This Week in B2B Tech: 13-17 July 2026

A Google Android order, a $1.5 billion Fireworks round and a nine-second database deletion set the week in B2B tech. Buyers were not short of AI news, but the sharper story was control: who gets into the operating system, who pays for the compute, who owns the training data and who is liable when an agent acts faster than a human can stop it. The week made AI feel less like a feature race and more like a permission fight.

In parallel, influencer discussions moved closer to the revenue line. Sarah Evans argued that PR, SEO and sales now share one AI-mediated funnel, while David C. Edelman warned that the next customer may be filtered by their AI before they ever reach a website. Daniel Rijo pointed to higher-converting AI search leads. The thread running through those posts was simple enough: visibility now has to be machine-readable, measurable and credible, or it will not show up at the buying moment.

Regulators are prising open the platforms AI needs

Android phones and Google branding used in coverage of European AI assistant access rules

Europe's platform fight moved from app stores into the operating system. The Hacker News reported the European Commission ordering Google to open Android camera, microphone and screen access to rival AI assistants, with changes aimed at Android 18 and a broader deadline of 1 August 2027. The Verge framed the same order as a direct intervention in Android and Search data. For any vendor building an assistant, the message is useful and uncomfortable: access is opening, but only on terms regulators can inspect.

Search is becoming regulated content as well as regulated distribution. The Decoder reported Germany putting Google AI Overviews and Perplexity under media law, a first-of-its-kind move that treats generated answers as something more than a neutral index. MediaPost covered the Commission's preliminary finding against Meta under the Digital Services Act, focused on addictive design features rather than a single piece of harmful content.

The UK added a more operational angle. Finextra reported Ofcom's proposed fines for scam adverts, and Diginomica tracked new direct oversight for AWS, Google, Microsoft and Oracle as critical third parties. The common thread is that default settings, uptime, data access and fraud controls are now public-interest questions. B2B buyers should expect more contracts to carry regulatory assumptions that used to live outside procurement.

AI capital is chasing compute, but the bill is showing

Fireworks AI team photo used in coverage of its enterprise AI model hosting round

The capital market still wants AI exposure, but this week's strongest deals were about capacity rather than slogans. Tech Funding News reported Fireworks AI raising $1.5 billion at a $17.5 billion valuation, after annualised revenue passed $1 billion and token volume grew from 15 trillion to more than 40 trillion a day. Quartz said Anthropic was lining up investor meetings for a possible autumn IPO. The market is rewarding the companies that can turn model demand into infrastructure leverage.

China's side of the compute story was just as direct. Silicon UK reported DeepSeek seeking fresh funding after a reported $7 billion round, while The Information said memory chipmaker CXMT was seeking an $8.6 billion Shanghai IPO. TechCrunch covered Reflection AI's $1 billion compute deal with Nebius. None of that reads like optional experimentation. It reads like the cost of entry.

The strain is starting to show in places investors prefer to ignore. The Register reported New York pausing large data-centre buildouts, Finextra covered a BIS warning that the AI investment boom could turn to bust, and the Financial Times linked IBM's valuation hit to timing and reprioritised capex. Buyers should watch this closely. If compute becomes scarce or expensive, vendors will pass that reality into pricing, limits and service quality.

Agent security turned into a live operating risk

Robot illustration used in coverage of Gemini CLI being abused as a hacking agent

The security beat stopped treating AI agents as awkward assistants and started treating them as tools attackers can operate. Bleeping Computer reported Google Gemini CLI being used as a hacking agent and malware botnet operator, over more than 200 sessions. DevOps.com covered HalluSquatting attacks that compromise coding agents through hallucinated packages. The weakness is not only model behaviour. It is the combination of model behaviour, package trust and workflow permissions.

Anthropic and OpenAI both appeared in the same risk frame. Dark Reading reported a Claude flaw that could automatically send malicious prompts to AI agents, while AI Business covered OpenAI's GPT-Red model for automated red teaming. A defensive model that creates prompt-injection attacks is useful, but it also proves the market now needs machine-speed testing for machine-speed failures.

Damage moved from theoretical to physical systems and repositories. The Independent reported OpenAI warning that Codex could delete files when granted broad computer access, SDxCentral said xAI's coding agent exposed entire repositories, and Infosecurity Magazine reported a controlled test where ChatGPT executed a full cyber-attack chain. The procurement test is changing. Buyers need to ask not just what the agent can do, but what it cannot do even if instructed.

AI at work is becoming a design problem, not a headcount slogan

Google office signage used in coverage of worker demands for layoff protections during the AI boom

Workforce AI stories looked contradictory because the labour market is contradictory. Fortune quoted Esther Perel warning executives that AI can deepen workforce social atrophy, while The Guardian reported thousands of Google workers asking Sundar Pichai for layoff protections during the AI boom. Forbes asked whether mass AI-driven layoffs could become a boomerang. Cutting people and demanding more human judgment from the people left behind is not a stable strategy.

The enterprise implementation argument is getting more serious. The AI Journal argued that companies need decision systems, not just language models. That is the right distinction. A chatbot can answer a prompt; a decision system has to carry context, accountability, monitoring and a way to recover when it drifts.

Smaller companies were offered the optimistic version of the story. Forbes listed ways small businesses can use AI to grow and even hire. Microsoft supplied the vendor tension, with Quartz reporting Satya Nadella criticising Anthropic's model restrictions in front of engineers. The best read is that adoption is no longer blocked by awareness. It is blocked by operating design: rights, roles, quality checks and the social fabric around the tool.

The AI content fight is turning into discovery risk

Illustration of AI music used in coverage of Suno training data allegations

The rights fight around AI widened again. The Financial Times reported Apple sending legal letters to former OpenAI employees, after its trade-secret lawsuit over alleged hardware plans. TechCrunch carried OpenAI's pushback against the Apple case. That is not a routine talent dispute if it slows a company trying to move from model access into devices.

Training data claims kept coming. The Verge reported hacked Suno data suggesting songs were scraped from YouTube, Genius and Deezer, and BetaNews covered publishers suing Google over books allegedly used to train Gemini. The pattern is familiar: AI companies want broad ingestion, rights holders want proof and payment, and courts are being asked to decide what the product market already raced past.

The consumer-data side is just as relevant for B2B visibility. ZDNet reported Google expanding AI training on user data unless people opt out, while Startup Daily covered Meta backing down after privacy concerns around its Muse image generator. If AI discovery depends on content access, brands will have to decide what they want machines to read, what they want to withhold and which third-party sources carry enough authority to survive the filtering.

AI governance split between blocs and boardrooms

International flags and AI imagery used in coverage of a China-backed world AI body

Global AI governance is becoming less global. Computerworld reported China, Russia and 27 other countries creating a World AI body without the US. Euronews covered Xi Jinping's call for global AI cooperation as US restrictions squeeze China's technology access. The politics are clear enough: standards are now a route to influence, not a neutral technical exercise.

Supply chains are being pulled into the same argument. TechCrunch reported Apple Intelligence winning approval in China through Alibaba's Qwen and Baidu, while Bloomberg covered Washington's concern that Chinese firms may train models from US systems. The result for multinational buyers is awkward. AI capability, compliance and market access can point in different directions.

Risk governance is also moving from speeches into proposed mechanisms. Business Chief covered Demis Hassabis calling for a US-led frontier AI testing network, The Guardian reported the Bank of England governor calling for global cooperation on AI threats, and Forbes wrote about China planning recall rules for AI agents. Boards should read that as a timing signal. Voluntary policies are being overtaken by jurisdiction-specific controls.

Agents are getting payments, browsers and legal wrappers

Linux Foundation and Google imagery used in coverage of native payments for AI workflows

The agent economy is starting to acquire the boring infrastructure that makes it real. InfoWorld reported a new Linux Foundation project to make payments native to AI workflows. The Decoder said Google is rebranding NotebookLM as Gemini Notebook and giving notebooks a cloud computer for code execution. Those are not small interface tweaks. They make agents more able to transact, run code and act across apps.

Visibility into those actions is already contested. The Decoder reported OpenAI's Codex encrypting instructions between AI agents, leaving some developers unable to read delegated session history. AI News covered Cloudflare making AI agent crawlers seek permission by default on ad-supported pages. One side of the market wants agents to operate smoothly. The other wants to know who is knocking and why.

Law and incident response are catching up. Fortune reported Delaware testing a legal entity for AI agents, The New Stack collected destructive agent incidents under the warning that instructions are only suggestions, ITPro described a rogue agent deleting a production database and backups in about nine seconds, and The AI Journal showed how sales agents can pressure vulnerable customers. The lesson is blunt: autonomy without gates is not a strategy. It is outsourced risk.

What the influencers are discussing

The most useful influencer discussions this week came from people treating AI as a buying system, not a content trick. Sarah Evans called it the one-funnel era: the agent discovers, verifies and buys in one motion, while PR, SEO and sales teams still sit in separate reporting lines. Her point lands because it gives comms teams a commercial test. If an AI intermediary is doing the first pass, evidence has to be structured, third-party and useful before a human ever asks for a demo.

David C. Edelman made the same warning from the customer-strategy side, arguing that the next buyer may not discover a brand directly because their AI will do the comparing first. Daniel Rijo added a harder metric: AI search leads are already being discussed as higher-converting than Google in some funnels. His Microsoft Clarity post also matters because it shows the tooling race around AI visibility moving from niche dashboards into mainstream analytics.

On agentic AI, David Linthicum's practical scepticism was sharper than the usual vendor language. He framed agents as a familiar automation pattern made more dangerous by looser claims, then followed with a warning about the AI subscription trap. a16z's line that companies have hired a million bad employees was deliberately abrasive, but it captured the same worry: cheap agents still need supervision, correction and boundaries.

The human-work thread supplied the counterweight. Lenny Rachitsky's discussion of tech-worker sentiment picked up the anxiety underneath all this, where productivity systems are arriving before organisations have rebuilt trust. The best creator work this week refused the easy split between marketing visibility and operations. Machines may find the brand, qualify the vendor, run the workflow and trigger the payment. People still need to decide what proof, guardrails and human relationships make that tolerable.

The unresolved thread is delegation. AI now wants access to devices, data, code, payments, publishers, legal status and regulated workflows. Every gatekeeper is responding at once: regulators, workers, publishers, cloud providers, security teams and courts. The vendors that win the next quarter will not just offer more capable agents. They will show where agency stops, who can intervene and what evidence a buyer gets when something goes wrong.

Meet Agentcy

The Agentcy Difference