configuration for that protection.

The important limit is scope. The workflow hold currently covers public repositories on GitHub.com, not GitHub Enterprise Server, and GitHub hasn’t disclosed its detection criteria. We therefore can’t yet judge how many dangerous workflows it will miss or how much legitimate work it may delay.

Separately, Dependabot can now use OpenSSF malicious-package advisories across npm, PyPI and additional ecosystems. That broadens warnings when a dependency matches known malware, but those expanded alerts are automatic only where Dependabot malware alerting has been enabled.

For maintainers, the practical move is to nominate trusted workflow reviewers and confirm that malware alerts are active. The workflow hold adds a useful barrier before suspicious automation reaches CI credentials, but only one of these defences arrives without configuration. Human approval is now part of the security boundary, so the quality and speed of that review will matter.

Public Claude Links Were Searchable

There’s also a quieter trust failure here: a link can feel private even when the system treats it as public.

Publicly shared Claude chats and Artifacts appeared in search results, according to reporting published on July 27. The affected material came from Claude’s public-sharing feature; ordinary chats remain private by default.

Anthropic said shared links aren’t guessable and aren’t submitted through a public directory. But if somebody posts one on a page that a crawler can reach, a search engine may index it. TechCrunch could no longer reproduce the Google results by Monday afternoon, suggesting some remediation, although the number of exposed links and the persistence of cached copies weren’t established.

A shared snapshot can include messages and Artifacts created before the link was made. Uploaded files and raw MCP tool-call data are excluded, but that still leaves plenty of room for personal, commercial or confidential material to travel further than intended.

Anyone who has used public Claude links should review the Shared Chats list in privacy settings and revoke anything sensitive. More broadly, an unguessable URL is not a durable privacy control. AI products should make indexing and archival risk visible at the moment of sharing, because users naturally interpret friction to discovery as a promise of confidentiality.

School Photos Become Source Material

Now bring that same discoverability problem into a school community, where the consequences become much more personal.

Australia’s eSafety Commissioner issued guidance on July 28 urging schools to reconsider public images of students and staff. Accessible AI tools have lowered the effort required to create face swaps, fabricated videos and sexualised deepfakes from ordinary photographs.

Between January and March 2026, eSafety received more than 100 reports involving anonymous accounts that targeted schools or staff. Almost all used images taken from school websites or social accounts. During 2025, the agency also recorded twice as many reports from people under 18 concerning digitally altered intimate images, including deepfakes.

Those figures count reports received, not the prevalence of abuse across every Australian school. The publication is guidance rather than a new binding rule, and eSafety notes that some harmful non-sexual deepfakes may sit outside its legal removal thresholds.

Even with those limits, the operational response is concrete. Schools should audit old galleries, remove unnecessary identifying details, reconsider who can see new posts and update incident plans for manipulated media. The useful shift in thinking is to treat every public portrait as a reusable source asset for impersonation, not merely a harmless communications photo.

Nvidia Backs Safe Superintelligence

At the frontier end of AI, one new partnership is notable partly for what remains undisclosed.

Nvidia and Safe Superintelligence, or SSI, announced a long-term agreement on July 27. It includes an investment of unspecified size, access to Nvidia’s Vera Rubin computing platform and collaboration on current and future Nvidia systems.

SSI claims the arrangement will increase its available compute by an order of magnitude. There’s no disclosed deployment timetable, research output, safety evaluation or description of the specific alignment techniques that additional capacity will support.

For SSI, the deal provides capital and substantially more computing infrastructure. Nvidia gains close technical feedback from a frontier-model laboratory. My reading is that the partnership further concentrates advanced AI research around a dominant infrastructure supplier. It may accelerate SSI’s work, but increased compute is an input, not evidence that the company’s safe-superintelligence objective has been achieved.

AI Enters the Test Bench

A more immediate form of generation is arriving in tools that interact with real engineering systems.

Emerson released enhancements to its NI Nigel assistant on July 28. Nigel can now generate LabVIEW code and TestStand test sequences, and it has been extended across InstrumentStudio, FlexLogger and VeriStand as well as the existing LabVIEW and TestStand workflows.

Emerson says an internal benchmark found reductions of up to 50 per cent in test-system development and deployment effort across representative workflows. That figure hasn’t been independently validated, and the announcement doesn’t provide model details, accuracy results or pricing.

For test engineers, embedded generation could remove repetitive work without forcing them out of familiar measurement and validation environments. The risk is unusually concrete, though: faulty test logic can pass into hardware validation and affect decisions about physical systems. The claimed productivity gain is therefore less important than whether teams can trace, review and reproduce every generated change before it reaches a critical test sequence.

What Changes for You

One last shift is immediately useful if you’re building agents in .NET.

Microsoft has released experimental support for discovering and loading Agent Skills from authenticated Model Context Protocol servers. It’s available through the Microsoft.Agents.AI.Mcp package, which currently has to be installed with the prerelease flag.

For working .NET agent teams, this replaces copied skill folders with a central distribution point. A platform or governance team can update a remotely hosted playbook once, and connected agents can pick up the revision during their next discovery cycle without being rebuilt and redeployed individually. Microsoft’s implementation supports standalone skill-markdown resources as well as ZIP, TAR and gzip-compressed TAR archives.

There are sensible controls: scripts contained in downloaded archives are never executed automatically, archive size and file-count limits can be configured, and skill tools require approval by default.

The limitation is maturity. Microsoft labels both the API and the underlying MCP skills specification experimental, so interfaces and interoperability may change. And centralisation creates its own security boundary. My recommendation for builders is to treat the MCP server like a software-supply-chain service: require authentication, retain tool approvals and enforce archive limits. Remote skills can make policy updates far easier, but a compromised distribution point could spread a bad instruction just as efficiently.