1. April is here, and that means most of the analysts on staff — myself included — will be logging lots of miles to attend a wide range of events. One of the things I am always looking forward to is pre-event announcements to see if we can pick up on some themes, and last week IBM provided some of this fodder. IBM’s new agentic IDE, called “Bob,” went GA after being originally announced at the company’s TechXchange event last fall. Overall, it’s a market-competitive IDE for a company that is coming into the AI-assisted IDE race fairly late. But Bob is an important signal in terms of how IBM is trying to bring its disparate software landscape (and ecosystems) together into harmony. An interesting note is that when I first “met” Bob at IBM TechXchange, IBM suggested I participate in a series of briefings about its broader IT Automation platform. Over the past few months, I got to know the IT Automation team and IBM’s thoughts on market opportunities and possible futures in that area. Those sessions led to me releasing this new research on our site.

2 One thing that I am cautious about with respect to Bob is the pricing model, which is activity-based and abstracted into something called Bob Coins. (I can’t help but think “Bob Bucks” would have been so much better.) IBM is not alone here; other vendors are developing new monetization options for non-deterministic solutions as well. Like many observers in this space, I am not very excited about this. My main issue here is the lack of a consistent anchoring effect. Enterprises have always had the chance to peg their spend off of something besides activity, whether it was seats, hardware infrastructure, or time. This lack of an anchor has bred a degree of mistrust, particularly when definitions for AI activity are still kind of in their early stages. The other challenge is that agents can be tricky because they are non-deterministic. So one run of a process may cost twice as much when you run it the next day. The good news is that I think some vendors are acutely aware of the challenge and issues. And while they may not have all the answers, they are at least trying to make their pricing models more understandable and transparent. I had a briefing from one leading vendor in this space last week, and I do see progress. Stay tuned on this, because I will be following up on my agent pricing research series from last year soon.

3. Microsoft has announced a new Researcher Agent in Microsoft Copilot 365 (see my video short here about it). I am becoming a fan of these agents as they are fairly powerful in terms of data gathering and structure. That said, what I am finding is that the typical output of a report may not be as valuable as the steps and data gathered along the way. And this is where Microsoft has done something really interesting: Instead of locking Researcher to one AI model, it uses multiple models to provide the best insights and answers. As a frequent user of AI for research, I can say that I find the most value in the data rather than the final report, and often I am switching between models to see what different ones have to say and where the data comes from. Maybe Microsoft can save us some time and deliver better accuracy using this new approach.

4. Finally, let’s talk about stepping out of your comfort zone for the sake of learning. Moor Insights & Strategy has had a relationship with Synopsys for many years, and the company invited me to its Converge event to hear about its AI and agentic strategies. I walked away with a far better understanding of what it will take for AI to be useful in physical settings, and also how Synopsys is preparing for the challenge. This Analyst Insight is worth a read, especially if you spend all your time thinking AI is only a cloud thing.

1. April is here, and that means most of the analysts on staff — myself included — will be logging lots of miles to attend a wide range of events. One of the things I am always looking forward to is pre-event announcements to see if we can pick up on some themes, and last week IBM provided some of this fodder. IBM’s new agentic IDE, called “Bob,” went GA after being originally announced at the company’s TechXchange event last fall. Overall, it’s a market-competitive IDE for a company that is coming into the AI-assisted IDE race fairly late. But Bob is an important signal in terms of how IBM is trying to bring its disparate software landscape (and ecosystems) together into harmony. An interesting note is that when I first “met” Bob at IBM TechXchange, IBM suggested I participate in a series of briefings about its broader IT Automation platform. Over the past few months, I got to know the IT Automation team and IBM’s thoughts on market opportunities and possible futures in that area. Those sessions led to me releasing this new research on our site.

2 One thing that I am cautious about with respect to Bob is the pricing model, which is activity-based and abstracted into something called Bob Coins. (I can’t help but think “Bob Bucks” would have been so much better.) IBM is not alone here; other vendors are developing new monetization options for non-deterministic solutions as well. Like many observers in this space, I am not very excited about this. My main issue here is the lack of a consistent anchoring effect. Enterprises have always had the chance to peg their spend off of something besides activity, whether it was seats, hardware infrastructure, or time. This lack of an anchor has bred a degree of mistrust, particularly when definitions for AI activity are still kind of in their early stages. The other challenge is that agents can be tricky because they are non-deterministic. So one run of a process may cost twice as much when you run it the next day. The good news is that I think some vendors are acutely aware of the challenge and issues. And while they may not have all the answers, they are at least trying to make their pricing models more understandable and transparent. I had a briefing from one leading vendor in this space last week, and I do see progress. Stay tuned on this, because I will be following up on my agent pricing research series from last year soon.

3. Microsoft has announced a new Researcher Agent in Microsoft Copilot 365 (see my video short here about it). I am becoming a fan of these agents as they are fairly powerful in terms of data gathering and structure. That said, what I am finding is that the typical output of a report may not be as valuable as the steps and data gathered along the way. And this is where Microsoft has done something really interesting: Instead of locking Researcher to one AI model, it uses multiple models to provide the best insights and answers. As a frequent user of AI for research, I can say that I find the most value in the data rather than the final report, and often I am switching between models to see what different ones have to say and where the data comes from. Maybe Microsoft can save us some time and deliver better accuracy using this new approach.

4. Finally, let’s talk about stepping out of your comfort zone for the sake of learning. Moor Insights & Strategy has had a relationship with Synopsys for many years, and the company invited me to its Converge event to hear about its AI and agentic strategies. I walked away with a far better understanding of what it will take for AI to be useful in physical settings, and also how Synopsys is preparing for the challenge. This Analyst Insight is worth a read, especially if you spend all your time thinking AI is only a cloud thing.

OpenAI closed its biggest funding round ever, ending it with the acquisition of a popular business and tech podcast TBPN. This controversial round comes in at about $122 billion dollars and includes investors such as SoftBank, Microsoft, NVIDIA, and Amazon. As a reminder, this follows OpenAI’s decision to kill off its Sora video platform, which led Disney to pull out of its $1 billion investment in OpenAI. OpenAI is surely focused on staying competitive with the likes of Google Gemini and Anthropic Claude, but I don’t think it needed to kill off an entire class of model to accomplish that.

Google released its Gemma 4 open model, and all of the major silicon vendors have already said that they support it on their chipsets. Gemma 4 will come in four sizes — 2B, 4B, 26B, and 31B — with 26B being a mixture-of-experts model and 31B being a dense model. Offering this range of models clearly means that Google wants them to be able to run locally on-device at lower cost and with less lower power consumption. That said, given that these are open models, the expectation is that they’ll power some third-party applications’ offline modes or hybrid AI paired with Google Cloud.

Generating long videos with AI models that are pre-trained on short clips can cause visual degradation because of two out-of-distribution (O.O.D.) issues: (1) frame-level relative position O.O.D., and (2) context-length O.O.D. To address this, researchers have just launched FreeLOC, a training-free, layer-adaptive framework designed to resolve these challenges without requiring additional fine-tuning.

FreeLOC uses two main techniques:

• Video-based Relative Position Re-encoding (VRPR) — a multi-granularity strategy that realigns temporal positions with the model’s pre-trained distribution.
• Tiered Sparse Attention (TSA) — a method to preserve local details and long-range dependencies by structuring attention density across temporal scales.

A key innovation is its layer-adaptive probing mechanism, which identifies the specific sensitivity of each transformer layer to O.O.D. issues. This permits a selective, efficient application of the researchers’ methods. A number of experiments have demonstrated that FreeLOC is able to achieve results that outperform other training-free approaches in both temporal consistency and visual quality. Check out the project code on GitHub.

NVIDIA’s $2 billion investment in Marvell and Marvell’s support for NVLink in its XPUs is less about Marvell specifically and more about where NVIDIA is placing its bets. The company is steadily positioning NVLink as the connective tissue of the AI stack, regardless of whose silicon is actually doing the compute.

Step back and consider the company’s moves. Besides Marvell XPUs, Intel Xeon gained NVLink support last year. NVIDIA has also locked in key optical and networking partners like Coherent and Lumentum. This isn’t an acquisition strategy, but ecosystem consolidation. NVIDIA understands heterogeneity in silicon is inevitable. The real battle is controlling the fabric and software layer that ties those systems together.

I think this is a smart bet. When considering inference across an enterprise, this heterogeneity is somewhat inevitable. The value is shifting toward how those components are integrated and orchestrated at scale.

There is, of course, a competing interconnect standard in UALink, backed by a broad industry consortium (including Marvell, Astera Labs, Intel, and others). But it really hasn’t seen market adoption yet. And it notably does not natively support NVIDIA GPUs. This is the strange dynamic where openness is optimal, but if it doesn’t support the chips that command more than 90% of the market, its path to wide adoption is challenged.

For enterprise IT, the practical takeaway is straightforward. You won’t be building these fabrics yourself, but your OEM or cloud provider will. And whichever interconnect model wins will directly influence your ability to mix vendors, manage cost, and maintain flexibility over time.

OpenAI’s latest $122 billion finance round and $852 billion valuation are impressive. But the way the company is structuring its cloud and silicon partnerships is maybe even more impressive and points to something more fundamental than simple expansion. The company is reshaping how infrastructure is sourced, scaled, and coordinated at the highest end of the market.

By committing to massive, multi-year compute agreements across cloud providers, OpenAI is effectively dictating what gets built, right down to silicon choices, while maintaining an abstraction layer above it. The strategy reflects a focus on capacity, resilience, and ensuring that the end-user experience (whether enterprise or consumer) is consistently available and performant. This is why we see multi-cloud and multi-silicon partnerships.

Speaking of silicon, there was some eyebrow-raising when OpenAI entered into an agreement with AMD a couple months back, with some observers not understanding the importance of supplier diversity. And in this round, we see a continued reliance on NVIDIA, alongside growing support for alternative silicon (Cerebras is another partner). It is important to remember that NVIDIA remains the backbone for frontier-model training and large-scale inference, even while emerging options are being integrated as complementary capabilities to expand capacity and improve efficiency.

For enterprise IT, this isn’t a blueprint to replicate directly, but it does signal where things are heading. As AI moves into production, standardizing on a single cloud or hardware stack becomes increasingly difficult. The control point shifts from owning infrastructure to coordinating across it. And this is the dynamic enterprises will need to plan for over time.

One of the longest-running and most well-known VR gaming apps, RecRoom, has announced that it will be shutting down. The developer of the game says that this is purely a business decision and that the company simply cannot move forward given the current VR climate and the state of monetization in VR today. While I’m not saying that VR is dead, I do believe that we’re in a phase of the market that doesn’t bode well for many smaller companies in the near term. The market was buoyed heavily by Meta and Valve, yet both companies have had to delay their own headsets and have pulled back on funding. Apple remains strong and a solid enterprise player in XR, but it is mostly shipping MR and AR products — not VR.

NVIDIA DLSS 4.5 has finally launched, right after the uproar of the company announcing DLSS 5.0 at GTC 2026. DLSS 4.5 introduced dynamic frame generation, which allows for up to 6x higher frame generation depending on the need for the scene or game. Early benchmarks show that it significantly improves average FPS while sometimes also improving 1% frames as well. In general, 4.5 continues to deliver on the promise of DLSS and enhance performance without taking too much of an image-quality hit.

MediaTek
On April 1, I participated in MediaTek’s annual Analyst Day in San Francisco. The company’s AI-savvy strategy and crisp execution brought it $19.1 billion of revenue in 2025, with healthy year-over-year growth. With this as context, I sat down with Sameer Sharma, general manager of MediaTek’s IoT business in the U.S. and Europe, to get his perspective on the rapid transition of the Genio edge/IoT product line from incubation to execution. His team is firing on all cylinders with the Genio Pro 5100, announced recently at Embedded World 2026. This chip is the flagship of the Genio fleet, combining scalar and AI performance with connectivity and I/O, engineered for AI-first products such as drones, robotics, agriculture, industrial automation, edge AI, machine vision, and transportation.

MediaTek has proven its ability to integrate heterogeneous IP blocks on advanced TSMC process nodes — the core competency behind the GB10 SoC co-developed with NVIDIA for the DGX Spark. In my view, this gives the company a structural advantage as edge AI chips become increasingly complex mashups of CPU, NPU, GPU, memory, comms, and application-specific I/O. The NVIDIA relationship extends beyond silicon co-development to include 150 pre-hardened AI models in the NVIDIA TAO catalog, giving Genio developers a fast path from prototype to deployment across MediaTek’s edge product line.

I’ll be writing more about MediaTek’s edge strategy and developer enablement over the next few weeks. Meanwhile, many thanks to Sharma and the MediaTek executive team for a delightful and informative event.

Matter 1.5.1
The big smart-home news last week was the Connectivity Standards Alliance’s release of Matter 1.5.1. Version 1.5 introduced support for cameras, and this update adds multi-stream capabilities, allowing a single camera to deliver multiple optimized streams in the same session. For instance, cameras can now deliver multiple resolutions simultaneously. The update also enables new media formats. HEIC images, familiar to Apple users, provide better image quality at lower bandwidth. HLS and DASH support improves HTTP streaming uploads using the CMAF Interface-2 profile, enabling greater compatibility with the technologies widely used by modern cloud and media platforms. Beyond that, Matter 1.5.x allows customers to use cameras from multiple suppliers with a single standards-based application — which has immediately become one of Matter’s most compelling use cases.

I attended Zoom Perspectives 2026 in Half Moon Bay, California, last week, which allowed me to spend a couple of days with Zoom’s leadership discussing the company’s strategic direction. While much of the content was under NDA, I can tell you that the company outlined plans to grow channel contribution from 40% to 80%. Zoom’s leadership also discussed aggressive SMB and solopreneur GTM strategies. Marketing initiatives will focus on the conversation-to-completion framework across meetings, calls, and chats. It was also notable to see alignment on recent M&A activity, particularly Workvivo by Zoom. Although not new, its integration into the broader platform continues to grow. I expect to see more strategic M&A, given Zoom’s approximately $8 billion in cash to accelerate this platform journey.

My main takeaway is that Zoom Revenue Accelerator is positioned to be a breakout product as Zoom executes on its strategy. Zoom already has the infrastructure in place. Tools include AI Sales Assist for real-time call coaching, automated scorecards, agentic prospecting, and Deal Explorer for pipeline analysis. The company has built an AI-powered decisioning layer that connects conversations, enterprise data, and third-party applications through AI Companion 3.0’s federated AI platform with no-code orchestration. AI Companion monthly active users have tripled year-over-year. And all top-10 deals in its most recent quarter (ending January 2026) included paid AI solutions. Revenue Accelerator could become a meaningful growth driver if Zoom continues building agentic capabilities and deeper workflow integration to realize the conversation-to-completion vision.