Last week was all about AWS re: Invent, and while we did see the typical hundreds of announcements from AWS and partners, we also encountered a different tone from the show. In previous years, people often complained about AWS’s “shock and awe” approach to keynotes. But this year was an improvement, where the most important announcements got the most time. And frankly, it was all about agentic AI and developers. My technical highlights were the bulleted items below — but I also have a list of new things to get brushed up on, including Nova Act and Kiro Powers.

• AgentCore — Can you build agents faster and with more control using existing frameworks? Yep! In fact, Signal65 and Moor Insights & Strategy just released new research showing that AgentCore can deliver agents 2.1x faster by augmenting existing frameworks. But it’s not just about speed; the new Policy and Evaluation features seem certain to lead to better and safer agent results. I also appreciate the building-block modularity of the solution.
• Nova Forge — AWS just dropped the barriers for building custom models in a very meaningful way. By leveraging Nova models and tools from Bedrock and SageMaker, companies can now securely train models based on their own data, versus fine-tuning or content reinforcement. It’s pretty amazing, especially if you are already on board with the Amazon AI model tool chain. If you’re not there yet, this may be enough to get you to commit to an AWS-derived and -managed model.
• Kiro and Frontier Agents — I am going to stick my neck out here and say that for the past year, I have not seen anything that has approached what GitHub has been doing with autonomous coding agents … until now. The combination of Frontier Agents, Kiro, and spec-based development is finally a complete, alternative solution that should prompt serious consideration.

Last week I also had the pleasure of spending some time with Stripe to hear about its open source Agent Commerce Protocol. The best way to think about the framework is if you are using ChatGPT (or Perplexity as well) to evaluate different options for a purchase, wouldn’t it be great to just click “Buy It Now”? Well, if you try to do that without ACP there are a ton of latency and security challenges, even if you use an agentic browser. ACP is a protocol like MCP or A2A that easily connects a client to a set of APIs and services. While I have been cautious about agentic payments because of the non-deterministic aspects of agents and AI clients, ACP is a bridge between the non-deterministic elements and the very deterministic world of payments processing — via Stripe, which is already a major player in the space. I have a podcast with Six Five coming soon, but if you want to read about it now, here’s a link.

Lastly, we are starting to see a trend of some AI-specific acquisitions being announced, such as ServiceNow’s planned purchase of Veza. (Melody Brue goes into detail about that specific acquisition in the “Enterprise Apps” section of this page.) As a trend, I think this has been a long time coming, and I think it will continue into 2026. If you don’t believe in an AI bubble, there is still a lot of money out there at the moment. Even the most prominent players must mitigate the risks of AI in production to maintain revenue growth. And for those who do see a bubble, sometimes a well-timed acquisition can distract from less-than-favorable financial results vis-à-vis reduced startup valuations. I don’t know which way it goes and when but, either way, expect more M&A activity.

Last week was all about AWS re: Invent, and while we did see the typical hundreds of announcements from AWS and partners, we also encountered a different tone from the show. In previous years, people often complained about AWS’s “shock and awe” approach to keynotes. But this year was an improvement, where the most important announcements got the most time. And frankly, it was all about agentic AI and developers. My technical highlights were the bulleted items below — but I also have a list of new things to get brushed up on, including Nova Act and Kiro Powers.

• AgentCore — Can you build agents faster and with more control using existing frameworks? Yep! In fact, Signal65 and Moor Insights & Strategy just released new research showing that AgentCore can deliver agents 2.1x faster by augmenting existing frameworks. But it’s not just about speed; the new Policy and Evaluation features seem certain to lead to better and safer agent results. I also appreciate the building-block modularity of the solution.
• Nova Forge — AWS just dropped the barriers for building custom models in a very meaningful way. By leveraging Nova models and tools from Bedrock and SageMaker, companies can now securely train models based on their own data, versus fine-tuning or content reinforcement. It’s pretty amazing, especially if you are already on board with the Amazon AI model tool chain. If you’re not there yet, this may be enough to get you to commit to an AWS-derived and -managed model.
• Kiro and Frontier Agents — I am going to stick my neck out here and say that for the past year, I have not seen anything that has approached what GitHub has been doing with autonomous coding agents … until now. The combination of Frontier Agents, Kiro, and spec-based development is finally a complete, alternative solution that should prompt serious consideration.

Last week I also had the pleasure of spending some time with Stripe to hear about its open source Agent Commerce Protocol. The best way to think about the framework is if you are using ChatGPT (or Perplexity as well) to evaluate different options for a purchase, wouldn’t it be great to just click “Buy It Now”? Well, if you try to do that without ACP there are a ton of latency and security challenges, even if you use an agentic browser. ACP is a protocol like MCP or A2A that easily connects a client to a set of APIs and services. While I have been cautious about agentic payments because of the non-deterministic aspects of agents and AI clients, ACP is a bridge between the non-deterministic elements and the very deterministic world of payments processing — via Stripe, which is already a major player in the space. I have a podcast with Six Five coming soon, but if you want to read about it now, here’s a link.

Lastly, we are starting to see a trend of some AI-specific acquisitions being announced, such as ServiceNow’s planned purchase of Veza. (Melody Brue goes into detail about that specific acquisition in the “Enterprise Apps” section of this page.) As a trend, I think this has been a long time coming, and I think it will continue into 2026. If you don’t believe in an AI bubble, there is still a lot of money out there at the moment. Even the most prominent players must mitigate the risks of AI in production to maintain revenue growth. And for those who do see a bubble, sometimes a well-timed acquisition can distract from less-than-favorable financial results vis-à-vis reduced startup valuations. I don’t know which way it goes and when but, either way, expect more M&A activity.

AWS re:Invent last week was quite the event. With more than 60,000 attendees, 250 industry analysts, 25 AWS product announcements, and hundreds of announcements from ecosystem partners, I’m still trying to make sense of it all. While most of what enterprise customers think about are the services they consume (e.g., Bedrock, SageMaker) and even the AI models, it is the underlying technology that drives differentiated performance. On this front, AWS generated considerable noise around its infrastructure and silicon portfolio. Below are four of what I consider the more significant announcements.

1. AI Factories

Most enterprises begin their AI journey in the cloud and then bring on-prem to achieve a hybrid (AI) environment eventually. This on-premises deployment can be quite tricky and complex. Enter AWS AI Factory. This is the AWS AI stack (Bedrock and/or SageMaker) built on AWS-provided infrastructure, powered by either NVIDIA or the company’s own Trainium chips for acceleration.

If this sounds a little like the company’s Outpost offering for AI, you would be drawing a directionally accurate parallel. However, where Outpost delivers a comprehensive set of AWS services, AI Factory focuses solely on setting up AI. Finally, this is fully AWS-managed, with the customer simply consuming it.

My take on AWS AI Factory: I recognize the value that AWS is delivering. I believe an enterprise IT organization will undoubtedly find value in what AWS is delivering. It is a valuable tool for existing AWS customers that have already started their AI journey. However, I don’t believe this is a solution that will, for instance, attract customers from competitors. Azure (local zones), Google Cloud (Gemini on GDC), and OCI (Dedicated Regions) have been playing in this space for some time and are well established. Nonetheless, AWS AI Factory will be a good fit for AWS’s at-scale customers that require a hybrid environment to satisfy regulatory or latency requirements.

2. Interconnect

In today’s multi-cloud world, moving data from cloud A to cloud B is a heavy lift. You must build the circuits and architect a networking solution that can take several months to activate. And even after this hard work is completed, you will undoubtedly need to turn knobs regularly to ensure optimized (and secure) connectivity.

AWS thinks it has the answer with its jointly developed multi-cloud interconnect specification. I refer to it as jointly developed because both AWS and Google claim a leadership role in the development of this specification. Regardless, this is essentially a straightforward reference network architecture that enables any cloud provider to connect to any other cloud provider that has adopted and implemented an architecture in line with this specification. From a user’s perspective, deploying fast, secure cloud connectivity should be quite simple — literally just pointing and clicking.

Initially, AWS and Google Cloud have activated this interconnect, with Azure in the queue. This is all good for enterprises to treat their multi-cloud environment as a true extension of the datacenter. However, the value is only realized when this spec is adopted en masse. A lot of what enterprise IT organizations struggle with is moving data (and apps) from specialized clouds for training on, say, AWS. I would really like to see how AWS plans to expand adoption beyond the big players and into more specialized and regional markets.

One last note. OCI has been doing this for years with Azure and Google Cloud. What this interconnect specification means for OCI’s dedicated interconnect is a bit of a head-scratcher. However, I think it’s important to acknowledge OCI as a first mover in this space, as well as acknowledge the reckoning the OCI team needs to undertake.

3. Trainium

While CEO Matt Garman announced Trainium3 at re:Invent 2024, we are seeing its move into production, with some quite compelling performance numbers. For training, the company is claiming a 3x to 4x faster time-to-train (due in part to its new UltraServer). Likewise, for inference, AWS claims a 3x to 4x improvement in QPS (queries per second) and a materially better cost-per-token. (All of this is based on FP8 precision.)

In case you are wondering why I reference inference in a Trainium overview, it’s a good question. Turns out that Trainium is pretty darned good at inference as well. Not surprising, given that training and inference share many computational characteristics. And just as NVIDIA and AMD are eager to showcase the versatility of their high-end GPUs, AWS aims to demonstrate a platform that is relevant across the entire AI lifecycle.

This leaves a lot of questions for Inferentia — questions AWS leadership didn’t directly answer. However, if I were to read between the lines, I would suggest looking for a single platform where some shaping of what we know as Trainium is announced to be fully optimized for both training and inference.

My take on Trainium3: I think AWS has led the CSPs in terms of how it designs, deploys, and positions its custom silicon. It is the standard by which others are now deploying. It is not overly aggressive, but rather positions these solutions as a “choice” for its customers. And before the market even knows it, we realize that this silicon is responsible for more than half of the deployments over the last couple of years.

What I struggle with a bit is how Trainium in particular applies to the enterprise with 2,000 servers and 10,000 employees looking to activate GenAI. Everything I hear AWS say refers to the largest of large organizations. And in fact, both Trainium2 and Trainium3 are fully subscribed.

This is not a criticism of AWS, really. Or not only of AWS. Every tech provider seems to be focused solely on the whales, rather than the overall market volume. And I’m certain this has to do with who is buying today. But if I were a product marketing manager at AWS, I’d start building that connective tissue for the volume market segment. That commercial enterprise CIO is arguably more dependent on AI for success than any other market segment.

4. Graviton5

OK, one of my favorite topics — Graviton. I’ve been covering this chip since it was unveiled and have been quite bullish. Some would say too bullish. But before getting into what’s new in Graviton5, it’s worth seeing how much Graviton has taken hold at AWS.

• There are more than 90,000 Graviton customers, from really small startups to the largest of large enterprise customers.
• Over the last two years, more than 55% of new EC2 CPU capacity has been powered by Graviton.
• This includes over 90% of Redshift Serverless.
• And over 65% of Amazon Aurora.
• And over 50% of DocumentDB.
• And finally, over 40% of the compute powering Amazon Prime Day? Graviton.

With Graviton 5, AWS has become bolder, speaking directly to how the chip competes against its x86 rivals, head-to-head. Not some more nuanced “performance-per-watt” or “performance-per-dollar” kind of measurement or positioning. With 192 cores, 12 DDR5 memory channels, 5x the L3 cache of Graviton 4, and support for PCIe Gen 6, the company can make these comparisons.

This places a significant amount of pressure on the x86 players. Perhaps more on AMD. Why? While both Intel and AMD generate significant revenue from their respective AWS businesses, Intel does so primarily through highly customized Xeon chips — Xeons that have been tightly tailored for specific workload and power-performance requirements. This is not simply some fused cores, but chip customization that requires a separate manufacturing process. Conversely, AMD is selling its volume EPYC chips to AWS, specifically where the AWS team is targeting Graviton.

Either way, Arm has not only arrived in the datacenter; it is thriving. Game on!

Box has announced the general availability of Box Shield Pro. This is an AI-driven extension of its content security and governance capabilities that applies agentic automation to protect files and manage risk across the Box Content Cloud. The offering uses AI to interpret content and context in order to automatically classify and label sensitive data, recommend or enforce sharing controls, and assist security teams with faster investigation and remediation. The stated goal is to minimize friction for end users as they collaborate.

For Box, Shield Pro appears to represent a meaningful step in positioning its platform as an intelligent control plane for data security and governance rather than just a repository for unstructured content. This approach embeds AI directly into core workflows instead of treating it as an add-on. For customers, the move suggests a shift toward more proactive, continuous, and context-aware protection that aims to scale governance without overloading security teams or disrupting everyday work. This strategy ties models to policy, permissions, and content context rather than prioritizing generic productivity features. More broadly, it highlights the focus Box is placing on applying agentic capabilities to specific enterprise problems.

Adobe and Amazon are deepening their collaboration, primarily leveraging AWS infrastructure to accelerate AI-driven creativity and customer experience orchestration. The partnership uses AWS compute resources, specifically EC2 P5 and P6 instances along with S3 and FSx storage, for training and deploying Adobe’s generative AI models such as Adobe Firefly. Integration points extend across multiple core products. Adobe Express incorporates an AI assistant for conversational editing built on AWS, and Adobe Acrobat Studio uses Amazon Bedrock to access various foundation models for PDF-focused AI assistants.

On the marketing technology front, Adobe Experience Platform (AEP) applications are available on AWS. This enables enterprises to centralize data unification and audience activation. New applications, such as GenStudio for Performance Marketing, enable the direct activation of display ads via Amazon Ads. Additionally, the Adobe Real-Time CDP Collaboration provides a secure environment for advertisers and publishers to jointly utilize first-party data with Amazon Ads for media measurement. Both companies have indicated they are exploring the adoption of multi-agent systems using AWS capabilities, including Amazon Bedrock AgentCore.

Google has announced the general availability of Google Workspace Studio, a platform for designing, managing, and sharing AI agents within Workspace. The agents leverage the reasoning and multimodal capabilities of Gemini 3. They are intended to automate everyday work, ranging from basic tasks to complex and end-to-end workflows, without requiring specialized coding skills.

The system aims to enable employees, regardless of their developer status, to create custom agents that address unique business processes. These agents are presented as a flexible alternative to traditional and rule-based automation tools because they are capable of reasoning through problems and adapting to new contexts. Integration with core Workspace applications like Gmail, Drive, and Chat provides agents with work context. This enables them to apply company policies and generate personalized content.

For development, users can start from templates or describe the desired automation in natural language. For more technical requirements, agents are extensible. They connect to enterprise apps like Asana and Salesforce via pre-built actions. They also offer capabilities for custom steps via Apps Script, which enables integration with internal tools and proprietary models via Vertex AI. Customers in the Gemini Alpha program reportedly used these agents for more than 20 million tasks in the past 30 days. The feature is scheduled for rollout to business customers over the next few weeks.

ServiceNow has announced its intent to acquire Veza, an AI‑native identity security platform provider. This move aims to deepen ServiceNow’s security and risk portfolios and strengthen its position as a “system of action” for enterprise security and governance. The deal is designed to address a common control gap in modern environments, which is the fragmented and inconsistent management of identities and permissions across applications, data, cloud services, and AI systems.

Veza’s core technology is its patented Access Graph. This tool maps and analyzes access relationships for human, machine, and AI identities across heterogeneous environments. By normalizing and correlating who or what has access to specific resources and under what conditions, Veza supports identity governance workflows and the enforcement of least‑privilege access. This function is increasingly relevant as enterprises move from traditional automation to agentic AI where autonomous or semi‑autonomous agents make and execute decisions across business and IT systems.

For ServiceNow, the strategic importance appears to be threefold. First, identity and authorization data become a primary signal in ServiceNow’s broader security operations and risk workflows. Integrating Veza’s identity context into products such as Vulnerability Response and Incident Response can help security teams move from static lists of issues to prioritized and contextual actions that reflect actual access risks. Second, as customers deploy AI copilot and agent capabilities on the Now Platform, Veza’s ability to reason over machine and AI identities provides a governance layer to secure AI artifacts within the same platform where workflows run. These artifacts include models, prompts, actions, and data connections. Third, the acquisition advances ServiceNow’s narrative around converged security, IT, and risk. By embedding identity security directly into the operational fabric, ServiceNow may help enterprises unify detection, access governance, and remediation in a single system.

Veza was founded in 2020 and serves approximately 150 enterprise customers. The company brings a specialized capability set into ServiceNow’s portfolio. The transaction is subject to customary regulatory approvals and closing conditions. It suggests an intent by ServiceNow to compete more assertively in identity‑centric security and to differentiate its AI strategy with an emphasis on permissioning, control, and governance rather than only productivity and automation.

Like a number of my colleagues, I was in Las Vegas at AWS re:Invent 2025 last week. AWS did a great job and delivered a full set of announcements across its core data and AI services, and the theme this year was lowering cost while tightening performance. Updates like faster S3 operations, the scale-out design of S3 Vectors, GPU-accelerated vector search in OpenSearch, and incremental database optimizations should all make life easier for teams already building on AWS. The agent stack also expanded with AgentCore, new frontier agents, and the Nova model family supported by Nova Forge, which gives developers more room to customize and operationalize AI inside the AWS environment. These announcements together ought to improve AWS’s capability to handle high-volume AI workloads and the data infrastructure required by modern applications.

The open question I have is whether these stronger components start to add up to something that feels like a unified enterprise data platform. AWS still largely hands customers the pieces and expects them to build the architecture: data governance, semantics, workflow integration, and the connective tissue across ERP, supply chain, CRM, and other operational systems. For companies running complex environments, capability alone isn’t enough; they need these services to work together in a way that reflects how their businesses actually operate.

Read more in my latest LinkedIn article discussing how AWS is improving its capabilities, but enterprises are still waiting to see whether those improvements evolve into a platform that reduces integration overhead and supports more natural cross-system workflows.

Google Cloud’s latest agentic AI updates show how fast the data cloud is moving from an analytics layer to supporting real-time operations. The new agents focus on consuming tasks that usually slow down team productivity, such as data preparation, pipeline design, and repetitive analysis. The goal is to bring those capabilities into BigQuery, AlloyDB AI, and Vertex AI so teams can move from a question to actionable results with less manual work.

What stands out is the broader strategy. Google is treating AI as a built-in layer across its data environment rather than something bolted on. Gemini Enterprise fits into that approach by tying together data, AI, and everyday productivity workflows while integrating with tools like Microsoft 365 and Salesforce.

Governance, oversight, and data quality will decide whether these agents deliver results you can trust. The guardrails matter just as much as the technology. I’m seeing data platforms shift toward helping teams take on more of the daily work, and Google Cloud’s move toward agent-driven workflows is one example of that direction. Read more in my latest Research Note; Google Cloud’s AI Agents and the Next Phase of the Data Cloud.

AMD’s approach to AI infrastructure is decidedly different than NVIDIA’s, focusing on ecosystems and open standards to drive scale and adoption. I really like its continued partnership with HPE for AI rack-scale infrastructure. At HPE Discover Barcelona last week, the two companies announced an industry-first AMD Helios scale-up Ethernet switch.

This product is notable given that it uses a standardized, open-fabric design that avoids vendor lock-in and provides massive bandwidth to support AI workloads. From my perspective, it provides enterprise customers a highly performant and less costly option compared to NVIDIA.

As enterprises head into 2026 and continue to modernize, system integration will remain a focus. Older systems are being phased out, more workloads are shifting to the cloud, and organizations are working to streamline how data moves across their operations. The market outlook reflects this shift, with system integration expected to reach about $600 billion by 2030 — which gives that market a CAGR of 6.1%. That matches what I’m seeing across vendors and customers as they work through cloud adoption, ERP updates, and the ongoing need for dependable data foundations.

A driver behind this is the reality that most enterprises still do not have data ready for AI, as shown, for example, in a report from ERP Today saying that 92% of organizations lack AI-ready data. Many of these organizations also do not have a clear data strategy or ownership model in place. Data quality issues, fragmented sources, and inconsistent governance slow down modernization and make AI programs difficult to scale. This is exactly why integration has become a big deal. Without reliable data foundations and consistent data flow, organizations cannot move past experimentation or unlock the real operational benefits they expect from AI.

These observations line up with how ERP is evolving. ERP is shifting from a system that mostly records past activity to one that supports real-time decisions and actions. Industry 5.0 concepts reinforce this change because they depend on people and technology working side by side, which requires cleaner and more immediate data moving across the enterprise. None of this is possible without a strong integration layer. Organizations in manufacturing, finance, government, and telecom feel this most because their operations rely on multiple systems interacting with minimal friction.

Integration has become a core part of modernization rather than something you address after new systems go live. If data and workflows do not connect, the rest of the transformation rarely works as intended. For organizations planning ERP or supply chain modernization, cloud migrations, or early AI automation, integration is often the determining factor in whether they can move toward more responsive and human-centered operating models that align with the direction of Industry 5.0.

As 2026 approaches, your organization needs to design systems with integration and data readiness in mind. Identify where data actually lives, determine which systems need real-time connections, and address the friction points that slow teams down. The companies that make the most progress are the ones that simplify their architecture early and create a dependable data foundation before scaling AI, automation, or next-generation ERP capabilities.

Stripe is acquiring usage-based billing startup Metronome for roughly $1 billion. This reportedly doubles Metronome’s valuation less than a year after its Series C funding round. This acquisition indicates how critical sophisticated consumption billing has become in the current AI era. The deal positions Stripe to evolve from primarily a payments processor into a broader monetization infrastructure provider. It combines global payment rails with Metronome’s real-time metering and billing engine for usage-heavy businesses such as AI platforms, APIs, and infrastructure services.

For competitors like Chargebee and newer specialists such as Orb, the acquisition may tighten pressure from an integrated payments-plus-billing stack. At the same time, SaaS vendors including Salesforce, Microsoft, and Zoom are bundling AI features into flat or seat-based pricing, which effectively commoditizes basic AI usage. This dynamic leaves complex and high-value agentic workloads as a primary area of competition for advanced and usage-based monetization.

As AI companies scale from research and development into production workloads, and as they monetize through token usage, compute, or API call metering, Stripe aims to position itself as the infrastructure backbone for their financial operations. This is particularly strategic as enterprises evaluate how to charge for AI-powered services.

Meta’s hiring of Apple’s highest-level UI design staff is an indication of both Meta’s realization of its UI challenges across its many platforms and the company’s desire to improve. Meta has for a long time struggled with good UI and cohesion across the apps that it owns. As we march towards an AI-centric world with strong ties to spatial computing, companies are going to have to rethink their UIs and build with AI, much like Google has done with Android XR. I think this is just the beginning.

Antigravity, a new drone company backed by Insta360, has announced the Antigravity A1 — an 8K 360 first-person view (FPV) drone that I believe will massively change the FPV drone market. You can now see virtually everything around the drone in a full 360-degree view, while also seeing it from the drone’s perspective. I haven’t had a chance to try it myself yet, but conceptually it makes a ton of sense and really just depends on a reliable camera feed, low latency, and some flying skills. I do wonder how long it will be before Antigravity draws the attention of the U.S. government, but until then these could prove to be really exciting 360-degree drones.

Quantum Teleportation
Teleportation research is always interesting to me, and there has been plenty of it. In the most recent experiment, scientists were able to teleport the properties of one photon to another photon in a different location. This experiment does have usefulness because it’s what we will eventually need to secure the quantum internet when it becomes available — making it impossible to hack. Until this experiment, entangled photons had to come from the same source. Not so for this one, which makes it more interesting to think about.

So how was it accomplished? The teleportation took place between two quantum dots about 300 meters apart. That means the information was exchanged between devices that were separated by the length of about three football fields. The experiment was far more accurate than I expected it to be, with an accuracy of 82%.

The next experiment will have a greater degree of difficulty. The research team wants to exchange data with multiple quantum communication nodes, which would essentially make it the first practical quantum relay network over long distances. But I’m not going to get excited to see the results yet. This little experiment took three years.

Nobel Prize–Winning Quantum
I’m glad to see more information coming out about Qolab, a company founded by Professor John Martinis, one of the winners of the 2025 Nobel Prize in Physics for his groundbreaking work in quantum.

The Israeli Quantum Computing Center (IQCC) is the first to install a quantum computer designed by Qolab. While Qolab doesn’t have any revolutionary quantum technology, it does represent a leap forward in terms of reliability and scalability; it is designed to be more stable and less prone to errors, and scalable enough to support larger machines. The new Israeli computer will also be accessible through ICCC’s cloud.

The IQCC, located at Tel Aviv University, hosts many types of quantum computers in addition to supercomputers. This approach lets users pick and choose the type of quantum machine — or combination of classical and quantum computers — needed for an experiment. In short, the IQCC appears to have transformed Nobel Prize-winning science into practical use for the community of global researchers.

According to Adobe Analytics data, the 2025 holiday shopping forecast projects that total online expenditures will reach approximately $253.4 billion, representing a 5.3% year-over-year increase. The data suggests the increasing centrality of mobile in the digital commerce landscape, as mobile devices are expected to generate 56.1% of online revenue and account for seven out of every ten retail site visits.

A significant development is the accelerated integration of generative AI into the consumer’s path to purchase. Traffic referred from AI sources, such as LLMs, is projected to surge by more than 500% compared to the previous holiday season. Additionally, the adoption of flexible payment models continues its upward trajectory. Buy now, pay later (BNPL) spending is forecasted to grow 11% year-over-year, reaching a total of $20.2 billion. These metrics collectively indicate that e-commerce optimization for mobile experiences, flexible payment options, and leveraging AI for discovery are key technological imperatives for the season and beyond.

Recently, a group of researchers at UC San Diego and the University of Maryland revealed a startling discovery about a satellite security hack that employs an inexpensive satellite receiver system made with off-the-shelf components — not purpose-built spyware. Following three years of data collection and analysis from a vantage point in La Jolla, California, the team was able to collect a myriad of confidential data from geosynchronous satellites.

As mobile network operators including AT&T, T-Mobile, Vodafone, and others embrace non-terrestrial networks to extend coverage, this study shines a bright light on the need for stronger security provisions. I will be providing my insights in a future research note on the subject.

Micron has announced that it will sunset its consumer DRAM and storage brand Crucial after nearly 30 years of operations. It says it is doing this to better supply its B2B customers with memory and storage, whether for consumer or enterprise applications. With the current price of DRAM skyrocketing by 3x in the last few months, gamers are already upset about the shortage of memory caused by the AI boom. With Crucial leaving the market, there are sure to be even more upset consumers, given that Crucial was one of the better brands in terms of competitive pricing.

I recently reviewed some of Crucial’s latest products, and the company truly had competitive products in both memory and storage, and its T710 SSD was the fastest Gen5 drive on the market. Additionally, the company was attempting to rebuild the Crucial brand around gaming, plus investing in faster and more expensive memory kits — which are now no longer available. I will admit that I’m a little disappointed that Micron has done this, but it’s clearly in favor of Micron’s shareholders, who know that selling DRAM and SSDs for AI is far more profitable than for consumers or gaming.

At AWS re:Invent 2025, what stood out in this area were the updates about how the PGA TOUR and AWS have moved to the use of agentic AI. The focus is on solving long-standing coverage and workflow gaps.

The multi-agent content engine built on Amazon Bedrock AgentCore can generate full articles for every player in the field, giving fans meaningful insights on players who usually get little broadcast attention. The PGA also expanded its AI-supported commentary, an evolution of what launched in TOURCAST at The Players Championship; this means that broadcasters can get near-real-time context and shot analysis without adding overhead.

The “Fan Experience” zone made this more concrete. Real-time birdie probability tools, immersive visuals, and a data-aware simulator showed how data can make the game easier to follow and more engaging without forcing fans to be experts.

What I also noticed is the impact of how AI agents are reducing manual content work and improving consistency, which matters as the PGA TOUR ramps up the volume and pace of digital coverage. The organization is using AWS’s AgentCore, Trainium, and Bedrock to create experiences that are more tailored to each person and based on data, all while managing the complexity of production.

In ongoing legal battles that started with T-Mobile suing AT&T over a deceptive advertising campaign, AT&T has fired back recently with a lawsuit of its own. It accuses T-Mobile of using AI bots to access and scrape AT&T customer data to facilitate the use of an “Easy Switch” onboarding tool that aims to poach AT&T subscribers. T-Mobile claims that it removed the scraping feature prior to its launch on December 1, prompting AT&T customers to provide account information manually. However, it appears that the scraping feature is activated for Verizon customers.

It is an interesting development — one that shines a light on the power of agentic AI applications to automate complex functions. However, T-Mobile could have avoided scrutiny by simply requiring subscriber opt-in for use of the tool.