Monthly AI Briefings

August was the month when the AI story moved another layer up the stack.

While the headlines were still full of new models, some of the most important developments had little to do with benchmark scores. Instead, they were about how AI connects to the systems around it.

The clearest example came from Anthropic, which previewed its Model Hardware Standard (MHS). MHS is designed to give AI agents a common way to interact with physical equipment such as microscopes and lab machines.

What MCP did for Agentic AI, MHS intends to do the same for Physical AI. MCP gave an agent a common interface to external software and tools. MHS is aimed at something similar for physical equipment. Today, connecting a robot, microscope, or lab instrument to an AI workflow or an AI agent can involve weeks or months of custom integration. A common interface could eventually reduce that to hours or minutes. If this works, physical equipment will become more programmable by AI. Labs and factories could start moving away from fixed automation toward agentically configured, operated, and adapted systems. Physical AI needs an integration layer to scale, and MHS can expedite that integration.

A shift also happened in the enterprise SaaS space. Salesforce and Anthropic launched Claudeforce, allowing Claude to interact directly with Salesforce data and workflows. The immediate benefit is obvious for users. Once employees can access business applications through an AI agent, why would they log on to the SaaS’s own interface? Although Salesforce declared it a win, it will put the traditional SaaS model under pressure. While the data, permissions, and business logic still sit inside Salesforce, the user experience could increasingly sit somewhere else.

So, the question every enterprise software company should now be asking: if AI becomes the way employees interact with applications, will SaaS, relegated as a backend, be able to keep a high rate of customer retention?

Now, it doesn’t mean that models will win it all. In fact, we are seeing evidence that the models themselves are becoming less important as harness engineering is taking over the mindspace.

Meta, in an experiment, matched Claude Opus 4.5 on complex workflows by combining its 8B model with a sophisticated “harness” that provided execution feedback, state tracking, and error recovery. DeepSeek also moved in the same direction with an open-source agent harness alongside V4-Pro, its model designed for agentic workloads. Both are pointing to a trend that a smaller model with a good agent architecture can outperform a much larger model operating without one. The implication is significant. The competitive advantage will come from the combination of model + harness + tools + context + workflow, rather than the model alone.

This further backs the rapid commoditization of models, paving the way for model routing as a new offering.

Stripe’s $7 billion (reported) acquisition of OpenRouter, a company whose core proposition is making it easier to switch between AI models, gave this proposition even more credibility. Following it, Ramp, an AI company, also launched its own model router, allowing customers to dynamically select between models based on their requirements. NVIDIA’s NeMo Switchyard is taking a similar approach for agent workloads, routing tasks between specialized and frontier models.

So, we see an emerging trend in the AI space where organizations will not rely on one or two models but operate a portfolio of models. This is exactly where AI Tokenomics becomes an enterprise architecture issue. Model selection, routing, token consumption, and workload placement will determine the economics of AI at scale.

While all this was happening, the cybersecurity story became even more complicated this month.

Google used AI agents to identify and fix 1,000+ Chrome vulnerabilities, including ones that existed for a decade, in 60 days. With Chrome used by billions of people, the scale of that capability matters. It was not an isolated experiment. Zhipu’s (Z.ai) GLM-5.3 also claimed to have identified thousands of vulnerabilities in the CyberGym benchmark. If Gemini and Z.ai could identify them, so could hackers.

This is connected to Agentic AI governance. Anthropic’s multi-agent research showed agents given conflicting objectives could interfere with one another and escalate their behaviour, including attempts to deploy malware. As enterprises move from individual copilots to fleets of agents, agent-to-agent behaviour will become another part of the security architecture.

Meanwhile, investments in foundational AI and data centers only grew in August.

• Nvidia committed ~$105 billion in financing for OpenAI’s Ohio data-center project, for 4.25 gigawatts of capacity, extendible up to 8 gigawatts.
• Anthropic also agreed to spend ~$45 billion on computing capacity from Nscale.
• Mistral is also planning to build ~1 gigawatt of European compute capacity by 2030 to support regional inference and third-party open models.

On the acquisition side, besides Stripe’s $7.5B purchase of OpenRouter, Nvidia also acquired Hugging Face for $12.9B.

So, even though AI models are getting more power/dollar, the infrastructure required to run them is enormous. At some point, the economics of those two trends will have to meet.

Until that happens, enterprise demand for AI is real and powering the growth of both Anthropic and OpenAI, both up for IPOs soon. For now, Anthropic seems to be winning the enterprise AI market, and OpenAI is determined to catch up.

There was also an important regulatory development on the last day of August.

The EU Commission designated ChatGPT as a very large search engine, bringing it under the Digital Services Act and imposing additional systemic-risk and transparency obligations for OpenAI. Soon it will cover other AI models.

The AI landscape is changing fast. While models are still important, they are not as important as they used to be.

The agent harness determines how capability is executed. The model routing determines which model performs the task. The integration layer determines which enterprise systems and machines the agent can control. And the governance layer determines what it is allowed to do.

The most important takeaway from August is that AI is becoming the operating layer through which software, machines, data, and enterprise workflows will connect. The companies that recognize this early will have a very different AI architecture by the time everyone else realizes the model was never the whole story.