Port.io Is Bringing the ROI Story to Agentic Control Planes
Port.io is pushing the agent control plane conversation past governance talk and into hard numbers: Rather than leading with the standardization-and-risk pitch most platform vendors use, the company ran a controlled test across 12,000 AI queries. It found that its “Context Lake” pre-integrated catalog cut per-query costs by 56% to 61% versus raw MCP connections, with an optimized skill file pushing total savings to roughly 80% across Claude’s model tiers. (Opus dropped from $1.76 to $0.35 per query) It’s a notable reframe for a category that’s mostly sold on trust and audit trails rather than ROI a CFO can bank on — though it’s worth flagging that this is Port’s own study, on its own benchmark, without independent validation yet. The real test will be whether the findings hold up under outside scrutiny, and whether rivals answer with their own cost data. You can learn more about this in my recently published Field Note.

Baz.Ai Wants to Make Sense of the AI Coding Bottleneck
Last week I had a chance to sit down with Guy Eisenkot, CEO of Baz (baz.ai), an AI software governance platform addressing two functions that enterprise IT organizations are going to need much more of very soon. Baz operates across two core product lines: AI Code Review, which deploys agents that learn each team’s conventions and tribal knowledge through developer feedback, and Planner, which performs inline code review during the agent planning phase rather than after the fact. For enterprise IT, the case for caring is straightforward: Most large organizations are now running multiple coding agents simultaneously — Claude Code, Cursor, Kiro — each with incompatible instruction and rule files. This is a fragmentation problem Baz is purpose-built to unify and govern; as agents gain autonomy, the auditability gap is becoming a genuine liability. What I find most compelling is the similarity to GitHub, which became indispensable by governing the process of managing a process versus being the process itself. Baz is staking out the same kind of durable, process-layer position for the agentic SDLC.

Not All Credits Are Created Equal
This spring and summer I have been researching pricing strategies for agents, and while I do have a major piece of research coming soon, I wanted to tease a very fundamental learning before publication. In Episode 11 of the IT Talk podcast, I spoke about how not all credit-based pricing models are the same. As people select where they build and host agents, how those agents are priced will become a key decision. But right now, comparisons between vendors are almost impossible. As an example, the difference between Microsoft and Salesforce is very telling: Microsoft prices across the entire agent request, and there’s some variability in the decisions an agent may take. Salesforce charges a single flat price per tool call, which is more granular but may get more expensive for more complex agents. To illustrate this better, I also posted this infographic to help.

Port.io Is Bringing the ROI Story to Agentic Control Planes
Port.io is pushing the agent control plane conversation past governance talk and into hard numbers: Rather than leading with the standardization-and-risk pitch most platform vendors use, the company ran a controlled test across 12,000 AI queries. It found that its “Context Lake” pre-integrated catalog cut per-query costs by 56% to 61% versus raw MCP connections, with an optimized skill file pushing total savings to roughly 80% across Claude’s model tiers. (Opus dropped from $1.76 to $0.35 per query) It’s a notable reframe for a category that’s mostly sold on trust and audit trails rather than ROI a CFO can bank on — though it’s worth flagging that this is Port’s own study, on its own benchmark, without independent validation yet. The real test will be whether the findings hold up under outside scrutiny, and whether rivals answer with their own cost data. You can learn more about this in my recently published Field Note.

Baz.Ai Wants to Make Sense of the AI Coding Bottleneck
Last week I had a chance to sit down with Guy Eisenkot, CEO of Baz (baz.ai), an AI software governance platform addressing two functions that enterprise IT organizations are going to need much more of very soon. Baz operates across two core product lines: AI Code Review, which deploys agents that learn each team’s conventions and tribal knowledge through developer feedback, and Planner, which performs inline code review during the agent planning phase rather than after the fact. For enterprise IT, the case for caring is straightforward: Most large organizations are now running multiple coding agents simultaneously — Claude Code, Cursor, Kiro — each with incompatible instruction and rule files. This is a fragmentation problem Baz is purpose-built to unify and govern; as agents gain autonomy, the auditability gap is becoming a genuine liability. What I find most compelling is the similarity to GitHub, which became indispensable by governing the process of managing a process versus being the process itself. Baz is staking out the same kind of durable, process-layer position for the agentic SDLC.

Not All Credits Are Created Equal
This spring and summer I have been researching pricing strategies for agents, and while I do have a major piece of research coming soon, I wanted to tease a very fundamental learning before publication. In Episode 11 of the IT Talk podcast, I spoke about how not all credit-based pricing models are the same. As people select where they build and host agents, how those agents are priced will become a key decision. But right now, comparisons between vendors are almost impossible. As an example, the difference between Microsoft and Salesforce is very telling: Microsoft prices across the entire agent request, and there’s some variability in the decisions an agent may take. Salesforce charges a single flat price per tool call, which is more granular but may get more expensive for more complex agents. To illustrate this better, I also posted this infographic to help.

Two governance items moved last week. First there was Teradata pushing to make agents work in locked-down environments, and then there was Databricks putting hard limits on what an agent is allowed to spend.

Teradata joined the Agentic AI Foundation, the open-standards group for AI agents, and it’s pushing to shape those standards for the sovereign and air-gapped case. I think that the world of disconnected systems gets underrated in all the agent hype, but it happens to be exactly where Teradata operates. The Foundation’s agenda is set mostly by the big cloud and model providers, whose agents operate based on the assumption that they can always reach their own services. But Teradata’s customers are banks, government agencies, and hospitals, whose agents often can’t reach the internet at all. Inside those walls, you still run plenty of agents, tools, and data systems that have to work together, and you’d rather they run on shared open standards than one vendor’s proprietary wiring. Teradata is one of the few members that actually works in that world, so it’s staking out ground that the cloud-native players are least equipped to serve. It’s got the footprint to back it up, since its on-prem Factory platform already runs agents right next to regulated data on models the customer brings and controls. That puts Teradata in a real position to shape how agents work in the one place its cloud-native rivals can’t easily follow — the disconnected enterprise.

That handles where an agent runs. What it costs you once it’s running is the next problem, and Databricks has been chipping away at it. I’ve made the point before that Databricks wants to govern agents from the harness above them rather than inside the model, and cost is where that bet gets real. Agents spend money faster than anyone’s watching. They call models in loops, and a misconfigured one can run up a serious bill before a human notices. Most cost tooling only tells you after the fact, once the money’s already gone. What Databricks added to its Unity AI Gateway last week is a real stop on usage. You set a budget per user, per use case, per workspace, or account-wide, and once an agent hits it, the gateway blocks further requests until you raise the limit or the next billing cycle starts. That’s the difference between a dashboard and a circuit breaker, and it’s the kind of protection I’ve been waiting for Databricks to ship. For anyone using agents in production, a runaway agent is a real financial risk, and a hard cap lets you turn them loose without watching every minute. This is governance over what an agent actually does. That’s the harder half of the problem.

I’ve been saying for a while that what really drives AI cost is how much useful work you squeeze out of the GPUs you’ve paid for, and WEKA’s launch last week lands right on that target. When a model rebuilds context it already worked out, the GPU burns cycles redoing that work, and you pay for it twice. It gets worse with every user you add. But WEKA’s new Augmented Memory Grid keeps the key-value cache in fast storage instead of throwing it away, so the GPU spends more time generating tokens and less redoing what it already finished. WEKA published its own benchmark numbers, which I’ll leave to them. Either way, the payoff is more usable output from GPUs you’ve already bought, and that’s the cost math that every inference team is running right now. The rest of the release is built for the GPU-cloud and neocloud crowd that has to isolate a lot of tenants on shared infrastructure, which is where WEKA tends to win.

VAST is going after the same cost problem, and it’s not alone. Storage vendors have been pulling the key-value cache off the GPU and into their own systems for months. What I didn’t expect was the hardware: VAST built last week’s release on AMD, using EPYC CPUs and Instinct GPUs, even though it’s been almost entirely an NVIDIA shop for most of its life. A company standing up AMD infrastructure for AI now has a data platform built for it, and to me that says VAST sees enough AMD demand to get ahead of it, or else it already has a customer pulling it there. They’re all betting that inference economics will get settled in the memory layer. If the memory-layer vendors start building for AMD, that’s a small crack in how much of that layer NVIDIA gets to keep.

NVIDIA Expands Vera Beyond Agentic AI
When NVIDIA launched its first-generation CPU (codenamed Grace), it was designed and built to support the company’s GPUs. And as the company launched its latest-generation chip — Vera — there was some hangover. Despite being designed for standalone performance, many observers were still associating it only with NVIDIA’s Rubin GPUs. The company has since put a big effort into changing that market perception, and that’s what led to this past week’s repositioning exercise. NVIDIA is pointing Vera at a broader set of enterprise workloads, including analytics and HPC, as well as the control-heavy work surrounding agentic AI and reinforcement learning.

The architecture explains the pitch. Vera combines 88 custom Olympus cores and 176 threads on a monolithic compute die, avoiding the cross-chiplet traffic NVIDIA argues can introduce variable latency under load. A shared coherency fabric connects the cores, cache, memory, and I/O, while the LPDDR5X subsystem provides up to 1.2 TB/s of bandwidth and 1.5 TB of capacity. NVIDIA says this gives each core roughly three times the memory bandwidth of a traditional datacenter CPU and enables the socket to sustain more than 90% of peak memory bandwidth under load.

The competitive claim is equally direct. NVIDIA published early SPEC CPU 2026 results showing a dual-socket Vera system scoring 925 on SPECrate integer, about 3% ahead of a dual-socket AMD EPYC 9755 system at 898. The margin is modest, but the result is notable because Vera achieves it with substantially fewer cores. NVIDIA also claims up to 1.8 times the performance of EPYC on Python and much larger gains on selected memory- and graph-intensive workloads.

There are caveats. NVIDIA compared Vera against AMD’s current-generation Turin processor rather than the newly launched Venice, and most of the broader workload results remain vendor tests. Still, Vera should no longer be viewed simply as the CPU attached to Rubin. NVIDIA is making a standalone play for workloads where per-core performance, memory bandwidth, and predictable latency matter more than maximum core density. That puts Vera directly into the enterprise CPU conversation, not just the AI-factory discussion.

I do like NVIDIA’s push here — and the way it has managed this delicate situation where it has to make the argument for Arm versus x86 and Vera versus other Arm-based chips.

AMD Leans Forward at Advancing AI
As if on cue with the NVIDIA announcement, AMD came to its Advancing AI 2026 event ready to demonstrate that the AI infrastructure market remains competitive. The announcements spanned EPYC CPUs, Instinct accelerators, Helios rack-scale systems, networking, software, and an expanding set of partnerships. The larger message was straightforward — NVIDIA may define much of the market today, but the AI game is far from over, and AMD is firmly in it.

“Venice” was central to that argument. The sixth-generation EPYC family scales to 256 cores and 512 threads per socket, supports 16 memory channels with up to 1.6 TB/s of bandwidth, and adds PCIe Gen 6. AMD claims that it has up to 1.8 times the performance of the preceding Turin family. And more provocatively, the company used NVIDIA’s own Vera data to claim that Venice delivers 20% higher per-core performance and 2.2 times the per-socket throughput. Those comparisons should be read carefully, because AMD is comparing its next-generation processor with Vera, while NVIDIA benchmarked against the current-generation EPYC. Even so, AMD’s willingness to name Vera rather than Intel’s Xeon says a lot about where it now sees the competitive center of gravity.

Instinct and Helios completed the infrastructure story. Helios combines 72 MI455X GPUs, 18 Venice CPUs, and Pensando networking. AMD claims it handles up to 30% more inference tokens per dollar than the leading competitive system, while MI455X delivers up to 34 times the token throughput of MI355X on selected tests.

The partnerships made AMD’s openness argument tangible. Cerebras is pairing its wafer-scale inference technology with Helios, showing that AMD is willing to integrate a differentiated accelerator rather than insist that every workload run exclusively on its silicon. Cisco is combining AMD inference systems with its networking, observability, and security portfolio for hybrid and local agentic AI. Anthropic, OpenAI, Meta, and Microsoft add deployment credibility.

AMD still has work to do, particularly around software maturity and execution at rack scale. And it still has a significant competitor to fight against. For those who want to make comparisons to Intel when AMD launched EPYC — don’t. NVIDIA is not making the missteps Intel made with Xeon. And the inertia is real.

But Advancing AI showcased a company with products, customers, and a credible architectural alternative.

Intel Crushes Earnings, but Is Foundry the Big Story?
Intel’s second-quarter results were far stronger than the market expected. Revenue reached $16.1 billion, up 25% year over year, while non-GAAP earnings of $0.42 per share roughly doubled consensus expectations. Intel also guided third-quarter revenue to between $15.8 billion and $16.8 billion, ahead of prior Wall Street estimates.

The DCAI business unit killed it. Revenue increased 59% to $6.3 billion as AI infrastructure demand drove renewed interest in Xeon. That matters because the CPU continues to sit beneath accelerator-rich systems, running orchestration, data services, agent execution, and the enterprise applications that AI ultimately touches. Intel’s product story also gained some momentum with Xeon 6+, its first server-class product on Intel 18A, and new rack-scale inference partnerships involving SambaNova and Foxconn.

The more interesting signal, however, may be coming from Intel Foundry. Revenue grew 31% to $5.8 billion, although most of that remains internal Intel business, and the segment still posted a roughly $2.1 billion operating loss. This is not yet proof that Intel has built a successful external foundry; it is evidence that manufacturing execution is improving. Intel also said that the 18A-P node entered risk production, Panther Lake moved into high-volume manufacturing, yields and cycle times improved, and customer engagements are underway for the 14A node.

That distinction matters. DCAI can drive near-term revenue and margin, but the foundry business determines whether Intel becomes something structurally different. A competitive external manufacturing business would give Intel relevance well beyond its own CPU portfolio and make it a strategic supplier to companies that might otherwise compete with Intel Products.

There is still a long way to go. External foundry revenue remains small, losses remain significant, and 14A commitments must become real designs and volume production. But this was one of the first quarters where the foundry discussion felt less like an aspiration and more like measurable operating progress.

The middle of the data market continues consolidating. Domo has agreed to sell its platform to Progress Software for $400 million in cash, and the structure of the deal is worth a look. It’s an asset sale, so the product and the staff go to Progress, while the public company stays listed under a new name and ticker, keeping its cash, its tax assets, and its CEO, Josh James. The deal is signed but not closed yet, with completion expected later this year pending regulatory approval. Assuming it closes, the platform belongs to Progress, which framed the deal in terms of trusted data and context for AI, and called Domo’s analytics and cloud architecture complementary to the data platform it’s been building around MarkLogic. That reads as a company buying Domo to grow a data platform, not to park it. I’ve been following Domo’s agentic push for a while, and earlier this year I wrote about their App Catalyst, the tool that spins up governed apps on your data. That’s exactly the capability Progress points to when it talks about improving the security, governance, and cost of enterprise AI. So if you run Domo, the signals point to continued investment and support, with the usual caveat that any acquisition takes a while to settle.

Denodo went the other way. It’s still independent, and instead of selling, its latest release doubles down on context. “Context” is the word just about every data vendor reaches for now, and I made the case in a recent Forbes piece that each one defines it to fit whatever they already sell. Denodo’s a clean example of that. For Denodo, context means federation. You get to the data where it already lives instead of copying it somewhere central first. The release itself is mostly semantic and metadata work, turning definitions like “revenue” or “customer value” into governed objects that you set once and reuse everywhere. That’s Denodo leaning on federation and semantics it’s had for years and pointing them at the catalog fight, not breaking new ground. The real question for me is whether a metric defined once over live, federated data holds up against the tools built specifically for that job. If it does, not having to copy anything is a genuine reason to pick it.

Samsung expanded its foldable smartphone lineup to fit a three-tier strategy, introducing the Galaxy Z Fold8, a new form factor that slots in between the Fold8 Ultra and the Flip8. The new passport-style foldable brings many of the Fold8 Ultra’s advancements, but with a dual 50MP camera configuration (no telephoto). The new Fold8 Ultra takes the previous generation’s Fold7 design and builds on it with a faster, better chip and a bigger, longer-lasting battery, in a thinner 4.1mm unfolded size. Samsung also returned to using Qualcomm Snapdragon chips exclusively on its foldable line with the Snapdragon 8 Elite Gen 5. Samsung also introduced two new smartwatches that, for the first time, run on a Snapdragon Wear Elite chip rather than an Exynos chip. The new Galaxy Watch9 and Watch Ultra2 both use the latest Qualcomm chip, adding more consistency across the Google Wear OS market.

T-Mobile and AT&T reported earnings, both beating on EPS. T-Mobile revenue was slightly lighter than expected at $22.79 billion, and postpaid net account adds were 277,000, down 13% year over year. The stock sold off after hours. AT&T, on the other hand, beat on revenue with $31.6 billion and posted 430,000 postpaid net adds and 370,000 fiber adds, both huge drivers of growth for the company. AT&T’s CEO also pushed back on Starlink rumors of a direct-to-consumer mobile offering and reminded people of the AST SpaceMobile partnership. I believe that both carriers have pretty solid businesses, but also have room for improvement, especially as the industry ramps up for 6G.

IBM has announced an agreement to acquire HRL Laboratories, a research joint venture previously held by Boeing and General Motors. HRL has expertise in silicon-spin qubits, advanced cryogenics, quantum sensing, and microelectronics that can be integrated into IBM’s long-term hardware roadmap. By fusing silicon-spin capabilities with its existing superconducting architectures, IBM aims to accelerate progress toward its fault-tolerant systems, including the planned 2029 “Starling” platform. Financial terms were undisclosed, but Boeing and GM are set to maintain collaborative research partnerships with IBM. This transaction highlights an industry shift toward consolidating advanced quantum physical IP, cryogenic infrastructure, and specialized silicon manufacturing capabilities to solve fault-tolerant scaling bottlenecks. In fact, I have been expecting IBM to announce a move toward spin qubits and am working on an article with more complete details about this approach to quantum.

German quantum startup SaxonQ is taking orders for its SXQ128 and SXQ512 systems, which operate at room temperature. The company’s diamond nitrogen-vacancy (NV) centers don’t require the massive liquid-helium dilution refrigerators normally required for superconducting processors. It says that fabrication yields are above 85%, and that the systems have single-qubit gate fidelities of 99.92%.

Continuing the focus on practical implementation, the units run in standard server environments using ordinary wall power. SaxonQ says they are also available for mobile, edge, and high-density localized computing, again without heavy cryogenic overhead.

PsiQuantum secured a $125 million DARPA benchmarking contract, which is its largest U.S. government contract to date. The company is part of the DARPA Quantum Benchmarking Initiative (QBI). The award further supports PsiQuantum’s efforts to validate and deploy utility-scale, fault-tolerant photonic quantum computers. The company uses photonic qubits that can be manufactured in standard semiconductor foundries, with the objective of scaling to millions of physical qubits. The DARPA initiative rigorously benchmarks whether fault-tolerant architectures can deliver verifiable computational advantages over classical supercomputers on high-impact scientific problems.

Intel’s earnings were a huge boon for the company, showing the best growth in 15 years. Specifics include $16.1 billion in revenue, a 25% year-over-year improvement, with the Client Computing business growing by 13% to $8.9 billion and the DCAI unit growing by 59%, with yields from the Intel 18A chip manufacturing node improving as well. Intel also guided upward and announced a partnership with Fortinet to manufacture Fortinet’s Security Processor 6 at Intel Foundry.

Samsung teased its upcoming AI smart glasses partnership with Warby Parker and Gentle Monster at its Unpacked launch event, where it once again trickled out some details about the new glasses. All of these glasses are built by Samsung in partnership with Google and the fashion brands, with the goal of delivering high-quality Android XR experiences that heavily leverage AI. At the event, I got to see these glasses up close; to me, they look competitive with Meta’s Ray-Ban and Oakley glasses already on the market. That said, I was not given a chance to put them on or use them, as that was forbidden — and which likely confirms that the Gentle Monster and Warby Parker models won’t launch until the fall.