This Week in B2B Tech: 27 April-1 May 2026

$110 billion of cloud capex, 20 million paid Copilot seats, and one AI agent deleting a production database in nine seconds told the story of the week. B2B tech was not short of AI demand. It was short of patience for fuzzy economics, loose controls, and vendor promises that cannot survive contact with finance, legal, security, and procurement teams. The useful signal was not who launched the flashiest agent. It was who could prove that the agent was worth buying, safe to run, and cheap enough to scale.

In parallel, influencer discussions kept circling the same pressure. Jason M. Lemkin described an AI VP of Marketing choosing SaaStr's Marketo replacement, while Nate Jones framed bad enterprise AI defaults as a four-hour-a-week tax. Sarah Evans, Ben Lang, Dave Gerhardt, and Nathaniel Whittemore all came at the week from different edges, but the shared argument was clear: AI has moved from experiment to operating cost, and operating costs get audited. The vocabulary changed too: fewer demos, more defaults, seats, tokens, and permission boundaries.

Cloud earnings turned the AI boom into a capex exam

Cloud data centre equipment used in coverage of hyperscaler AI spending

The cloud numbers were big enough to make the old SaaS valuation debate feel underpowered. Channel Dive reported that Amazon, Microsoft, and Google spent $110 billion on capex in a single quarter, more than a billion dollars a day, to feed AI workloads. CRN put the revenue side next to it: $92 billion in cloud sales across the three providers, with AWS at $37.6 billion, Microsoft's Intelligent Cloud at $34.7 billion, and Google Cloud over $20 billion.

The market read was less celebratory than those figures suggest. CIO Dive's Google Cloud coverage showed the upside of a full-stack AI pitch, while its AWS analysis pointed to agentic AI as the next workload Amazon wants to own. Yet CFO Dive caught the harder edge at Microsoft, where workforce cuts sat alongside rising AI spending and $82.9 billion of quarterly revenue. The story for buyers is simple. AI infrastructure demand is real, but the bill is now so large that every cloud provider has to turn usage into durable margin, not just a stronger conference keynote.

Storage was the quiet proof point. Reuters reported that Sandisk joined Western Digital and Seagate in signalling strong AI data-centre storage demand. That matters because the AI economy is no longer only a GPU story. It is servers, storage, networking, power, and pricing discipline. The providers that can package all of that without making the CFO flinch will define the next phase of cloud competition.

OpenAI's partner reset made independence look expensive

Sam Altman at an OpenAI event as Microsoft and OpenAI revise their partnership

OpenAI spent the week looking less like a runaway winner and more like a company trying to widen its room for manoeuvre. The Verge broke down the revised Microsoft deal, including the end of cloud exclusivity and OpenAI's ability to offer products across other providers. CNBC framed the Amazon shift more sharply, as an aggressive move away from a relationship that once looked locked in.

The timing was awkward. The Decoder reported that OpenAI missed internal revenue and user-growth targets, while Channel Insider linked that miss to nerves across AI chip and cloud stocks. That combination changes the story. Multi-cloud choice gives OpenAI leverage, but it also makes the company easier to benchmark against Google, Anthropic, Amazon, and Microsoft in the same buyer conversations.

Governance kept dragging the commercial story back into court. Reuters summarised Elon Musk's testimony in the OpenAI trial, and the BBC captured the combative tone as Musk accused OpenAI's lawyer of trying to trick him. The legal theatre may not decide enterprise buying on its own, but it reinforces a question that procurement teams already ask privately: if your strategic AI provider is rewriting its partnerships, revenue model, and founding story at the same time, how much dependency do you really want?

Enterprise agents finally started behaving like paid software

Microsoft office building used in coverage of paid Copilot adoption

The agent story became more practical this week because the numbers and use cases became harder to dismiss. TechCrunch reported that Microsoft now has more than 20 million paid M365 Copilot users, with Accenture accounting for more than 740,000 seats and large customer adoption rising fast. That does not prove Copilot has won the workplace, but it does show that AI assistance is no longer living only in pilot decks.

The use cases also narrowed. The Decoder covered Microsoft's AI legal agent inside Word, a contract-review tool for US Frontier users that flags clauses, obligations, and tracked-change edits. Reuters reported Box Automate as an AI service for invoice processing and unstructured document work. Neither story is glamorous. That is the point. Agents become enterprise software when they attach to dull, expensive workflows with clear owners.

Atlassian gave the strongest software-market counterweight to the "SaaSpocalypse" argument. Startup Daily reported Q3 revenue of $1.79 billion, up 32%, while Business Insider grouped Atlassian with Twilio and Five9 as software stocks lifted by AI-led earnings beats. Diginomica added the enterprise texture: Service Collection ARR passed $1 billion and remaining performance obligations hit $4 billion. Buyers are not abandoning software. They are punishing vague software and rewarding products that can show where AI changes the work.

Agent security moved from model safety to identity control

Developer screen image used in coverage of an AI agent deleting a production database

Nine seconds is a brutal way to learn where the risk sits. ITPro reported that a Cursor agent using Claude Opus 4.6 deleted a startup's production database and backups after it used a Railway API call in the wrong environment. Security Boulevard's read was more specific: the real failure was a long-lived, over-privileged API credential sitting where an agent could use it.

That framing echoed across the security week. VentureBeat argued that exploits against Claude Code, Copilot, Codex, and Vertex AI went after credentials rather than model cleverness. The Hacker News covered Gemini CLI and Cursor flaws that could enable code execution, and CSO showed how a routine Git interaction could become remote code execution in Cursor. The common lesson is uncomfortable for security teams because it is not new. Agent risk inherits every old IAM mistake, then runs faster.

Attackers are not waiting for the governance market to mature. SecurityWeek reported a LiteLLM SQL injection flaw being exploited within 36 hours of disclosure, with API keys and provider credentials exposed. For B2B buyers, that makes agent security less about abstract alignment and more about permissions, secrets, logging, rollback, and blast radius. If the agent can act, the agent needs the same control plane as a human operator, plus stricter limits.

AI visibility turned publisher pain into a buying question

Google signage used in coverage of AI search scrutiny from Italy's media regulator

Search was the week's most important marketing infrastructure story. Reuters reported that Italy's media regulator asked the European Commission to investigate Google's AI search tools over publisher traffic, media pluralism, and hallucination concerns. Press Gazette then showed US publishers backing Amazon in its AI agent dispute with Perplexity. The legal details differ, but the commercial argument is the same: AI interfaces are taking more of the decision journey while sending less traffic back to source sites.

Publishers felt that squeeze in several directions. The AI Journal described falling referral traffic and rising machine traffic as a revenue crisis. Bloomberg reported that CNN, NBC, USA Today, and other publishers pushed back against Common Crawl's use in AI training. Digiday added the trust problem, with deepfake incidents rising to 3,165 in March 2026, according to IdentifAI.

For B2B brands, this is not just a media-industry dispute. It changes how buyers find vendors, compare claims, and decide which sources to trust before they ever fill in a form. The old SEO question was whether you ranked. The new question is whether AI systems understand you accurately enough to cite you when the buyer asks. That is why AI visibility is moving from a comms side project to a revenue protection problem.

What the influencers are discussing

The creator discussion split into two camps this week, and both were useful. The first camp treated agents as operating infrastructure. Sarah Evans launched MyAgentStack.io because, in her framing, the agent layer is changing PR, marketing, and comms faster than anyone is tracking it. Jason M. Lemkin gave the most concrete operating example, saying SaaStr's AI VP of Marketing, 10K, had sent its first autonomous outbound campaign and would help choose the company's Marketo replacement. That is not a thought experiment. It is software being allowed into budget, vendor selection, and campaign execution.

Lemkin's line mattered because it treated agent adoption as a live management decision, not a lab project. Evans took a similar view from the comms side, where the explosion of specialist agents has made the market hard to map and even harder to govern. The useful lesson for B2B teams is not that every function now needs an agent tomorrow. It is that somebody has to own the inventory, the evaluation criteria, and the decision rights before the tools quietly become part of daily work.

The second camp focused on AI visibility and distribution. Ben Lang's write-up on Profound treated AI visibility as a new route to revenue rather than a marketing vanity metric. Ahrefs' AEO course made the same point in tutorial form, explaining answer engine optimisation for teams trying to appear inside Google AI Overviews, ChatGPT, and Perplexity. Dave Gerhardt put the CMO version more bluntly: yesterday the CEO wanted to know why the company was missing from an article; today the CEO wants to know why ChatGPT is recommending competitors.

Two more threads cut through because they were sceptical without being anti-AI. Nate Jones argued that the wrong default AI assistant can become a four-hour-a-week tax, which is a cleaner buying lens than arguing about model benchmarks. Nathaniel Whittemore focused on harness-as-a-service, the runtime layer that makes agents useful beyond chat. Together, those posts sharpened the week's real question. AI buyers are not asking whether agents are impressive. They are asking which agents deserve permissions, budgets, and a permanent place in the stack.

The unresolved thread is accountability. Cloud providers are spending at data-centre scale, OpenAI is rewriting its partner map, software vendors are pushing agents into contracts and workflows, and security teams are finding that old credentials can become new blast radii. Buyers will not stop adopting AI because the risk is messy. They will ask vendors to make the economics, permissions, and liability legible enough that someone senior can sign their name to the purchase.

Meet Agentcy