Two weeks ago we saw some big product announcements from Microsoft regarding agentic AI, and last week the company followed up with some changes to its Copilot leadership team. I got to speak with Paul Barker about this for a Computerworld article last week. Overall, I think it’s an interesting pivot. Over the past three years, vendors have been moving at a torrid pace for product delivery; when that happens, it’s common for less attention to be paid to cross-product innovation and incorporating the feedback of early adopters. Basically, the teams are so focused on the competition that they let other things slide. The irony is that those “other things” are often the most important capabilities required for enterprise adoption. When responding to new market opportunities, vendors will eventually get the memo that the focus needs to shift from raw innovation to monetization and — almost invariably — product consolidation. So I think that Microsoft’s leadership announcement is a signal of that shift in thinking. It’s a good thing in terms of AI maturation, and I expect we will see more changes like this from other vendors this year.

While I did not attend NVIDIA GTC last week, I was watching intently from the sidelines. You can hear Matt Kimball and me discuss the highlights of the conference in Episode 4 of the IT Talk Podcast. In addition to that, I have been watching the local agent movement gaining momentum. Whether it’s NVIDIA NemoClaw, Claude Cowork, or even Snowflake’s new offering, local AI agents are getting hot. But while I respect the progress, I am still a bit wary. I dug a bit deeper into this topic last week in an article on LinkedIn.

Lastly, I have been a keen observer of Amazon Bedrock for almost two years now. It’s a solid solution for building custom, high-scale agents. And when I say custom and high-scale I’m talking about massive multi-department business processes with multiple humans in the loop, specialized agents, models, and tools. These solutions tend to be at the leading edge of new innovation and ultimately are where a lot of best practices get created. So, I was happy to see this new customer case study about Bedrock. Previous deterministic solutions for these types of applications used to take years to deliver, versus the 16 weeks it took Lendi. It’s a sign that agents and complex AI are not only being deployed, but also having a real business impact.

Two weeks ago we saw some big product announcements from Microsoft regarding agentic AI, and last week the company followed up with some changes to its Copilot leadership team. I got to speak with Paul Barker about this for a Computerworld article last week. Overall, I think it’s an interesting pivot. Over the past three years, vendors have been moving at a torrid pace for product delivery; when that happens, it’s common for less attention to be paid to cross-product innovation and incorporating the feedback of early adopters. Basically, the teams are so focused on the competition that they let other things slide. The irony is that those “other things” are often the most important capabilities required for enterprise adoption. When responding to new market opportunities, vendors will eventually get the memo that the focus needs to shift from raw innovation to monetization and — almost invariably — product consolidation. So I think that Microsoft’s leadership announcement is a signal of that shift in thinking. It’s a good thing in terms of AI maturation, and I expect we will see more changes like this from other vendors this year.

While I did not attend NVIDIA GTC last week, I was watching intently from the sidelines. You can hear Matt Kimball and me discuss the highlights of the conference in Episode 4 of the IT Talk Podcast. In addition to that, I have been watching the local agent movement gaining momentum. Whether it’s NVIDIA NemoClaw, Claude Cowork, or even Snowflake’s new offering, local AI agents are getting hot. But while I respect the progress, I am still a bit wary. I dug a bit deeper into this topic last week in an article on LinkedIn.

Lastly, I have been a keen observer of Amazon Bedrock for almost two years now. It’s a solid solution for building custom, high-scale agents. And when I say custom and high-scale I’m talking about massive multi-department business processes with multiple humans in the loop, specialized agents, models, and tools. These solutions tend to be at the leading edge of new innovation and ultimately are where a lot of best practices get created. So, I was happy to see this new customer case study about Bedrock. Previous deterministic solutions for these types of applications used to take years to deliver, versus the 16 weeks it took Lendi. It’s a sign that agents and complex AI are not only being deployed, but also having a real business impact.

The AI industry may soon be shaken up by a new AI technology. Turing Award winner Yann LeCun has officially launched Advanced Machine Intelligence Labs with a record-breaking $1.03 billion seed round, which is the largest seed round in European history. This values AMI at $3.5 billion. The company was founded because LeCun — who until recently served as the top AI scientist at Meta — believes that large language models lack “common sense” and fail to understand physical reality. Instead of using LLMs to predict the next text token, AMI Labs uses the Joint Embedding Predictive Architecture (JEPA) to build world models.

World models can learn by observing high-dimensional sensory data such as video and lidar. This allows them to create abstract representations of the physical world. By predicting outcomes in a representation space rather than pixel-by-pixel, JEPA allows AI to understand cause-and-effect and spatial relationships. The key distinction with JEPA is that it is a non-generative architecture. Generative AI technologies such as autoregressive transformers, generative adversarial networks (GANs), and variational auto-encoders (VAEs) learn by reconstructing data in the original input space of pixels or words. JEPA, by contrast, predicts higher-level representations rather than raw data.

AMI is backed by NVIDIA, Samsung, and Jeff Bezos, among other high-profile investors. There is speculation that AMI’s technology may be able to achieve AGI at a future date.

Super Micro co-founder Yih Shyan “Wally” Liaw was charged with trying to smuggle large quantities of NVIDIA GPUs into China through Southeast Asia. According to U.S. officials, Liaw worked with others to set up “dummy” servers in a neighboring country and would route Blackwell- and Hopper-class GPUs into China. In the wake of the charges, Liaw resigned from Super Micro and its board of directors.

This is the second go-round with the law for Liaw. When Super Micro was fined by the SEC and had its stock delisted for a year in 2018 for misreporting revenues, Liaw was at the center of that scandal; he resigned for three years before being rehired. It’s worth noting that Super Micro was also in the news about 18 months ago because of failures in timely financial reporting. This led its auditor, Ernst & Young, to resign — a highly unusual act. 

I have a couple of thoughts on this. First, I do not believe that NVIDIA itself was involved in this scandal. My belief is that Liaw and others at Super Micro perpetrated this fraud on their own. Based on his past behavior, this would make sense. And I think this warrants Super Micro being scrutinized more closely by the U.S. government and law enforcement.

My second thought: We, as an industry, need to start thinking more critically about the companies we deal with when we try to accelerate time-to-market or the introduction of new technologies and products. Both NVIDIA and AMD lean heavily on Super Micro because the company focuses on being faster and cheaper. And they sell into hyperscalers and CSPs for the very same reasons. However, it seems clear that Super Micro achieves this by cutting corners and loosening operational discipline. Surely there is another ODM or white-box vendor that can work with the industry to do these things ethically.

I don’t blame any of Super Micro’s partners or customers for what the company has become. But I will be disappointed to see continued enabling of this behavior.

Enterprise AI is starting to run into a structural constraint. The infrastructure that made AI training possible is not the same infrastructure that will scale inference. And this is becoming more visible as organizations move from pilots to production. Inference workloads, especially during token generation, are serialized workloads that are latency-sensitive and heavily dependent on memory bandwidth. These workload characteristics expose the “inefficiencies” in the parallelized design of GPU-centric architectures. This is where many enterprise deployments start to break down.

At GTC 2026 last week, NVIDIA addressed this with the introduction of the Groq 3 LPX, a rack-scale inference accelerator designed to pair with its Vera Rubin NVL72 platform. The key shift here is in separating inference into prefill (handled by GPUs) and decode (handled by LPUs optimized for bandwidth and latency). This disaggregated approach aligns hardware with workload behavior rather than forcing a single architecture to handle both.

The broader signal is more important than the product itself. NVIDIA is acknowledging that inference is no longer a secondary workload and that no single architecture will address the full range of enterprise requirements. As similar approaches emerge, enterprise AI infrastructure is moving toward greater heterogeneity. We can expect to see different accelerators from different vendors deployed to provide the most performant and efficient inference for the right task on the right device and in the right location.

This highly distributed, heterogeneous AI infrastructure will require a sophisticated control plane to manage performance, cost, and operational complexity at scale — a control plane that doesn’t yet exist. NVIDIA has the beginnings with its Dynamo inference control plane layer.

The company that can successfully build a universal control plane that works up the inference stack and across the inference environment (i.e., to optimize inference on any device at any location with any model) will own the inference market.

Could that be NVIDIA with Dynamo?

The move toward disaggregated inference is not unique to NVIDIA’s LPX architecture. Last week I published an article about the new partnership between Cerebras Systems and AWS, where Cerebras is positioned to accelerate the decode phase while other silicon (AWS Trainium) handles prefill. The underlying shift is consistent: Inference is no longer treated as a single workload, and separating prefill from decode allows each phase to be optimized independently.

Where these approaches diverge a little is in what they prioritize. Cerebras is built for throughput at scale, delivering high tokens-per-second across large models and high-concurrency environments. Groq’s LPU, as deployed in LPX, is optimized for deterministic, low-latency token generation. Both are addressing the same constraint, but from different directions. The decision is not about which is faster in absolute terms, but which constraint defines the workload — throughput or latency.

RSAC 2026 starts this week in San Francisco. As usual, there are a lot of announcements. I want to point out a few that address resilience, data security, and how recovery is changing as data continues to spread across systems.

Commvault shared a couple of updates. One is an expansion of Threat Scan in Commvault Cloud. Most organizations have threat intelligence, but applying it to backup data before restoring is still a gap. Commvault’s new offering adds a way to scan for known indicators and a deeper inspection layer to catch what doesn’t show up as an exact match. It also ties directly into recovery, which should help reduce the back-and-forth between tools. The challenge, as usual, is trust. Teams still hesitate to rely fully on scan results when recovery decisions carry real risk.

Commvault’s second update extends its data security posture into structured data through its Satori acquisition. Data security pros know that sensitive data is no longer in one place. It’s spread across structured, semi-structured, and unstructured data repositories, as well as vector databases tied to newer workloads. Most organizations don’t have a clean view of where all that data sits or who has access to it. Expanding visibility helps, but it also adds complexity. More coverage means more policies to manage and more opportunities for gaps. The Satori-based functionality is a good example of the direction Commvault is taking to address this. Detection, governance, and recovery are being brought closer together instead of handled separately. This makes sense to me, and it could even help address organizational issues where security and IT are not aligned.

I’ll be at RSAC this week to see how others are approaching these issues and handling security in real environments. I expect a lot of focus on identity, data protection, and managing access as things get more distributed. There will no doubt be plenty of talk about automation, but I’m more interested in what’s truly working versus what still needs heavy human oversight. RSAC is usually a good place to separate what sounds good from what teams can actually run day to day.

Last week I attended Microsoft FabCon and SQLCon 2026. Arun Ulag, president of Azure Data at Microsoft, kicked things off with his opening keynote. The theme was how databases and Microsoft Fabric come together. The job is to integrate data, bridge transactional and analytical systems, and make AI usable in day-to-day operations. And the real test of how well it works comes in what actually changes for customers.

This year, Microsoft is pushing Fabric beyond a unified data platform. Fabric IQ is taking shape as a semantic layer that connects data, business context, and operational systems. In concept, this allows you to define your business entities and relationships once, then reuse them across analytics, apps, and AI. That’s a big deal if Microsoft wants AI to be consistent and usable across the enterprise. There’s also more focus on real-time data, tighter SQL integration, and agent-driven workflows. We’re seeing less about dashboards, but more about systems that interpret signals and help with decisions.

This builds on a conversation I had with Ulag last year about simplifying data management for AI and cutting friction between systems and insights. Unifying data is one step. Making it usable and actionable across the enterprise is the harder one. Look for my thoughts on the entire event coming soon.

Snowflake recently announced Project SnowWork, and I’ve been thinking about what it really means. Snowflake is moving beyond analytics and starting to close the loop on execution. Instead of stopping at dashboards or queries, SnowWork takes a business question and produces an actual result using the data that’s already in Snowflake.

For years, the handoff between business users and data teams has been the slow part of the analytics equation. You would ask a question, wait on a report, try to make sense of it, and then decide what to do. SnowWork is trying to pull all of that into one step to help get work done faster. To do this, Snowflake is not trying to replace core systems like CRM or ERP, but rather sit across them. That could help cut friction as routine reporting, planning, and analysis move faster while staying aligned to the same governed data.

The current research preview of SnowWork isn’t yet running across complex enterprise environments. Generating an answer is one thing, but executing actions across operational systems is a much bigger challenge. Governance will matter more, too. Once insights start turning into actions, the stakes rise. Enterprises will need to decide how much autonomy they’re comfortable with. And while data teams won’t disappear, their focus is likely to change, with less energy spent on building reports and more on governance, modeling, and oversight. Not every organization is ready for that yet.

There’s real upside here if Snowflake can execute. A tighter connection between data and what actually happens in the business is on the table. Ideally, it could go beyond just managing data and start to change how work gets done day to day.

Snowflake is moving in the right direction, and I don’t doubt the quality of the vision. The question now is how quickly companies feel comfortable letting Snowflake play a role in real operations.

At last week’s SAP Concur Fusion event, SAP shared some key updates that I see as good steps toward connecting travel, expense, and payment activity. These crucial processes are still disconnected in many environments. Booking, expensing, approvals, and reconciliation don’t always work together, which can lead to excess effort and delays. SAP wants to fix this by improving integrations and ensuring that data flows between travel systems, finance platforms, and partners. This could help close an operational gap.

There’s a focus in the announcements on control, which is where finance teams spend most of their time. Knowing where money is going, staying within policy, and cutting down on manual work all matter — a lot. The difference is that when those pieces are handled as a single process instead of a series of disconnected steps, reports can move faster, costs can come down, and compliance can become easier to manage.

Where I see this going is pretty straightforward. Progress isn’t coming from changing ERP itself, but from tightening what sits around it. Because travel and expense processes are high-volume and repeatable, SAP has chosen a practical area for improving consistency and bringing more discipline to how money is managed without making things more complicated.

The most controversial announcement at GTC 2026 last week was NVIDIA’s decision to tease its DLSS 5.0 image software. This was surprising in part because 5.0 isn’t scheduled to arrive until autumn 2026, yet it was announced before DLSS 4.5 has even rolled out. The more surprising part was the overall reaction to NVIDIA’s demos and screenshots. People were visibly angry and upset that DLSS 5.0 was using neural shading to relight some older games. Admittedly, some of these games do not look remotely the same after DLSS 5.0 has been turned on, but using 5.0 — just like every previous version of DLSS — is optional. I believe that when DLSS 5.0 is deployed alongside a new game that has been built in anticipation of its relighting capabilities, we will probably see a lot of people giving it credit for enhancing realism. I found that most of the improvements in DLSS 5.0 were net positives, and the areas I didn’t like seemed to have exposed character design or lighting issues that never accounted for the level of realistic lighting enabled by the new software. I suspect that people’s tempers will calm down by the time the launch happens, and that most people will forget about the current controversy. Even those who don’t forget may find themselves impressed with the final result rather than continue shouting angrily at something just because it uses AI.

Meta announced that it would be shutting down Horizon Worlds after spending billions of dollars in the Metaverse and unknown amounts of money on the Horizon platform. Ultimately, I believe that Meta’s biggest mistakes were that its own expectations in this area didn’t match market realities and that it didn’t take the most growth-focused approach, for example by buying some of the most successful XR apps on the Meta Store.

PTC’s Pivot: From Connectivity to Lifecycle Governance
On March 16, PTC finalized the sale of its ThingWorx and Kepware businesses to TPG for $523 million. This divestiture marks a strategic exit from the industrial IoT identity that defined the company for a decade. While physical connectivity remains a leverage point for the broader industry, PTC is stepping away from the operational data infrastructure — factory floors, field assets, building systems — to focus on higher-value data within the digital product definition.

Shifting the Center of Gravity
The evidence suggests a transition toward what CEO Neil Barua calls the intelligent product lifecycle (IPL). This move re-centers PTC on its core strengths: CAD (Creo, Onshape), PLM (Windchill, Arena), ALM (Codebeamer), and SLM (ServiceMax). PTC is also embedding AI across each of these — Creo AI, Windchill AI, Onshape AI, Arena AI, Codebeamer AI, ServiceMax AI — to build a portfolio-wide intelligence layer rather than a standalone AI product.

By de-emphasizing the “as-operating” sensor data of the IoT era, PTC is betting on higher-value data in the “digital thread” — the interconnected flow of product data from ideation through design, manufacturing, service, and retirement. The goal is to become the definitive system of record for a product’s digital DNA, rather than the connectivity provider for the physical assets that product data describes.

AI-Readiness as the New Foundation
PTC is taking a data-foundation-first approach to AI rather than leading with agentic architecture. Across the portfolio, PTC is embedding task-specific AI directly into each product’s workflow. Windchill AI Parts Rationalization detects duplicate parts and drives consolidation through change workflows — the clearest example of AI improving data quality at the source. Codebeamer AI flags requirements quality issues against INCOSE standards and generates test cases from requirements. Arena AI automates document review and version comparison for PLM and QMS compliance workflows. Onshape AI Advisor provides real-time design guidance and troubleshooting inside the CAD workspace. The pattern is consistent: AI at the point of use, improving both data quality and practitioner effectiveness within each domain.

PTC is not ignoring the agentic trajectory. The company demonstrated an Onshape-to-NVIDIA Isaac Sim workflow for robotics simulation at GTC 2026. But the sequencing is deliberate: make the data trustworthy and the workflows intelligent before attempting orchestration across them.

Financial Realignment
PTC is channeling approximately $375 million in net after-tax proceeds from the divestiture into share repurchases. Post-divestiture, the company is guiding for constant-currency ARR growth of 7.5% to 9.5%, signaling that the remaining portfolio — now entirely movable to SaaS/cloud delivery — is the growth engine PTC wants investors to see clearly, without the connectivity businesses diluting the profile.

My take: PTC’s pivot removes its ownership stake in the industrial connectivity layer. The new portfolio generates value from the data payloads, regardless of who provides the plumbing. In 2026, the winners may not be companies that provide connections to machines, but companies that govern product digital lifecycles. However, as companies like PTC move up the stack to pursue higher-margin opportunities, industrial customers may find fewer well-resourced options to address persistent connectivity and data management challenges.

NXP’s New Radar Transceiver: 4-D Imaging Radar Gets Real
NXP has announced the TEF8388, a third-generation RFCMOS automotive radar transceiver integrating 8T8R on a single monolithic die. The device supports configurations up to 32T32R, enabling a 576-channel virtual antenna array at a power profile that NXP says is comparable to today’s 3T4R devices. The strategic headline is channel density and power efficiency, enabling higher-resolution imaging radar that can ship in volume.

Practical High-Resolution Radar Imaging
High-resolution imaging radar has been technically possible but economically impractical for mass deployment. Higher channel counts imposed a system tax — more power, higher BOM cost, complex thermal management. The TEF8388 integrates the RF front-end into a single die, improving channel isolation and signal quality while shrinking the physical footprint.

The competitive context is important. Texas Instruments’ AWR2944 family integrates the DSP, MCU, and radar accelerator on the same die as the transceiver — a single SoC for maximum footprint efficiency. NXP separates the transceiver from the radar processor (in the S32R4 series), giving OEMs the flexibility to scale sensing and processing independently. Infineon takes a third path: Its RASIC transceivers emphasize best-in-class SNR and pair with AURIX microcontrollers, targeting OEMs that prioritize safety certification and long-range reliability. Three architectural philosophies. Choose wisely.

FORVIA HELLA has already selected the TEF8388 with the S32R47 processor for a high-resolution radar program with a major OEM. The company expects to begin production in mid-2028. That design win has a production date, not just a roadmap placeholder.

Beyond Automotive
The same radar technologies apply in industrial environments where optical sensors struggle — poorly lit factories and warehouses, dusty construction sites, outdoor locations at night or in fog. 4-D imaging radar captures range, azimuth, elevation, and velocity in a single modality. Unlike lidar or camera systems, which calculate motion from successive frames, radar delivers velocity data intrinsically via Doppler effects. For robotics and physical AI applications that need persistent spatial models with real-time motion data, imaging radar is a natural fit.

Industrial demand is materializing across the radar ecosystem. NXP Gold Partner smartmicro already deploys NXP-based radar in industrial robotics, last-mile delivery, and traffic management. Au-Zone, also working with NXP, has shipped a production-ready radar-vision-AI fusion system for autonomous machines. On TI silicon, Zadar Labs demonstrated “Free Path” at CES 2026, a free-space mapping product for autonomous tractors, cranes, and mobile robots. The emerging pattern is multi-domain applications (automotive and industrial) supported by rapid advancements from competing silicon vendors.

My take: The TEF8388 moves high-resolution imaging radar from technically impressive to production-feasible. The immediate beneficiary is ADAS. The longer-term significance is industrial. When high-fidelity 4-D perception gets cheap enough for every car, it gets cheap enough for every robot and every 4-D digital twin modeling camera.

At GTC 2026 last week, NVIDIA announced its GRID initiative with partners including Cisco, Comcast, T-Mobile, and Charter. This initiative seems like a rebirth of what was previously called Mobile Edge Computing (MEC) and adds more powerful compute within networks — somewhere between the RAN and the cloud — that accelerates AI. Considering that many of these deployments plan to leverage RTX PRO 6000 GPUs, there may also be a media or XR application for this AI compute. Ultimately AI will be the workload that makes most edge computing successful, but it’s unclear which workloads will leverage AI RAN versus something like GRID, which is more powerful and capable but sits further back in the network with more latency.

(For more on NVIDIA’s announcements at GTC 2026, see my “Semiconductors” entry below.)

Infleqtion has delivered the U.K.’s only quantum computer with 100 physical qubits to the country’s National Quantum Computing Centre (NQCC) in Oxford. The intent is to enable the U.K. to transition from experimental prototypes to useful real-world applications for quantum. Like other neutral-atom quantum computers, it uses laser arrays to trap and manipulate individual atoms. The system has achieved a 99.73% two-qubit gate fidelity. That is an important metric that will reduce the number of physical qubits required to create a stable, error-corrected logical qubit.

Infleqtion’s roadmap projects that the company will build a processor using more than 30 logical qubits by late 2026 and 100 logical qubits by 2028. The delivery coincides with the U.K. government’s broader £2 billion “Quantum leap” investment, aimed at making the country the first country to make and deploy quantum computers at scale by the early 2030s.

At GTC 2026 last week, NVIDIA made many announcements that spanned semiconductors and more. One of the biggest was the company’s revision of its projected revenue opportunity in AI to $1 trillion by the end of 2027, up from $500 billion by the end of 2026 in its projections last year. I believe that NVIDIA wouldn’t have guided so far upward without some significant visibility into even more accelerated growth, for example the announcement last week of AWS’s intention to order 1 million GPUs over that same period. That’s $30 to $40 billion in GPUs from a single deal, and it’s not even taking into account any potential revenue from networking or other chips from AWS. I don’t think NVIDIA necessarily needed to announce any new products at GTC, but it did need to show accelerating momentum — and it did.

AT GTC, NVIDIA also launched the GB300 DGX Station, a product that it had previously announced along with the DGX Spark, but hadn’t previously given an official date or price. Using the DGX Station is like having an entire GB300 node from a server on your desk, and that includes the power draw: it sucks down 2400W and requires a 20A circuit (more than a standard 120V, 10A plug). I believe that this is the perfect workstation for someone who is regularly working on fine-tuning AI models or needs desk-side inference for testing models. I think this could save some people a lot of money and time in the long term if they are currently chewing through cloud-based tokens. I don’t know if this machine is really for me, although I think I would appreciate it more than the DGX Spark, which has a limited capability mostly because of its SoC. You could theoretically use the DGX Station to power your own OpenClaw deployment and have an extremely powerful agentic computer at your command.

There was also an expansion of NVIDIA’s platforms and products that leverage the new Vera Rubin chip architecture, most notably the introduction of the DSX AI Factory and Space 1 platform. The DSX AI Factory blueprint gathers up all the different things that NVIDIA has already built for its AI Factory offering, but with additional models and software capabilities to round out the solution. This means incorporating a family of models for developing and deploying the AI Factory as well as using things like Omniverse for digital twins to help ensure stable and reliable datacenters. The DSX AI Factory is specifically for larger gigawatt-scale deployments, and so it serves as yet another way for NVIDIA to remove friction from implementing larger GPU-accelerated compute deployments. Meanwhile, the Space 1 is an update of NVIDIA’s rad-hardened capabilities, moving them from the scale of Jetson and H100 chips up to Vera Rubin, enabling as much as 25x more AI compute for inferencing. This also puts a bit of wind in the sails of companies trying to deploy AI datacenters in space, although so far the economics — and the risks — of launching datacenters in space don’t make sense to me. NVIDIA and its partners will still have to address issues such as cooling and radiation, which are inherently some of the most limiting factors for having any compute in space. Consider me extremely bearish on this outside some niche small-scale applications.