The Serious Insights State of AI 2026 May Update: Capital concentrates as trust and infrastructure lag

The Serious Insights State of AI 2026 May Update

May 2026 did not overturn the core findings from the State of AI 2026 report. It reinforced them. The month made clear that AI has moved beyond model competition as the primary story. The sharper questions now involve operational control, institutional trust, energy access, regulatory maneuvering, and the ability of organizations to absorb AI into work without turning every deployment into another unmanaged experiment.

The April update argued that the AI market had shifted from capability announcements to infrastructure, governance, energy, and capital concentration. May pushed that argument further. Google’s I/O announcements put AI deeper into search, personal productivity, agents, smart glasses, and the consumer operating environment. Anthropic’s Opus 4.8 and planned Mythos rollout continued the separation of public frontier models from restricted high-capability systems. OpenAI’s foundation committed $250 million to economic-disruption work, signaling that labor-market impact is no longer a secondary social-policy issue but part of the AI platform narrative itself. At the same time, U.S. federal AI oversight stalled at the White House, reinforcing the fragmented governance environment tracked throughout this report series.  

Key Takeaways

• AI moved further into the interface layer. Google’s AI search changes, Gemini updates, and smart-glasses push show that AI is becoming the default mediation layer between people and information, not just another application category.
• Agentic AI moved from aspiration to product architecture. The May news around Google, Anthropic, Microsoft, and a number of research all point to the same conclusion: agents are becoming embedded in work systems faster than organizations are learning how to govern them.
• The governance gap widened. Grant Thornton’s 2026 AI Impact Survey found that 78% of organizations lack strong confidence that they could pass an independent AI governance audit. That finding aligns with other research showing widespread AI use, weak policy maturity, and limited trust in systems already being operationalized.
• AI trust became a product feature. Anthropic’s emphasis on Opus 4.8’s honesty and uncertainty handling, Morning Consult’s AI Trust Report, and HireVue’s findings on AI in hiring all point in the same direction: trust is no longer brand reputation language. It is becoming a measurable constraint on adoption.
• Search became a strategic risk surface. Research on Google AI Overviews found that 11% of atomic claims were unsupported by cited pages, even as AI summaries increasingly shape what users see before they click. That matters for publishers, brands, researchers, and any organization that depends on being discovered through the open web.  
• The labor-market debate became more explicit. OpenAI’s $250 million foundation commitment to workers and economic disruption should be read as more than philanthropy. It acknowledges that AI labor displacement, wage pressure, and workforce transition are becoming central to public legitimacy.  
• The U.S. regulatory posture remains unstable. A postponed AI executive order shows that federal AI policy remains caught between national-security review, industry pressure, state-level action, and the geopolitical imperative to compete with China.  

AI Moves Into the Interface Layer

Google I/O 2026 produced the month’s clearest indicator that AI is moving from tool to interface. Google announced a major AI-centered overhaul of Search, new Gemini capabilities, information agents, and a renewed push into smart glasses through partnerships with Samsung, Warby Parker, and Gentle Monster. The strategic implication is not that Google added AI features. Google is repositioning AI as the mediation layer across search, personal data, video, commerce, mobile, and wearable computing.  

That move reinforces a position from the earlier 2026 reports: AI adoption will not occur only through deliberate enterprise procurement. It will arrive through the normal refresh cycle of search, office suites, operating systems, browsers, collaboration tools, phones, and consumer devices. Organizations that think of AI governance only as a procurement-control issue will miss the larger problem. AI will enter through the interface, not just through the contract.

The emerging risk is that interface-level AI changes the organization’s information diet without appearing as a traditional system of record. Google AI Overviews research published in May found that AI Overviews were activated in 13.7% of measured trending queries overall and 64.7% of question-form queries. The study also found that 11% of decomposed claims were unsupported by cited pages. That is not simply a search-quality issue. It is an epistemic infrastructure issue.  

For Serious Insights readers, the implication is important: every organization now needs an AI visibility strategy. Not SEO with AI terminology pasted on top, but a strategy for being represented accurately in AI-mediated answers, summaries, recommendations, and agent-driven workflows.

Frontier Models: Capability Still Matters, But Behavior Matters More

Anthropic’s May release of Claude Opus 4.8 points toward a subtle but important shift in frontier-model competition. Anthropic emphasized improved performance on advanced reasoning and coding tasks, but also framed the release around more reliable behavior, including better calibration, transparency, and a greater willingness to acknowledge uncertainty. That positioning matters because frontier competition is no longer only about benchmark leadership. It is also about whether models can act as trustworthy components in governed enterprise workflows. Reuters separately reported that Anthropic plans to roll out Claude Mythos, tied to Project Glasswing and cybersecurity-oriented access, in the coming weeks.

Anthropic is not moving in a vacuum. OpenAI’s official release of the GPT-5.5 Frontier Model this month set new baseline milestones for agentic coding and computer use, immediately followed by an aggressive sunsetting of older architectures like o3 and GPT-4.5. The market is no longer tolerating model fragmentation; providers are forcing transitions to high-capability, high-supervision tiers.

May’s security research added a further layer to this split. Reporting from the UK AI Security Institute, AISI indicates that two frontier models, Anthropic’s Claude Mythos Preview and OpenAI’s GPT‑5.5, cleared a 32‑step end‑to‑end cyber‑attack range within a single month, and that frontier cyber-offense capability is now doubling roughly every four months, up from a seven‑month doubling rate at the end of 2025. That reinforces the case for restricted, high‑supervision tiers and for treating frontier models as dual‑use assets whose deployability is constrained as much by security and auditability as by raw performance.

That reinforces the April update’s point about restricted high-capability systems. The frontier market is splitting. One tier optimizes for broad commercial availability. Another tier will be restricted, monitored, and aligned with specific use cases such as cybersecurity, national security, or high-risk scientific work.

The competitive message is no longer just “more capable.” It is “more governable,” “more honest,” “more transparent,” “more controllable,” or “available only under structured access.” This does not make the systems safe by default. It does mean that frontier AI vendors are beginning to compete on behavioral claims that CIOs and CTOs can translate into governance requirements.

Those claims still require verification. Organizations should not treat vendor-provided statements about honesty, transparency, or safety as controls. They require internal testing, contractual obligations, audit rights, and incident-response procedures.

Agentic AI: The Readiness Gap Is Now the Main Story

Our analysis confirms that AI agents are becoming part of organizational infrastructure before most organizations have the operating model required to manage them. Microsoft’s 2026 Work Trend Index frames the issue in terms of agents, human agency, and the need for organizations to become learning systems. ServiceNow’s agentic-business brief emphasizes orchestrated work. Futurum’s research on agentic AI platforms points to multi-agent collaboration, orchestration, governance, security, and observability as core requirements for enterprise platforms.

Grant Thornton’s 2026 AI Impact Survey makes the risk more concrete. It found that 78% of organizations lack strong confidence that they could pass an independent AI governance audit. The same report states that most organizations are scaling AI that they cannot explain, measure, or defend. That is the AI proof gap, and it should become one of the central operating concepts for CIOs and CTOs in 2026.

The “proof gap” aligns with the original State of AI 2026 report’s argument for Agent Ops. Agents require monitoring, role-based authority, escalation paths, behavior versioning, audit trails, and defined boundaries for tool use. A chatbot can be wrong. An agent can be wrong and act. That distinction changes the governance burden.

May’s agentic AI news, including Google’s information agents and Microsoft’s reported plan for new task-specific models, confirms that agents are becoming the normal architecture of AI-enabled work rather than a specialized automation category.  

Parallel May coverage of headless enterprise architectures and “agent‑native” runtimes, including Salesforce’s API‑first shift and Stripe’s and Cloudflare’s co-developed protocols for agents that can autonomously provision accounts and deploy applications, shows how quickly the surrounding software ecosystem is restructuring itself to make agents first‑class operational entities.

Salesforce’s headless push and Agent Fabric work show one path: turning CRM into an API‑first data and workflow layer where agents operate directly against records under tight governance and observability controls, with humans increasingly consuming agent‑curated views rather than hand‑driving UIs.

Stripe is tackling a different layer by building an “agentic commerce” stack with wallets, shared tokens, machine payments, fraud signaling, and micropayment billing that lets agents safely hold scoped payment authority and transact across existing card and stablecoin rails. Cloudflare, in turn, is going after the runtime, rolling out isolate‑based Dynamic Workers and Project Think to execute agent‑generated code quickly and cheaply at the edge, which makes agent workloads operationally feasible without the overhead of traditional container stacks.

This architectural shift is hitting the desktop environment. Microsoft’s active document-editing rollout in Office and its positioning of Windows as an agent platform at Build 2026 highlight the enterprise dilemma. To contain IT admin concerns about autonomous agents on corporate networks, Microsoft is leveraging Windows 365 Cloud PCs as isolated “virtual cubicles” — ephemeral, task-scoped environments where agents check out compute, execute a task, and check back in, with Entra Agent IDs, audit logs, and human-in-the-loop controls built into the stack.

The practical recommendation remains unchanged but has become more urgent: organizations need agent inventories. They need to know which agents exist, which systems they can touch, which actions they can take, which data they can access, who owns them, and how they can fail.

Work, Skills, and the AI Absorption Problem

Multiple industry reports, presentations and events confirm one of our key tenets: AI adoption is not the same as AI absorption. Adoption means a tool is available. Absorption means the organization has changed its practices, controls, incentives, workflows, and learning mechanisms enough to create durable value.

Microsoft’s Work Trend Index argues that every firm must become a learning system. Randstad’s AI capability-gap report argues that scaling AI requires continuous upskilling, not just better tools. IBM’s 2026 CEO Study reports that CEOs see talent and technology leadership converging, with 77% saying those roles are coming together. These findings point to the same conclusion from different angles: AI strategy cannot sit inside IT alone.

The scale of the absorption challenge is clear in the massive enterprise footprints emerging this month. KPMG’s deployment of Claude to 276,000 employees globally represents a major shift. Rather than providing simple chatbot access, the firm integrated managed agentic workflows straight into its core client delivery platform, prioritizing structural trust and governance over simple deployment speed.

This is where many organizations will underperform. They will license AI broadly, count seats or prompts, and declare progress. Then they will discover that people do not know when to trust the output, managers do not know how to redesign work, HR does not know how to update roles, legal does not know how to audit agent behavior, and finance does not know how to measure value beyond activity.

The absorption problem is also visible in hiring. HireVue’s 2026 Global AI in Hiring Report found that 77% of HR teams use AI weekly or daily, but only 41% trust those systems. It also found that 71% of candidates use AI to write resumes. Hiring has become an AI-on-AI environment before trust, transparency, and evidence have caught up.

For organizations, that should trigger a reassessment of talent processes. The resume is losing evidentiary value. AI-generated candidate materials, AI-screened applications, and AI-assisted interviews create a closed loop that can amplify sameness, obscure authenticity, and weaken judgment if humans do not redesign the process around demonstrated capability.

Trust Is Now a Deployment Constraint

The AI Trust Report 2026, based on Morning Consult data, adds a consumer-facing dimension to the enterprise governance story. Trust in AI brands is not just a marketing issue. As AI becomes embedded in search, hiring, customer service, public services, financial advice, and health-related workflows, trust becomes a condition for adoption.

The recent government data reinforces the same point from another sector. The Granicus 2026 State of Digital Government report found that 55.7% of government organizations now use AI, but only 42.9% report having formal AI policies. That mismatch proves critical because government AI touches services, benefits, permits, public information, and citizen trust.

This finding is not an argument against government AI. It is an argument against unmanaged government AI. Public-sector organizations face the same absorption challenge as enterprises, but with less room for reputational failure. If AI tools improve service speed while weakening transparency, the long-term result may be lower trust rather than a better public experience.

The trust theme also connects to Anthropic’s May positioning around Opus 4.8. “Honesty” and uncertainty-handling are becoming commercial differentiators because users and organizations have learned that fluent answers are not the same as reliable ones. That shift should be welcomed, but not overread. A more honest model still needs external validation when the stakes are material.

GoTo also provides evidence that AI risk has already moved from hypothetical to operational. Around one-quarter of IT leaders say AI mistakes have affected customers, clients, or the company’s bottom line, 16% say AI errors have disrupted business operations, and 91% worry AI could make a mistake that negatively affects the organization. Those numbers support the May update’s argument that AI governance has to become an operating model, not a policy artifact.

Regulation: The U.S. Still Does Not Have a Stable Federal Center

The postponed U.S. AI executive order became one of May’s clearest governance signals. Reporting indicates that the draft would have established a voluntary federal review process for advanced AI systems prior to release, with particular attention to national security and cybersecurity risks. President Trump postponed the signing, saying he did not like aspects of the order. Additional reporting pointed to pressure from Silicon Valley leaders concerned that voluntary review could evolve into de facto licensing.  

This event reinforces the regulatory-fragmentation theme from February, March, and April. The U.S. remains caught between federal restraint, state-level activism, national security concerns and the industry’s argument that regulation will weaken competition with China. The result is not deregulation. It is uncertainty.

For CIOs and CTOs, uncertainty is not an excuse to wait. Internal AI governance needs to assume that external policy will remain fragmented. That means mapping AI use cases by risk, documenting controls, preparing for audit, and watching state-level obligations as closely as federal action.

The EU remains the more structured regulatory environment, but even there, the direction is not simply toward stricter rules. The April update discussed the Digital Omnibus proposal and the potential delay of some high-risk AI obligations. May did not resolve that tension. The broader pattern remains: governments want AI investment, but they also want risk control. Those goals often collide once regulation moves from principle to implementation.

Infrastructure and Energy: The Constraint Has Not Gone Away

May did not produce a single infrastructure event as dramatic as April’s data-center power warnings, but the underlying constraint remains central. Industry research continues to illustrate that AI infrastructure has become a power-system issue, not just a cloud-capacity issue.

Recent research on AI data-center locations found that projected AI compute capacity is concentrated in North America, Western Europe, and Asia-Pacific, which together account for more than 90% of the total. The same paper identifies regional stress risks in places such as Oregon, Virginia, and Ireland, where concentrated demand can create local grid vulnerability.  

May also underscored that open‑weight and smaller‑scale models are not an escape hatch from these constraints. Several May releases highlighted increasingly capable open‑weight audio, video, and multimodal models that can run on a single GPU with licenses designed for production, which shifts some workloads to enterprise or edge infrastructure but does not eliminate energy, governance, or proof obligations; it simply redistributes them from hyperscalers to organizations that may be even less prepared to manage them.

If there is any doubt that compute capacity is a power-grid issue, look at the capital moving into the dirt. NVIDIA’s $2.1B deal with IREN to unlock 5 gigawatts of infrastructure—anchored by a staggering 2-gigawatt facility in Texas—proves that AI availability is completely tethered to massive regional energy plays. Sourcing AI is no longer a software procurement task; it is a heavy-industrial supply chain strategy.

This supports our original report’s position that energy architecture must be a component of AI strategy. AI cost cannot be reduced to tokens, GPUs, or cloud invoices. It must also include power availability, location, latency, cooling, water, sustainability exposure, and contractual access to capacity.

The next strategic shift will likely involve workload placement based not just on cost and performance, but on power availability, carbon intensity, regulatory exposure, and resilience. That is where AI infrastructure begins to look less like traditional IT sourcing and more like supply-chain strategy.

Capital Concentration and the Social License Problem

OpenAI’s $250 million foundation commitment for workers and economic disruption is notable because it places the labor-market transition within the AI narrative. OpenAI is not alone in acknowledging the social risks of AI, but the scale and framing of the commitment make the issue harder for the industry to treat as external.  

May’s finance headlines showed how quickly capital is compounding around a few platforms. Industry reporting suggests that Anthropic is closing a funding round structured around a long‑term compute commitment that could push its valuation toward or beyond the levels previously associated with OpenAI, even as OpenAI itself reportedly lines up an IPO window for later in 2026. Those moves matter not because IPOs are novel, but because public‑market expectations can harden short‑term growth and deployment goals at the very moment when regulators, infrastructure planners, and affected workers need providers to slow down long enough to demonstrate proof, governance, and trust

The capital-concentration story from April remains intact. The largest AI firms continue to attract capital, talent, infrastructure, and political attention at a scale that few markets have seen. The emergent May concept is the social license problem: AI firms may win customers, developers, and investors while losing public confidence if labor disruption, publisher harm, energy strain, and opaque governance accumulate faster than visible benefits.

That risk will not be solved by foundations alone. It will require evidence that AI improves work rather than merely extracting it, improves public services rather than masking understaffing, improves access to knowledge rather than enclosing it, and improves decision quality rather than automating institutional bias.

Emergent Concepts Not Fully Reflected in the Previous Reports

AI visibility management

AI search and answer engines create a new organizational need: managing how an organization is represented inside synthesized AI answers. This is adjacent to SEO, but not identical. It requires authoritative content, structured evidence, third-party validation, clear product and policy language, and monitoring of AI-generated summaries.

The AI proof gap

Grant Thornton’s phrase deserves adoption because it names a widespread condition: organizations deploying AI faster than they can explain, measure, audit, or defend it. The proof gap may become the practical bridge between AI governance, risk management, and ROI. (Grant Thornton 2026 A Impact Survey).

Agent inventories

Organizations will need formal inventories of AI agents, not just models or applications. An agent inventory should include owner, purpose, connected systems, permitted actions, data access, escalation rules, evaluation results, version history, and incident logs.

Interface governance

AI governance needs to expand beyond model and application governance. Search boxes, taskbars, browsers, office suites, phones, and wearables are becoming AI-mediated work environments. The interface should now be considered a governance surface.

Synthetic trust markets

As hiring, search, reviews, media, and customer interactions fill with AI-generated content, trust will increasingly depend on provenance, verification, identity, reputation, and controlled channels. This will create new markets for validation, auditing, watermarking, and content authentication.

AI-on-AI work loops

Hiring already shows the pattern: candidates use AI to create resumes; employers use AI to screen resumes; managers use AI to summarize interviews; candidates use AI to prepare answers. Similar loops will appear in procurement, sales, customer service, compliance, and research. These loops can improve throughput while degrading evidence if not redesigned.

What Organizations Should Do Now

Serious Insights suggests that organizations take the following actions based on our May 2026 research:

• Treat May as confirmation that AI has entered the absorption phase. The strategic question is no longer whether AI will matter. It already does. The question is whether the organization can absorb it without losing accountability, trust, and coherence.
• Build an agent inventory before agents proliferate beyond visibility.
• Create an AI proof standard that defines what must be measured, explained, logged, and audited before a system moves from pilot to production.
• Update governance for interface-level AI, including search, browser, collaboration, and productivity environments.
• Treat AI visibility as part of brand, communications, knowledge management, and risk.
• Tie AI investments to work redesign, not just seat deployment.
• Evaluate trust as an operating constraint. If users, employees, customers, candidates, or citizens do not trust the system, technical success will not translate into adoption.
• Assess energy and infrastructure exposure for any AI roadmap that depends on high-volume inference, model fine-tuning, or data-intensive automation.

---

Bottom Line

May 2026 confirms that the center of gravity in AI continues to shift. Models still matter, but the strategic story now sits around absorption, governance, infrastructure, trust, and the interface layer. The organizations that succeed will not be those that buy the most AI or announce the broadest deployments. They will be the ones that prove where AI works, govern where AI acts, redesign work around what AI changes, and maintain enough institutional discipline to know when not to automate.

The May update reinforces the original State of AI 2026 position: AI is no longer a technology trend moving toward organizations. It is already inside them. The work now is to make it accountable.

Sources

Note: some embedded sources are not restated here.

Serious Insights, “The Serious Insights State of AI 2026 April Update: How Power, Capital, and Governance Will Shape the Next Wave of AI”
https://www.seriousinsights.net/state-of-ai-2026-april-update/.

Serious Insights, “The Serious Insights State of AI 2026 March Update: The Capabilities, Infrastructure, and Deployment Gaps”
https://www.seriousinsights.net/state-of-ai-2026-march-update/.

Serious Insights, “The State of AI 2026 February Update: Autonomy, Regulation, Synthetic Data and Security”
https://www.seriousinsights.net/the-sstate-of-ai-2026-february-update/.

AI Security Institute (AISI) Our evaluation of Claude Mythos Preview’s cyber capabilities. https://www.aisi.gov.uk/blog/our-evaluation-of-claude-mythos-previews-cyber-capabilities.

GoTo. Pulse of Work 2026. https://www.goto.com/resources/pulse-of-work-2026.

Google. 100 things we announced at I/O 2026. https://blog.google/innovation-and-ai/technology/ai/google-io-2026-all-our-announcements.

Grant Thornton. 2026 AI Impact Survey Report. https://www.grantthornton.com/services/advisory-services/artificial-intelligence/2026-ai-impact-survey.

Reuters, “Anthropic to roll out Claude Mythos in coming weeks, launches Opus 4.8.”
https://www.reuters.com/business/anthropic-roll-out-claude-mythos-coming-weeks-launches-opus-48-2026-05-28/.

Reuters, “OpenAI Foundation commits $250 million to help workers, economies navigate AI disruption.”
https://www.reuters.com/business/openai-foundation-commits-250-million-help-workers-economies-navigate-ai-2026-05-27/.

Reuters, “Microsoft to release new coding model next week, The Information reports”
https://www.reuters.com/business/microsoft-release-new-coding-model-next-week-information-reports-2026-05-28/.

The Guardian, “Google announces glasses are back, and search is getting an AI makeover.”
https://www.theguardian.com/technology/2026/may/19/google-glasses-search-ai.

The Verge, “Claude’s new model is more ‘honest’ when it messes up.”
https://www.theverge.com/ai-artificial-intelligence/939094/anthropic-claude-4-8-opus-honesty-effort.

New York Post, “Trump says he postponed signing AI order because he didn’t ‘like’ it.”
https://nypost.com/2026/05/21/us-news/trump-says-he-postponed-signing-ai-order-because-he-didnt-like-it/.

Washington Post, “Last-minute lobbying by tech industry officials led Trump to cancel AI order.”
https://www.washingtonpost.com/politics/2026/05/22/last-minute-lobbying-by-tech-industry-officials-led-trump-cancel-ai-order/.

Business Insider, “Read the AI executive order Trump didn’t sign.”
https://www.businessinsider.com/trump-ai-oversight-executive-order-draft-2026-5.

Xu, Iqbal, and Montgomery, “Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact”
https://arxiv.org/abs/2605.14021.

Churilov, “The Range Shrinks, the Threat Remains: Re-evaluating LLM Package Hallucinations on the 2026 Frontier-Model Cohort”
https://arxiv.org/abs/2605.17062.

Chen et al., “Concentrated siting of AI data centers drives regional power-system stress under rising global compute demand.”
https://arxiv.org/abs/2604.06198.

For more serious insights on AI, click here.

All images via ChatGPT from a prompt by the author.

Did you enjoy The Serious Insights State of AI 2026 May Update? If so, please like, share, or comment. Thank you.