I’m in the camp that believes writing policy for AI agents is the easy part. But when it comes to figuring out which agent is actually acting and whether anyone set it up on purpose, that’s a lot harder. And most of what I see sold as agent governance assumes that somebody registered the agent first.

Nobody registered the two OpenAI models that got out of a sandboxed test a few weeks back. They found a zero-day in a package-registry proxy and ended up running their own code inside Hugging Face’s production systems. No human was steering them. A week later, NVIDIA formed the Open Secure AI Alliance, and Red Hat, Elastic, and ClickHouse were among the 37 partners. (If you’d like more details, I wrote a Field Note on it.) I’d worry less about which labs didn’t join. The models got loose from a lab’s own setup, and I don’t see anything the Alliance is building that would have kept them inside it. That said, the Alliance does stand to help everybody downstream of those labs, which is most of us, and I think that’s worth doing. If I’m running security right now, I’m finding out whether my tooling can even look at attack traffic that a model generated. An agent nobody set up won’t have an identity you can check.

I’d argue that verifying an agent’s identity is the hardest piece of this, and Snowflake just shipped a real chunk of that functionality. Agent Verified Identity is generally available now, along with AI risk posture and sensitive data protection. The Cortex AI Gateway around it is still on its way to public preview. None of it works unless it plugs into what you already run for identity, so I was happy to see Okta sign on alongside SailPoint and Saviynt. Most companies already run Okta, so that’s the one that decides whether this shows up in production. Once it’s running, an agent moves across your data, your identity systems, and your APIs on somebody else’s permissions. The governance you already have wasn’t built for that. Agent governance has filled up fast, and I counted three vendors making claims about it inside six days. Snowflake’s angle is running it on the same platform that already governs the data the agent reaches for, and that pairing would be harder to assemble from parts. I’d expect some version of this to end up standard, showing up inside a platform that companies already run rather than something they buy separately.

I’m in the camp that believes writing policy for AI agents is the easy part. But when it comes to figuring out which agent is actually acting and whether anyone set it up on purpose, that’s a lot harder. And most of what I see sold as agent governance assumes that somebody registered the agent first.

Nobody registered the two OpenAI models that got out of a sandboxed test a few weeks back. They found a zero-day in a package-registry proxy and ended up running their own code inside Hugging Face’s production systems. No human was steering them. A week later, NVIDIA formed the Open Secure AI Alliance, and Red Hat, Elastic, and ClickHouse were among the 37 partners. (If you’d like more details, I wrote a Field Note on it.) I’d worry less about which labs didn’t join. The models got loose from a lab’s own setup, and I don’t see anything the Alliance is building that would have kept them inside it. That said, the Alliance does stand to help everybody downstream of those labs, which is most of us, and I think that’s worth doing. If I’m running security right now, I’m finding out whether my tooling can even look at attack traffic that a model generated. An agent nobody set up won’t have an identity you can check.

I’d argue that verifying an agent’s identity is the hardest piece of this, and Snowflake just shipped a real chunk of that functionality. Agent Verified Identity is generally available now, along with AI risk posture and sensitive data protection. The Cortex AI Gateway around it is still on its way to public preview. None of it works unless it plugs into what you already run for identity, so I was happy to see Okta sign on alongside SailPoint and Saviynt. Most companies already run Okta, so that’s the one that decides whether this shows up in production. Once it’s running, an agent moves across your data, your identity systems, and your APIs on somebody else’s permissions. The governance you already have wasn’t built for that. Agent governance has filled up fast, and I counted three vendors making claims about it inside six days. Snowflake’s angle is running it on the same platform that already governs the data the agent reaches for, and that pairing would be harder to assemble from parts. I’d expect some version of this to end up standard, showing up inside a platform that companies already run rather than something they buy separately.

Agents forget everything when a session ends. This means they rebuild the same context over and over, and you pay for it every time. I think that’s one of the most expensive unsolved problems in agent infrastructure right now, and nobody agrees on where the memory should actually live.

I dug through Google Cloud’s annual AI infrastructure report and wrote a LinkedIn item on it. One of the gaps that leaders named most often was not having a vector database fast enough for what agents demand. I’d push back on that. A lot of teams are describing one problem and shopping for another. Vector search finds things that look similar to what you asked, which is genuinely useful, but it doesn’t know what’s true right now. There’s no sense of time or order in a vector index, and nothing marks a record as replaced. Change your shipping preference, for instance, and the old embedding is still sitting there scoring as relevant, so the agent could answer with the preference you already abandoned. Teams read that as a retrieval quality problem and go tune the chunk size or swap the embedding model. The gains stay small, because none of those moves gives you a record of what happened and in what order. Agents need a transactional database or a graph sitting next to the vector store holding decision history and entity state.

MinIO just announced AIStor Memory and put its version of that record in storage. Memory becomes a native data type next to objects and tables, so an agent keeps context across sessions and picks up work without rebuilding state. The placement is smart. The memory stays on infrastructure the customer owns, under its own keys, instead of inside somebody’s managed service. MinIO frames it as an ownership question. What an agent learns becomes company knowledge, so it belongs on hardware the company controls. That’s the right argument to be making, and MinIO is better placed than most to make it. MinIO is being ambitious about consolidation, with one layer meant to sit underneath the vector store, the metadata service, the secrets manager, and the sync pipelines. That’s real work, and a general-purpose store has to hold up next to tools built only for this. I like that somebody is arguing that memory belongs to the enterprise rather than to whoever happens to be hosting the agent.

Buying a platform used to mean taking whichever AI came bundled inside it. That setup is coming apart now, and I think buyers come out ahead. The model, the retrieval layer, and the app underneath are turning into things you source separately.

Oracle plans to put Google’s Gemini models inside Fusion Applications, NetSuite, and its AI Agent Studio, with Google Cloud announcing it alongside Oracle. No date attached, so I’d read it as directional rather than specific. Gemini has been reachable through Oracle’s cloud since the two struck an arrangement last year. Now it moves up into the business applications. Oracle sells its own cloud directly against Google Cloud and is embedding Google’s frontier models one layer above that fight. Most vendors won’t do both. This approach holds together if you believe the application and the data are what retain the customer, and the model is a piece you swap out. Oracle runs on the same logic, putting its database on AWS, Azure, and Google Cloud. Model choice has come loose from cloud choice. If your architecture still assumes that picking a cloud means picking a model, that’s worth revisiting.

Almost every “context” pitch I hear is about data a company already owns. Documents, tables, tickets, code. But nobody owns a copy of the live web, and the teams that need it have been building that plumbing themselves. Nimble just made Web Search Agents generally available through an API, an SDK, and a hosted MCP server, selling the whole retrieval stack instead of a piece of it. Today you stitch together a search API, a browser, a parser, and validation, then have to maintain all of it forever. That’s engineering time nobody put in the AI budget. I’m a fan of Nimble’s Search Plans. Every run comes back showing what was searched, where, and why. Agents don’t belong in production if nobody can tell what one of them did, and in banking or insurance “Where did this come from?” is the first question anyone asks. Most agent token spend goes to reading things the agent never should have read, so cutting the bill starts with retrieval.

Qualcomm
Qualcomm made three edge-related announcements on July 29, 2026: a BMW deal, Q3 results, and the closing of the Modular acquisition. Here’s the upshot from my edge vantage point.

Qualcomm and BMW Group co-developed Snapdragon Ride Pilot, which launched commercially last November in the iX3, the flagship of BMW’s Neue Klasse program. It’s safe to say that BMW is happy with the results, because the company just named Qualcomm as its lead compute silicon provider for next-generation digital cockpit and ADAS/AD systems. This deal covers the Snapdragon Digital Chassis portfolio: Snapdragon Cockpit, Snapdragon Ride, the Snapdragon Elite automotive SoCs, and dedicated AI accelerators. Qualcomm’s announcement said deliveries start “with model programs starting next decade.” Given silicon and automotive lead times, the next-generation parts at the center of the agreement have not yet taped out.

Qualcomm dropped its Q3 FY2026 results the same day. Automotive revenue reached a record $1.59 billion, up 61%, the 23rd consecutive quarter of double-digit growth. IoT revenue grew 9% to $1.83 billion, with industrial, networking, and robotics offsetting declines in tablets and consumer categories. Handset revenue fell 20% to $5.1 billion on difficult memory pricing industrywide. Qualcomm reported an industrial design-win pipeline above $7 billion, with more than $3.5 billion secured this fiscal year.

The Modular acquisition also closed on July 29, 2026. I covered the strategy behind it in these pages on July 3, 2026. Modular compiles trained models across heterogeneous silicon, giving Qualcomm a cross-vendor answer to CUDA. Chris Lattner, Modular’s CEO, is now Qualcomm’s EVP of advanced AI software and platforms.

My take: The BMW agreement shows that a “design win” for an edge AI platform is a lot more than a socket sale. Edge AI is now a long-term platform relationship. This means that supplier selection at this scale is no longer just a procurement task, but a strategic architectural commitment. Competition among silicon providers moves from speeds, feeds, features, and roadmaps to architecture, software, integration, and long-term support.

The BMW agreement is a paradigm example of a customer buying an architecture, not a chip. BMW realizes that many models have to run well on edge silicon without heroic effort. This requires tools that simplify and automate model optimization for heterogeneous AI engines. Qualcomm’s engines evolve rapidly and vary across product lines, so each chip is a separate target. The technology from Modular addresses that diversity by taking in a model in neutral form and automating much of the silicon optimization work. Most AI development starts on NVIDIA silicon, so advanced porting capabilities determine model availability.

AMD
AMD provided details about its first complete physical AI platform on July 23 at its Advancing AI 2026 event. The Ryzen AI Embedded X100 Series, announced in January with few specifics, now has full specs, joined by three new elements: the Kria AI SOM, the Kria AI Robotics Developer Platform, and a Robotics Partner Network.The X100 combines Zen 5 CPU cores, RDNA 3.5 graphics, and an XDNA 2 NPU, all addressing up to 128 GB of LPDDR5X from a single coherent pool. AMD guarantees supply for ten years, with production slated for Q4 2026. The Kria AI SOM packages the processor on an open-standard COM-HPC module. Earlier Kria modules integrated an FPGA; this one does not. The FPGA moved to the developer platform’s carrier board, a Spartan UltraScale+ device that aggregates up to eight automotive camera inputs and drives CAN-FD and EtherCAT/TSN.

My take: AMD is selling a module, carrier, toolchain, partner network, and ten-year supply commitment — not just a chip. The ten-year window, which extends to 2037, targets robotics and industrial automation buyers, for whom lifecycle commitments are essential. Moving the FPGA to the carrier makes sense because physical AI needs customer-configurable sensor and actuator interfaces with deterministic timing, plus headroom for multiple sensors. Unified memory is also important, but it’s becoming table stakes for this product class. The software story is thinner than the hardware, and that’s the gap to watch as production nears.

NVIDIA
NVIDIA introduced Cosmos 3 Edge in Tokyo on July 15 and extended the Cosmos Coalition to Japan. The coalition launched six weeks earlier at GTC Taipei with six members, mostly AI labs. Japan adds 22 industrial names, including FANUC, Yaskawa Electric, Kawasaki Heavy Industries, Fujitsu, Hitachi, Sony, SoftBank, and Honda R&D.

Cosmos 3 Edge is a 4-billion-parameter open-world model built on Nemotron, the on-device tier of a family that includes 16B and 64B versions released in May. It deploys across RTX GPUs, DGX systems, and Jetson, including the newly announced T2000 and T3000 modules, which are scheduled to ship in Q1 2027. NVIDIA says developers can adapt it to a specific robot or sensor rig in about a day, but I think this is a best-case estimate. Fujitsu is exploring a collaborative control platform with FANUC, Yaskawa, and Kawasaki.

My take: The coalition changed its makeup in six weeks, from AI labs to companies that build robots already deployed in factories. That’s a key takeaway. I have argued that vertically integrated stacks serve greenfield robotics well and brownfield plants poorly. Existing factories run a broad mix of automation platforms and replace equipment on decade timescales. Recruiting the incumbents addresses the brownfield problem because each new machine carries a copy of the unified stack onto factory floors. Japan is not an isolated case. ABB, KUKA, Universal Robots, Doosan, Boston Dynamics, and Figure AI have all announced NVIDIA stack integration since January.

The models are free, but the silicon is not. Open weights get Cosmos into robots and factories that would never license a proprietary model, and every deployment assumes Jetson, RTX, or DGX underneath. One model family running across all three means a developer can build on the workstation already on the desk, deploy to the robot, and retrain in the datacenter without porting anything.

Two cautions: (1) All 22 companies intend to join, which is not the same as joining; and (2) a coalition built on shared world models needs the proprietary factory data those members might not want to share.

Farewell, Mbed!
On July 31, 2026, Arm shut down the Mbed online tools, ending its support for the developer program that has been in place for 16 years. This was not a surprise — the company began the EOL process two years ago. Mbed is open source, and the community fork of Mbed OS (Mbed CE) is still alive. But without support from Arm, I don’t expect anyone to use it for new projects.

When Mbed launched in 2009, it was the first serious multivendor development environment for Arm 32-bit Cortex-M MCUs. The project introduced thousands of people to these powerful microcontrollers with a full set of software and tools that “just worked” right out of the box. The CMSIS hardware abstraction layer ensured hardware portability; the online IDE enabled C++ programming with fast cloud compilation and zero setup; and all Mbed devices supported drag-and-drop device flashing without the complexity of JTAG.

Mbed was ahead of its time by more than a decade, but it’s time to move on. Zephyr is becoming the de facto industry standard for MCU development, and Arduino now enables rapid prototyping on 32-bit chips. Mbed paved the path that these systems follow today.

Mbed’s EOL reminded me of all the brilliant people at Arm who created the platform and cared for it since 2009. I’m proud to have played a small role in its long history during my time at Arm. Mbed was free, open, and groundbreaking. The people who worked on it did so with enthusiasm and altruistic dedication, often on their own time.

AT&T finally closed on its acquisition of EchoStar’s spectrum, totaling $23 billion. This represents about 50 MHz of low- and mid-band spectrum that should help AT&T catch up with T-Mobile on speeds and coverage. Considering that EchoStar wasn’t doing a very good job of deploying this spectrum, it should now be put to use in a meaningful way as part of AT&T’s large network. The aspect of this deal that bothers me is that I feel like EchoStar was bullied into it by the U.S. government. It’s really not the way that spectrum should change hands, and I would’ve much preferred an auction.

Apple reported strong earnings, with revenue up 16% thanks to the iPhone business seeing a 22% increase. This was Tim Cook’s last earnings call as CEO and John Ternus’s first; both spoke during the call. Apple’s total revenue was $109.4 billion with a gross margin of 50.1%, mostly buoyed by tariff refunds. Apple’s stock still had a sell-off after the earnings call, likely due to the lighter-than-expected services revenue of $30.7 billion. Services has been one of Apple’s strongest growth areas, but I believe that Apple is near saturation of its existing user base and would need to grow the Apple user install base across markets to continue to grow services.

One of the biggest standout items from Apple’s earnings was that Mac revenue grew 29% YoY to $10.3 billion in a single quarter. This is a significant validation of Apple’s silicon strategy and, I think, reflects how using Apple silicon has become a major strength for the company; the M5 lineup has been one of its most effective. However, I also believe that Apple has benefited from lower prices for quite some time, and now it’ll be interesting to see if Apple’s recent price increases affect its volume or market share.

For several years I have been following the progress of organizations pursuing quantum advantage (QA) — the points at which quantum computers outperform classical computers. IBM and three separate partners have recently achieved historic QA milestones. These partners — the University of Chicago, Qedma Quantum Computing, and Algorithmiq — have announced a suite of landmark demonstrations confirming that quantum computing has crossed the QA threshold. IBM is now calling this the era of trusted quantum advantage. (In an article at the end of 2025, I talked more about IBM’s efforts to achieve and track quantum advantage.) What follows are the results for each partner.

1. The University of Chicago and IBM addressed one of quantum computing’s persistent issues. As quantum computations scale beyond classical simulations, verifying that the hardware returned a correct answer becomes classically intractable. Random circuit sampling (RCS) can be used to demonstrate quantum speedups, but to prove fidelity, assumptions must be made about hardware noise. In a joint paper that used IBM Heron-class processors, “Sampling hard circuits with verifiably high fidelity,” researchers created an alternative to RCS that maintains high computational hardness while embedding error detection directly into the encoded circuit structure. Here are some of the key details:

• The experiment used 70 logical qubits, one of the largest logical quantum computations demonstrated to date. It ran 2,415 logical two-qubit operations and 468 logical T gates.
• Quantum error correction (QEC) encoding achieved effective logical error rates 10 times lower than the underlying physical hardware error rates.
• The IBM Heron quantum computer completed the sampling task in about 15 minutes. That compares to extremely long runtimes of classical tensor network and Clifford-expansion algorithms.

2. Qedma Quantum Computing and IBM worked together to demonstrate quantum advantage in a simulation of complex quantum material dynamics using Qedma’s QESEM software platform with IBM hardware. Simulating how quantum materials change over time is difficult for standard supercomputers. Even powerful ones like Fugaku quickly become overwhelmed when a large number of quantum particles become entangled and complex. When that happens, classical computers run out of memory, which produces inaccurate results almost immediately. While full fault tolerance in quantum computers is likely still a few years away, algorithmic error mitigation can be used to suppress noise on near-term hardware in the meantime. This happened when the team modeled non-equilibrium physics in a two-dimensional Floquet Ising system. The experiment’s objective was to determine how magnetic properties evolve under periodic external drive pulses.

QESEM’s error-mitigation protocols were validated against supercomputing runs where classical calculations were still able to produce accurate results. This allowed the researchers to establish a valid baseline of trust before the quantum hardware tests moved into areas where classical approximations failed. This experiment proved that quantum processors can serve as trusted scientific tools in condensed-matter physics.

3. Algorithmiq and IBM simulated a programmable heterogeneous quantum material on IBM Heron processors that demonstrated how information flows, localizes, and interferes across irregular structural boundaries. The team developed a novel trust framework via intentional noise manipulation. By systematically injecting controlled noise, altering gate calibrations, and executing across different physical chips, they showed that the result was stable and invariant to hardware noise. According to the findings, the quantum system delivered trusted, accurate solutions faster and more efficiently than top classical methods. The claim has also been unchallenged on the Quantum Advantage Tracker for over eight months.

Qualcomm reported earnings in a fairly tumultuous time for the smartphone market, with its smartphone revenues falling 20%. However, the company did deliver revenue at the high end of guidance thanks to a gangbuster automotive quarter that saw 61% growth. This helped to offset some of the smartphone revenue losses, resulting in revenue being down just 4% YoY. Qualcomm also announced that prices for its chipsets will be going up in September, marking another point of pressure on smartphone OEMs when they want to ship new phones. For example, Samsung could be significantly impacted because it has just launched a bunch of new Qualcomm-based phones and smart watches. Qualcomm also reported that it expects Apple’s modem shipments to taper off faster than anticipated, likely due to the success of Apple’s C1/C1X modems and their successors.

Samsung itself posted record earnings yet again thanks to the industrywide memory crunch that has driven up prices for its memory chips. Samsung reported $118.7 billion in revenue, an increase of 130%. The company also gave its outlook on the memory shortage, saying it expects things to remain constrained through 2027. (Some industry leaders suggest constraints will continue into 2028.) Samsung’s device experience division lost money, but the main reason is that it sells itself the same memory that everyone else is struggling to get. I believe that Samsung and its investors know that this is a short-term trend, and that eventually the company will be profitable in the mobile business once again.