My Thoughts on IBM’s Earnings Pre-Announcement and Stock Fallout
IBM’s 25% single-day stock drop on July 14 — the worst in the company’s history — followed disappointing Q2 revenue guidance and a rare admission from CEO Arvind Krishna that “We did not adapt and move quickly enough, and numerous large deals failed to close.” My take: That acknowledgment isn’t a strategy, and IBM’s path back runs through three different playbooks, not one. The Software unit needs a more forward-looking narrative as it continues to integrate its various acquisitions. Go-to-market needs a more ruthless triage: Too many offerings are draining resources that should go toward what’s actually winning. And Infrastructure should stop selling long-term roadmaps and start selling near-term execution — supply and delivery speed, not vision decks. The thread underneath all three: IBM’s AI positioning currently reads as defensive, when the underlying technology is genuinely interesting. Read the full article: What Would I Say to IBM Leadership Today?

IBM Bob Matures into a More Complete Agentic Development Platform
Five days before the stock drop, IBM quietly shipped the kind of update that could be a part of its bigger AI story: Bob, its agentic dev platform, picked up parallel model-native tool calling, subagents that run in isolated context to control token costs, and a new “Bobalytics” layer for tracking AI consumption and spend across the SDLC. That last piece matters more than it sounds — cost governance is the single biggest challenge facing agentic application development, and IBM is answering it directly. The sharper move, though, is the inclusion of three prebuilt modernization packages for IBM Z (COBOL/PL/I/JCL), IBM i (RPG), and Java 25 migration. That’s the differentiated ground IBM should be standing on: nobody else has the mainframe and midrange install base to make “Let an agent modernize your COBOL” a credible pitch. And it turns the AI migration story on its head by providing a low-cost means for customers to stick with IBM infrastructure. Read IBM’s full article about it here: IBM Advances Enterprise AI Software Development with Multi-Agent Capabilities and Specialized Modernization Workflows

Kiro Turns One: AWS’s Spec-Driven Coding Agent Goes Model-Agnostic
Kiro — Amazon’s AI IDE built around writing specs before code — hit its first anniversary this week, and the adoption numbers are worth noting: 100,000-plus developers in the first five days of preview, adoption doubling by October, and a Discord community that now has more than 15,000 community-built extensions. The more competitively interesting move is on the model side: Alongside its Anthropic and open-weights lineup, Kiro just added OpenAI’s GPT-5.6 in three tuned variants (Sol for complex multi-step work, Terra for standard tasks, Luna for fast/low-cost throughput) across IDE, CLI, and web. That’s AWS explicitly positioning Kiro as a model-agnostic coding agent rather than a showcase for any one model partner. Set against IBM’s Bob news above, it’s a useful contrast in strategy: IBM is winning by maximizing its key value propositions that no one else can touch, while AWS is winning by maximizing developer choice at the model layer. Both are valid plays — the market will decide which moat holds up better. Read the full article from the Kiro team: Kiro’s One-Year Anniversary

My Thoughts on IBM’s Earnings Pre-Announcement and Stock Fallout
IBM’s 25% single-day stock drop on July 14 — the worst in the company’s history — followed disappointing Q2 revenue guidance and a rare admission from CEO Arvind Krishna that “We did not adapt and move quickly enough, and numerous large deals failed to close.” My take: That acknowledgment isn’t a strategy, and IBM’s path back runs through three different playbooks, not one. The Software unit needs a more forward-looking narrative as it continues to integrate its various acquisitions. Go-to-market needs a more ruthless triage: Too many offerings are draining resources that should go toward what’s actually winning. And Infrastructure should stop selling long-term roadmaps and start selling near-term execution — supply and delivery speed, not vision decks. The thread underneath all three: IBM’s AI positioning currently reads as defensive, when the underlying technology is genuinely interesting. Read the full article: What Would I Say to IBM Leadership Today?

IBM Bob Matures into a More Complete Agentic Development Platform
Five days before the stock drop, IBM quietly shipped the kind of update that could be a part of its bigger AI story: Bob, its agentic dev platform, picked up parallel model-native tool calling, subagents that run in isolated context to control token costs, and a new “Bobalytics” layer for tracking AI consumption and spend across the SDLC. That last piece matters more than it sounds — cost governance is the single biggest challenge facing agentic application development, and IBM is answering it directly. The sharper move, though, is the inclusion of three prebuilt modernization packages for IBM Z (COBOL/PL/I/JCL), IBM i (RPG), and Java 25 migration. That’s the differentiated ground IBM should be standing on: nobody else has the mainframe and midrange install base to make “Let an agent modernize your COBOL” a credible pitch. And it turns the AI migration story on its head by providing a low-cost means for customers to stick with IBM infrastructure. Read IBM’s full article about it here: IBM Advances Enterprise AI Software Development with Multi-Agent Capabilities and Specialized Modernization Workflows

Kiro Turns One: AWS’s Spec-Driven Coding Agent Goes Model-Agnostic
Kiro — Amazon’s AI IDE built around writing specs before code — hit its first anniversary this week, and the adoption numbers are worth noting: 100,000-plus developers in the first five days of preview, adoption doubling by October, and a Discord community that now has more than 15,000 community-built extensions. The more competitively interesting move is on the model side: Alongside its Anthropic and open-weights lineup, Kiro just added OpenAI’s GPT-5.6 in three tuned variants (Sol for complex multi-step work, Terra for standard tasks, Luna for fast/low-cost throughput) across IDE, CLI, and web. That’s AWS explicitly positioning Kiro as a model-agnostic coding agent rather than a showcase for any one model partner. Set against IBM’s Bob news above, it’s a useful contrast in strategy: IBM is winning by maximizing its key value propositions that no one else can touch, while AWS is winning by maximizing developer choice at the model layer. Both are valid plays — the market will decide which moat holds up better. Read the full article from the Kiro team: Kiro’s One-Year Anniversary

What scares people about agents is the confident wrong answer. Maybe worst of all is when nothing breaks: An agent picks up a definition that broke three years ago, acts on it, and nobody notices until someone goes looking. Every governance vendor has pivoted to agent governance this year, so the launches tend to blur. But Alation’s AIOS has a feedback loop, and last week I wrote a Field Note on it. Thanks to the feedback loop, when an agent gets something wrong, the correction goes back to whatever caused it, and every other agent using that context picks up the fix.

Most context layers hand you information and never learn what happened next. Alation’s catalog has long ranked definitions by trust and usage, which surfaces the popular answer rather than the right one. Certification catches those, but only while somebody keeps stewarding it. Most data teams are underwater already, but a feedback loop that learns from its own corrections doesn’t ask anyone to stay on top of it. Diagnosis is the catch, since something still has to decide whether the fault was stale data, a bad definition, or the agent itself. If you’re buying a context layer, ask what happens to the correction history. Open standards make the semantic models portable; none of them export the record of what your business meant and who fixed it.

With that said, meaning is only half of it. An agent can read every definition right and still cause trouble in how it chains actions together. Attacks rarely look like attacks one step at a time. Each action clears its own check, and you only see the problem across a whole session. Databricks added contextual policies to Omnigent, the open-source layer it runs above agent systems. Now it weighs the whole session instead of judging one action. Control belongs in the harness above the agents rather than inside the model — the same bet behind Unity AI Gateway. Bury your policies in individual agent prompts, and every new agent becomes another place for the rules to drift.

Running analytics on your data has always meant copying it somewhere else first, then paying to secure and govern that copy all over again. NetApp bought DataPelago to go after that problem. Nucleus, the DataPelago engine, spreads work across CPUs and GPUs and sorts out where each piece of a job runs best. It slots underneath Spark and Trino, so the tools that data teams already run get faster without anybody rewriting a query. NetApp has done processing on the storage itself for a while, but that work serves NetApp’s own pipeline. This is broader. How far NetApp takes this determines how big it gets. Pointing Nucleus only at NetApp’s own pipeline should mean customers get a faster NetApp. But let it accelerate the engines people already run, and NetApp ends up underneath a lot of analytics that has nothing to do with storage. I’d bet on the second.

The same bottleneck bites harder on the GPU side. Companies have spent a fortune on compute and then left it idling because of slow data access. Cloudera and VAST Data have now partnered to go after this problem, and I wrote a Field Note on where I’d point buyers. VAST keeps the GPUs fed from one fast data layer, and Cloudera’s governance follows the data wherever it goes. Together, they let a company run an AI factory in its own datacenter. That’s what regulated buyers have been waiting on. Plenty of them can’t move data off their own infrastructure for compliance or privacy reasons, and private AI has mostly meant assembling it yourself.

A funding round can tell you about more than just valuation. Databricks has agreed terms on about $3 billion of new funding at a $188 billion valuation — up from $134 billion a quarter ago — and the round hasn’t closed yet. The money is pointed at Unity AI Gateway, Genie, and Lakebase: the governance layer, the analytics coworker, and the operational database for agents. None of it goes to the core lakehouse. That’s a company telling you what it thinks it sells now. The rationale underneath this is that customers stopped being impressed by how much they consume and want to see what it bought them instead. A $54 billion step-up in Databricks’ valuation in three months says that investors like it.

Regatta Data wants the same job Lakebase does, except it’s the primary focus of the whole company — with $68 million behind it — rather than one product line. Its RegattaDB hit general availability last week running transactions, analytics, and vector search on a single system. There’s a real benchmark behind the launch: a distributed join across 20 billion rows on 50 nodes, holding 50,000 transactional updates a second. And Synergy Logistics willing to put its name on it as a customer. Where I’d push back is the framing. Calling this a database built for agents drops Regatta in alongside every giant making the same claim, and it undersells what the company has built. The hard part it solved is running mixed workloads without a pipeline in the middle. I’ll be keeping an eye on it.

Siemens, Databricks, and FFT
First, a quick note on timing. This entry covers a mid-June Siemens announcement. It came back onto my radar while I was digging into Rockwell’s July MES survey, described below. I found a broader pattern — Siemens has now struck several edge-to-cloud data partnerships in the past few weeks, each addressing the same integration problem. That pattern, not any single press release, is the interesting story, and it doesn’t have an expiration date.

On June 16, Siemens announced an edge-to-cloud integration with Databricks and longtime automation partner FFT Produktionssysteme (which, like Siemens, is headquartered in Germany). Siemens Industrial Edge streams contextualized shopfloor data through FFT’s new DataBridge connector straight into the Databricks platform, where it trains models and then pushes them back down to the edge for execution. Rainer Brehm, Siemens’ automation COO, called it a way to make sure “data, context, and execution come together.” FFT’s Volker Stark pitched it as a gateway for “more than 30,000 potential customers.”

It’s certainly a good idea, and it’s not Siemens’ only one this year. Weeks earlier, Siemens paired Industrial Edge with HighByte’s Intelligence Hub for the same kind of job — connecting, contextualizing, and prepping OT data for AI. Siemens is also running an Industrial Data Fabric arrangement with AWS. Three partners, three cloud destinations, and one workflow: Supply plant-floor data somewhere useful without building a new pipeline every time.

These options reflect flexibility, not indecision, and it’s a practical response to a problem nobody has completely solved yet. Rockwell’s newly released “Scaling MES Across the Enterprise” survey backs this up: 44% of manufacturers name integration as the top MES buying requirement, and only 23% report full integration across ERP, PLM, quality, and OT. That’s an industry-wide gap, not a Siemens gap, and no single partnership — Siemens’, Rockwell’s, or anyone’s — closes it in one step. Meeting customers where the cloud stack already runs, rather than forcing everyone onto one path, is a sensible move.

My take: the most critical component of Siemens’ data flow isn’t the pipe to Databricks — it’s what Siemens’ Industrial Information Hub does before the data leaves the edge: semantic modeling, asset structuring, and turning raw data into something an AI platform can actually use. That’s the hard part, and it’s reusable no matter which cloud a customer picks. I would advise Siemens and FFT to drop the “eliminates IoT middleware” tagline because it undersells the work. I’d say, “We built the contextualization layer once, so you can plug it into whichever analytics platform you already trust” is more accurate and more compelling. Given the extreme diversity and long lifetimes of industrial systems, I’m not sure “eliminating middleware” is a useful goal. But simplifying OT-IT connectivity and making it more “plug and play” with a lot less customization is real progress. I’m focusing on that.

Nokia released some details around the upcoming AI-RAN platform that it has been building with NVIDIA. It gave more detail on things like upgrades, software solutions, compatibility with ORAN, and increased spectral efficiency. On the topic of spectral efficiency, Nokia is claiming that its anyRAN, paired with NVIDIA’s Aerial AI-RAN, will deliver a more than 100% improvement in spectral efficiency, doubling the capacity of existing spectrum assets. Nokia says the AI-RAN platform has already shown 20% spectral efficiency gains thanks to AI radio innovation, and says that the company is on track to deliver 50% gains by 2027 and 100% gains by 2028. This is especially important because operators worldwide are spectrum-constrained, and any way to make the most of the limited spectrum they have could be massively impactful on the bottom line. This GPU-based platform will also support 5G initially and be software-upgradeable to 6G — in line with what most people expected, given how each G transition usually happens.

AST SpaceMobile has once again delayed its commercial launch, this time to 2027, as it is struggling to get enough satellites into orbit to support a commercial service. The company also said that it will pivot toward acquisitions in the near term to help build out its satellite network in the interim. Because AST SpaceMobile relied heavily on Blue Origin as its launch partner, it has experienced delays stemming from the Blue Origin launchpad explosion at the end of May, which pushed back all launches by at least six months. Also, the most recent AST SpaceMobile launch didn’t reach a sustainable orbit and was considered a total loss. I believe that AST SpaceMobile will be fine in the long term, but as a publicly traded company, there’s much more scrutiny on what’s going on with it than there would be for a private company.

OnePlus is officially dead as a brand after Oppo confirmed that it would be pulling out entirely from North America and Europe, followed by India and China next year. I believe that this is a consequence of the OnePlus brand becoming too strong and doing too good a job compared to the Chinese OEMs it is owned by. The OnePlus brand’s scope has been fairly narrow, with much of its audience in India, especially for the mid-range phones that have driven most of the brand’s volume. The problem, I believe, is that when the OnePlus brand was performing well, it competed too much with the Oppo and Realme brands. OnePlus genuinely made great products for Android enthusiasts, and I think it became apparent that it was getting too much attention over the larger brands within the BBK Electronics umbrella. This will mean fewer choices for consumers not only of smartphones but also of tablets and smartwatches, which the company also made extremely well. (You can read more of my thoughts on OnePlus in this article from last year.)

Microsoft doubled down on topological qubits with its latest Majorana 2 quantum chip, which increased qubit coherence times by 1,000x compared to Majorana 1. The company’s researchers have been working on this potentially game-changing technology for some time. Its goal is to eliminate the quantum industry’s biggest problem, which is the fragility of traditional qubits that limits scaling. Microsoft addresses that problem by using a topological architecture that makes qubits more stable and less error-prone.

Majorana 2 has been developed at the intersection of materials science and AI-assisted research. The new chip was designed with the extensive use of Microsoft Discovery, a new agentic AI platform capable of supplementing human researchers to accelerate research and materials design. If Microsoft’s AI-assisted research succeeds, it could change quantum computing by providing an alternate path to developing quantum systems based on topological design and a revised material stack. It has already led Microsoft to move up its timeline for delivering a fault-tolerant quantum computer from 2033 to 2029.

Some critics say Microsoft has not provided enough evidence to support its topological-qubit claims. But despite past scientific controversies, Microsoft asserts the validity of its progress, citing DARPA validation and its own active computation with the qubits.

Microsoft believes that the next breakthrough in quantum computing will come from the continued pairing of experimental physics with software-driven discovery and analysis. The company also believes that its hardware-protected, single-chip solution offers a scalable path to a million-qubit system, while differentiating it from competitors.

For more details, read my Forbes article here.