This Week in B2B Tech: 22-26 June 2026

$50 billion for MGX, a £1 billion Google app-store trial and 26,000 compromised AI agents told the story of the week. B2B tech buyers weren't short of product news, but the real argument sat elsewhere: who pays for frontier AI, who controls the platforms around it and who carries the risk when autonomous systems start acting on bad instructions. The market is still excited by Friday afternoon. It is also starting to ask for receipts.

In parallel, influencer discussions were circling the same pressure point from the buyer side. Allie K. Miller framed it bluntly: buyers now ask ChatGPT, Claude, Gemini and Perplexity before a brand gets a human visit. David C. Edelman pushed the B2B angle, warning that companies need to be a specific answer to a buyer problem, not a louder version of their feature list. That was the week's useful tension: more automation, but less room for weak proof.

Frontier AI hit the procurement desk

Illustration used in coverage of enterprise AI spending pressure

Enterprise AI spend finally met the awkward part of the finance meeting. Quartz reported pullbacks from OpenAI and Anthropic as customers chased cheaper routing and clearer ROI, while The Decoder said OpenAI may wait until 2027 for an IPO if Sam Altman wants a valuation above $1 trillion. Those two stories belong together. They show the same buyer instinct from different ends of the market: no one wants to fund a premium model bill without knowing what durable advantage it buys.

Competition also stopped looking like a clean US platform race. Silicon Republic tracked DeepSeek's planned hiring surge as the company reportedly eyes a $7.4 billion raise, while TechCrunch pointed to a 75% jump in Claude consumer payments since January. Even that growth sat beside harder politics: Forbes described Anthropic's distillation fight with Alibaba as a new front in the US-China AI dispute. The judgment for buyers is simple enough. Frontier AI is becoming a portfolio decision, with cost, jurisdiction and model provenance all sitting in the same spreadsheet.

Google Play Store imagery used in coverage of the UK app developer class action

Platform dependency had a rough week. UKTN reported that Google will face an 11-week UK trial over more than £1 billion in alleged Play Store overcharging, and Techzine covered Google's new payment options as developers get more room to steer customers outside its billing rails. The legal argument is old. The buyer implication feels fresher: software economics can change when a platform owner changes a fee model, a default or a payment path.

Microsoft faced the same scrutiny from different directions. IT Brief UK covered the CMA's probe into browser choice across Microsoft business software, while Reuters reported Italy's investigation into Microsoft 365 pricing tied to Copilot and Designer. The cloud layer is no quieter. Retail Gazette said the European Commission is preparing tougher gatekeeper treatment for AWS and Azure under the Digital Markets Act. Buying committees should read this less as regulatory theatre and more as a warning about switching cost. The more AI is bundled into office, cloud and app distribution, the harder procurement has to test the exit door before signing.

AI governance moved from principles to liability

Cybersecurity risk image used in coverage of proposed AI incident reporting law

AI oversight is turning into an incident process. CSO reported on a proposed US AI Incident Reporting Act that would require covered model developers to report major safety and security events within seven days, and within 48 hours where serious harm is imminent. Penalties could reach $2 million. That is a different mood from voluntary principle papers. It says advanced model operators may soon need board-level escalation paths that look more like cyber breach response.

Courts and public bodies filled in the rest of the picture. Security Boulevard wrote about a German ruling on Google AI summaries that treated generated output as part of Google's own business activity, and TechTarget covered weak oversight of VA generative AI tools used by thousands of health staff. Copyright pressure kept building too, after MediaPost reported that The New York Times amended its OpenAI and Microsoft complaint. Even moderation showed the labour trade-off, with Social Media Today reporting Meta's plan for AI to handle about 90% of review work by the end of 2026. The week made a hard point: AI governance is no longer a policy PDF, it is operational liability.

Agent security had its proof-of-concept week

Reddit and research imagery used in coverage of poisoned deep-research agents

The most useful security stories this week were small enough to be unsettling. Search Engine Land covered Cornell Tech research showing that a 13-word edit on a public page can steer deep-research agents, with fake entities appearing in 38% to 51% of reports and up to 62% when multiple pages are targeted. Security Boulevard reported a malicious AI skill reaching more than 26,000 agents after passing security checks. These are not sci-fi failures. They are supply-chain failures wearing an AI badge.

The browser and endpoint stories widened the gap between demo and deployment. Infosecurity Magazine described BioShocking attacks that pushed agentic browsers and plugins toward credential leakage across six products, while Bleeping Computer detailed macOS malware that embeds fake errors to mislead AI analysis tools. Funding followed the anxiety: Silicon Angle reported Patronus AI's $50 million raise to stress-test agents in simulated environments. Security teams don't need another generic AI risk category. They need to know which tools can be manipulated by retrieved text, external instructions and plausible-looking context.

The AI build-out kept asking for more power

Bloomberg video image used in coverage of MGX raising one of the biggest AI funds

AI infrastructure money kept arriving at a scale that makes normal venture cycles look quaint. Bloomberg reported that MGX has raised close to $50 billion for AI infrastructure and technology, and AI Business said Amazon will invest another $13 billion in India for AI infrastructure and cloud services. The build-out is no longer just a hyperscaler story. It is an industrial policy, supply-chain and energy-capacity story.

The constraint is power, memory and interconnect, not ambition. IT Brief UK covered Capgemini research finding that 77% of electricity executives expect AI data-centre demand to outrun grid expansion, while Forbes said Qualcomm is targeting $15 billion in data-centre chip sales by 2029. Qualcomm also moved on the software layer, as AI Wire covered its Modular acquisition for a more open AI stack. The geopolitical layer stayed close by, with Reuters reporting EU participation in the US-led Pax Silica chip-supply initiative. Buyers should expect AI capacity to carry geography, grid and supplier-risk assumptions, not just a price per token.

AI search turned visibility into a trust problem

Search did not just become more automated this week. It became easier to game and harder to audit. Search Engine Land warned about paid brand mentions in GEO as low-quality placements try to influence AI answers, while Search Engine Roundtable tracked Google's June spam update and wider AI reporting access. The race to appear in machine answers is already attracting the same bad incentives that bent parts of SEO.

Measurement is improving, but the rules are still moving. Search Engine Roundtable covered Google's documentation on generative AI controls for AI Mode, AI Overviews and Discover, and its reporting on AI performance impressions showed how narrow the counting rules can be. Media and trust are being pulled into the same argument. Adweek reported Jellyfish data showing YouTube creator content in more than 25% of AI answers, while Influence tied declining news trust to earned media's value. The smart read is not that every brand needs more content. It is that credible, structured, cited information is becoming part of the buying interface.

What the influencers are discussing

The Media Copilot image used on its AI discovery post

The influencer thread was less interested in model launches than in who gets chosen by AI before a buyer opens a vendor site. Allie K. Miller made the sharpest commercial case, arguing that ChatGPT, Claude, Gemini and Perplexity now sit in front of many purchase decisions and citing ChatGPT referral traffic converting at 7.1%. Her point was not that websites are dead. It was that websites built only for human browsing are a weak source of truth for machines that summarise, compare and recommend.

David C. Edelman brought the B2B strategy view, saying AI search is already moving faster in B2B than B2C and pushing suppliers to become a specific solution to a buyer's need. That was a useful correction to the cheap version of answer optimisation. If a company merely stuffs more mentions into the web, it may win a citation and lose the sale, because buyers still need evidence that maps to a real problem.

Greg Kihlström widened the issue into a workflow problem: teams can now create far more than they can review, and click data is losing some of the explanatory power it used to have. HubSpot Marketing put the same idea into buyer language: when buyers ask AI for recommendations, a brand either shows up or it doesn't. The answer is not more dashboards. It is better evidence, better governance and a tighter link between what a brand publishes and what a buyer is trying to decide.

Media Copilot framed AI search as a new incentive system for media, where high-quality, bot-readable journalism can matter more than raw traffic. Daniel Rijo gave it a budget signal, reporting that AI search is pushing 44.8% of brands to spend more on PR. D S Simon's AI visibility episode carried the practitioner version, with the warning that some companies appear in less than 5% of AI-generated answers. Together, those posts describe the new middle ground for B2B marketers: AI can scale production and change discovery, but the market is starting to reward proof, taste and trust more than volume.

The unresolved thread is whether buyers will get better signals or just faster noise. Regulators are turning AI incidents into reportable events, platforms are being forced to justify their defaults, and security researchers keep showing how brittle agentic systems can be. At the same time, capital is still pouring into compute and vendors are still promising cheaper, stronger models. Next week, the interesting question is not whether AI adoption slows. It is whether the people paying for it get more leverage, or merely inherit a longer list of obligations.

Meet Agentcy

The Agentcy Difference