Moving Beyond AI as Buyer Research Tools
This week Salesforce dropped a number of agentic releases, including making the Slack Bot a more powerful agentic frontend for business processes. (This is similar to what we are seeing with tools like Amazon Quick, which I wrote about last week.) But most intriguing to me were the new Agentforce Commerce capabilities. This is a broad set of selection and purchasing agents for both B2C and B2B interactions. I like the notion of how you can execute a shopping experience within common chat tools like ChatGPT or Gemini. But initially I think the more meaningful impact will be on the B2B side, where companies have richer and often more complex relationships. What also stands out on the B2B side is its headless nature, which means you can execute tasks like opening a PO across multiple channels including WhatsApp, a commerce site, or an AI chat client. Agentforce Commerce also has backend agents to optimize order fulfillment, along with a more unified catalog.

When Traffic Falls 75%, What Do You Do?
One thing that sometimes has been missing from the AI discussion is the dramatic decline of web search. It’s easy enough to see this in action when you Google something: The AI answer is often good enough, so the user moves on. This has been a disaster for companies that used Google Ads and SEO as a primary route to market. Over the past 24 months, search and digital advertising has effectively been rendered moot by AI. So I was very intrigued to read this piece in the New York Times last week about the big media company People Inc. and how it is dealing with this change. (People Inc. publishes well-known periodicals including Entertainment Weekly, Food & Wine, Southern Living, and — no surprise — People.) The TL;DR version is that People Inc. found new ways to engage users, doubling down on expertise and aiming to provide the best experience possible to those who are willing to pay for it. But the reality is that there was a lot of disruption in the process. And despite being disrupted by AI, People Inc. does license content to AI players. While this story was enlightening, there is no happy ending for it yet. The real test will be whether People Inc. can continue to zig and zag as AI continues to improve.

Why AI Pricing Is Starting to Make Me Mad
Over the past three months I have sat through a number of vendor briefings about AI pricing. Last week I attended one that finally pushed me over the edge — enough to do a deeper dive into the topic. Arbitrary pricing has been in the technology industry as long as there has been a technology industry. But AI represents an interesting pricing challenge because it’s been very hard to measure. If you were to charge per-seat pricing for an agent, a vendor’s margins could easily evaporate thanks to a few power users. So that approach is too coarse, while the flipside of measuring token consumption is far too granular. (I dug into some of these issues in my initial research on agentic AI pricing last autumn.) The more current compromise position is this new tranche of action-based or consumption-based pricing, which I feel is flawed. Here’s why:

1)    It’s not standardized. How one vendor classifies an action is not the same compared to another vendor.

2)    It has no benchmark to compare against. For example, in cloud infrastructure one could somewhat easily compare cost in-cloud versus cost on-premises. But what’s the cost of building a dashboard or a report? That is totally subjective.

3)    Where is the shared incentive? If the cost of inference drops every year, why should I pay you the same next year as I do this year?

So, while I appreciate the complexities vendors face, as a customer, I would be hard pressed to sign a long-term enterprise contract until we all have a better idea of the value proposition offered from these new models. It’s a topic that will be getting a lot of my cycles this summer.

Moving Beyond AI as Buyer Research Tools
This week Salesforce dropped a number of agentic releases, including making the Slack Bot a more powerful agentic frontend for business processes. (This is similar to what we are seeing with tools like Amazon Quick, which I wrote about last week.) But most intriguing to me were the new Agentforce Commerce capabilities. This is a broad set of selection and purchasing agents for both B2C and B2B interactions. I like the notion of how you can execute a shopping experience within common chat tools like ChatGPT or Gemini. But initially I think the more meaningful impact will be on the B2B side, where companies have richer and often more complex relationships. What also stands out on the B2B side is its headless nature, which means you can execute tasks like opening a PO across multiple channels including WhatsApp, a commerce site, or an AI chat client. Agentforce Commerce also has backend agents to optimize order fulfillment, along with a more unified catalog.

When Traffic Falls 75%, What Do You Do?
One thing that sometimes has been missing from the AI discussion is the dramatic decline of web search. It’s easy enough to see this in action when you Google something: The AI answer is often good enough, so the user moves on. This has been a disaster for companies that used Google Ads and SEO as a primary route to market. Over the past 24 months, search and digital advertising has effectively been rendered moot by AI. So I was very intrigued to read this piece in the New York Times last week about the big media company People Inc. and how it is dealing with this change. (People Inc. publishes well-known periodicals including Entertainment Weekly, Food & Wine, Southern Living, and — no surprise — People.) The TL;DR version is that People Inc. found new ways to engage users, doubling down on expertise and aiming to provide the best experience possible to those who are willing to pay for it. But the reality is that there was a lot of disruption in the process. And despite being disrupted by AI, People Inc. does license content to AI players. While this story was enlightening, there is no happy ending for it yet. The real test will be whether People Inc. can continue to zig and zag as AI continues to improve.

Why AI Pricing Is Starting to Make Me Mad
Over the past three months I have sat through a number of vendor briefings about AI pricing. Last week I attended one that finally pushed me over the edge — enough to do a deeper dive into the topic. Arbitrary pricing has been in the technology industry as long as there has been a technology industry. But AI represents an interesting pricing challenge because it’s been very hard to measure. If you were to charge per-seat pricing for an agent, a vendor’s margins could easily evaporate thanks to a few power users. So that approach is too coarse, while the flipside of measuring token consumption is far too granular. (I dug into some of these issues in my initial research on agentic AI pricing last autumn.) The more current compromise position is this new tranche of action-based or consumption-based pricing, which I feel is flawed. Here’s why:

1)    It’s not standardized. How one vendor classifies an action is not the same compared to another vendor.

2)    It has no benchmark to compare against. For example, in cloud infrastructure one could somewhat easily compare cost in-cloud versus cost on-premises. But what’s the cost of building a dashboard or a report? That is totally subjective.

3)    Where is the shared incentive? If the cost of inference drops every year, why should I pay you the same next year as I do this year?

So, while I appreciate the complexities vendors face, as a customer, I would be hard pressed to sign a long-term enterprise contract until we all have a better idea of the value proposition offered from these new models. It’s a topic that will be getting a lot of my cycles this summer.

Two areas of agentic adoption came up last week, and both of them reflect governance problems that arise once agents start acting on live data. First, you have to know what an agent did after the fact, and second, you have to control which data agents can access before they run.

Monte Carlo recast itself as an agent trust platform in the spring, and two releases from the company last week put more product behind the branding. Agent Lineage ties a bad agent output back to the upstream data that caused it. It presents both in a single view — the part that’s genuinely hard to copy, because watching agents and tracing data have always lived in separate products. Cost Agent reuses that same lineage to surface waste — the dead tables and oversized warehouses quietly running up a company’s bill — and it runs inside Claude Code and Cursor instead of giving users one more dashboard. Reliability sells to data engineers, but cost gets the budget owner’s attention. The real market pressure comes from the big data platforms, which own the data and are now building the agent runtime on top of it. It’s a crowded race, with the platforms and a stack of observability players reaching for the same thing. Monte Carlo’s bet is that the lineage data it already collects gets it there with less new building.

Governance also has to reach back earlier in the process, before an agent ever runs. CData’s Connect AI Developer Edition, with a new Python SDK and command-line tool, aims to put governed data access in developers’ hands from the start. When developers wire an agent to enterprise data, they tend to grab whatever connector is fastest and bolt on governance later, if at all. An agent is only as trustworthy as the data feeding it, so a connection layer that enforces access rules by default beats relying on every developer to remember them. The question to put to your teams is how the agents you’re building actually reach your systems, and who decided what they’re allowed to see.

Two infrastructure moves stood out last week. One is about keeping data recoverable once agents start acting on it. The other is about what it costs to keep data around at all.

First, Commvault and Microsoft have signed a multi-year partnership that embeds Commvault as a native service inside Azure, built to keep AI workloads resilient. The plain read is that resilience is becoming part of the cloud operating model. This fits where Commvault has been heading, making cyber resilience something you run inside the platform rather than alongside it. This matters even more as agents start acting on data on their own. When an agent can change or delete data at machine speed, recovery can’t remain a manual job performed in a separate tool that someone gets to after the damage is done. It has to be native to where the workload already lives. For regulated industries especially, building that functionality into the cloud they already run lowers the friction of trusting agents with production data.

The second angle — about cost — showed up at pgEdge. Postgres databases keep growing with data that nobody reads and nobody’s allowed to delete, and the bill climbs on data that’s gone quiet. The new ColdFront offering from pgEdge pushes that cold data to cheap object storage but keeps it writable under the same table name, so an application keeps querying one table and never knows the old rows moved. Most cold archives go read-only, which means that reaching back into the data requires a costly round trip. Keeping it open and writable in standard Iceberg is the real differentiator, and it’s a smart edge for a small player to pick instead of chasing the big platforms on breadth. Consistency is the hard part. Once a hot tier and a cold archive both take writes, staying in sync is what is likeliest to break under load. The instinct is right, though: Pick one thing, do it well, and target a cost that keeps growing.

Qualcomm used its investor day to do more than launch a chip. While the headline was Dragonfly, the company’s datacenter CPU family, the larger argument was that inference economics, rather than training scale, will decide who wins the next phase of AI infrastructure. That’s the right framing. Agentic workloads generate far more tokens than the chat-style prompts enterprises started with; Qualcomm pegs the increase at 50x to 100x the volume, and that shift moves the conversation from peak benchmark performance toward efficiency per watt and per dollar.

The portfolio aligns to this position. Dragonfly arrives in three specifically tuned variants of the same IP for the jobs an agentic rack pulls apart: an agentic variant, a general-purpose chip, and a lean head-node for orchestration. More interesting than the silicon are three supporting pieces: the High Bandwidth Compute memory architecture, aimed at the decode phase where the memory wall bites hardest; the $3.9 billion Modular acquisition, which makes models portable across mixed hardware; and the Alphawave interconnect portfolio, already in production.

While Qualcomm was able to report revenue commitments from Meta (Dragonfly) and Azure (HBC-based accelerators), I think customer validation is what’s most important for the company and its portfolio rollout. This vote of confidence should lower the barrier for other hyperscalers and neoclouds looking for silicon partners to address their at-scale inference requirements.

The caution is timing. Most of the marquee silicon Qualcomm announced is slated to ship in 2027 and 2028, so today this roadmap is backed by commitments. If HBC delivers and Modular stays open, Qualcomm could become a credible third pole in inference infrastructure. Either way, the inference market gains a serious new entrant.

I must admit, I didn’t foresee the partnership between OpenAI and Broadcom launching on the same day as Qualcomm’s datacenter play. Nonetheless, OpenAI and Broadcom unveiled Jalapeño on June 24. This is OpenAI’s first custom processor and a clear statement that it intends to own more of the stack beneath its models. The chip is a purpose-built inference ASIC (rather than a dual-purpose chip) that puts OpenAI in more direct competition with NVIDIA for training. This chip is designed around OpenAI’s own understanding of how large language models behave at serving time. The companies say it went from initial design to tape-out in nine months, which they describe as the fastest ASIC development cycle achieved in high-performance silicon, with OpenAI’s models helping accelerate parts of the work.

Interestingly, this chip looks a lot like Google’s TPUv8 chip, with memory residing very close to compute to drive best absolute inference performance. And while no benchmarks have been released, the company is claiming substantially improved performance per watt relative to the most competitive chips on the market.

Is this a sign of times to come? Will all large model builders engage in custom chip development to serve tokens in the most performant and economical way? What does this say about the relationship OpenAI has with the silicon, system, and cloud providers in the AI ecosystem? I think the only certainty is that there is no certainty in this market. And those who think that inference will be served by a few chip vendors are either naive or simply unwilling to see the coming diversity that will populate every datacenter — from clouds to the enterprise.

Data platforms had a busy week across a few live issues: what agents cost to run, which part of the stack they belong in, and how to make an agent good at a single industry. Snowflake, Databricks, and Dataiku each took on the industry question, shipping NVIDIA-backed blueprints for life sciences and manufacturing. The bigger moves were on cost and on where agents live.

Start with pricing. Salesforce’s new Agentforce Help Agent charges a flat $2 per resolution, so you pay when the agent closes out a support issue — and pay nothing when the customer escalates or walks away. Per-resolution billing has been around for a while; the thing that matters in this case is a platform the size of Salesforce adopting it, which changes outcome pricing from a startup experiment into something every enterprise software buyer has to reckon with. It sounds great until you try to forecast it. You’re trading a predictable per-seat line item for a bill that scales with how much the agent does, which is harder to budget and easy to underestimate. The teams that get ahead of this will model agent spend the way they already model cloud consumption — before the first invoice teaches a hard lesson.

EDB’s Postgres AI update pushes agent intelligence, analytics, and governance down into the database itself, all running on infrastructure the customer owns. It answers a question I keep getting from buyers: In an agentic stack, is the front door the operational database or the lakehouse? EDB is betting on the database, and on owning the whole thing rather than renting it. Most AI setups still haul data to wherever the models run, but keeping the work on the data, on hardware you own, is exactly what the big managed platforms struggle to match. The governance layer that makes that bet credible is still in preview, and it’s what a regulated buyer will pick at hardest, so that’s where EDB has to prove its approach first.

The other mark of maturity is where companies are willing to put agents. Google Cloud landed Gemini agents in Nokia’s autonomous-networks software and extended its work with fintech provider Jack Henry on AI-driven security for roughly 7,400 banks and credit unions. Both are still early-stage announcements, but the targets matter, because network operations and bank security are about as unforgiving as enterprise workloads get. Google’s habit of pulling partners onto its stack is steering them into regulated, mission-critical territory, which is where agentic AI has to work to mean anything.

Onsemi to Acquire Synaptics
On June 25, 2026, onsemi announced an agreement to acquire Synaptics in an all-stock deal valued at nearly $7 billion. The financial summary is “A larger company that generated $1.4 billion in free cash flow in 2025, even with revenue falling for a third straight year, buys a smaller company poised for accelerated growth.” Assuming the deal is completed, onsemi gains a growth vector and an edge AI compute platform, while Synaptics gains scale, financial strength, and a high-bandwidth channel into onsemi’s automotive and industrial customers. Synaptics shareholders would own about 12% of the combined company. The deal is expected to close in mid-2027.

The product lines are complementary, not overlapping. onsemi brings power, motor control, and sensing; Synaptics brings Astra edge AI processors, wireless connectivity, touch and human-machine interface controllers, and multimodal sensing. onsemi maps the combined operation to four pillars: power, sensing, connected compute, and control. The company expects these pillars to support a $243 billion addressable market by 2030.

Synaptics’ Astra edge AI processors anchor the acquisition strategy. Synaptics reports more than 35 customers in the robotics design pipeline, including a leading generative AI OEM, a first humanoid robotics design win, and a Google Research partnership that integrates Google’s Coral NPU. The traction is real, but the volume is not there yet. Synaptics expects an initial uptick in the second half of calendar 2026, scaling through 2027. Core IoT drives current growth, up 31% year over year in the quarter ending March 2026.

Meanwhile, onsemi is profitable but shrinking. Revenue fell from $8.3 billion in 2023 to $6.0 billion in 2025, yet the company reached a record 24% free cash flow margin through disciplined cost control. Synaptics is growing, with revenue up 12% to $1.07 billion in fiscal 2025 and double-digit gains continuing through fiscal 2026. The company posted a GAAP net loss but $144 million in non-GAAP net income in fiscal 2025. However, these numbers don’t reveal the strategy.

My take: This acquisition is a physical AI sense-think-act play, and its long-term strategic value is greater than the sum of the parts. The simple story is “Synaptics compute meets onsemi’s sensing and power.” Onsemi’s automotive and industrial chips address functional and safety standards and ship into program lifecycles of 10 to 15 years. Synaptics’ edge AI silicon has shorter development cycles that keep pace with edge AI development. The technologies, processes, and timelines differ, so the immediate focus should be high-level systems integration and interoperability.

The good news is that interoperability at this level is a tractable problem that can deliver benefits quickly, and it’s why I’m bullish. Longer term, deliberately integrating sense, think, and act across robotics, automotive, industrial, energy, and consumer applications — with reach into medical, aerospace, and defense over time — compounds the acquisition’s value-add. The combined company has the portfolio and the cash to ship interoperable products now while building deeply integrated ones for longer-term upside.

Samsara Beyond
Samsara, a leading supplier of physical operations solutions, held its annual Beyond customer conference in Las Vegas last week. I attended in person and came away with many new insights from the presentations and meetings. However, my most significant findings came from informal conversations with Samsara customers. The company went all-out to provide networking opportunities, including hosting the closing event at Allegiant Stadium, so I took advantage of that and spoke with dozens of the 4,000 customers in attendance.

My most significant finding reinforced impressions from past Beyond events. I heard similar stories from truckers, fleet operators, and companies with large supply chains: “Samsara provides effective, single-source solutions that address my real-world problems. And the company listens to me.” Samsara turns a surprisingly large number of users into evangelists.

Here’s why I’m opening with this observation instead of diving straight into the product announcements: Samsara leads with customer outcomes, not technology. The Beyond presentations skipped the vehicle gateways, the cellular and Wi-Fi links, the Bluetooth connectivity, and the cloud platform. I heard no hype about AI strategy. Instead, I heard the results: more than 25 trillion data points collected last year, 99.99% uptime, more than 250,000 accidents prevented, 340 million workflows digitized.

Years of solving customer problems helped Samsara build a practical private network and a solid operations database. This foundation enables Samsara to innovate simultaneously at the edge and in the cloud. Customers wanted to build custom applications faster, without costly consultants. Agent Studio, announced at Beyond this year, lets operations staff do exactly that: Specify a trigger and a prompt describing the request, and the tool builds the agent automatically. One example produced a daily fleet briefing in minutes, with no consultants, contractors, or IT staff involved. Because the agent runs on managed data inside Samsara’s cloud, governance and security apply automatically. Although Claude Code and commercial agentic frameworks are more flexible and better at integrating enterprise data, horizontal tools cannot match Samsara’s built-in identity, data context, operational workflows, governance, and ease of use. Agents are gradually replacing dashboards across all industries, and Agent Studio accelerates this trend.

One other announcement at Beyond stands out — an RF tracking label, positioned to address cargo theft that costs U.S. businesses roughly $35 billion annually. It’s an inexpensive, disposable, printable, Bluetooth-connected sticker that shippers can slap onto any package. It’s like an AirTag, but it connects at surprisingly long range with millions of Samsara gateways in nearby trucks, trailers, vans, and cars. Phones with Samsara apps can also connect. Shippers can see where a package is with no manual scanning, no cellular, and no costly RFID infrastructure. My sample tag tracked my briefcase on my trip home, including points along my Uber rides. This product amplifies the value of the company’s proprietary network, and I predict strong sales.

My take: Some suppliers lead with AI and struggle to find ROI. Samsara leads with customer outcomes and uses AI to deliver them. And most importantly, Samsara knows how to build relationships that create a product flywheel. The telematics and fleet management space is crowded, with strong competitors like Motive and Geotab. However, the product flywheel — customer engagements, devices, network, operations data, and agents — is what sets Samsara apart. Future earnings and ARR numbers will measure the flywheel’s acceleration.

The biggest news of the week was that Apple formally announced the price increases that it warned about the previous week. These increases applied across the board to all Apple products except the iPhone. I suspect that the price changes will come in September during the next iPhone launch. The increases range from 14% up to 54%, depending on the device. For example, the iPad Pro saw a $200 price increase, which is 20% of the original $999 price. The MacBook Air also saw a $200 increase (18%) from $1,099 to $1,299, and the newest MacBook Neo saw a $100 increase, from $599 to $699. These are all significant changes that I believe will have a significant impact on Apple’s sales and level the company’s pricing with its competitors. I also believe it makes the MacBook Neo less attractive to its original customer base and will, overall, push more customers to hold onto devices for longer, or else finance upgrades. There was a lot of blame being put on Apple for these price increases, but I believe the company held off as long as it could and was forced to raise them by the memory suppliers’ continued price increases. Apple has arguably the most sophisticated supply chain on earth, and it has planned for all kinds of contingencies, but memory pricing is such a prolonged and global issue that even Apple couldn’t escape it. To deal with the price hikes, I believe we may see more users opting for cheaper machines and leveraging cloud computing and storage.

There seems to be a trend for both large incumbents and smaller specialists in the quantum industry to announce fault-tolerance as their next commercial milestone. IBM, QuEra, Pasqal, Quantinuum, and D-Wave have all announced plans for fault-tolerant systems sometime within the 2028–2030 window. The entire industry seems to have shifted from raw qubit counts to error-corrected usefulness.

IBM provides the best example of how a company translates research into a formal roadmap. It recently announced that it plans to invest more than $10 billion in quantum computing over the next five years. That is reflected in its roadmap, which includes plans to build a large-scale fault-tolerant quantum computer (FTQC) in 2029, centered on IBM’s Starling quantum processor.

Among smaller companies, QuEra has been one of the most aggressive in announcing its plans for a fault-tolerant machine. Toward that goal, QuEra and AWS recently announced a plan to bring an FTQC to Amazon Braket in 2028. Those plans include QuEra’s Libra system, which is projected to have more than 256 error-corrected logical qubits and about one million reliable operations.

In 2024, Pasqal was one of the first to announce plans for an FTQC. In 2028, it plans to deliver 20 to 128 high-fidelity logical qubits through advanced quantum error correction (QEC) coding.

Quantinuum’s modality is trapped-ion with a QCCD architecture. Its roadmap indicates plans to implement a universal FTQC by around 2030 by scaling its high-fidelity, fully connected physical qubits into reliable logical configurations.

D-Wave recently updated its gate-model roadmap with plans to build a commercial FTQC by 2032. Rather than raw physical qubits, D-Wave prioritizes aggressive hardware-level error reduction using a proprietary superconducting dual-rail qubit architecture.

A few years ago, plans for fault-tolerance were rare. Today, those plans are a baseline requirement if a quantum hardware provider is serious about commercial viability. The quantum industry has become more interesting since it pivoted to an aggressive pursuit of high-fidelity logical qubits. By this point, I believe everyone agrees that commercial utility depends more on fault tolerance and scalable quantum error correction instead of larger arrays of noisy hardware.

Qualcomm announced its new Dragonfly CPU family and roadmap at its investor day and outlined how the company plans to expand into the datacenter and further broaden its TAM. It was also interesting to see the company state how many wafers it buys each year and how many billions of chips it ships from those wafers. I believe that Qualcomm did this because it wanted to show the world (and its potential customers) the scale it can offer, which few companies on earth can match. Qualcomm also does more than 75 chip tape-outs per year, and while that shows Qualcomm’s breadth, the company can also get a leading-edge handset chip from tape-out to launch in eight months. It can also ramp new nodes from 0 to 100,000 wafers in just six months. Qualcomm also touted its ecosystem, which includes 12 partners each across foundries, outsourced semiconductor and test (OSAT) providers, and substrate vendors. Qualcomm also touted 19 memory and 12 contract manufacturing partners, showing an extremely diverse ecosystem. The stock popped 15% before the sector started to take a tumble.

Intel’s Arc G3 finally launched, with early reviews showing that it wins on performance across the board against AMD’s Z2 family of handheld chipsets. This release has also led many to realize that Intel’s drivers are extremely mature now, and that the company has a very competitive gaming solution in this space. I expect we’ll see more OEMs besides MSI and Acer adopt the Arc G3 family and offer better performance and battery life than AMD’s current generation of products. I have an MSI Claw 8 Ex AI in hand for review against the ROG XBOX Ally X and will have my review published shortly to discuss the experience. Overall, though, it’s good that we now have healthy competition in this space. Gamers should benefit from that, even though pricing for virtually all handhelds has gotten out of control due to the global memory, storage, and chip shortage.

Meta launched its own house brand of smart glasses, which offer a new lower entry point of $299 and don’t include any EssilorLuxottica brands such as Oakley and Ray-Ban. I believe that with this move Meta may be looking to expand its design expertise and save money to access a broader customer base. As Meta’s Oakley and Ray-Ban Display glasses start to approach the higher end of consumer pricing, it makes sense for the company to offer new designs that build on Meta’s learnings from previous generations to deliver a more accessible, differentiated product. That said, the company’s new partnership with Kylie Jenner sounds like a way to circumvent glasses brands and use celebrities to build credibility. But even so, style is still very important, and I think Meta understands what it needs to do on the glasses front nowadays.