We are fast approaching fall tech event season and are now less than 30 days out from Dreamforce. In the run-up to the event, I have had the chance to attend a few briefings, but last week Salesforce pre-empted Dreamforce with an update on its headless strategy. Salesforce announced that strategy last spring at its TDX event, which I also attended and wrote about. The headless strategy is interesting in that it challenges some of the more widely held pillars of SaaS platforms. It’s notable, and if Salesforce continues down this path, the approach could prove promising. Check out my latest Field Note on the topic.

OpenAI also made a couple of announcements this week as it prepares for its IPO. I had a chance to speak with journalist Evan Schuman about it in this CSO article. Overall, I think OpenAI is partly positioning itself to overcome perception issues that have mounted over time. But the reality is that both announcements, regarding model releases and customer privacy, are more than just good PR. Both OpenAI and Anthropic have had remarkable growth trajectories over the past two years, but by going public they will not only need to keep that revenue trajectory high but also do it consistently. To do that, both companies will need to cater more and more to the enterprise, which can deliver that revenue scaling better than any of us individually paying between $20 and $200 per month. But winning in the enterprise also requires solid guardrails and assurance. These OpenAI announcements are part of proving the company’s enterprise credentials.

Last week we also saw the announcement that Stripe is planning to acquire OpenRouter. This is an interesting move, since Stripe is really good at routing money and now will be in the business of routing tokens. Hyperscalers have been offering model routing for over a year now, but those offerings are generally in the service of an individual cloud. The idea of a multi-cloud, multi-premises model is a good starting point in optimizing AI and agent costs. As enterprises are becoming more interested in negotiating their own deals with different model providers, and in some cases mixing in their own models, this trend seems bound to continue. If you’d like to learn more, earlier this month I wrote a piece about OpenRouter competitor NVIDIA Switchyard.

Lastly, I got to speak with journalist Beth Pariseau about GitHub’s recent outages. To say the least, it was a frustrating day last week for the many customers who suffered an almost eight-hour GitHub outage. This is the second major outage this summer, and it comes on the heels of recent pricing increases for GitHub agents. The good news is that the GitHub team appears to have nailed down the cause and has been executing on its plans to reinforce GitHub as usage has exploded in the past year. However, I am also seeing an emerging trend where different developer tools and agents are beginning to expand their aspirations across the software development lifecycle. While I don’t foresee anyone that can disrupt GitHub’s massive market lead today, one does have to wonder whether competition will begin to heat up in 2027 or 2028.

We are fast approaching fall tech event season and are now less than 30 days out from Dreamforce. In the run-up to the event, I have had the chance to attend a few briefings, but last week Salesforce pre-empted Dreamforce with an update on its headless strategy. Salesforce announced that strategy last spring at its TDX event, which I also attended and wrote about. The headless strategy is interesting in that it challenges some of the more widely held pillars of SaaS platforms. It’s notable, and if Salesforce continues down this path, the approach could prove promising. Check out my latest Field Note on the topic.

OpenAI also made a couple of announcements this week as it prepares for its IPO. I had a chance to speak with journalist Evan Schuman about it in this CSO article. Overall, I think OpenAI is partly positioning itself to overcome perception issues that have mounted over time. But the reality is that both announcements, regarding model releases and customer privacy, are more than just good PR. Both OpenAI and Anthropic have had remarkable growth trajectories over the past two years, but by going public they will not only need to keep that revenue trajectory high but also do it consistently. To do that, both companies will need to cater more and more to the enterprise, which can deliver that revenue scaling better than any of us individually paying between $20 and $200 per month. But winning in the enterprise also requires solid guardrails and assurance. These OpenAI announcements are part of proving the company’s enterprise credentials.

Last week we also saw the announcement that Stripe is planning to acquire OpenRouter. This is an interesting move, since Stripe is really good at routing money and now will be in the business of routing tokens. Hyperscalers have been offering model routing for over a year now, but those offerings are generally in the service of an individual cloud. The idea of a multi-cloud, multi-premises model is a good starting point in optimizing AI and agent costs. As enterprises are becoming more interested in negotiating their own deals with different model providers, and in some cases mixing in their own models, this trend seems bound to continue. If you’d like to learn more, earlier this month I wrote a piece about OpenRouter competitor NVIDIA Switchyard.

Lastly, I got to speak with journalist Beth Pariseau about GitHub’s recent outages. To say the least, it was a frustrating day last week for the many customers who suffered an almost eight-hour GitHub outage. This is the second major outage this summer, and it comes on the heels of recent pricing increases for GitHub agents. The good news is that the GitHub team appears to have nailed down the cause and has been executing on its plans to reinforce GitHub as usage has exploded in the past year. However, I am also seeing an emerging trend where different developer tools and agents are beginning to expand their aspirations across the software development lifecycle. While I don’t foresee anyone that can disrupt GitHub’s massive market lead today, one does have to wonder whether competition will begin to heat up in 2027 or 2028.

Waymo announced that its new Ojai vehicle, which it has built for its next-generation autonomous taxi service, is powered by its own custom 5nm ASIC. The company intends to give more details at the Hot Chips event underway at Stanford. The one detail it did offer in advance is that the chip is capable of 1,000 TOPS of AI compute; most of the computation these autonomous vehicles perform today is for advanced driver assistance systems (ADAS). One of the greatest challenges of autonomous driving, whether it’s Level 3, L4, or even L5, is the amount and cost of compute and sensors necessary to enable autonomy. Waymo has shown that most of the computing for the Ojai vehicle appears to fit in the trunk of the Zeekr-brand EV it is modifying for the taxi service. Waymo has also optimized the Ojai with fewer sensors than its predecessors to reduce cost and computing needs, and I expect that the company isn’t done tinkering with the design. While I haven’t ridden in an Ojai yet, the company officially introduced it this month in a few markets, and it is weeks away from launching the service in San Diego (where I live) using only Ojai vehicles.

There are two moments when you can hold an agent accountable. You can read what it’s about to do before it runs, or you can trace what it touched after the fact. Not many vendors are building for both of those moments yet.

An agent runs up a bill, and somebody has to figure out which query did it. Bigeye just shipped cost anomaly detection in its Agent Trust Hub, and it points right at the tables. I told TechTarget that the attribution is the uncommon part. It costs Bigeye almost nothing, because the access records it already keeps for policy enforcement have the answer in them. It works on older systems too, where a lot of regulated companies keep their data. Bigeye already sells a control that can block a request outright, and wiring an anomaly straight into it seems like the natural next step. That turns a quick alert into a real spending limit. The figures are estimates today, so a team can spot a trend without charging back the spend yet. I’d love to see chargeback land, too, because that would bring finance into the conversation.

Reviewing an agent’s work gets easier when the automation is just code you can open. Serval made that bet early. Everything it automates gets written in TypeScript, so a new workflow is something you can read, version, and audit. Its setup agent stages every change for a person to approve. when Victor Dey interviewed me about this for a piece on Forbes, I gave the company credit for that when — and I still would. Nobody in this category has landed on who owns all that code a year in. When an agent writes your automation, you end up with an estate of it, and estates need owners and review cycles. I’d like to hear how the teams building this see it going.

Platform engineering is being asked to absorb a job it wasn’t designed for, and the Platform Engineering 2.0 framework published in June is the clearest attempt yet to describe that job. It’s a directional argument from VMware, Broadcom, and the platform engineering community rather than a standard or a settled architecture. Parts of it are aspirational. The pressures behind it aren’t.

AI turned a lot of smart people into developers. It didn’t make them engineers. A finance manager can now stand up a working application, which is genuinely good, and it’s the same democratization we saw with spreadsheets and low-code apps. But testing, architecture review, data governance, cost controls, and resource policy don’t show up just because the code compiles. Somebody has to supply them, and that used to be the person writing the code. The bottleneck moved from writing code to shipping it. An organization that shipped four releases a year can now plausibly ship forty, even if the pipeline and the governance layer are still sized for the old rate.

Agents turn that judgment problem into a scale problem. An inexperienced developer writes inefficient code once. An agent writes it continuously, across thousands of invocations, landing on whatever silicon happens to be free. Nobody notices until the bill arrives and the CIO freezes coding assistants.

The new framework’s five pillars are AI-native infrastructure, multiple personas, embedded FinOps, security enforced at the infrastructure layer, and composable architecture. Almost none of that represents a new capability. GPU provisioning lives in MLOps. Spend control lives in FinOps tooling. Policy-as-code lives in DevSecOps. The actual claim is that platform engineering should become the layer that unifies them into one operating model, and that’s an organizational question rather than a technology selection.

There are two things here enterprise IT should act on. First, embedded FinOps is the most immediately useful pillar, because surfacing cost at provisioning time rather than reporting it afterward is how every corporate travel policy already works. Second, plan for the expanded user base explicitly, because the citizen developer and the agent are both arriving whether the platform is ready or not.

Where I’d push back is scope. The framework treats accelerator provisioning as a GPU and TPU question. That’s where enterprises are today, but not where they’re headed. My Field Note covers this, and I’ll be digging into the implementation side at VMware Explore next week.

The gap that Platform Engineering 2.0 describes at the platform layer gets wider one level down, at the estate, and that’s where I keep landing after ModCon 2026 last week. Modular solved a real problem: A lot of engineering has gone into making AI infrastructure fast, but comparatively little has gone into making it manageable.

Three things go wrong in a mature AI estate. First, cost is opaque, because cloud opex, depreciating on-premises capacity, and unmetered edge hardware never roll up to a model, a team, or a business unit. Second, the operational picture is fragmented, since every vendor exposes telemetry in its own shape, and collecting metrics isn’t the same as maintaining a normalized view. Third, there’s no control surface, so policy lives in documents while workloads land wherever capacity exists.

“Control plane” is a loose term right now. Anyone with an AI dashboard uses it. A management layer earns the name when it can do three things across the whole estate: observe what exists, decide what should happen based on policy and business objectives, and act. Most AI orchestration handles the third and part of the first, inside a single stack. That’s automatic rather than autonomous.

The order of assembly matters more than the feature list. You get a system of record first, then intelligence, then control. Nobody accepts placement advice from software that’s been watching for a week, and no enterprise lets software act on production infrastructure before that. Several vendors appear likely to skip ahead anyway.

Here’s the competitive problem. Every obvious candidate has a structural reason to stop short of addressing the entire problem. Silicon vendors can’t credibly hold the record an enterprise uses to negotiate against them. OEM tooling anchors to OEM hardware. Hyperscaler management surfaces stop at the cloud boundary. Observability vendors observe rather than act, and AI cost attribution looks more like asset and contract management than tracing. Serving layer companies own the workload, not the environment.

For enterprise IT, cost attribution is the entry point, because a CFO funds it without needing a strategic argument. Do that work, and you’ve built the record as a byproduct. And with memory supply tight into 2027, using what you already own stops being an optimization and looks more like a necessity.

Modular used its ModCon event to make heterogeneous inference look like something an enterprise could actually do, and I came away thinking that the company is further along than most people realize. The platform already ran on NVIDIA and AMD GPUs, plus Apple Silicon for local development. The keynote added AWS Trainium, Google TPUs, and Qualcomm’s Dragonfly accelerators to that list. None of the additions are GPUs. Each brings its own toolchain, memory model, and execution style, which makes this the hard version of the portability problem rather than the familiar one.

Most of the coverage of ModCon is likely to lead with benchmarks. The number that matters more is what it costs to bring a new accelerator onto the platform. Modular says five engineers enabled Trainium with full serving support in under five months, against an industry norm it puts at hundreds of engineers over two years. No methodology came with that, so treat it as the company’s framing. More persuasive was the outside evidence. A services firm brought up TPU support in roughly four months with a handful of engineers and little prior TPU knowledge. A portability layer that only its author can extend is a roadmap. But one that others can extend is a foundation.

Then there are the economics. Modular showed a frontier model on AMD MI355X delivering about 90% of B200 throughput at roughly half the cost per accelerator-hour, netting more than 45% lower TCO. Same caveat applies, since precision, batch, concurrency, and pricing source all move that number — and none were disclosed. But it’s the first time I’ve seen this argument expressed as a figure a CFO would react to. Optionality never closed a budget conversation. A validated 45% delta would.

Competitively, this puts pressure on the assumption that software availability dictates hardware selection. If enablement gets cheap, then second- and third-source silicon would have a path to enterprise relevance it hasn’t had since CUDA locked things down. For enterprise IT, the practical benefit is that a refresh decision stops being a one-way door. The open question is neutrality, since Qualcomm now owns the layer that decides where work lands.

If you’re running more than one model, something has to decide which one answers each request. That decision is called routing. Everybody had agreed that enterprises would run a mix of models, but nobody had shipped the piece that picks among the models. Four platforms suddenly have one, inside eight days.

Databricks keeps putting its agent controls in the layer above the model instead of inside it, and for routing it makes the same move again. Smart Routing hit beta in the Unity AI Gateway, where the spending caps already live. So, the thing enforcing your budget now picks your model. I like the logic a lot. It’s still in beta, so the cost figures in circulation are early internal benchmarks rather than customer results, which is normal at this stage. It aimed the first use case at code and data engineering. I’d love to see customer-facing inference next, since that’s where token spend concentrates.

Snowflake, Google, and NVIDIA all have their own takes. Snowflake’s Dynamic Model Routing is in private preview, Google’s API Gateway routing is in public preview, and NVIDIA put NeMo Switchyard on GitHub as an open-source library. They’ve all landed on the same idea, that cheap models should answer the cheap questions, and I think they’re right. NVIDIA’s is the one you can grab without being anybody’s customer, which I give the company credit for. It does put the company selling the compute in charge of the logic picking the model, and I’d like to hear how NVIDIA thinks about that.

Routing exists now, and it opens up a number of questions: How much money did it actually save? Did the cheap model give a worse answer? What happens when a model changes underneath you? Can your team see the decision the router made? Those are just some of the observability questions that a lot of folks are asking.

Last week, I attended NXP’s Analyst Day and Developer Days in Santa Clara. Here’s the upshot: NXP is doubling down on physical AI.

I came away impressed with NXP’s edge focus and its expanding product portfolio, and convinced that the company can deliver the efficient productization support that developers need to deploy physical AI securely and at scale. Here are my top three takeaways from the event:

1. “The Edge is Our Home” — That’s how Charles Dachs (EVP and general manager for edge and industrial strategy at NXP) characterized NXP’s focus. The edge is a large market with diverse architectures that span a wide range of use cases across multiple vertical industries. NXP builds what it calls the “Neural Axis,” a real‑time physical AI nervous system that combines MPUs, MCUs, AI acceleration, sensing, connectivity, power, safety, and networking into a coherent portfolio.
2. Core Technologies “Blur” Across Verticals — Physical AI technologies like motion control, perception, and deterministic networking span robotics, factory automation, and other verticals. Dachs reckons that 60% to 70% of NXP’s edge revenue already comes from mass-market customers rather than key accounts. To support a larger number of smaller, more diverse customers, NXP is moving up-stack into platform-based software for simplifying development and productization for customers of all sizes — not just the big players.
3. “Right-Sized” AI — Ravi Annavajjhala (VP and general manager for neural processing, AI, and physical AI) stressed that customers buy business outcomes, not raw TOPS. Annavajjhala is pictured below advising us to “right-size” edge AI rather than indulging in overkill for it. Right-sizing requires a hardware continuum, and NXP offers a portfolio spanning low-power MCUs, MPUs, and Ara NPUs, supported by software (eIQ and AI Hub) designed to get models up and running quickly on scalable silicon.
   
   

NXP’s Ravi Annavajjhala at NXP Analyst Day 2026 (Credit: Bill Curtis)

I also did a deep dive on the MCX A5 MCU with 10BASE-T1S industrial networking that was announced at the event. My Field Note about this chip explains how it advances industrial Ethernet.

Google’s new Pixel lineup has become available to consumers after Google announced it earlier in the month. There’s been a lot of attention on the Tensor G6 chip as the core of the new Pixel experience, and on improving the phone’s on-device AI capabilities to enable new uses. That said, the CPU and GPU still aren’t quite up to par with Apple’s or Qualcomm’s latest chips, which serve the vast majority of the smartphone market. The raw numbers for the new Pixel Watch 5’s performance also don’t lead this market, but I’ve found that the user experience is still great — and I think that’s very much the core of what Google focuses on.

Qualcomm has teased that it will launch a dual-flagship strategy at this year’s Snapdragon Summit, with a Snapdragon 8 Elite Gen 6 chip as well as a Pro version, which may allow it to address more of the market. In previous years, the company had one flagship and one almost-flagship chip, so it will be interesting to see how it markets these two chips at launch and how OEMs adopt them. If Qualcomm’s rumored price increases are accurate, the new strategy might also give OEMs an opportunity to still use a flagship chip without paying the top-end price.