Last week I attended the Synopsys Converge event. This is not my typical beat in terms of event participation, but I learned a great deal more about AI in the design of something physical versus digital. I’ll have a more detailed analyst note available soon, but here are my three big takeaways from the event overall.

1. It seems that the mega-acquisition of Ansys has worked pretty well. A number of cross-product integrations already are underway. And while I am not an expert on chip design and systems engineering, the narrative holds together very well.
2. The new Electronics Digital Twin Platform stands out to me as a key future pillar of not only Synopsys’s stack but also anyone attempting to do anything in the physical AI space. I also like the partnership with NVIDIA on Omniverse as part of the overall story. Those of us in the AI space need to start thinking more about AI beyond the multi-gigawatt datacenter.
3. The agentic story is very sound. What is obvious is that Synopsys has been investing and maturing this technology well. In the near future your engineering agent could collaborate with a Synopsys expert agent to co-develop new products is a very real (and closer to reality than you may think) solution.

Microsoft made a major announcement around its agentic strategy, and it’s very interesting. While there is some great technology there, the packaging and release approach are reinforcing how Microsoft is actively targeting enterprises and wants to win the battle to be a broad and complete AI platform. Here’s why I think that.

The notion of Microsoft releasing in waves is very interesting since (a) enterprises need to manage versions, and (b) by launching functionality together, you can provide a more cohesive set of tools. AI has been moving fast, but maybe it’s time the technology’s velocity showed some respect for adoption rates. Waves is all about delivering platform value to customers.

Agent 365 is a much needed set of tools for the user side of agents. Microsoft Foundry is already doing a good job with security and observability, so extending it to the business side is smart. I was able to have a deeper conversation after the announcement, and one thing Microsoft is very focused on is what I would call a more guided experience. So instead of being asked what could be an open-ended question from Claude or Gemini, you are getting something more like a menu with specific choices and options from integrated products.

I am super interested in the Cowork integration with 365 Copilot. There’s been some real progress in LLM/human arbitrage when it comes to editing artifacts, but it’s mostly been through a Google stack, and even that is limiting. So, this collaboration could be quite interesting. However, I am trying to understand how different the Cowork experience will be with Microsoft — as compared with the Claude experience. My thinking is that initially it will probably be the curated experience I mention above, but not a lot else. That said, I could be wrong, and we need to keep our eyes on it. Meanwhile, I do want to give Microsoft credit for backing up last year’s claims regarding its use of the best model for the job.

Databricks made two big announcements last week that reinforce my notion that data platforms are evolving into context platforms. First was the acquisition of Quotient.AI, which makes an evaluation model. Evaluation models are going to be very important as agents begin to scale. In many ways you can think of an evaluation model like fraud detection on a credit card: the model is part of the overall agentic flow, providing a first line of defense in terms of response quality. As we see agents generating more work, humans will not be able to keep up, so we will need to give some of that work to models.

The second announcement was the release of Genie Code, which is a new data science agent. Much like the coding agents we see from AWS and Microsoft, Databricks is bringing its own agent to market to reduce the friction of performing data science actions on its platform. Actions like these are similar to what we have seen with Snowflake and MongoDB as they try to attract new users and simplify the use of their platforms to win over new customers.

Oh, and one more thing: Twenty years ago last week, an online bookstore started a revolution by getting into an additional line of work. Happy anniversary to AWS!

Last week I attended the Synopsys Converge event. This is not my typical beat in terms of event participation, but I learned a great deal more about AI in the design of something physical versus digital. I’ll have a more detailed analyst note available soon, but here are my three big takeaways from the event overall.

1. It seems that the mega-acquisition of Ansys has worked pretty well. A number of cross-product integrations already are underway. And while I am not an expert on chip design and systems engineering, the narrative holds together very well.
2. The new Electronics Digital Twin Platform stands out to me as a key future pillar of not only Synopsys’s stack but also anyone attempting to do anything in the physical AI space. I also like the partnership with NVIDIA on Omniverse as part of the overall story. Those of us in the AI space need to start thinking more about AI beyond the multi-gigawatt datacenter.
3. The agentic story is very sound. What is obvious is that Synopsys has been investing and maturing this technology well. In the near future your engineering agent could collaborate with a Synopsys expert agent to co-develop new products is a very real (and closer to reality than you may think) solution.
   
   

Microsoft made a major announcement around its agentic strategy, and it’s very interesting. While there is some great technology there, the packaging and release approach are reinforcing how Microsoft is actively targeting enterprises and wants to win the battle to be a broad and complete AI platform. Here’s why I think that.

The notion of Microsoft releasing in waves is very interesting since (a) enterprises need to manage versions, and (b) by launching functionality together, you can provide a more cohesive set of tools. AI has been moving fast, but maybe it’s time the technology’s velocity showed some respect for adoption rates. Waves is all about delivering platform value to customers.

Agent 365 is a much needed set of tools for the user side of agents. Microsoft Foundry is already doing a good job with security and observability, so extending it to the business side is smart. I was able to have a deeper conversation after the announcement, and one thing Microsoft is very focused on is what I would call a more guided experience. So instead of being asked what could be an open-ended question from Claude or Gemini, you are getting something more like a menu with specific choices and options from integrated products.

I am super interested in the Cowork integration with 365 Copilot. There’s been some real progress in LLM/human arbitrage when it comes to editing artifacts, but it’s mostly been through a Google stack, and even that is limiting. So, this collaboration could be quite interesting. However, I am trying to understand how different the Cowork experience will be with Microsoft — as compared with the Claude experience. My thinking is that initially it will probably be the curated experience I mention above, but not a lot else. That said, I could be wrong, and we need to keep our eyes on it. Meanwhile, I do want to give Microsoft credit for backing up last year’s claims regarding its use of the best model for the job.

Databricks made two big announcements last week that reinforce my notion that data platforms are evolving into context platforms. First was the acquisition of Quotient.AI, which makes an evaluation model. Evaluation models are going to be very important as agents begin to scale. In many ways you can think of an evaluation model like fraud detection on a credit card: the model is part of the overall agentic flow, providing a first line of defense in terms of response quality. As we see agents generating more work, humans will not be able to keep up, so we will need to give some of that work to models.

The second announcement was the release of Genie Code, which is a new data science agent. Much like the coding agents we see from AWS and Microsoft, Databricks is bringing its own agent to market to reduce the friction of performing data science actions on its platform. Actions like these are similar to what we have seen with Snowflake and MongoDB as they try to attract new users and simplify the use of their platforms to win over new customers.

Oh, and one more thing: Twenty years ago last week, an online bookstore started a revolution by getting into an additional line of work. Happy anniversary to AWS!

Google AI
Unlike Anthropic, Google is in good standing with the DoD. It has expanded its partnership by introducing the Agent Designer tool on the Pentagon’s GenAI.mil platform. The tool is powered by Google’s Gemini and is part of the company’s Gemini for Government offering, which is available to more than 3 million military personnel and civilians for unclassified tasks.

Agent Designer has a no-code/low-code interface for creating custom AI agents. The agents can perform tasks such as autonomously summarizing documents, drafting reports, and tracking complex workflows. This is a significant milestone for Google that strengthens it as a key provider at a time when Anthropic faces difficulties with the DoD. Despite the military’s fondness for Google — with more than 1 million DoD users already active on GenAI.mil — the company has also faced internal frictions similar to those seen during the 2018 Project Maven protests there.

AI Safety
Max Tegmark has issued a serious warning about human-level AI. Tegmark is a well-known physicist and AI safety researcher. He recently warned that human-level artificial intelligence could occur sometime this decade. Tegmark, who is a co-founder of the Future of Life Institute, believes that our development of LLMs has reduced estimates of the time to create superhuman AI, and that society is dangerously unprepared for such systems. Based on this, Tegmark believes everyone should halt large-scale AI experiments to give ourselves time to develop strong safety standards and guardrails. He believes we should direct our efforts toward projects that benefit humanity instead of building autonomous systems that could become uncontrollable.

It is difficult to ignore Tegmark’s concerns because there is strong evidence that supports taking such a pause. For example:

• Deceptive Tendencies — Recent studies on frontier models show that even 1 to 5% data contamination can trigger “cross-domain dishonesty,” where models learn to manipulate users to achieve internal goals.
• Autonomous Agentic Risk — We are seeing the first instances of AI viruses and worms capable of self-propagation across ecosystems like Gemini and ChatGPT, performing multi-stage attacks at speeds no human can counter.

Safety Filter Decay — The industrialization of “jailbreaking” has made current guardrails largely performative, with attackers using “flip tokens” to bypass filters that legitimate defenders are still bound by.

AI accelerator Cerebras has entered into a partnership with AWS whereby the Cerebras CS-3 system will work with AWS’s Trainium3 to deliver what the companies claim as the fastest cloud-based AI inference.

How does it work? AWS has taken a disaggregated approach to inference serving via Bedrock, its managed generative AI service. The Trainium3 chip, with its large HBM footprint, will serve the prefill stage of inference, where a query is loaded along with all the required context for servicing that query. CS-3, built around the wafer-scale WS-3 processor with an extremely large SRAM footprint, feeds 900,000 AI cores across an extremely low-latency fabric, delivering decode (token serving).

Trainium3 and CS-3 are connected via the AWS Nitro EFA (high-speed interconnect) and consumed through Bedrock in a process that should be seamless from the user’s perspective. Users will be able to choose a standard offering (all Trainium3) or premium (Trainium3 + CS-3), with Bedrock making the appropriate API call to instantiate the service.

While details are pretty scarce (other than availability in the latter half of 2026), it appears that the deep engineering work has been completed, and the two companies are working on early previews with select customers. Pricing has not yet been disclosed.

I think this is a super interesting play for a few reasons:

• First, it demonstrates AWS executing on a customer-choice strategy that its leaders have said will be the CSP’s approach to AI. By offering to serve up non-Trainium (or Trainium-plus, if you will) inference to meet its customers’ wants and needs, it is a very strong signal.
• Second, this partnership is essentially a playbook for abstracting the complexities of standing up AI infrastructure for enterprise customers. Imagine an enterprise IT organization trying to architect a disaggregated inference environment to deliver, say, real-time fraud protection. Firing up Bedrock Premium — or whatever the name will be — should be as simple as making a selection in a UI.
• Third, it gives inference-accelerator silicon companies a playbook for breaking the new duopoly (NVIDIA and AMD) rising in the inference market.
• Fourth, it demonstrates how early in the inference market we are. We, as an industry, are still figuring out how to design, build, and deploy the infrastructure that is going to serve the needs of the market for the next generation of computing.
• Finally, it demonstrates how open and big the inference market is. There is so much opportunity for innovators who are not simply iterating on classic concepts, but rather reimagining how to deliver platforms and services to the market.

It was kind of a quiet news week leading up to NVIDIA GTC in San Jose. Other than the bold move from AWS and Cerebras discussed above, it feels like every other tech company has been keeping its powder dry for big fireworks at GTC. It is almost like it’s 1998 and we are heading into Comdex or NetWorld+Interop, where every tech vendor was making their biggest announcements.

I note this to demonstrate just how big and influential GTC is — because of how influential NVIDIA is. I have never witnessed a single company with the influence on the industry that NVIDIA has today. Not Microsoft. Not Intel. Nobody. The company is setting the standard in terms of narrative, product cadence, launch timing, and more from silicon to systems to software.

From any perspective, this isn’t just impressive — it’s historical.

I will be at GTC for the week and will be attending meetings with partners and competitors for 12 hours a day. Look for updates and analysis throughout the week.

Integration platforms are starting to play a bigger role in how enterprises prepare data for AI. Boomi recently announced new capabilities in its Enterprise Platform focused on what it calls data activation, with the goal of making enterprise data more available in real time across applications, APIs, and workflows.

A few things about this announcement. First, the focus is on real-time data availability, with the goal of ensuring that information from systems including ERP, CRM, and supply chain can move across the enterprise as events occur rather than sitting in those isolated systems.

Second, the concept ties closely to the growing conversation around AI agents. Agents need consistent, contextual data across systems to perform tasks and automate workflows. If enterprise data remains scattered or difficult to access, AI tends to stop at providing insights rather than enabling operational action.

Third, Boomi’s platform already sits in the middle of many environments connecting applications and APIs, which naturally positions it as an orchestration layer for how data moves across the enterprise.

AI agents and automation systems only work well for customers when they can access trusted data across multiple applications and processes. Making that data available across workflows allows organizations to move from isolated analytics toward systems that can actually execute actions across the business.

This strategy reinforces Boomi’s role as a data and integration backbone for enterprises adopting AI. Rather than organizations building separate data pipelines for every AI initiative, Boomi is positioning its platform as the governed foundation that connects applications, prepares data, and supports analytics, automation, and AI-driven processes across the enterprise.

Teradata recently updated its Enterprise Vector Store with new capabilities that allow companies to build AI applications directly on top of the data platforms they already run. This gives customers an opportunity to potentially simplify architecture and governance as AI applications retrieve the context they need without copying large amounts of data across multiple systems.

A couple of important points about this update. One that stands out to me is hybrid search, which combines traditional keyword search with semantic search that understands the meaning behind a query. The other is support for multi-modal embeddings, allowing AI to work with text, images, and audio instead of just structured database data. In practice, this means that companies can bring in documents, PDFs, images, or audio files, convert them into vector embeddings, and search them alongside traditional enterprise data.

The bigger theme here is where AI workloads run. As companies experiment with AI assistants, semantic search, and agent-driven applications, those systems need context from enterprise data. Teradata’s approach is to keep those capabilities inside the data platform rather than requiring customers to move data into separate vector databases or AI tools.

Oracle’s Q3 FY 2026 results were positive with total revenue reaching $17.2 billion, up 22% year-over-year, and cloud revenue growing to $8.9 billion, up 44%. This main driver continues to be infrastructure: Oracle Cloud (OCI) grew 84% as demand for AI training and inference capacity expanded across industries. Remaining performance obligations reached $553 billion this quarter, up 325% YoY, which shows just how much contracted cloud business Oracle has sitting in the pipeline.

While infrastructure growth is getting most of the attention right now, the applications side still tells an important part of the story when thinking about Oracle’s longer-term enterprise strategy. Oracle Fusion Cloud ERP generated $1.1 billion in revenue, up 17%, while NetSuite also reached $1.1 billion, growing 14%. Together they showcase Oracle’s two-lane ERP strategy, with Fusion landing with larger global companies modernizing finance, supply chain, and planning, while NetSuite keeps gaining ground with mid-market organizations that want a quicker rollout and a simpler operating model.

A few of the customer wins this quarter give a good sense of where Oracle’s ERP and supply chain platforms are being used in real operations. For example, easyJet selected Fusion ERP and EPM to reduce manual data work and strengthen planning. J.M. Huber Corporation recently went live on Fusion ERP and SCM to standardize processes as it scales. Louis Vuitton expanded its use of Oracle retail and inventory systems to support store operations and inventory visibility.

I talk often about how ERP and supply chain transformations can be complex for customers, especially for those moving off systems that have been in place for years. I think it’s important to point out that for many enterprises, modernization is less about chasing the newest technology and more about keeping operations adaptable as business conditions change. Ongoing adaptability is crucial, because the cost of doing nothing can build over time through disconnected processes, manual work, and slower decisions.

Many vendors in this space, including Oracle, are trying to meet customers where they are. Most organizations are moving in stages rather than making a single large jump. ERP and supply chain systems often sit right in the middle of that effort as companies modernize their operations, improve access to data, and introduce more automation over time. So infrastructure may be producing many of the larger growth numbers right now, but ERP and supply chain platforms continue to play a central role in how companies run their operations and move forward with modernization.

Companies have invested heavily in modernizing supply chains over the past several years, but confidence still hasn’t fully caught up. New research from Sage shows what it calls a growing “supply chain confidence gap,” describing a widening divide between organizations that feel prepared to handle disruption and those still reacting as events unfold.

A few things I noticed in these findings: Only about half of supply chain teams say they feel confident responding to disruption in 2026, despite the investments many companies have made in systems, analytics, and data platforms. AI also dominates supply chain discussions, yet adoption remains early, with only about 10% of companies running AI in live operations today. The report also shows sourcing strategies shifting, with around 44% of companies planning to move sourcing closer to home to improve resilience, quality, and compliance.

The issue for me behind the confidence gap comes down to data visibility across the supply chain. When data flows across suppliers, logistics, inventory, and demand signals, teams can anticipate disruption and respond earlier. When that visibility is fragmented, organizations are left reacting, and that’s where the confidence gap really starts to show up.

Embedded World 2026 Summary
Over the past three years, the AI industry has proved that machines can reason about the digital world. Over the past 20 years, the embedded electronics industry has proved that machines can automate the physical world. At Embedded World 2026, those two trajectories demonstrated tangible signs of convergence — and the implications go far beyond making edge devices smarter.

What emerged from 1,262 exhibitors and record attendance in Nuremberg is early evidence that AI is extending beyond the digital world into the physical one, not just analyzing data and generating human-readable output, but sensing, deciding, and acting in real-world environments. Here are my top 10 takeaways:

1. Distributed intelligence was a unifying theme of EW26. Edge AI spans the continuum from cloud to endpoint.
2. EU Cyber Resilience Act (CRA) enforcement starts in September. Failure to comply means products can’t ship in Europe.
3. AI is now table stakes for embedded systems across the full power spectrum, from energy-harvesting sensors to high-performance edge appliances.
4. Edge AI platforms hit the streets, and Qualcomm’s VENTUNO Q is the most complete example
5. RISC-V shifts from evaluation to automotive design-in with certification-grade tooling.
6. Edge AI platforms emerge, abstracting hardware complexity to deliver integrated security, DevOps, and lifecycle management.
7. Rust is becoming a mainstream option for safety-critical embedded development as certification-grade tooling arrives.
8. Zephyr (MCU OS) may be approaching its “Linux moment,” with more than 900 boards and safety certification work underway.
9. Physical AI at scale is moving beyond the lab and into production with shipping silicon and production-targeted software stacks.
10. Agentic AI at the edge is becoming a practical embedded architectural pattern with distributed agents and no central controller.
    
    

As intelligence expands into the physical world, the embedded industry is building the compute and platform foundation to deliver it. What’s missing — and what barely showed up at EW26 — is the architecture that enables distributed AI endpoints to act autonomously, safely, and in coordination. Making AI think was the first era. Making AI act in the physical world — where every mistake is costly, some errors are irreversible, and physics won’t wait for round-trips to the cloud — is the next one. The companies, platforms, and standards that solve this will define it.

You can read my full analysis in my latest Forbes article.

Last week, researchers from Google Quantum AI published a new paper, “Measurement-Induced State Transitions in Inductively-Shunted Transmons.” It reflects some clever work to address a pressing problem in quantum computing.

Here’s the context. Developers continually work to build quantum computers that perform faster, deeper calculations. However, researchers are hitting a major roadblock: The more power you use to read a qubit’s state quickly, the more likely it is to cause an error. This is known as measurement-induced state transition (MIST). It is like trying to read a thermometer by blasting it with a heat lamp; the energy needed to achieve the desired results can ruin the measurement itself. To make it worse, errors are unstable. They can randomly drift around due to tiny, shifting electrical charges in the qubit’s environment. This can lead scientists to overcomplicate the system by keeping the qubit and the readout tool tuned very far apart or by using extra hardware to eliminate shifting charges.

Google’s research has now created an elegant but simple hardware fix through the addition of a small inductive shunt. The inductively shunted transmon (IST) acts as an electrical bypass for the qubit, stabilizing the system against environmental fluctuations. It makes the qubit immune to drifting offset charges that cause measurement errors to move around. This results in a more robust architecture that’s needed for accurate high-speed quantum error correction readings. In short, ISTs are promising candidates for quantum error correction. In the bigger picture, this approach appears to simplify the path toward stable, logical qubits that industry leaders are currently racing to define.

There were many announcements at Enterprise Connect in Las Vegas last week, so I started by unpacking just a couple. In my latest Forbes article, I look at how Zoom’s new AI canvases fit in with a shift toward collaboration platforms as true “systems of action” for getting work done. I also published a Moor Insights & Strategy research note that goes deeper into Zoom’s broader UC platform strategy and what it means for enterprise workflows and governance.

I also covered RingCentral’s launch of AIR Pro — a new voice‑first, agentic AI platform designed to let enterprises build autonomous, omnichannel voice agents that can reason over intent and execute multi‑step tasks across customer interactions. You can hear my perspective in the most recent Enterprise Apps podcast.

Expect more analysis from me about the announcements from Enterprise Connect soon.