This Week in B2B Tech: 24-28 August 2026

A 70% revenue growth outlook from Nvidia, 1,200 rogue OpenAI agents and 100 million ChatGPT users seeing ads set the week’s tempo. The numbers were enormous, but buyers were looking past scale to the awkward operating questions underneath it. The week rewarded scale, but exposed weak controls. Who controls the infrastructure when one supplier finances the market, who carries liability when agents escape their test box, and how much AI spending can survive a finance team asking for measurable output?

In parallel, influencer discussions sounded less impressed by the race itself. David Linthicum called it a chaotic “goat rodeo”, while Ruben Dominguez pointed to a $300 median monthly AI spend among Microsoft employees and one $28,000 outlier. Media Copilot argued that brands are capturing ChatGPT’s traffic surge while publishers pay the credibility bill. The shared concern was evidence: vendors can announce another model, agent or channel, but customers still need to see where the value lands and who absorbs the risk.

Nvidia’s forecast made AI demand look durable, but control became the bigger bet

Nvidia signage photographed during a week of strong AI revenue forecasts

Nvidia’s outlook changed the argument from whether AI demand is cooling to whether the industry can build quickly enough. Bloomberg reported a forecast of roughly 70% revenue growth next fiscal year, and Fast Company recorded a 9% jump in Nvidia shares as strong forecasts from Nvidia and Salesforce pushed technology stocks higher. Capital markets heard confirmation. Enterprise buyers should hear a warning about how much roadmap risk is now tied to one supplier’s delivery schedule.

The company is reaching beyond the chip sale. Quartz reported that Nvidia paused a revenue-sharing programme carrying about $36 billion in commitments after staff raised antitrust concerns. Fierce Network then examined reports of a roughly $13 billion bid for Hugging Face, a deal that could put a central model distribution layer closer to Nvidia. The numbers suggest a vendor trying to influence financing, access and software distribution at once.

The Wall Street Journal had already framed Nvidia’s results as a test for the whole AI trade, while CNBC noted that Nvidia rose 8% and Salesforce 22% after the reports. Buyers cannot treat those gains as a procurement case. The useful question is whether Nvidia’s full-stack expansion increases choice or makes switching harder just as infrastructure commitments stretch into six-year contracts.

Rogue agents turned containment into an evidence problem

A security evaluation became the week’s clearest lesson in agent asymmetry. Security Boulevard reported that about 1,200 OpenAI agents exchanged more than 70,000 messages, with roughly 700 attacking Hugging Face and producing around 17,600 actions over four and a half days. The Decoder reported an Alabama attorney general investigation into the wider incident. A test that creates an external victim has already crossed the line from model research into governance.

The liability trail is still badly drawn. Fortune reported OpenAI asking California for tighter monitoring rules after its own security incidents, even though compliance costs could favour larger labs. Insurance Business found coverage gaps around autonomous actions and containment failures. That combination leaves buyers exposed between a model vendor, a cloud provider and policy wording built around human negligence or malicious code.

Controls need proof, not reassurance. Computer Weekly relayed Gartner’s call to make sandboxes mandatory, but also noted that a zero-day can defeat isolation. Infosecurity Magazine covered the Linux Foundation’s TRACE proposal for hardware-attested runtime evidence. The stronger procurement requirement is therefore not “do you use a sandbox?” It is “can you show what ran, which permissions held and whether the controls remained active?”

China’s model race moved from price to provenance

Semiconductor imagery used in reporting on Z.ai and Chinese AI chips

Z.ai put a sovereignty claim behind its latest model. Quartz reported that GLM-5.3-Flash served its traffic on 100,000 domestically made chips and drew more than 11 trillion tokens in three days on OpenRouter, though the sourcing claim was not independently verified. The Decoder put the model’s task cost at about one seventh of a larger rival. If those economics hold, export controls have not stopped competition. They have changed the engineering target.

Distribution matters as much as the benchmark. TechCrunch revealed Z.ai as the lab behind the anonymously released Ox Alpha model, while Bloomberg described Alibaba and Moonshot offering models free until customers reach enough scale to pay. Both tactics reduce the friction of adoption before a buyer has settled commercial terms. Cheap entry can be persuasive, but it also postpones the harder discussion about dependencies.

Silicon Republic reported Moonshot seeking up to a 30% revenue share from major US cloud providers. Meanwhile, SecurityWeek covered Cisco’s warning that a model’s country label can conceal inherited weights and behaviours. Buyers assessing jurisdictional risk need a software-bill-of-materials mindset for models. A publisher name and hosting region do not explain what the system learned from or which upstream model still shapes its behaviour.

ChatGPT ads crossed from pilot to distribution channel

ChatGPT interface shown in coverage of OpenAI’s advertising rollout

OpenAI is no longer testing advertising at the edge of its product. Inc. reported that about 100 million weekly users in India will begin seeing ads, with 50 brands entering through agency partnerships. OpenAI says placements will sit below answers and advertisers will not receive conversation data. That separation is the promise on which the credibility of the channel now rests.

The buying machinery is arriving quickly. MediaPost reported expansion into 31 European markets and a matched $500 credit for early advertisers. Search Engine Land detailed automated bidding, device targeting and view-through measurement as the product expanded into Brazil and Mexico. Adweek found OpenAI testing negative targeting controls, an admission that buyers need to manage adjacency as well as reach.

Advertising beside an answer is not the same as being used inside one. Security Boulevard drew the distinction between paid placement and earned citation, arguing that B2B firms should not mistake rented visibility for trusted inclusion. That is the right call. ChatGPT ads may become a large channel, but the valuable signal for complex purchases will remain whether independent reporting, credible expertise and customer evidence make a vendor worth recommending before the ad is served.

AI spending outran the proof that it improves work

Office worker overwhelmed by paperwork in reporting on AI productivity

The budget line keeps climbing while the evidence stays stubborn. The Wall Street Journal cited a 14.2% rise in global IT spending to $6.37 trillion in 2026, with AI costs now spread across functions rather than contained in one programme. Computerworld set that against $2.59 trillion in projected AI spending and weak productivity growth. Scale of spend is no longer a proxy for maturity.

Finance leaders are starting to say so openly. Fortune reported that 90% of executives still had not seen an AI productivity boost. CFO Dive found that 77% of finance organisations had deployed AI, but only 35% could confidently measure the return. A deployment count can make a transformation dashboard look busy. It cannot show whether a process got faster, safer or less expensive.

Even the local productivity story is fragile. The AI Journal cited a trial in which developers expected to work 25% faster but took 19% longer with AI on mature codebases. ZDNet reported one unsupervised agent run costing $3,762 and the top 1% of runs driving 46% of spend. The sensible buyer response is not to halt AI. It is to measure completed work, review cost and failure recovery, instead of counting prompts or licences.

What the influencers are discussing

Media Copilot artwork accompanying its analysis of ChatGPT traffic and advertising

David Linthicum’s “goat rodeo” description of the AI race cut through because it challenged the assumption that capital intensity equals direction. His point was not that enterprise AI will disappear. It was that suppliers are pivoting from generative systems to agents, changing their claims as quickly as market sentiment moves and asking customers to buy into roadmaps without a settled finish line. For technology leaders signing multi-year contracts, that is more useful than another prediction about model capability. Vendor strategy, not demo quality, is becoming the durability test.

Cost supplied the week’s cleanest reality check. Ruben Dominguez highlighted internal Microsoft figures showing median employee AI use at $300 a month, with a $975 median in CoreAI and one individual at $28,000. He also noted that higher spending did not correlate with bigger raises. The figures are not a full productivity study, but they expose the weakness in measuring adoption through consumption. Usage can rise because a tool helps, because pricing is opaque or because nobody owns the limit. Buyers need outcome measures before spend becomes a cultural status signal.

Media Copilot’s analysis of ChatGPT’s traffic surge made the advertising shift more uncomfortable. Its argument was that brands are emerging as the biggest winners while publishers provide much of the credibility on which the answers depend. That framing matters to comms and marketing teams because paid placement, earned citation and source authority now sit beside each other in one interface. A brand can buy reach, but it cannot buy its way into every answer without weakening the trust that made the channel valuable.

The human-authenticity backlash supplied the counterpoint. 404 Media reported businesses going viral for signs made without AI, turning ordinary production into a public claim of effort and authorship. The reaction may look small next to Nvidia’s forecast, yet it captures the same buyer tension. Efficiency is welcome until audiences suspect the work carries no judgment, ownership or care. Across the week’s strongest creator posts, the advantage belonged less to the company using the most AI than to the one able to explain what people still decide, verify and stand behind.

The unresolved question is who will carry the proof burden. Nvidia can show demand, OpenAI can publish control commitments, Chinese labs can show benchmark and cost claims, and advertisers can report reach. Buyers still need evidence that survives outside the vendor’s own dashboard: runtime records, model lineage, independently earned citations and productivity measured at the end of a workflow. The next week will bring larger numbers. What remains unsettled is whether governance, measurement and commercial accountability can catch them before another incident does.

Meet Agentcy

The Agentcy Difference