In preparation for a briefing with Google last week, I took another look at its Antigravity development tooling and found that in the previous two weeks there had been a number of smaller announcements that reasonably count as a 2.1 release. Many of these features are common with other available tools, and adding them is a sign that Google is pointing Antigravity at a wider range of developer personas and enterprise usage. There have been updates to version control, mobile access, and standards-based credentialing, as well as token management and overage caps. Antigravity also now has a full suite of tooling interfaces, which again is not unique; but what is unique is the number of IDE extensions it makes available. Most of the competition provides for VS Code and sometimes JetBrains, but Google is going beyond that to include Visual Studio, Zed, and Xcode. This news ties in well with the recent release of the new Gemini Flash 3.8 model, which remains a low-priced, high-speed alternative to other frontier models that can cost between five and 12 times Flash’s price per million tokens. It’s also notable that, in addition to the new capabilities, Google has enabled Antigravity for use with certain Gemini Enterprise accounts. Last spring, I walked away impressed with the Google Enterprise Agent Platform, and the recent additions and improvements only strengthen the case for using Google for development tasks.

My recent exploration of agent pricing continues to be a great source of dialogue, and one of the predictions I made heading into 2027 was that we would see more aggressive bundles from major vendors. It seems that Salesforce agrees, and in its run-up to Dreamforce it announced a new set of agent-driven bundles. The idea is to simplify pricing and make it easier to select solutions based upon different types of users. It starts with a core set of offerings that apply to all of the editions that include Agentforce, Slack, and Analytics Agents (but not full Tableau), plus data and security capabilities. Then there will be Sales, Service, and industry-specific variants that include customized agents and access to Salesforce applications as needed. The pricing for each of these is tiered for different amounts of usage, and very similar to what you see from the likes of OpenAI and Anthropic. I really like the approach since it is simple and also competitive in the sense that these are specific to user types or industries where Salesforce already has a great deal of data and process leverage. While we have seen some of this from frontier companies, it’s been very limited, and I think Salesforce is on to something. I am also wondering what will be announced at Dreamforce next week if this is one of the warm-up announcements.

On the bundling and development tooling front, I also wanted to mention an interesting angle from IBM’s Bob team. This summer they rolled out a set of services known as Bob Premium, and I think that, as a pattern, this may be quite an intriguing business model for SIs and other services firms. Bob Premium is not a single solution but a set of offerings that are tuned to specific migration offerings. And while we have seen other migration solutions out there, mostly they are tailored toward upgrading and migrating into a specific target such as a cloud vendor. Additionally, more complex migrations require highly specialized agents and context. For example, a Java upgrade and a COBOL migration are quite different. Also, many times these migrations are about more than just the code, and infrastructure and architecture changes are also part of the project. IBM is handling that complexity by keeping each migration toolset and its agents separate to accommodate their unique challenges.

 The business model is also interesting. The additional cost of the Premium offerings is specific to the expertise that is built into the agents themselves. The execution of the migration is paid for with Bob Coins (IBM’s agent-metering credits) from the core offering. This is a critical idea for services engagements. One of the complaints historically associated with big migration projects is that customers have to pay for both the expertise and the execution to get a successful outcome. IBM is decoupling the two. So, expertise can be priced based on demand and complexity, which is where the highest value and margins for services companies sit. While projects of this nature may become smaller, there is still an incentive for companies to innovate and compete for high-end services. And customers should be able to apply the non-differentiated work to an AI coding agent.

In preparation for a briefing with Google last week, I took another look at its Antigravity development tooling and found that in the previous two weeks there had been a number of smaller announcements that reasonably count as a 2.1 release. Many of these features are common with other available tools, and adding them is a sign that Google is pointing Antigravity at a wider range of developer personas and enterprise usage. There have been updates to version control, mobile access, and standards-based credentialing, as well as token management and overage caps. Antigravity also now has a full suite of tooling interfaces, which again is not unique; but what is unique is the number of IDE extensions it makes available. Most of the competition provides for VS Code and sometimes JetBrains, but Google is going beyond that to include Visual Studio, Zed, and Xcode. This news ties in well with the recent release of the new Gemini Flash 3.8 model, which remains a low-priced, high-speed alternative to other frontier models that can cost between five and 12 times Flash’s price per million tokens. It’s also notable that, in addition to the new capabilities, Google has enabled Antigravity for use with certain Gemini Enterprise accounts. Last spring, I walked away impressed with the Google Enterprise Agent Platform, and the recent additions and improvements only strengthen the case for using Google for development tasks.

My recent exploration of agent pricing continues to be a great source of dialogue, and one of the predictions I made heading into 2027 was that we would see more aggressive bundles from major vendors. It seems that Salesforce agrees, and in its run-up to Dreamforce it announced a new set of agent-driven bundles. The idea is to simplify pricing and make it easier to select solutions based upon different types of users. It starts with a core set of offerings that apply to all of the editions that include Agentforce, Slack, and Analytics Agents (but not full Tableau), plus data and security capabilities. Then there will be Sales, Service, and industry-specific variants that include customized agents and access to Salesforce applications as needed. The pricing for each of these is tiered for different amounts of usage, and very similar to what you see from the likes of OpenAI and Anthropic. I really like the approach since it is simple and also competitive in the sense that these are specific to user types or industries where Salesforce already has a great deal of data and process leverage. While we have seen some of this from frontier companies, it’s been very limited, and I think Salesforce is on to something. I am also wondering what will be announced at Dreamforce next week if this is one of the warm-up announcements.

On the bundling and development tooling front, I also wanted to mention an interesting angle from IBM’s Bob team. This summer they rolled out a set of services known as Bob Premium, and I think that, as a pattern, this may be quite an intriguing business model for SIs and other services firms. Bob Premium is not a single solution but a set of offerings that are tuned to specific migration offerings. And while we have seen other migration solutions out there, mostly they are tailored toward upgrading and migrating into a specific target such as a cloud vendor. Additionally, more complex migrations require highly specialized agents and context. For example, a Java upgrade and a COBOL migration are quite different. Also, many times these migrations are about more than just the code, and infrastructure and architecture changes are also part of the project. IBM is handling that complexity by keeping each migration toolset and its agents separate to accommodate their unique challenges.

 The business model is also interesting. The additional cost of the Premium offerings is specific to the expertise that is built into the agents themselves. The execution of the migration is paid for with Bob Coins (IBM’s agent-metering credits) from the core offering. This is a critical idea for services engagements. One of the complaints historically associated with big migration projects is that customers have to pay for both the expertise and the execution to get a successful outcome. IBM is decoupling the two. So, expertise can be priced based on demand and complexity, which is where the highest value and margins for services companies sit. While projects of this nature may become smaller, there is still an incentive for companies to innovate and compete for high-end services. And customers should be able to apply the non-differentiated work to an AI coding agent.

Every agent reaches a system through a connection somebody has to approve. That approval usually lives inside whatever tool the agent was built in. This means that a company running agents in five places is enforcing five different sets of rules, and no single view covers all of them.

I wrote about Bigeye catching cost anomalies a few weeks back. The natural move, I said, was wiring that into something that can block a request. Boomi just put the blocking at the gateway. Its AI Gateway came out of the Lunar.dev acquisition, which closed at the end of July and reached product about five weeks later. It combines two things most teams run separately: a gateway for model calls and a gateway for the connections that agents make into business systems. One policy now decides which systems an agent can reach and how much it can spend getting there. It runs in a public cloud, in a customer’s own cloud, or on-premises. That covers the data residency question regulated buyers ask first. 

An another front, VAST and CrowdStrike built something neither one could ship alone. It matches VAST’s record of who touched which data against CrowdStrike’s threat intelligence. An account pulling training data at three in the morning looks ordinary to a storage system, and ordinary to a security tool. The anomaly only appears when both records are read together. The partnership also puts a Falcon sensor inside the VAST operating system. Falcon Guardian points at data preparation pipelines, where sensitive information tends to land in places nobody intended. Most AI security still watches the model and the endpoint. The data is where the exposure actually sits.

Twenty-eight years is a long time to go without a customer event. Equinix finally held one, and I spent two days in San Francisco for the event and the connected analyst summit. The company’s story has not changed. Neutrality, control, proximity, more metros than almost anyone. The market caught up to it. So now Equinix is working its way up the stack to add even more value. And I’m truly excited for the company.

I wrote in July that Equinix had turned neutral interconnection into a control plane for multi-provider AI. For most of those 28 years that meant renting space, power, and a cross-connect. Fabric One makes connectivity any-to-any and intent-driven, so a customer describes the connection they want instead of provisioning each leg. Inference Exchange puts NVIDIA reference architectures and Together AI’s 200-plus open models inside Equinix metros. Inference should run close to where the data already sits, in a building that does not care whose model you picked. I went deeper in my recent LinkedIn post.

At the conference, I sat through the sustainability session, and the read I left with applies well past Equinix. Sustainability almost never decides a datacenter deal. Buyers care. Demand runs far ahead of supply, and a new site takes roughly 36 months to become a running AI factory. A customer who doesn’t like a site’s water or power numbers can walk away. They just won’t find capacity somewhere else. So, they don’t walk. The one piece of leverage a buyer keeps is refusing to get locked into whoever hands them the capacity. That holds as long as supply stays tight. Equinix is building as fast as everyone else, and every new site hands some of that leverage back. The race to meet this demand is what caps how far any of them ride it.

VMware Explore wrapped in Las Vegas last week, and the announcement list came down to three things: VCF 9.1, the VMware AI Factory, and VMware Private AI Cloud.

VCF 9.1 shipped in May and got its first real field airing here. The fleet lifecycle management functionality moves patching and drift from a per-host job to a per-estate one. Live patching takes the planning exercise out of the maintenance window. Kubernetes comes in the bundle, FinOps tooling adds chargeback, and certification now reaches older server generations and tier-two OEMs.

Memory tiering is the feature I come back to — because it’s a real differentiator. It presents DRAM and NVMe as a single pool inside ESX, keeping hot pages in DRAM and pushing cold ones out to flash without the guest ever knowing. Why does this matter? Because servers run out of memory long before they run out of cores.

Tiering raises the effective capacity of each host and pushes out a purchase you probably can’t get delivered on your timeline anyway. This is both a cost lever and a supply lever at once, which is rare. This is also differentiated — VMware’s two main competitors don’t have a similar story.

So you read this and think, “What does this have to do with AI? Isn’t everything about AI?” You would be right — none of these are AI features. But they are all prerequisites for AI.

AI Factory and Private AI Cloud sit on top of VCF 9.1. A unified model gallery provides IT and data science with a single interface for deploying inference and RAG across VMs, containers, and pooled GPUs, with claimed support for more than 150 open-source and commercial models. AI Guardrails inspect requests in both directions with PII masking. Budgets attach per business unit. And AgentMinder gives every agent an identity and authorizes its actions against declared intent. Finally, through integrated FinOps, token consumption rolls up per team, visible to finance.

Most of what Private AI Factory provides addresses an operating-model problem rather than an engineering one. Who owns what, who approves it, who pays. VMware’s pitch is that enterprise IT already has the model and the team, so it doesn’t need to hire GPU scheduling and inference runtime expertise. It’s effectively an org-chart argument dressed as a technology argument. And it’s a good one. But it’s also a pitch that has to be made to folks who aren’t traditionally using VCF or VVF day in and day out.

One area where I still struggle: Broadcom’s survey data puts production inference at 56% private and 41% public. The numbers say “hybrid” for a company that has been maniacally focused on private cloud.

The AI Gateway is the answer to this tension, and it’s due to ship this fall. But until it governs calls to hosted model APIs, not just access to local ones, observability and policy stop at the VCF boundary.

Once this ships, the story gets a lot stronger, as enterprise IT can enable and manage inference at scale without hiring an army of data science professionals. 

One final note on this. I’d really like to see VMware embrace the AI ecosystem more … enthusiastically? The enterprise AI environment is going to be a collection of a lot of technology. While these organizations are more conservative in their adoption of new technology, AI is going to force the enterprise outside its comfort zone — and VMware has the opportunity to be that bridge.

ClickHouse acquired RunReveal, a security data platform that was already built on ClickHouse. I liked ClickHouse’s Langfuse acquisition earlier this year, and this is the same play. A database company is buying the applications that grew on top of it. RunReveal keeps its bring-your-own-database model and customers keep their contract terms. ClickHouse also went out of its way to say that it will stay a neutral foundation for the other security vendors building there. To be fair, it has to say that. If you’re building a product, you pick your database early and you live with it for years. Plenty of companies built their products on ClickHouse. Langfuse and RunReveal were two of them, and ClickHouse now owns both. So, anyone starting out like that today has to wonder whether they’re building a company or building something that ClickHouse will pick up later.

Qdrant just published a new benchmark for vector search, and it dwarfs what the category has been running on — ten billion vectors, where most published results use a few million. I spent years doing performance engineering and benchmark work, so the caveat is automatic: A benchmark a vendor builds serves that vendor, and most people reading one already know it. That doesn’t mean the scale should go unnoticed. At a few million vectors, everything fits in memory and everything looks fast. At 10 billion, everything doesn’t fit, and you start trading off what stays in memory, what moves to disk, and how much accuracy you give up to keep queries quick. None of that shows up in a small test. Feed a company’s documents into an AI system and you hit billions of vectors quickly. Most teams pick a database well before that, using numbers that never described the size they end up running.

AI for Physical Assets — A Pattern Emerges (John Deere, Samsara)
Last week John Deere announced JD, an AI assistant that answers plain-language questions about farm operations. When is the optimal time to harvest based on historical trends? Which sprayer operator covered the most acres per hour? Answering those used to mean digging through years of reports. JD runs inside Operations Center, the cloud service that holds more than a decade of a farmer’s field, machine, and operational records. But here’s the interesting part. The company announced the Farmer Data Commitment in parallel — 10 principles covering data control, transparency, and choice. Knowing that customers won’t use JD unless they trust it with their most sensitive farm data, the company wisely got out in front of this issue on day one. Other companies deploying AI should take note: no trust, no data. My full Field Note is here.

Samsara reported Q2 FY2027 results on September 3. Annual recurring revenue (ARR) reached $2.125 billion, up 30% year over year, the third consecutive quarter at that rate. Large enterprises continue to standardize on the platform: 242 new customers above $100,000 in ARR and 20 above $1 million, both quarterly records, with the $1 million-plus cohort now past $500 million in ARR and growing more than 50% for the third straight quarter. Annual data collection passed 30 trillion data points, up more than 40%. CEO Sanjit Biswas said adoption of the company’s AI features is up more than 4x in the past two months.

My take: These two companies are running very similar plays. Both host new AI assistants within cloud services that customers already use, where identity, authentication, and access control are established and a decade of operational history is already in place. This gives companies like these two a huge advantage in generating ROI with new AI capabilities. There’s nothing new to deploy, secure, or learn, and no AI experts to hire. That’s an enviable position for any company trying to offer new AI features.

Turning to Samsara’s numbers, what makes the quarter significant isn’t the absolute growth rate but the fact that it hasn’t decayed. Growth typically moderates as the base grows, because the pool of untouched large accounts shrinks. Samsara’s ARR grew 30% for the third consecutive quarter, now above $2 billion, while the company set records for large-customer additions. Retention around 115% says existing customers spend about 15% more than they did a year ago, but that is table stakes for a platform like this. The signal is that Samsara is still landing new enterprises at the rate it did when it was half this size. Physical operations digitization is just getting started.

The 4x jump in AI feature adoption covers two months, so it’s early days. It follows the launch of Agent Studio, which I covered in detail, and it tells you that customers who already trust a platform will try, and often buy, software that extends it. Samsara and Deere start from a natural advantage: The operational context is already there, sitting in the cloud, with identity and security baked in. And few farmers or logistics companies have a staff of AI and IT experts to apply horizontal ISV solutions. But the clock is ticking. Competition in agentic AI is fierce, and a head start is only worth what a company does with it.

UK National Cyber Security Centre Advisory
The UK National Cyber Security Centre published an advisory on August 27 warning of increased targeting of operational technology (OT) systems across multiple sectors globally by a range of threat actors, with some limited real-world disruption. The advisory warns organizations not to assume OT systems are isolated from the Internet, and to check for inadvertent exposure through misconfigurations, connections with legacy systems, rogue networks, or other unexpected conditions. It points back to the eight Secure Connectivity Principles for Operational Technology published earlier this year by NCSC-UK with CISA, the FBI, and agencies in Australia, Canada, Germany, the Netherlands, and New Zealand. Principle 3 is to centralize and standardize network connections.

Separately, CISA’s July bulletin on water utility attacks noted that exposed operational technology may include undocumented cellular modems installed by operators, vendors, or system integrators.

My take: In my recent Forbes article, I wrote that the water utility attacks had no throat to choke because manufacturer guidance addressed plant operators, not the party that designed the network connections and installed the equipment. Although the advisories are beginning to catch up to the actual problem, I still don’t see a viable resolution. CISA names three parties that could have installed the modem but doesn’t assign responsibility. NCSC tells operators to find unknown connections, a task that points right back to the installers. A cellular modem behind a control cabinet has no relationship to the network diagram, which is why “verify you are not exposed” is harder advice than it sounds. Principle 3 is the right answer, but it’s costly and not typically in a utility’s budget. “Centralizing connections” requires an owner for every link into the plant, which these utilities don’t have and can’t afford.

Lenovo had a very busy IFA, announcing availability for its Qira AI on 16GB systems, Motorola devices with Android 17, and additional software integrations. The company showed off a ThinkBook concept called Project Swan with a rollable display that expands horizontally. It also detailed its new RTX Spark designs for the Yoga Pro 9n and Yoga 9n 2-in-1 laptops. Lenovo also announced the ThinkCentre X Ultra, which uses AMD’s latest Ryzen AI Max+ Pro 495, supports up to 128GB of unified memory, and allows configurations of up to four units together for local AI compute. Lenovo also announced its own Apple MacBook Neo competitor, the IdeaPad Vibe, which comes in 14- and 15-inch sizes and seven colors, as well as Snapdragon configurations and matching keycaps and mice.

NVIDIA announced a multitude of products, including DLSS 5 with NBA 2K27. This was a highly anticipated release, which came right on the heels of Gamescom. In addition to DLSS 5, NVIDIA also had some local AI announcements, such as NVIDIA PAIR (Personal AI Router), which enables users to build a network of devices for local AI compute. The PAIR software takes advantage of any NVIDIA compute on the network, whether Linux or Windows, as well as any Apple Macs. I think this is a recognition by NVIDIA that a lot of its Spark and other hardware is leveraged by Mac users as additional compute for AI, and allowing both to coexist on PAIR makes a lot of sense.

Dell had a few announcements at IFA as well, with the Dell 14S being the company’s newest mainstream laptop. This laptop leverages Intel’s new Core Series 3 processors, code-named Wildcat Lake, giving it strong performance and battery life for the price. It is designed to complement the fairly recently announced XPS 13, slotting in just below it. Because this is targeted at a younger audience, the company is also offering it in four colors. The Dell 14S is also one of Dell’s ways of offering even more MacBook Neo competitors, with the XPS 13 at the high end of those offerings. Dell also announced a new gaming display, the 32-inch AW3226Q, which uses the latest RGB Stripe Tandem OLED technology to improve overall image quality and text sharpness. The new 25-inch AW2527HX is Alienware’s fastest OLED monitor to date with a 560 Hz refresh rate on a QD-OLED display, offering better response times, image quality, and refresh rate than traditional high-refresh TN-based gaming monitors.

Microsoft announced Project Zenith in partnership with AMD and on Ryzen Halo. This targets systems with 64GB or more of unified memory and 250 GB/s or more of memory bandwidth. It is designed to be a more bare-bones, higher-performance version of Windows for local AI workflows, and it aims to ensure that Windows stays relevant in the AI compute era as Linux and Mac gain traction in AI. This is still about running models with 30 billion or more parameters locally, but with more optimizations from Microsoft for better out-of-the-box coding performance. That’s why AMD’s Ryzen AI Halo platform makes so much sense, and I will absolutely be giving this a try on the Ryzen AI Halo machine I am testing. It should increase the value of the platform as a competitor to NVIDIA’s DGX Spark platform.

The American Quantum Competitiveness Act
This bill (H.R. 10163) is another U.S. government action designed to protect and advance U.S. leadership in quantum technology by shielding its commercial supply chain from foreign adversaries. The Act designates the Secretary of Commerce as the lead advisor for scaling, manufacturing, and deploying quantum systems.

The legislation identifies trusted supply chains and suppliers and codifies standards that require quantum hardware, software, and raw materials to originate domestically or from trusted allies (NATO, EU, and OECD nations). Suppliers affiliated with foreign adversaries (such as China, Russia, Iran, and North Korea) are excluded. For stack-wide vulnerability mapping, it requires comprehensive assessments across the entire quantum hardware and software ecosystem. That includes such things as dilution refrigerators, laser subassemblies, control electronics, rare isotopes, and EDA software.

Defining On allied and and public-private partnerships partnerships, it authorizes cooperative agreements with private industry, universities, and international partners with the aim of building domestic manufacturing capacity and spurring private investment. And congressional For congressional strategy reporting reporting, it requires biennial assessments of global quantum competition while helping safeguard proprietary trade secrets and confidential business data submitted by private firms.

Issues and Consequences for U.S. Quantum Companies

1. Decoupling Pressures and Supply Bottlenecks — While the act is intended to harden supply chains, quantum computing relies heavily on global secondary and tertiary inputs, such as refined minerals, specialized optics, and cryogenic wiring. Forcing a rapid pivot away from non-allied sources would pressure hardware developers (e.g., IBM, IonQ, Rigetti) to re-engineer their bills of materials. This risks driving up unit costs and extending lead times until allied production scales.

2. Federal Procurement Advantage — Companies that proactively comply with “trusted supplier” definitions would gain preferred access to future federal, defense, and allied procurement pipelines. The act also positions domestic vendors favorably within multilateral trade frameworks like AUKUS Pillar II.

3. Risk of Architectural Favoritism — The Department of Commerce’s task of designating “critical components” may inadvertently favor mature qubit platforms (e.g., superconducting and trapped-ion systems) over emerging modalities (e.g., neutral-atom, photonic, or topological qubits), potentially skewing private venture capital and public resources toward incumbents.

4. The Gap Between Strategy and Direct Capital — Unlike the CHIPS and Science Act, H.R. 10163 functions primarily as a strategic coordination and supply-chain monitoring framework rather than a multi-billion-dollar direct subsidy vehicle. Without accompanying capital infusions, the costly burden of onshoring specialized fabrication and hardware packaging may disproportionately strain smaller startups.