This Week in B2B Tech: 6-10 July 2026

$1.36 trillion from SK Group, a 54pc efficiency claim from OpenAI and a £18 million Ofcom fine threat set the tone for B2B tech this week. The market is no longer debating whether AI will touch work. It is asking who funds the compute, which tools get trusted inside the company and how much control regulators will demand once those systems meet customers, capital markets and public infrastructure. For buyers, the question shifted from pilots to dependability.

In parallel, influencer discussions stayed closer to the buying moment. Greg Kihlström pointed to research that only 8.4% of brands show up consistently across ChatGPT, Claude and Gemini, while Sarah Evans argued that visibility has to become believability before it affects a sale. Ruben Dominguez pushed the founder-as-media-engine argument. The useful thread was not more AI enthusiasm. It was whether teams can prove presence, trust and operating leverage before the budget gets harder.

AI infrastructure became the week’s biggest balance-sheet story

SK Telecom signage used in coverage of AI chip and data-centre investment

AI demand showed up in the part of the market buyers rarely see: chips, memory, power and financing. Total Telecom reported SK Group’s $1.36 trillion plan for AI chips and data centres, with SK Hynix earmarking $706 billion for memory and SK Telecom aiming for 15GW of AI data-centre capacity by 2035. That is not a product roadmap. It is an industrial policy bet dressed as corporate capex.

Washington made the same pressure visible from the other side. Bloomberg reported Howard Lutnick pressing Samsung and SK Hynix to boost US memory output, while Samsung’s profit beat was tied to runaway AI memory demand. Quartz covered Apple extending its custom-chip partnership with Broadcom through 2031, a signal that buyers with enough scale want more control over the silicon underneath AI features.

The financing bill is arriving fast. TechRepublic reported Amazon’s $25 billion bond sale to fund AI infrastructure, and Data Centre Magazine linked GM and Micron through the same data-centre squeeze. Enterprise buyers should not treat compute as a hidden supplier problem. Capacity, memory and regional power are becoming part of the price, latency and resilience of every serious AI deployment.

OpenAI pushed agents from chat box to work surface

ChatGPT app artwork used in coverage of GPT-5.6 and work agents

OpenAI’s week was about turning ChatGPT into a work surface rather than another window beside work. SmartCompany reported ChatGPT working across Slack, Gmail and Google Drive, and Bloomberg described ChatGPT Work agents fielding tasks for hours. The claim is clear: the interface is moving from asking to delegating.

The model story backed that up. The Decoder paired GPT-5.6’s rollout with the new workflow agent, while Silicon Republic reported Sam Altman’s 54pc token-efficiency claim. Quartz said GPT-5.6 is becoming the preferred engine for Microsoft 365 Copilot, despite recurring talk about strain in the OpenAI-Microsoft relationship. The important detail is that efficiency and workflow control are now selling points together.

Silicon Angle framed ChatGPT Work as an agentic tool for business workflows. That framing is right, but it also raises the bar. If agents can read mail, pull files and act in work systems, enterprise trust cannot stop at model quality. Buyers need permission design, audit trails and clear failure modes before a productivity promise becomes a procurement win.

AI adoption is turning labour into an implementation argument

Office workers and AI graphics used in coverage of technology layoffs

The labour story around AI stayed messy because the evidence points in two directions at once. Computerworld reported Microsoft betting that enterprise AI needs engineers, not bigger sales teams, while TechCrunch tracked Microsoft layoffs across Xbox and commercial sales. That looks contradictory only if AI is treated as a single force. In practice, it is cutting some work and creating demand for people who can make deployments real.

TechRepublic cited data from 21,000 firms suggesting AI spending is creating jobs, not killing them. Yet City AM reported London workers are especially exposed to AI-driven job pressure. The difference is skills, not sentiment. Companies need engineers, data operators and process owners, while roles built around repeatable hand-offs are being marked for redesign.

The CFO angle now matters. CFO Dive described AI adoption as a high-stakes finance challenge, and Inc. argued the productivity argument is over. The more useful question for buying committees is whether the vendor can show the operating model behind the tool: who owns the process, what changes in headcount planning and how the benefit survives contact with real systems.

Agent security moved from awkward demo risk to attack surface

Cybersecurity dashboard imagery used in coverage of AI agent manipulation

The week’s security coverage made one point hard to ignore: agents now have enough autonomy to be useful and enough access to be dangerous. Ars Technica reported researchers showing popular AI tools could be used to assemble botnets, and The Hacker News covered HalluSquatting attacks against AI coding assistants. Hallucinated dependencies used to be a quality flaw. In an agentic workflow, they can become a delivery channel.

Development tooling supplied the sharpest examples. Silicon Angle reported GitLost leaking private repositories through GitHub AI workflows, while ITPro covered Friendly Fire attacks that trick defensive agents into running malicious code. CSO reported agents falling for indirect prompt-injection traps. That is not a prompt hygiene issue. It is identity, permissioning and endpoint control with an AI label.

SecurityWeek’s HalluSquatting coverage put the botnet risk in plain terms. Buyers should read these stories as a warning against agent pilots that sit outside normal security architecture. Once a tool can code, browse, trigger workflows or touch repositories, its instructions are part of the control plane.

AI policy became a continuity risk, not a side debate

Data-centre servers used in coverage of Chinese AI models and US cybersecurity concerns

Geopolitics kept intruding on AI procurement. The Decoder reported China weighing export curbs on its top AI models, leaving Europe caught between access, competitiveness and sovereignty. Fortune covered Brad Smith warning about Washington’s incomplete AI policy rules. Buyers should not treat that as policy noise. Model access, compute location and data routing are now continuity issues.

Financial regulators are moving faster too. Sky News reported the Bank of England warning that AI could threaten financial stability, while Computer Weekly covered a landmark UK regulatory review of AI in financial services. City AM said the FCA is eyeing tougher rules as consumers turn to chatbots for financial advice. AI has left the innovation lab and entered regulated decision-making.

Diginomica asked whether sovereign agentic AI can be secured through a Kore.AI and Atos partnership. The answer is not a slogan. Sovereignty has to mean identifiable controls: where the model runs, who can inspect it, what happens if a state restricts access and which human has authority when the system needs stopping.

Publishers are forcing the AI web to declare its terms

Cloudflare office signage used in coverage of AI crawler controls

The relationship between publishers and AI companies looked less like a technical argument and more like a pricing negotiation. Procurement Magazine covered Cloudflare’s new AI partnerships and crawler controls, while Adweek reported publishers considering the once-unthinkable step of opting out of Google Search. That is a serious change in leverage, even if only a few publishers follow through.

The crawlers themselves are becoming harder to categorise. Digiday broke down declared good bots, mixed-use crawlers and gray scrapers, and MediaPost argued AI is fast-tracking the open web’s collapse. Euronews reported news outlets seeking sanctions against OpenAI in their copyright fight. The commercial fight is now about access, proof and payment.

B2B visibility sits inside that fight. Marketing Dive argued unpaid media is now essential to AI visibility, and Search Engine Land separated being used by AI from being cited by AI. If AI answers become the first shortlist, brand credibility will depend less on owned claims and more on external sources that machines can verify and humans can trust.

Platform regulators are attacking defaults, not just misconduct

A person using a phone in coverage of Ofcom scam-ad proposals

Regulators spent the week looking at the small design choices that shape markets before users notice. The Independent reported Ofcom’s proposed scam-ad code with fines of up to £18 million or 10% of global revenue, while BBC News also covered the plan to make big tech deal with scam ads. Paid fraud is no longer being treated as an unfortunate edge case of platform scale.

Europe kept pressure on the operating systems and app stores underneath digital markets. Silicon UK reported the EU’s top court dismissing Google’s appeal over a €4.1 billion Android fine, and Bloomberg reported Apple losing a dispute over EU App Store and iPhone rules. CIO Dive covered SAP loosening ERP maintenance and support rules after EU competition pressure. The target is default power: the rules that decide what customers can choose before procurement begins.

Safety regulation is travelling the same route. Bloomberg reported the EU escalating its Meta probe over addictive design features for children. The message for B2B vendors is plain. Regulators are no longer waiting for harm to become spectacular. They are asking how the system is designed, who benefits from friction and whether customers can exit without being punished.

What the influencers are discussing

The influencer discussions with the most bite this week were about proof. Greg Kihlström pointed to the 8.4% figure for brands showing up consistently across ChatGPT, Claude and Gemini. That number is useful because it cuts through the easy optimism around answer engines. If buyers are researching inside AI systems and only a small minority of brands appear reliably, visibility is becoming a supply problem, not a messaging tweak. It also gives comms teams a harder benchmark than ranking screenshots.

Sarah Evans made the trust point sharper: getting into the AI answer is not the same as getting bought. Her framing matters for comms teams because it joins machine visibility to human credibility. A cited brand still needs third-party proof, recognisable expertise and enough corroboration for the buyer to believe the answer. Rand Fishkin asked the harder measurement question, whether AI answer presence creates demand, traffic or sales. That is the right test.

Creators on the operating side were less patient with old content models. Ruben Dominguez argued that one founder can now outpublish an entire content team, which is provocative but directionally right for lean B2B teams that can encode voice, memory and distribution into repeatable systems. HubSpot Marketing’s AEO tools video, meanwhile, showed how fast the market is productising the same anxiety: brands want a way to be cited, measured and fixed.

The publisher side supplied the counterweight. Adweek highlighted publishers considering blocking Google’s crawler, while Daniel Rijo framed the open web’s crawler math as turning against Google. Those takes matter because they show AI visibility and publisher control are not separate conversations. If content owners restrict access and AI systems become choosier about sources, reputation work gets harder and more valuable.

The best reading of the week is practical: answer visibility is real, but it is not magic. Teams need source quality, analytics discipline and a point of view strong enough to survive being compressed by a machine. Anything less is just a prettier search ranking report. The creators who sounded most useful were the ones refusing to separate distribution from evidence.

The unresolved thread is permission. AI needs more chips, more power, more data, more workflow access and more publisher content, but every one of those inputs now has a gatekeeper asking for better terms. The next phase of B2B tech will favour vendors that can prove control without killing usefulness. That is a harder test than shipping another agent button.

Meet Agentcy

The Agentcy Difference