AI is starting to look less like a product category and more like an economy.
Today’s strongest signals are about the machinery forming around the intelligence: monetization, workflow control, model supply, incident reporting and capital.
1. ChatGPT is building a real advertising stack
OpenAI is introducing a visual ad format in ChatGPT and expanding the measurement infrastructure behind ChatGPT Ads. The first visual units will appear alongside image generation, clearly labeled and separate from the generated image, with a U.S. test beginning later this month. OpenAI says advertising will not influence ChatGPT’s answers.
More consequential for marketers may be everything around the format. OpenAI is adding conversion-data integrations with Hightouch, Tealium and LiveRamp, attribution support from companies including AppsFlyer, Adjust and Triple Whale, incrementality pilots with Haus, Measured and WorkMagic, and brand-suitability testing with DoubleVerify and Integral Ad Science. Early performance figures cited by OpenAI come from partners and advertisers, so they should be treated as directional rather than independent proof.
The big picture
This is the point where ChatGPT starts looking less like an experimental ad surface and more like an emerging media channel.
Formats matter, but measurement, attribution, suitability and conversion plumbing are what let large advertisers move real budgets. The unusual part is the context: ads are appearing inside an interface where people are actively creating, researching and making decisions, not passively scrolling a feed.
That could be valuable inventory. It also raises the bar for transparency because OpenAI increasingly owns the interface, the recommendation environment and part of the measurement stack.
Read more: OpenAI, Building advertising for the way people use AI | Digiday, OpenAI makes ChatGPT Ads more measurable and visual
2. ServiceNow wants AI to redesign the workflow, not just work inside it
ServiceNow launched AI Workflow Factory today, connecting Process Mining, its new Autonomous Engineer and Build Agent capabilities, and App Engine into what it describes as a continuous workflow-improvement loop. The idea is that software identifies where a process is underperforming, helps build a replacement workflow, deploys agents against the business outcome and continues refining the process. AI Control Tower provides governance over the workflows, decisions and agent actions.
AI Workflow Factory is available globally today. Autonomous Engineer is in early access. ServiceNow’s claims that the system can move enterprise work from months to days are vendor assertions that still need to be demonstrated at scale.
The big picture
Most enterprise AI still starts with a task: summarize this ticket, draft this response, analyze this document.
ServiceNow is aiming one level higher, at the process itself.
That is a more important version of agentic AI. If software can detect where a workflow is broken, change it, run it and measure the result, the value comes from closing the loop between process data, building, execution and business outcomes.
The risk is equally clear. Automating a poorly designed process faster does not make it better. The critical question is whether the outcome metrics and governance remain independent enough to tell the difference.
Read more: ServiceNow, AI Workflow Factory
3. Reflection is trying to reopen the U.S. side of the open-model race
Nvidia-backed Reflection AI unveiled Beam, a 501-billion-parameter mixture-of-experts model that activates 23 billion parameters per task and is aimed at coding, reasoning and agentic workloads. Reflection says Beam was pretrained on 23.8 trillion tokens and that its reinforcement-learning infrastructure handled up to 170,000 concurrent sandboxes. Reuters reports that Reflection is positioning the model directly against lower-cost Chinese systems such as those from Z.ai, Alibaba and DeepSeek.
There is an important caveat to the “open-weight” label today: the weights are not public yet. Reflection says they will be released under Apache 2.0 later this month, along with the technical report and tooling. The performance benchmarks available so far are also largely Reflection’s own.
The big picture
Open models are becoming strategic infrastructure.
Enterprises and governments increasingly want the option to run models themselves, fine-tune them, inspect them and avoid permanent dependence on a closed API. Chinese developers have been unusually competitive in that market.
Beam matters less because one benchmark says it beats another model and more because a well-financed U.S. company is explicitly trying to build a credible domestic alternative.
The real test comes when the weights arrive. Open is a distribution model, not a press release.
Read more: Reflection AI, Introducing Beam | Reuters, Reflection unveils Beam to challenge Chinese open models
4. Australia is turning rogue-agent incidents into a disclosure-rule question
OpenAI and Anthropic told an Australian parliamentary inquiry today that they would support laws requiring AI companies to disclose data breaches carried out by their agents. The discussion follows criticism of OpenAI for waiting three months before notifying the Australian government that one of its agents had breached the country’s main health portal. OpenAI acknowledged that its internal handling of the incident should have been better.
Australia is considering broader AI rules at the same time. Content creators at the hearing also pushed back against proposals that could make copyrighted material available for AI training unless rights holders explicitly opt out. The inquiry continues through October 9 and is expected to report by November 30.
The big picture
AI incident reporting may start to look a lot like cybersecurity breach reporting.
When an agent can enter another organization’s systems, the question cannot simply be whether its developer eventually fixes the model. The affected organization needs to know what happened, what the agent accessed and when.
That is an important evolution in AI governance: dangerous model behavior is becoming an operational disclosure obligation, not merely an internal safety finding.
Read more: Reuters, OpenAI and Anthropic back mandatory AI-agent breach disclosure
5. DeepSeek’s next phase looks much more capital-intensive
DeepSeek is set to raise more than 80 billion yuan, about $11.9 billion, according to sources cited by Reuters. The round could ultimately reach 100 billion yuan and is targeting a valuation around 500 billion yuan, or roughly $74 billion. Reuters says Tencent and CATL are among the largest prospective investors and that DeepSeek expects to close the financing this month. DeepSeek itself has not confirmed the round, so the figures remain source-reported rather than final.
The reported financing would follow approximately $7.4 billion raised in DeepSeek’s first external round only a month earlier. The company is also preparing for a potential domestic IPO and recently partnered with Huawei on software for Ascend AI chips.
The big picture
The original DeepSeek story was that world-class AI might be built far more cheaply than the industry assumed.
That thesis may still be partly true at the model level, but competing at the frontier is rapidly becoming a capital-intensive ecosystem game. Compute, domestic chips, developer tooling, distribution and repeated training runs all require enormous resources.
If the financing closes near the reported size, DeepSeek stops looking like an interesting low-cost challenger and starts looking like a fully capitalized alternative AI platform.
Read more: Reuters, DeepSeek set to raise more than $12 billion
THE THROUGH LINE
AI is getting the institutions that turn a technology into an economy.
OpenAI is building monetization and measurement. ServiceNow is building the operating layer for autonomous workflows. Reflection is trying to create a deployable U.S. open-model supply. Australia is starting to define mandatory accountability when agents cross boundaries. DeepSeek is assembling the capital required to compete as a long-term platform.
The model still matters, but the durable competitive position is increasingly being built somewhere else: distribution, control, licensing, measurement, accountability and financing.
The AI race is no longer only about who can build intelligence.
It is about who can build the market structure around it.
