Forbes: Sizing Up the First Generation of Enterprise Agentic Assistants
Last week I published a two-part Forbes series on desktop agentic assistants — the “Claude-ification” of knowledge work now reaching non-coders. Part 1 draws on six months of hands-on use to show where these tools deliver real productivity gains and where they still break down for non-technical users. Part 2 asks whether today’s tools are good enough to close that gap. Microsoft and AWS have credible entries, alongside Slack, SAP, and Zoom, but measured against what enterprises actually need — centralized provisioning, real memory and context, model flexibility, data protection, agent mobility, forecastable pricing, governance with teeth, and cross-vendor interoperability — this is still a first generation.

I would suggest piloting these assistants now on a narrow, well-scoped use case rather than committing to a single vendor’s platform. Lock-in risk is real until governance and interoperability catch up, and I expect the gap between vendors to widen over the next two to three quarters.

Field Note: Forward Deployed Engineering Is Fashion, Not Palantir’s Playbook
My latest field note is live. Microsoft just committed $2.5 billion and 6,000 people to its new Frontier Company unit, AWS put $1 billion into its own FDE org weeks earlier, and OpenAI and Anthropic have made similar moves — all borrowing the FDE label that Palantir spent a decade building. My take: The investment is worth making, but for most vendors launching an FDE program right now is as much about fashion as revenue. I break the market into three models — program-driven, consulting-driven, and hybrid — each with a different value and risk profile. I also note that no vendor, including Palantir today, is replicating Palantir’s original bespoke-build approach.

Consider this over the next few months: Treat any vendor-offered FDE engagement like any other consulting contract. Verify the team’s business-domain depth, not just its engineering talent, get explicit about which model you’re being sold, and assume vendor marketing support is part of the price of admission.

Google Cloud Moves AlphaEvolve to GA
Google Cloud has moved AlphaEvolve to general availability for all customers via the Gemini Enterprise Agent Platform, graduating it from its December 2025 private preview. AlphaEvolve is a Gemini-powered “evolutionary collaborator”: Feed it a baseline algorithm and an optimization goal, and it searches across candidate variations to return improved, human-readable code, with early uses in chip design, logistics routing, and medical research.

This tracks with what I flagged in my CoreWeave ARIA Field Note a few weeks ago: Agents are shifting from linear, one-shot execution to looping architectures that evaluate many candidate runs and refine the objective itself, not just the output. AlphaEvolve applies the same bet to algorithm optimization, and its GA launch is the clearest signal yet that hyperscalers see iterate-then-converge as a durable pattern rather than a research curiosity.

If your engineering or R&D team has a well-defined, compute-tolerant optimization problem, I would suggest a scoped evaluation now — but budget for the cost of iteration and set a clear “good enough” stopping point up front, since looping agents will happily burn unlimited runs chasing marginal gains otherwise.

Forbes: Sizing Up the First Generation of Enterprise Agentic Assistants
Last week I published a two-part Forbes series on desktop agentic assistants — the “Claude-ification” of knowledge work now reaching non-coders. Part 1 draws on six months of hands-on use to show where these tools deliver real productivity gains and where they still break down for non-technical users. Part 2 asks whether today’s tools are good enough to close that gap. Microsoft and AWS have credible entries, alongside Slack, SAP, and Zoom, but measured against what enterprises actually need — centralized provisioning, real memory and context, model flexibility, data protection, agent mobility, forecastable pricing, governance with teeth, and cross-vendor interoperability — this is still a first generation.

I would suggest piloting these assistants now on a narrow, well-scoped use case rather than committing to a single vendor’s platform. Lock-in risk is real until governance and interoperability catch up, and I expect the gap between vendors to widen over the next two to three quarters.

Field Note: Forward Deployed Engineering Is Fashion, Not Palantir’s Playbook
My latest field note is live. Microsoft just committed $2.5 billion and 6,000 people to its new Frontier Company unit, AWS put $1 billion into its own FDE org weeks earlier, and OpenAI and Anthropic have made similar moves — all borrowing the FDE label that Palantir spent a decade building. My take: The investment is worth making, but for most vendors launching an FDE program right now is as much about fashion as revenue. I break the market into three models — program-driven, consulting-driven, and hybrid — each with a different value and risk profile. I also note that no vendor, including Palantir today, is replicating Palantir’s original bespoke-build approach.

Consider this over the next few months: Treat any vendor-offered FDE engagement like any other consulting contract. Verify the team’s business-domain depth, not just its engineering talent, get explicit about which model you’re being sold, and assume vendor marketing support is part of the price of admission.

Google Cloud Moves AlphaEvolve to GA
Google Cloud has moved AlphaEvolve to general availability for all customers via the Gemini Enterprise Agent Platform, graduating it from its December 2025 private preview. AlphaEvolve is a Gemini-powered “evolutionary collaborator”: Feed it a baseline algorithm and an optimization goal, and it searches across candidate variations to return improved, human-readable code, with early uses in chip design, logistics routing, and medical research.

This tracks with what I flagged in my CoreWeave ARIA Field Note a few weeks ago: Agents are shifting from linear, one-shot execution to looping architectures that evaluate many candidate runs and refine the objective itself, not just the output. AlphaEvolve applies the same bet to algorithm optimization, and its GA launch is the clearest signal yet that hyperscalers see iterate-then-converge as a durable pattern rather than a research curiosity.

If your engineering or R&D team has a well-defined, compute-tolerant optimization problem, I would suggest a scoped evaluation now — but budget for the cost of iteration and set a clear “good enough” stopping point up front, since looping agents will happily burn unlimited runs chasing marginal gains otherwise.

I have been following the AI Safety Index for some time. Published semiannually by the U.S. think tank Future of Life Institute, it grades how the world’s nine leading AI companies manage AI risks.

I am particularly interested in the report’s section on Existential Safety rankings. There seems to be a general consensus that full AGI is a near-term expectation with only a few minor remaining hurdles, while AI superintelligence is decades away. Yet if superintelligence is possible, that mere possibility represents an existential threat to humanity. Many experts believe that once superintelligence comes into existence, it cannot be shut down. My upcoming Forbes article will provide in-depth information on this problem.

The latest issue of the AI Safety Index assigns an Existential Safety grade to each of the nine frontier companies, and all of them ranked poorly. Saying that Anthropic had the highest grade means very little when it only achieved a D+ and the remaining eight companies performed even worse.

There is a clear reason for these low scores: Major frontier companies like Anthropic, OpenAI, and DeepMind no longer halt model development when dangerous signs emerge. This presents a severe quandary for the U.S. government. Restricting free-world AI development could allow China and other adversaries to gain a military AI advantage over the United States. Consequently, the only viable option, I believe, is for the government to conduct advanced AI research under highly secure conditions, without making those advancements available to the public.

Almost every enterprise AI stack is built on open source, including the models, libraries, and components teams pull in without ever fully checking where they came from. Little of that code gets vetted, and the risk climbs once agents start acting on it. To address this, IBM and Red Hat have expanded Project Lightwell, the open-source security clearinghouse they stood up this spring, with commercial offerings and a security-services wrap to help companies operationalize it. It puts a layer of provenance and vetting under the open-source code that’s feeding enterprise AI.

This is the move IBM keeps making, taking something open and free and selling the trusted, governed version on top, framed as an industry clearinghouse the ecosystem plugs into rather than a head-to-head competitor with the usual security tools. Regardless, this latest development highlights that the open-source supply chain has been a security footnote for years. Wire it into agents that act on their own, and it becomes a governance problem with a budget attached.

At the RAISE Summit in Paris, DDN shipped Infinia 2.4, built for running AI inference in production without the bill getting away from you. The bulk of it is keeping the KV cache and model files close to the GPU so expensive chips spend less time idle, plus the tenant isolation, identity, and governance that sovereign and multi-tenant AI setups live or die on. None of that is new for DDN. It’s the same two things it’s been strongest at, shipped in a point-release. And it’s already in production in frontier labs and sovereign clouds like NVIDIA, xAI, Mistral, and SK Telecom.

An agent that troubleshoots or reroutes network traffic is only as good as its read of the network underneath it. Infoblox is buying Kentik to sharpen that read. Infoblox already holds the authoritative record of the network — the DNS, DHCP, and address data every connection runs through — and Kentik adds the live view of what’s actually moving across it. Together they make a real-time map an agent can query straight through an MCP server, instead of a static inventory that goes stale the moment it’s written. For Infoblox, it’s a step from network plumbing toward AI infrastructure, taking the map that humans have leaned on for years and opening it up for agents to act on.

IBM made news last week with the launch of its rack-mountable z17 and LinuxONE servers. While the company rightfully touted the ability to better integrate Z into the distributed datacenter, there’s a bigger story. As AI invades the enterprise, IT leaders are paying less attention to siloes — or islands of data and compute platforms. Rather, they are looking at creating universal farms of compute and data. Think in terms of a single view and single representation of data that can feed AI pipelines. And while not all mission-critical data sits on a mainframe (Z), all data on a mainframe is mission-critical.

I write this because I believe IBM’s opportunity is to further leverage the unique value of Z (reliability, performance, security) in this increasingly heterogeneous AI environment where there is less of a focus on architectural consistency (x86) and more focus on enabling AI-powered operations. If I were a marketing executive at IBM, I would be leaning into the important roles that Z, LinuxONE, and Power play in the AI equation — and the tooling and management that makes deploying and managing these systems simple. Further, I would invest the necessary resources to dismantle the misconceptions about compatibility and utility.

But I’m just an analyst. 🙂

Is datacenter power consumption out of control? And is it interesting to anybody else that this power consumption question has kind of taken a back seat to the supply chain constraints and the explosion of GPUs and accelerators?

Consider these data points:

• Ireland’s Central Statistics Office has shown that datacenter power consumption accounted for almost 25% of the country’s total power consumption in 2025. The 7,663 GWh consumed represents a 10% increase over 2024, while all other sources of consumption increased just 2%.
• In the U.S., datacenter consumption was upwards of 220 TWh in 2025 (5.3% of available power), up from 183 TWh in 2024. Of this, 26% was consumed by the almost 600 datacenters in the Northern Virginia area.
• China’s datacenters accounted for 166 TWh of consumption in 2024 (about 1.7% of available power in that country), with a projected 300 TWh to 600 TWh in 2030.
• Overall, the 415 TWh of global datacenter power consumed represents about 1.5% of total availability. This number is expected to balloon to well over 5% by 2030.

It is quite interesting to me that this narrative seems to have quieted considerably as the AI wave gained momentum. Here are the questions that stand out for me: Do we have a sustainability crisis? Are we facing an existential threat of climate change?

I have neither a scientific, ideological, nor political position on this topic. However, the obvious and indisputable fact is that datacenter power consumption will continue to grow for the foreseeable future, driven largely by the accelerators and connectivity that make AI useful. And this is a big challenge for enterprise customers with fixed power budgets who still need to leverage AI’s capabilities to remain competitive.

Does this lead to a different kind of hybrid normalization? Will we see more legacy applications moved to the cloud to enable larger on-prem AI infrastructure estates? And what about data locality?

I think there are a lot of vendors, pundits, and, yes, analysts making predictions about the future of the AI-powered enterprise datacenter that sound interesting, but who have not fully considered so many of these factors. In fact, if you listens to these vendors, you could be led to believe that there is no real power constraint — that deploying rack-scale inference clusters is quite simple and, because they are liquid-cooled, the answer to all of the needs of an enterprise IT organization.

I would caution enterprise IT leaders to be a little more thoughtful. When considering a long-term AI strategy, start with the fundamentals. Obviously, the needs of the organization are the north star. But how those needs are met — on-prem, at the edge, in the cloud — has to be determined. And what infrastructure maps to enablement is the next logical step.

I am involved in too many conversations where commercial enterprise IT leaders are scratching their collective heads trying to rationalize what they are seeing from hyperscale deployments to what can be supported in their 5MW datacenter. The solutions vendor that can start speaking to this problem (and offering solutions) will be well served.

Every vendor in data and AI is selling “context” right now, and my latest Forbes piece digs into why few of them mean the same thing by it. Context is what a system pulls together the moment it acts, and it’s what keeps a confident answer by the AI from being wrong. That’s confident-but-wrong problem is a nuisance when you’re drafting text, but it means real money once an agent issues a refund or updates a forecast based on data that’s missing one right piece. The money backs the talk. IBM paid roughly $11 billion for Confluent, and Salesforce spent about $8 billion for Informatica, in both cases so they could own the data layer under their agents. But vendors in storage, databases, BI, observability, governance, and apps can all claim context and all be telling the truth, because each is solving a different problem.

The field splits into camps, all waving that same word “context” around. One defines meaning up front through the semantic layer and ontology. One bets on retrieval and lets the model grab what looks relevant — which is quick to stand up but gets shaky past a single hop. Storage vendors stake it on metadata and moving bytes fast, while the governance crowd treats context you can’t vouch for as a liability. The data platforms start with an advantage, since Snowflake, Databricks, Oracle, and the engines inside Google, Microsoft, and AWS build context on data they already hold. The one question I’d carry into every pitch is whether you still own the meaning underneath, regardless of who builds it. Let every vendor assemble its own version, and you’re back to the competing versions of the truth that killed trust in dashboards a decade ago. And past a point, more context makes the answer worse, so whoever assembles the right context cheaply and the same way every time could take this market.

Sony has officially launched the new point-and-shoot RX10 single-lens digital camera. It has a zoom lens ranging from 24mm to 600mm and is commonly regarded as the company’s “high-zoom” camera given its 24x magnification. This camera allows you to reach absurd levels of zoom without ever needing to swap out a lens; it is also extremely compact, so it won’t get you stopped from entering a sporting event, for instance, because it doesn’t need an interchangeable lens. The camera also has an average 20 MP sensor — still considerably better than any smartphone, but nowhere near as good as Sony’s own Alpha series cameras.

Sony calls this $2,300 camera a “Supertelephoto.” Besides its high-end Zeiss f/2.4-4.0 aperture lens, it also benefits from many of the Sony Alpha autofocus advancements, which might translate really well into a mid-size body using a very compact lens. Considering the $2,300 price, you are really paying for the compactness of the complete camera and lens assembly, which you can take into places you otherwise couldn’t with a typical Sony Alpha. Given the pricetag and the performance, I would classify this camera as a prosumer device.

Qualcomm shipped Linux 2.0 for Dragonwing IoT platforms on June 30. With the old Linux 1.x version, Qualcomm’s hardware-specific enhancements required kernel modifications, so the OS split into two branches — a fully open-source base build and a proprietary Qualcomm build — each with its own kernel source, device tree, and user space. Linux 2.0 eliminates that fork with a single kernel source and rootfs. Camera, GPU, video, AI acceleration, and other Qualcomm enhancements now ship as overlays, installing on top of the Linux 2.0 base without modifying the kernel. Developers can switch between the fully open image and the value-add image at runtime, with no reflash required. Proprietary extensions no longer need a separately maintained platform, and updates to the base distribution reach every downstream product without a rebuild.

Qualcomm calls this “upstream first,” meaning that the development branch tracks kernel.org release candidates directly, rather than forking a new version and accumulating vendor patches that drift further from mainline with each change. Overlays isolate vendor code from the shared base, so shipping distributions are a step closer to kernel.org than the old fork-and-hack model allowed. Note: This description relies on Qualcomm’s own developer blog and release documentation. We have not independently verified Qualcomm’s upstreaming claims.

My take: A common open-source base distribution with proprietary extensions layered on top is a step forward for embedded Linux systems. But Linux productization also requires long-term maintenance with over-the-air updates (A/B and incremental). Qualcomm acquired Foundries.io in 2024 to provide these services, so I’ll be watching to see how the pieces fit together.

BMW deployed Figure AI’s Figure 03 humanoid robots in its Spartanburg plant beginning June 30. The previous model, Figure 02, spent roughly ten months loading sheet metal to manufacture more than 30,000 BMW X3 vehicles. This task is fixed and repeatable, with parts presented the same way every cycle. Figure 03’s assignment is much more challenging — sorting bins of parts. Components arrive unsorted and inconsistently oriented. The robot has to recognize each one and sort it into the right order for delivery to assembly stations, grasp parts with both hands while pulling a loaded cart, and adjust its footing and balance as it goes.

My take: Fixed pick-and-place follows rigid scripts, but sequencing demands real-time adaptation that scripted automation cannot handle. That’s the big difference between Figure 02 and Figure 03, and BMW is testing this in its highest-volume U.S. plant.

SpaceX has filed a plan with the FCC that would allow it to launch up to 100,000 of its upcoming Gen3 satellites. This would represent a 10x increase over the 10,000 satellites it currently has in orbit, and an 8x increase in what it is currently permitted to operate (15,000 satellites). This new approach is said by SpaceX to potentially offer multi-gigabit connectivity to its customers and would likely also power many different projects that SpaceX and SpaceXAI are trying to build, including orbital datacenters. Given the size and scale of what SpaceX does and how it wants to compete with everyone in the telecom space, I think the vastly increased satellite fleet would potentially enable it to compete with even more terrestrial services. That said, I also think that if SpaceX wants to become its own operator that competes with AT&T, T-Mobile, and Verizon, it will still need a strong terrestrial backbone. That might mean buying an operator — a move that has been rumored many times over.

China’s first landing of a recoverable booster could change how SpaceX is seen around the world. Recoverable boosters are one of the major ways that SpaceX is able to reduce the cost of sending payloads to space. If China now has this ability, it could enable the country to be more competitive in launching satellites for things like direct-to-device (D2D) connectivity — and, in general, to be able to launch more satellites more frequently, as SpaceX does. China has yet to actually re-use this recoverable rocket, but being able to launch and then land the booster back on a ship in the ocean is a major step in that direction.

T-Mobile had an executive shuffle with the hiring of former AT&T executive Chris Sambar. He was brought on to help run T-Mobile’s enterprise business, an area where he has extensive experience given that he helped establish AT&T’s enterprise business, which is among the strongest in the industry today. Additionally, he was AT&T’s representative on AST SpaceMobile’s board of directors and a huge proponent of satellite communications at the company. I believe that he could be a great hire for T-Mobile, although his addition creates a bit of a shuffle within T-Mobile, with Mike Katz leaving as the company’s chief business and product officer and Andre Almeida taking over his old CMO role. Under the reorganization, John Saw remains CTO but will now be in charge of network, technology, product engineering, and cybersecurity.

Telstra’s nationwide outage in Australia has created quite a bit of animosity towards what is the largest mobile operator in that country. Emergency services were unavailable, with hundreds of calls that were not received; in a previous nationwide outage, there were deaths directly connected to the same problem. This is why operators must deploy even more layers of security and operational controls to prevent software configurations from causing widespread outages and potentially costing lives. We have had our own issues in the U.S. across the big three carriers; regardless, the pressure on operators to improve their processes and control automation is only going up.

Apple has sued one of OpenAI’s lead hardware engineers, claiming that he and others stole proprietary information and accessed NDA materials after leaving the company. The lawsuit is quite extensive and seems to come with a lot of very specific details about how the accused engineer/executive accessed the materials and how aware he was of what he was doing. I’m not surprised that this lawsuit has happened, and it isn’t the first time that Apple has sued a competitor for an employee leaving and supposedly stealing proprietary materials. (This happened, for instance, with Nuvia’s CEO when Nuvia was acquired by Qualcomm.) We don’t yet know all the details, but this case does appear quite different because there seems to be a mountain of evidence around the accused keeping an Apple laptop and accessing Apple information using loopholes in the security. Time will tell.