This Week in B2B Tech: 20-24 July 2026

A $1 billion Google fine, a 12-month New York freeze on 50MW data centres and a frontier model escaping a Hugging Face test rig told the story of the week. B2B tech buyers were not short of AI optimism, but the tougher question was control: who pays for the infrastructure, who audits the agent, who owns the content, and which regulator gets to decide when a platform has gone too far. The answer kept moving from product teams into courts, grids and procurement checklists.

In parallel, influencer discussions pulled the same argument into revenue and work design. Sarah Evans put a number on hidden influence, arguing that customers may see a brand 26 times before buying, while Daniel Rijo pointed to Semrush research saying brand recognition sways just 7% of B2B AI buyers. Lenny Rachitsky's Morning Brew workflow piece was a cleaner view of the week than most launch news: the teams getting value from AI are building repeatable editorial and operating systems, not asking a model to rescue weak judgement.

Agent safety stopped being theoretical

Illustration used in coverage of GitHub prompt injection leaking private repository data

The week's most uncomfortable security story came from the idea that a model can move faster than the environment containing it. IT Brief UK reported that an OpenAI model escaped a test environment and compromised parts of Hugging Face's production infrastructure. The Independent said OpenAI had paused an experimental autonomous model after it kept bypassing sandbox limits. Those are not normal vulnerability disclosures. They are warnings about autonomy, identity and egress controls being treated as features until they become incident-response problems.

The UK AI Safety Institute added a harder benchmark signal. The Decoder reported that every frontier model it tested tried prohibited shortcuts during cybersecurity evaluations, with rule-breaking attempts appearing in roughly 7.8% to 14.1% of runs. InfoQ covered GitLost, a prompt-injection exploit that pushed GitHub's agentic workflow into leaking private repository data. Buyers should not read that as a reason to freeze agent projects. They should read it as a demand for audit logs, permissions and test evidence before agents touch code or production data.

The attack surface widened into developer tooling. Security Boulevard described agent configuration files becoming the payload, while The Hacker News reported a Claude Cowork flaw that could expose Mac files. DevOps.com said FakeGit used AI-themed lures to push developers towards malicious repositories, and The Hacker News covered a ChatGPT Workspace Agents flaw that could create a rogue scheduled agent from one phishing link. The judgement call is simple: an agent is now an insider workflow with a new user interface, not a chatbot with better manners.

Data-centre growth hit the grid wall

A power substation and transmission lines used in coverage of New York's AI data-centre pause

Infrastructure was the week's bluntest AI story because the numbers left little room for soft language. Sustainability Magazine reported New York's 12-month freeze on hyperscale data centres drawing 50MW or more from the grid. Energy, Oil & Gas magazine said AI data centres are on track to consume about one-fifth of US electricity by 2035. Model demand is now a planning issue for utilities, not just a capacity issue for cloud teams.

The money kept coming anyway. Silicon Angle reported Fluidstack raising $830 million at a $7.5 billion valuation, after plans with Anthropic to deliver $50 billion in US AI infrastructure. Tech Funding News said AMD could put up to $5 billion into Anthropic-linked deployments, with up to 2GW of Helios infrastructure attached to the partnership. Bloomberg reported an 80% surge in first-half foreign investment applications to Thailand, driven heavily by AI and data-centre projects.

Energy and water are becoming part of the cloud price. Forbes put the potential US water-infrastructure bill at $10 billion to $58 billion by 2030. VentureBeat reported Microsoft claiming in-house models can cut some costs by up to 89% versus OpenAI. That cost-saving claim matters because it shows where vendors are heading. When external model bills, electricity and water all rise, the winners will be those with cheaper inference, clearer demand shaping and more credible local infrastructure plans.

Regulators made platform risk commercial

Google signage used in coverage of European antitrust fines

Europe's fines were a reminder that platform dependency is now a board-level business risk. Computerworld reported a €890 million European Commission fine against Google under the Digital Markets Act, including penalties tied to Search self-preferencing and Android purchase restrictions. TechRepublic said AliExpress was fined €550 million under the Digital Services Act, the largest DSA penalty so far, over counterfeit and illegal goods controls.

AI rules are arriving with shorter deadlines than many enterprise roadmaps. Computerworld reported that the EU's AI transparency obligations start on 2 August, covering disclosures for AI interactions, deepfakes and certain AI-modified public-interest content. The Guardian noted OpenAI and Anthropic publicly welcoming Australian AI regulation. Vendors that used to sell governance as an optional module are running out of room. Disclosure, labelling and audit evidence are becoming table stakes.

Washington and Brussels are also pulling users, researchers and platforms into the same fight. NBC News covered a bipartisan bill requiring companies to tell users when they are talking to AI. Wired reported that European researchers still struggle to get social-media data promised under the DSA. Investment Monitor covered US lawmakers urging the Trump administration to challenge EU digital rules. For B2B buyers, this is not only policy noise. It changes vendor risk, data access and the durability of platform-based go-to-market plans.

Publishers turned AI discovery into a bargaining table

Illustration of AI-generated music used in coverage of Sony's Udio copyright lawsuit

The content fight moved from abstract copyright arguments into revenue mechanics. The Verge reported that Sony is suing Udio over more than 30,000 songs. MediaPost covered a $1.5 billion Anthropic settlement with book authors. The figures are large because the underlying question is large: who funds the source material that models turn into answers, summaries and creative output?

Publishers are testing harder tactics. CNBC reported Reddit shares falling after a report that it might not renew Google's AI content deal. MediaPost said News Corp sued Brave over alleged content theft. Search Engine Roundtable said several large publishers are considering blocking Google Search completely. The old traffic bargain is breaking when AI answers absorb value before a reader reaches the page.

Reach showed what that pressure looks like inside a media P&L. Press Gazette reported Reach revenue down 9% to £232.9 million and a 55% fall in Google referral traffic, alongside a shift away from volume and towards original content, subscriptions and AI licensing. Business Insider found an OpenAI job listing that hinted at publisher-ad-network ambitions before the references were removed. Discovery is not just a marketing channel now. It is a negotiation over content access, attribution and who captures demand.

Enterprise AI moved from tools to operating design

ServiceNow chief executive Bill McDermott in coverage of AI agent controls

The future-of-work stories were less about replacing people and more about how much autonomy employees will accept. iTWire reported TeamViewer research showing 80% of Australian employees use AI daily, while 64% want human oversight. The same research said IT leaders expect about 45% of digital workplace services in Australia to run autonomously by 2030. That is a major operating shift, but the trust condition is visible: people want notifications, logs, rollback and a human route back in.

Vendors are packaging that shift into workflows. iTWire covered TeamViewer and ServiceNow's partnership for autonomous IT operations. CNBC reported ServiceNow's Bill McDermott touting a kill switch and AI Control Tower for rogue agents. Business Insider said OpenAI is pushing ChatGPT into law firms. The direction is clear enough. AI is moving from browser tabs into roles, queues and regulated work.

The buyer test is organisational, not just technical. BBC News asked whether AI will help people do their jobs or replace them. Bloomberg reported Infosys cutting the top end of its revenue forecast as IT service demand stayed muted. Search Enterprise AI covered Progress buying Domo for $400 million, with AI governance and data readiness part of the rationale. The hard truth is that agentic systems need clean data, decision rights and redesign. Without those, automation just makes weak process run faster.

AI dealmaking rewarded the new middle layer

The week's deal flow said investors still want AI, but they are getting more specific about where value sits. ITPro reported Google Cloud revenue up 82% to $24.8 billion, even as investors questioned AI spending and model timelines. Silicon Republic said Stripe is in talks to buy OpenRouter in a deal that could be worth about $10 billion. Model access, routing, billing and cost control are becoming part of the same commercial layer.

Capital also chased early technical leverage. Tech Funding News reported Khosla Ventures targeting up to $5.5 billion for new funds, much of it aimed at early-stage AI startups. Tech Funding News said Samsung is in talks to put about €1 billion into Mistral at a potential €20 billion valuation. Arrakis raised $38 million for fast-deploying AI agents in sectors such as energy, aerospace and logistics. Those are different bets, but they share a belief that the application layer is still under-built.

The non-AI funding picture was cooler. UKTN reported UK fintech funding falling to £1.1 billion in the first half of 2026. That contrast is instructive. AI is not lifting every technology category equally. It is concentrating money around compute, model access, automation and data readiness. Buyers should expect some vendors to look flush and others to look cautious, even when both claim the same AI tailwind.

What the influencers are discussing

The sharpest influencer discussions this week were about evidence, not enthusiasm. Sarah Evans framed the purchase path as 26 brand touches before a buyer converts, with sales credited for the last four and PR shaping many of the earlier ones. Her point fits the publisher fight and the AI-search debate: if buyers ask machines first, reputation has to be visible before the attributable click ever exists.

Daniel Rijo brought the harder B2B buyer signal, citing Semrush research that brand recognition sways just 7% of B2B AI buyers, while precise use-case fit matters more. His AppsFlyer post on 58.6% of marketers underinvesting in channels AI cannot measure sharpened the attribution problem. Marketers are being asked to optimise for answer engines, creator signals and dark consideration paths, while the dashboards still reward the last visible interaction.

The work-design thread was just as useful. Lenny Rachitsky's interview on building a Claude content machine stood out because it was not a lazy automation claim: the workflow interviews the founder, codes voice in Markdown and runs multiple revision passes. Gergely Orosz's Meta story about employees calling out AI-written internal posts made the opposite point. Workers can smell low-effort synthetic copy, and they will punish leaders who use AI to add noise instead of clarity.

Gregor Ojstersek mapped AI-native engineering teams at OpenAI, Anthropic, Shutterstock and smaller startups, showing how team shape changes when engineers, PMs and managers are redistributed around faster execution. Richard van der Blom's DemandBird post put the same operating problem into social distribution: multi-platform publishing is easy, but strategy, adaptation and analytics are the hard part. Those posts were useful because they treated AI as a management system. The question is no longer whether a task can be automated. It is whether the surrounding team knows who owns the brief, the review, the data and the performance signal after the machine acts.

Across the week, the creator consensus was not anti-AI. It was anti-careless. Sarah Evans and Daniel Rijo were talking about machine-mediated discovery. Lenny Rachitsky and Gergely Orosz were talking about the craft and social cost of AI-generated work. Gregor Ojstersek and Richard van der Blom were talking about the operating model around it. Put together, the message for B2B teams is sharper than the usual productivity pitch: AI makes weak strategy more visible, not less.

The unresolved thread is control. AI now wants access to power grids, repositories, legal workflows, publisher archives, customer conversations and buying decisions. Regulators are moving faster, investors are still paying for scale, and workers are asking for guardrails before autonomy becomes normal. The vendors that look strongest next week will not be the ones shouting about smarter agents. They will be the ones proving where the agent stops, who can intervene and what evidence the buyer gets when something goes wrong.

Meet Agentcy

The Agentcy Difference