Agent Infrastructure Matures

The enterprise AI conversation is shifting from capability to governance. Workday's expansion of its Sana agent system illustrates a fundamental truth: the bottleneck in deploying AI agents isn't model performance anymore. It's permissions, audit trails, and knowing who authorized what action on whose behalf.

This matters because most AI deployment discussions still focus on accuracy metrics and model selection. But in regulated environments like HR and finance, "almost right" creates liability, not value. The real infrastructure question is whether your system of record can verify that an agent acted within proper scope before executing a payroll change or scheduling decision.

The pattern extends beyond Workday. When companies offer free home cleaning in exchange for training footage, they're solving for embodied AI's equivalent problem: grounding models in physical reality. When AI replaces summer internships, it breaks the traditional pipeline that taught judgment and discretion, not just task completion.

Even Blue Origin's testing setback fits this frame. Mature infrastructure isn't about prototype success rates. It's about production readiness, verification systems, and knowing exactly what failed when things go wrong. The question isn't whether AI can do the job. It's whether the surrounding systems can prove it should have.

Deep Dive

Physical AI's data problem looks nothing like the text and image boom

The data bottleneck for robotics is fundamentally different from what powered the last AI wave. Companies offering free home cleaning in exchange for training footage signals that physical AI cannot scale the way language models did. There is no massive corpus of labeled physical interaction data sitting on the internet waiting to be scraped. The data has to be created, and that means paying people to generate it.

This creates a completely different economic structure. Text and image AI companies built moats by training on data that already existed, often without compensation. Physical AI companies have to construct data flywheels where they either pay workers directly to perform repetitive tasks while being filmed, or they offer services at a discount in exchange for footage. Shift is paying tens of thousands of people across 15 countries to record daily activities. Others are building what amount to data farms where workers fold towels and stack boxes repeatedly while sensors capture every movement.

The implications cut several ways. First, there is a real business in being a data intermediary for physical AI, which explains why companies are experimenting with everything from camera hats for gig workers to partnerships with home service platforms. Second, the capital requirements for robotics companies are higher than for pure software AI because training data is a direct ongoing cost, not a one-time scraping operation. Third, early movers who can collect proprietary physical interaction data at scale have a structural advantage that is harder to replicate than model architecture.

The wild card is whether companies that ship robots early, even when they still need human backup frequently, can use customer deployments as their data generation engine. That makes the unit economics work differently: you are getting paid to collect training data rather than paying to generate it. But it also means your customers are your data source, which introduces privacy and competitive concerns that text AI largely avoided.

The missing rung problem compounds faster than companies realize

Summer internship postings in tech have dropped 30% since 2023, but the second-order effects matter more than the headcount. Internships were never just about the work interns produced. They were the mechanism that taught judgment, professional norms, and domain context that AI cannot transfer. When you eliminate that rung, you do not just reduce current headcount. You break the pipeline that produces qualified senior people five years from now.

The immediate driver is clear: AI handles the structured, repetitive tasks that justified hiring interns. Research, first drafts, data entry, and basic analysis now cost tokens instead of supervision time. The economic logic favors AI for most companies. But what companies are optimizing for is this quarter's productivity, not whether they will have a functional talent pipeline in 2031.

The evidence that this creates a skills gap is already visible. McKinsey now tests candidates on how they collaborate with its AI assistant because AI fluency is table stakes. But AI fluency without domain experience produces what researchers call the Editor Problem: people who can prompt well but lack the judgment to evaluate output. That judgment historically came from doing the low-level work that AI now handles. The gap between "can use AI" and "knows when the AI output is wrong" is the difference between entry-level and experienced, and internships were how you crossed it.

The talent market is splitting. Senior AI roles pay $240,000 to $500,000 and require experience. Entry-level positions are disappearing. Some companies are pivoting to apprenticeships as a longer, more structured alternative, but apprenticeships require more corporate investment than summer internships and fewer companies are willing to make that commitment. The result is a smaller pipeline with higher barriers to entry, which particularly affects people without existing professional networks.

For founders, this is not just a hiring problem. It is a culture and succession problem. The next generation of senior talent is being trained by AI tools, not by experienced people. What that produces at scale is an open question, but the early evidence suggests companies are trading short-term productivity gains for long-term organizational capability they have not priced in.

Signal Shots

Microsoft Threatens Researcher Over Zero-Day Disclosures: Microsoft warned it may pursue criminal action against a security researcher who published unpatched vulnerabilities in Defender and BitLocker without coordinating disclosure. The researcher claims Microsoft revoked their reporting portal access, leaving public disclosure as the only option. This matters because it could chill security research if companies can credibly threaten prosecution for disclosure practices the industry debated and largely settled a decade ago. What to watch: whether Microsoft's threat creates a measurable drop in vulnerability reports and whether other enterprise software vendors adopt similar aggressive stances when researchers bypass their preferred disclosure channels.

Dell Posts 757% AI Server Revenue Growth: Dell reported AI server revenue of $16.1 billion, up 757% year over year, driving its stock up 32% in a single day for its best trading session ever. Total revenue grew 88% as GPU demand from companies building AI infrastructure showed no signs of slowing. This matters because Dell's results confirm that AI infrastructure spending remains concentrated at the hardware layer, not just cloud services, and that enterprise buyers are still racing to build internal capacity. What to watch: whether this pace sustains through 2026 or whether we see the typical infrastructure buildout pattern where early surge is followed by digestion period as utilization catches up to capacity.

BlackBerry Stock Surges on Robot OS Positioning: BlackBerry shares jumped 160% over three months as its QNX automotive software division pivots to positioning itself as an operating system for robots, partnering with Nvidia's robotics platform. QNX already has adoption in autonomous driving systems. This matters because it shows how companies with domain expertise in real-time embedded systems can potentially capture value in the physical AI stack, even if they missed the smartphone transition entirely. What to watch: whether QNX can demonstrate actual robot deployments beyond automotive or whether this remains a speculative repositioning that collapses when retail enthusiasm for AI stocks fades, as it has in previous BlackBerry rallies.

Okta Builds Kill Switch for Rogue AI Agents: Okta launched AI agent governance products that let enterprises directory their autonomous agents, set access policies, and revoke permissions when agents behave outside parameters. ServiceNow specifically requested the capability to sever agent access tokens at the authorization layer. This matters because 92% of enterprises report deploying AI agents but only 22% have tied identities to those agents, creating an exposure most companies have not yet addressed. What to watch: whether this becomes table stakes infrastructure that every identity platform must offer or whether the kill switch concept proves inadequate because most agent failures require quarantine and analysis, not just permission revocation.

Pinterest Cuts AI Costs 90% With Custom Vision Layer: Pinterest reduced AI inference costs by 90% by removing the vision encoder from Qwen's multimodal model and replacing it with proprietary embeddings that precompute image metadata offline. The customization also improved accuracy by 30% for their 620 million monthly users. This matters because it demonstrates that companies with unique data and use cases can achieve better economics and performance by rebuilding parts of foundation models rather than using them as-is. What to watch: how many other companies with scale and proprietary data follow this pattern versus continuing to pay frontier model API costs, and whether model providers respond by making their architectures more modular to facilitate this kind of customization.

Developer Sabotages AI Coding Agents With Prompt Injection: A developer added a hidden prompt injection to jqwik, a Java testing library, instructing AI coding agents to delete all test code. The instruction was concealed from human terminal viewers using escape sequences. This matters because it represents developer backlash against AI coding tools in a form that could cause real damage if agents follow the instructions, though major tools like Claude flagged it without executing. What to watch: whether this becomes a pattern where open source maintainers hostile to AI adoption inject adversarial prompts into widely used libraries, and whether that forces coding agent providers to implement more robust prompt injection defenses or sandbox execution environments.

Scanning the Wire

Groq raises $650M to pivot from chips to AI inference software: The AI chipmaker is shifting focus toward optimizing how models respond to prompts rather than competing on raw compute hardware, following Nvidia's $20B acquisition of rival startup engineers. (TechCrunch)

South Korean chip startup XCENA raises $135M betting memory, not compute, is AI's bottleneck: The company is targeting the data transfer and storage layer as the constraint that will define next-generation AI performance. (TechCrunch)

Nvidia teases N1X laptop processors at Computex: Microsoft, Nvidia, and Arm are all openly promoting the Arm-based laptop chip announcement expected this weekend, marking Nvidia's formal entry into the PC processor market. (The Verge)

Meta internal memo reveals AI pendant testing in 2027, new glasses next month: The company plans to launch updated AI glasses in June, begin testing a standalone AI pendant next year, and establish a new Wearables for Work division targeting enterprise deployments. (The Information)

Dell's $9.7B Pentagon contract raises questions after $6.25B Trump donation: The Department of Defense awarded Dell the hardware contract months after Michael and Susan Dell donated billions to fund Trump Accounts for 25 million US children. (CNBC)

Kentucky school district secures $27M in social media harms settlements: Meta paid $9M, the largest share, with Snap, TikTok, and YouTube contributing the remainder in settlements over claims the platforms fueled student mental health issues. (Reuters)

IBM and Red Hat launch $5B open-source security initiative: Project Lightwell uses AI to find and fix vulnerabilities in open-source software at industrial scale, backed by 20,000 engineers. (ZDNet)

Tesla Robotaxi fleet registers just 42 vehicles in Texas: The company's driverless service footprint is less than one-tenth the size of Waymo's fleet in the state, according to regulatory filings. (CNBC)

Mistral AI targets $1.17B revenue with industrial AI push and Paris data center: The French startup announced manufacturing partnerships with Airbus and BMW, a new inference facility south of Paris, and rebranded its consumer assistant to Vibe at its inaugural conference. (VentureBeat)

Anthropic releases Claude Opus 4.8 with 3X cheaper fast mode: The new model costs $5 per million input tokens and approaches the alignment quality of the restricted Mythos Preview model, while fast mode pricing drops from $30 to $10 per million input tokens. (VentureBeat)

Authorities dismantle 17 million device botnet tied to Russian proxy network: The takedown targeted infrastructure reportedly used for residential proxy services. (Ars Technica)

Proposed US funding rules would allow grant cancellation without cause: New guidelines would make peer review optional and let political staff screen research proposals for prohibited topics, with termination rights at any time. (Ars Technica)

Outlier

Peer Review as Optional Feature: The US government is floating new funding rules that would make peer review discretionary, let political appointees screen grant proposals for forbidden research topics, and allow termination of any grant without cause at any time. This is not about efficiency or reducing red tape. It is about replacing the scientific method's core quality control mechanism with political veto power at the individual project level. What this signals is a potential structural break in how knowledge production gets funded in the world's largest research economy. If implemented, it creates conditions where researchers optimize for political acceptability rather than scientific merit, universities lose the ability to plan multi-year projects with any confidence, and the US funding advantage that has drawn global talent for decades becomes a liability as researchers redirect to more stable environments. The second-order effect is what matters: when you make the rules arbitrary, you do not just change behavior at the margin. You select for different types of people entering the field entirely.

The infrastructure always matures slower than the demos suggest, which is why the most interesting companies right now are building the boring stuff: permission systems, data flywheels, kill switches. The robots will figure out how to fold towels. Whether anyone trusts them to is a different stack entirely.