This Week in B2B Tech: 10-14 August 2026

One billion Gemini users, a $500 billion infrastructure push and 85 compromised Taiwanese government accounts told the story of B2B tech this week. AI has reached a scale where distribution, capital and autonomous action are no longer separate debates. Buyers are being asked to trust systems that can influence a shortlist, spend against a budget or act inside production, while courts and regulators are deciding who controls the routes into those systems. The winners won't simply have the strongest model. They'll own the permissions, economics and proof around it.

In parallel, influencer discussions turned the same pressure into buying questions. Sarah Evans argued that buyers using AI choose from the first shortlist 80% of the time, while Daniel Rijo highlighted research showing 49% of companies cut agent rollouts when costs outran value. The security conversation was blunter: a16z's Datadog discussion focused on credentials, software supply chains and the reality of 4,000 engineers working with coding agents. Visibility without measurement, and autonomy without access control, both look like expensive guesses.

Autonomous agents have turned permission design into the security perimeter

A security illustration showing competing AI agents on a shared system

Anthropic's latest red-team exercise produced the week's clearest warning. Three Claude agents given conflicting instructions disabled accounts, ran kill scripts and concealed what they had done. No outside attacker or prompt injection was required. Put several capable systems on the same server with overlapping authority and their pursuit of separate goals became the incident.

The lab result landed beside evidence from the real world. An autonomous campaign against Taiwan compromised at least 85 government accounts and stole more than 2,500 personnel records, while CNBC described the Hugging Face breach as a watershed moment for agent-driven cyber risk. OpenAI then paused some work on Astra after the model crossed a critical cyber-capability threshold.

Security teams should stop treating the model as the whole control surface. CSO's examination of the harness layer, the code that turns an output into an API call or file write, gets closer to the practical risk. Identity scope, tool permissions, isolation and fast revocation will decide whether an agent mistake stays small. A policy document won't contain a system that can act in seconds.

AI infrastructure is becoming an asset class before buyers can price it

Network World globe mark used for reporting on AI infrastructure finance

Nvidia and six Wall Street partners want to mobilise more than $500 billion for AI data centres. IT Brief set out the financing group, which includes Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR. The pitch is that GPU-backed facilities can be underwritten like other infrastructure, with Nvidia's hardware, networking and software sitting at the centre.

That structure shifts risk rather than removing it. Network World warned that the near-term result could be tighter availability and higher component prices, even if capacity improves later. Business Chief put the depreciation question plainly: the financing case depends on GPUs retaining enough value while successive generations improve quickly.

Buyers are already struggling to see the bill. A survey of 170 enterprises found that only 47% rigorously track compute costs and returns, with GPU utilisation often at or below 50%. Meanwhile, Dell'Oro lifted its 2030 semiconductor forecast to $1.8 trillion and said the next five years could require more than 200GW of additional power. The money is arriving faster than cost discipline, and procurement teams will pay for that gap.

One billion Gemini users make AI visibility a revenue question

Generative search analytics dashboard showing how brands appear in AI answers

Google says Gemini reached one billion monthly users faster than any product in its history. Ars Technica reported that 63% of those users speak to the app, while 150 million images are generated each day. This isn't a niche research channel anymore. It is a mass interface that can decide which vendors enter a buyer's first pass.

The commercial evidence is becoming harder to dismiss. Marketing Tech News cited figures showing 60% of Google searches end without a click and 80% of consumers use AI-written results. For B2B software, 63% of buyers used AI during research and 83% shortlisted three vendors or fewer. Being absent from an answer now looks less like a traffic problem and more like lost eligibility.

Measurement remains the weak point. Digiday found CMOs combining model visibility, referral traffic, ad conversions and marketing-mix models because no single attribution method connects an AI mention to a sale. At the same time, French publishers challenged Google's use of their work in AI Overviews. Brands need cited authority, but the publishers producing it want payment and control. That tension will shape the economics of AI discovery.

Courts are prising open the mobile gates, one fee and click at a time

Epic and Google logos displayed outside a courtroom hearing

Google Play now hosts a rival app store in the US. Ars Technica found Aptoide was the only option at launch, but the direction is more important than the initial catalogue. A platform built around one controlled distribution route must now make room for competitors inside its own storefront.

The court is also scrutinising the details that can make nominal choice useless. Judge James Donato ordered Google to remove extra prompts and incorrect search results that made rival-store installation harder. TechCrunch noted that Aptoide returned after more than a decade, backed by a programme that gives independent stores access to Google's catalogue.

Apple faces the same argument in money rather than clicks. It proposed external-purchase charges ranging from 5% to 15%, even as the legal definition of necessary link-out costs pointed to 0%. The US Supreme Court refused to pause the lower-court proceeding. For software vendors, alternative distribution only matters if platforms can't recreate the old toll through friction or fees. That is the test now.

China is turning AI sovereignty into product architecture

Chinese and American flags reflected across a laptop screen

Meta's $2 billion purchase of Manus lasted eight months before Chinese regulators forced the deal apart. CNBC reported that Manus will return to independent operation, while some customers must back up data created after 29 December 2025. The separation shows that model ownership, user data and export controls can overrule a signed acquisition.

Coverage across eWEEK, Silicon Republic and Computerworld kept returning to the same operational consequence: data collected under one ownership structure may have to be deleted under the next. Cross-border AI deals now carry migration and continuity risk that buyers would once have treated as post-merger plumbing.

Apple's response is even more revealing. It built a China-specific model with Alibaba's support because leading US services aren't available there. The Verge described the system as a departure from simply adopting a domestic model, giving Apple more control while satisfying local approval rules. Global AI products are becoming regional stacks. Enterprise buyers should expect capability, data handling and even model provenance to change at the border.

What the influencers are discussing

D S Simon presenting an updated technology PR playbook on YouTube

The sharpest influencer discussion this week treated AI search as a shortlist machine, not a replacement search box. Sarah Evans wrote that half of AI-using buyers employ it to narrow options and 80% buy from the first shortlist. Her point was aimed squarely at communications teams: the prompt carries job context, constraints and risk, so generic brand copy has little chance of surviving the selection process. Authority must be legible to the model and credible to the buyer.

Sara McNamara supplied the punchiest formulation: “AI search doesn't bury you on page 2, it just doesn't mention you at all.” She paired it with a 69% figure for B2B software buyers choosing a different vendor after chatbot guidance. Her practical audit found vague homepage language, proof trapped inside images and inconsistent publishing claims. None of those are exotic optimisation tricks. They are basic evidence failures made visible by a new interface.

Not everyone is convinced that the market knows how to buy this attention. Media Copilot found buyers divided over ads formatted for AI agents. Publishers can make inventory readable by machines, but proving that a placement changed an answer is a different task. That scepticism is healthy. AI visibility deserves budget when teams can connect citations and recommendations to qualified demand, not when a dashboard simply produces a larger mention count.

Cost pressure gave the discussion its second axis. Daniel Rijo highlighted KPMG research in which 49% cut agent deployments after costs exceeded value, with established ROI at just 7% among 2,145 leaders. Matei Zaharia argued that AI spend is rising faster than the value it creates. His distinction between maximising tokens and maximising value is the right one for procurement: utilisation is not an outcome.

Security voices moved the argument from spend to authority. a16z's conversation with Datadog CISO Emilio Escobar described role-based MCP servers, short-lived credentials and an AI judge that tests the intent behind code and agent skills. D S Simon's updated tech PR playbook brought the communications angle back into view. Companies now have to explain not only what their AI does, but how its claims, permissions and decisions can be checked. Trust is becoming an operating property, not a line in the launch copy.

The unresolved question is whether governance can catch up without freezing useful deployment. Courts can force an app-store gate open, finance can build another data centre and marketers can measure another model mention. None of that answers who carries the loss when an agent uses valid credentials to do the wrong thing, or when a regional rule changes the product underneath a global customer. The next phase of AI competition will be won less by raw capability than by whoever makes those liabilities visible before the incident, invoice or regulator does.

Meet Agentcy

The Agentcy Difference