One of the more interesting bits of news last week was the rumored acquisition of Hugging Face by NVIDIA. And the operative word here is “rumored,” as no formal announcement has been made. My opinion is grounded in my belief that NVIDIA doesn’t necessarily get enough credit for its quite good open-weight and open-source AI contributions. One interesting aspect of those contributions is that NVIDIA has also been very good about posting its open-weight models and other testing tools on Hugging Face already. Given that NVIDIA now has some critical mass with both models and platforms (including its model router Switchyard, which I covered here), along with its servers and GPUs, it has a plausible and well-integrated AI stack. But if it wants to compete as an enterprise AI model vendor, it will certainly also need a marketplace similar to those offered by the cloud players. And instead of launching one itself, why not just buy the biggest marketplace out there?

So, this acquisition would make sense to me. NVIDIA gets a marketplace for its customers and access to a very large and vibrant community. Hugging Face gets the ability to continue the open-weight and open-data movement. Keep in mind that there are about 3 million models up on Hugging Face right now, so the big question for me is how NVIDIA would create a business model for third-party models and data sets. I can imagine demand for support services like those offered by Red Hat and others, because enterprise customers will probably want those as they build out their own AI estates.

Some people might argue that a big tech owner of Hugging Face could create issues for other large tech firms using the platform. My only thought on that point is that people said the same things about Red Hat and GitHub, and those acquisitions worked out pretty well for all parties. Once a platform like Hugging Face gets big enough, its own momentum can carry on, provided the new owner does not get in the way too much. And I suspect that NVIDIA is smart enough not to do that.

As a follow-up to last week’s Headless announcements, Salesforce introduced a deepening partnership with Anthropic that is being marketed as Claudeforce. And while enterprise players cozying up to Anthropic is not really newsworthy anymore, Salesforce has taken the idea of “deepening” further than most. While some enterprise vendors simply set up one MCP connector and allow customers to bring Anthropic APIs into the platform, Salesforce has launched a Claude-specific harness and a connector that can interoperate with the other 40-plus MCP services already available to Salesforce customers. There is also a suite of Claude skills available for developers and knowledge workers to automate processes using Salesforce’s tools. As I mentioned in my Field Note a couple of weeks ago, this is a smart move on Salesforce’s part to both offset the threat of people who think that AI can replace established SaaS platforms and possibly capture new customers by lowering the bar to Salesforce adoption.

Earlier this month, the AWS Kiro team launched a new open-source service called Kiro Crew. The idea behind Crew is that as agents get more robust, they can also become more autonomous — while still needing some help from the humans in the loop. But that human also needs help managing all of these autonomous sessions, and that is where Crew comes in. It’s a very pragmatic way to introduce autonomous agents into a development team, and while it is open-source, there are some peculiarities that you should be aware of before adoption. My field note on Kiro Crew is here.

And lastly, many clients are reaching out regarding my recent Forbes work on agent pricing. And while pricing issues persist, I have seen a couple of examples already of vendors making the pricing situation better. One of those vendors is Box, which has a very specific set of guidelines on how many credits certain actions consume. This is the type of simple, transparent action that other vendors can learn from; you can find that list here.

One of the more interesting bits of news last week was the rumored acquisition of Hugging Face by NVIDIA. And the operative word here is “rumored,” as no formal announcement has been made. My opinion is grounded in my belief that NVIDIA doesn’t necessarily get enough credit for its quite good open-weight and open-source AI contributions. One interesting aspect of those contributions is that NVIDIA has also been very good about posting its open-weight models and other testing tools on Hugging Face already. Given that NVIDIA now has some critical mass with both models and platforms (including its model router Switchyard, which I covered here), along with its servers and GPUs, it has a plausible and well-integrated AI stack. But if it wants to compete as an enterprise AI model vendor, it will certainly also need a marketplace similar to those offered by the cloud players. And instead of launching one itself, why not just buy the biggest marketplace out there?

So, this acquisition would make sense to me. NVIDIA gets a marketplace for its customers and access to a very large and vibrant community. Hugging Face gets the ability to continue the open-weight and open-data movement. Keep in mind that there are about 3 million models up on Hugging Face right now, so the big question for me is how NVIDIA would create a business model for third-party models and data sets. I can imagine demand for support services like those offered by Red Hat and others, because enterprise customers will probably want those as they build out their own AI estates.

Some people might argue that a big tech owner of Hugging Face could create issues for other large tech firms using the platform. My only thought on that point is that people said the same things about Red Hat and GitHub, and those acquisitions worked out pretty well for all parties. Once a platform like Hugging Face gets big enough, its own momentum can carry on, provided the new owner does not get in the way too much. And I suspect that NVIDIA is smart enough not to do that.

As a follow-up to last week’s Headless announcements, Salesforce introduced a deepening partnership with Anthropic that is being marketed as Claudeforce. And while enterprise players cozying up to Anthropic is not really newsworthy anymore, Salesforce has taken the idea of “deepening” further than most. While some enterprise vendors simply set up one MCP connector and allow customers to bring Anthropic APIs into the platform, Salesforce has launched a Claude-specific harness and a connector that can interoperate with the other 40-plus MCP services already available to Salesforce customers. There is also a suite of Claude skills available for developers and knowledge workers to automate processes using Salesforce’s tools. As I mentioned in my Field Note a couple of weeks ago, this is a smart move on Salesforce’s part to both offset the threat of people who think that AI can replace established SaaS platforms and possibly capture new customers by lowering the bar to Salesforce adoption.

Earlier this month, the AWS Kiro team launched a new open-source service called Kiro Crew. The idea behind Crew is that as agents get more robust, they can also become more autonomous — while still needing some help from the humans in the loop. But that human also needs help managing all of these autonomous sessions, and that is where Crew comes in. It’s a very pragmatic way to introduce autonomous agents into a development team, and while it is open-source, there are some peculiarities that you should be aware of before adoption. My field note on Kiro Crew is here.

And lastly, many clients are reaching out regarding my recent Forbes work on agent pricing. And while pricing issues persist, I have seen a couple of examples already of vendors making the pricing situation better. One of those vendors is Box, which has a very specific set of guidelines on how many credits certain actions consume. This is the type of simple, transparent action that other vendors can learn from; you can find that list here.

I keep getting asked whether zero trust survives agents. The question always assumes that you carve out an exception, but I feel like almost nobody is carving one out. Watch how an agent gets to production. It either gets a service account with broad standing permissions, or it runs inside the session of whoever launched it. Either way, it holds that access for a long time. Nothing re-checks it until the credential expires. That’s the access model that zero trust was built to kill, but with the “zero trust” name still on it. Organizations have always tended to over-grant access. It almost never showed because of how people go looking for answers. We fall back on the same kinds of data over and over. Someone holding far more access than the job warrants still only pulls what they think they need. By contrast, an agent will use all of it. I’d hang permissions on the job the agent was handed. Identity tells you which agent acted. The permissions should come from the task and expire with it.

The same problem sits in the file estate, and Nasuni just bought DryvIQ to address it. Nasuni sells the platform that holds the files. It has never known what’s inside them. DryvIQ reads the content itself, across more than 40 places where files live, including systems Nasuni doesn’t sell. It also acts on what it finds. It can quarantine a risky file. It can pull back permissions that got handed out too widely. Those wide permissions are the norm on a file-share. People get access to a whole folder at a time, and sensitive documents end up in the wrong folders constantly. An agent pointed at that folder reads everything in it. DryvIQ works today on its own, but Nasuni plans to fold it deeper into the platform over the coming months.

I’m working on a deeper write-up of Google Cloud’s recent agent news, because I don’t think it got the attention it deserved. Two of the announcements would each be a story on their own. Gemini Enterprise for Legal and for Financial Services both landed in preview, and Google calls them the first in a series, with Healthcare and Life Sciences named as what follows. Packaging an industry-specific offering is a different business from selling a general agent platform, where every customer builds the industry pieces themselves. Google has also started selling agent work by usage, with no upfront commitment and no base subscription fee. Per-seat pricing describes a person, but an agent does the work with nobody sitting in front of it. Project-spend caps went live too, so an agent’s calls to the model stop when a project hits its monthly ceiling. Neither vertical package carries a price yet.

In other news, last week I had a chance to speak with CIO.com about Broadcom expanding its relationship with Kyndryl. This reflects a couple of important realities. Enterprises moved onto VMware Cloud Foundation quickly, but modernizing on it is the harder stage. And plenty of companies have lost their deep VMware talent, so that work stalls. Broadcom putting real money behind training a few thousand Kyndryl consultants answers that scenario directly. It also helps Broadcom, because delivery capacity usually decides whether a platform gets used well. All of this is happening on estates enterprises already run. Private cloud isn’t new. Most large U.S. enterprises have had one for years. And the new AI work is now landing on this same infrastructure.

LEAP opens in Riyadh this week, and HUMAIN has the August 31 main stage to reveal what it’s calling its full stack, with AMD headlining alongside. If LEAP isn’t on your radar, it should be. Last year, it drew more than 200,000 people, which makes it one of the largest tech events anywhere. And yet it still gets almost no U.S. coverage.

Saudi Arabia designated 2026 its Year of AI. The timing isn’t accidental.

HUMAIN set the table last week with Microsoft. ALLAM (Arabic first language) models in Microsoft Foundry, and HUMAIN engineers paired with Microsoft’s forward-deployed teams. Details are a little thin, with both companies promising more at the show.

For HUMAIN, this is the move from building infrastructure to running a business. Two assets underneath it are hard to copy. First, power and land at a price that Europe can’t approach on this timeline. Second, geography, since Riyadh sits inside the latency envelope for a very large share of the world, including a European market with the AI demand but none of the spare power generation to serve it. Everything above the infrastructure is a longer build. Agentic products and an AI-native ERP pit a 15-month-old company against incumbents that already own those customer relationships.

For enterprise IT, the shift is that compute is turning into a portfolio decision. Inference is continuous opex, and it grows with adoption, so cost-per-token lands in the procurement model whether anyone planned for it or not. If the latency budget tolerates a round trip to Riyadh, cheap energized power starts winning on its own.

So can the Kingdom become the world’s AI hub? Announcements won’t tell us. Three things will. (1) A cost number with a method attached. (2) Megawatts energized and populated rather than allocated. (3) Movement on the Global AI Hub Law, because data embassies are what make any of this portable outside the region.

Mistral put real money into HUMAIN datacenters. That’s the signal I give the most weight.

VMware Explore runs this week in Las Vegas, and I think this year’s edition is the most consequential one VMware has had. The plenary leads on private AI cloud and agentic innovation, with 400-plus breakouts sessions behind it. And with VMware Cloud Foundation (VCF) 9.1 shipping in May, the sessions should be about how the platform behaves in the field rather than what’s coming.

Two things make the week matter more for me than the announcement list will.

The calendar is first. vSphere 8 hits its end of general support on October 11, 2027. It was also the last release you could license perpetually. So the path forward means subscription, and the destination is a bundle rather than a point upgrade. Migrations take months, not weeks. That puts the decision inside the next budget cycle for most IT shops.

The thing is, nobody is merely choosing an upgrade path here. They’re choosing a platform for the next ten years.

AI is the second thing, and it’s changing what enterprise IT expects the platform layer to do. Broadcom has done real work here. The Platform Engineering 2.0 framework it helped shape is a serious argument about who supplies governance, cost control, and policy once developers and agents both start shipping at volume.

My question is what shows up in product. Provisioning GPUs is the easy part. What does the VCF control plane do about attributing AI spend, normalizing telemetry across silicon and clouds, and enforcing policy where the workload actually lands? That’s where organizations stall.

Nutanix aimed straight at this window at .NEXT in April. Zero-copy migrations off vVols, bare-metal Kubernetes, an agentic AI platform. Real-world value. The migration pitch works because the deadline is real. The control plane pitch is the harder one, and nobody has closed it.

I’d like Broadcom’s answer to be framed in platform terms rather than feature terms. I’ll be on-site. More next week.

IBM disclosed a dual-architecture processor at Hot Chips 2026, with every core of its next-generation processor running both z/Architecture and Arm. No emulation, no translation layer, and no block of Arm cores sitting off in a corner of the die. Each core speaks both architectures and switches as the hypervisor dispatches work.

IBM says the part is 2nm, maintains north of 5.7 GHz continuously, and gives every core 36MB of private L2. Additionally, a new inference card succeeds Spyre with 16 active AI cores and 96GB of high-bandwidth memory. No date, no price, no Arm performance data. That’s normal for a Hot Chips disclosure — so don’t read too much into this early lack of information.

Cool technology. But does it matter? I think so. And it comes down to data.

Performance stopped being the mainframe’s problem a long time ago. Software availability is the problem. Consolidation projects don’t stall on the core application. They stall on the monitoring platform, the endpoint agent, the backup tool, and the database nobody ever ported to s390x. Arm inside the core rather than beside it keeps the consolidation math intact while opening the box to everything that already builds for arm64. That’s a leak closed for IBM.

It’s also a credibility milestone for Arm in regulated infrastructure, and I don’t think that aspect is getting enough attention.

Now for the enterprise argument. Mainframe-housed data is business-critical by definition: transactions, policies, accounts, and the record against which everything else reconciles. Any serious AI effort eventually needs it, yet today it gets copied somewhere else first. Meanwhile, the AI serving stack targets Arm by default. Inference servers, vector stores, orchestration, the Python long tail.

Native Arm means that stack runs next to the record instead of a hop away. No egress, no extra copy of regulated data, no network in the path.

It’s fair to press IBM about the rest. Timing. How a core running Arm code gets counted for third-party licensing. Distribution certification. Whether Arm guests get the same clustering and disaster recovery treatment as everything else. And the like. These are important details. But this is also a 2027–2028 buy. The inventory work is a now thing.

AWS just acquired DuckLabs. Most companies running DuckDB never chose it. I keep finding it inside products I cover, tucked into tools people bought for other reasons. It’s a free database that runs inside an application rather than on its own server. For years, that was mostly a developer convenience. Then agents turned it into a cost question. A cloud warehouse bills you for compute that sits running, no matter how small the question. That was fine when people were the ones asking. Analysts send a few hundred questions a day and think about each one. An agent doesn’t work that way. It pokes at small sets, throws most of the answers away, and asks again. Thousands of those questions add up to a bill for work nobody uses. By contrast, DuckDB has nothing to keep alive between questions. AWS was already a customer paying for some of it. Now the people setting the schedule work for AWS. I wrote up all the details in a Field Note.

Cloudera’s Anywhere Cloud is the other topic I wrote up last week. The recently launched platform runs Cloudera’s data and AI software on whatever infrastructure a customer already has. Could be a public cloud, a private datacenter, or a network with no internet connection at all. Services come from a marketplace and get picked one at a time. Need streaming and nothing else? Deploy streaming and nothing else. Buying a data platform has always meant buying its engines too, but Cloudera drops that connection. Bring your own, or opt for one of the partners it has certified. I like the timing on this move. Cloudera argued for years that serious enterprise data would stay put. Most of the market disagreed then, but has come around since. Anywhere Cloud shows up with that argument already won.

At Gamescom last week, NVIDIA released its much-awaited DLSS 4.5 Ray Reconstruction neural rendering technique, which uses a second-generation transformer-based denoiser, after talking about it since CES in January.What I think made people most excited about this release is that it supports at least three of its most important features across the entire GeForce series, from RTX 20 up to the latest RTX 50 series — something that won’t be possible with DLSS 5, which is coming soon. Features like Frame Generation will be available only on the RTX 40 and 50 series, with Multi-Frame Gen available only on the 50 series. DLSS 4.5 Ray Reconstruction should considerably improve both the performance and quality of the gaming experience for GeForce gamers and help them get more out of their gaming PCs. Given the current state of affairs in PC gaming, anything like DLSS 4.5 should make gamers happy, and especially mobile gamers, who benefit more from DLSS’s performance improvements. NVIDIA was already demoing DLSS 4.5 Ray Reconstruction at Computex 2026 back in June, running on RTX Spark systems, so that platform should benefit from this announcement as well.

Qualcomm Arduino VENTUNO Q
Arduino (a subsidiary of Qualcomm) opened worldwide pre-orders for the much-anticipated VENTUNO Q single-board computer on August 25, with the announcement made in Turin. The board pairs a Qualcomm Dragonwing IQ-8275 processor, rated at up to 40 dense TOPS, with an STM32H5 real-time microcontroller. The board comes with enough memory to be useful — 16 GB LPDDR5 and 64 GB eMMC, plus expandable M.2 NVMe storage. Connectivity includes Wi-Fi 6 tri-band, Bluetooth 5.3, and 2.5 Gb Ethernet, with three MIPI CSI camera connectors. The board ships preloaded with Ubuntu Linux, distributed by Canonical, and bundled with an Ubuntu Pro license that includes 10 years of security maintenance. Debian support is planned. The microcontroller runs Arduino Core on Zephyr RTOS for real-time control of motors, CAN-FD, PWM, and high-speed GPIO. Price is $299, which includes the memory. Initial pre-orders sold out immediately.

Arduino App Lab offers pre-integrated models including Qwen 3, Gemma 4, Whisper ASR, MeloTTS, MediaPipe gesture recognition, and YoloX object tracking. Developers can bring models from Hugging Face or Qualcomm AI Hub, or train custom models in Edge Impulse Studio. Qualcomm GenieX, an open-source on-device generative AI runtime, deploys LLMs and vision-language models (VLMs) in a few lines of code. App Lab is optional; the board can run standard Linux tooling.

Qualcomm announced the “Works with Arduino” program alongside the board. The program certifies third-party systems-on-modules as compatible with the Arduino ecosystem, and applications built in App Lab run on a certified module without code changes. SECO and Toradex are among the first participants, with Toradex building on the Aquila module architecture. Those modules are not yet available.

My take: VENTUNO Q made my Embedded World top-10 list in March, and the shipping product confirms my assumptions. Qualcomm isn’t just selling a chip and a reference design. Qualcomm is selling a multivendor chipset, a development environment, a model library, a training path through Edge Impulse, and a certified route to production silicon. Arduino supplies the developer base and the tooling, and Edge Impulse supplies model training. Both companies are Qualcomm subsidiaries. This is the first product where the acquisitions show up alongside Qualcomm’s AI-accelerated Dragonwing silicon in a platform that a developer can buy — and afford.

This board is also the clearest evidence yet for a thesis I have been tracking for several years. Chip companies are no longer just selling chips. The competitive product is now the platform, which includes operating systems, model libraries, development tools, and a documented path from prototype board to production module. And in this case, a large and active global developer community. I believe Qualcomm has assembled more of these stack components than anyone else selling edge silicon.

I’m watching two things. The modules from SECO and Toradex have been announced but are not yet available, so the production path is not fully mapped. And a sold-out first run suggests that thousands of developers will jump on this board right away and start building apps. These folks will be very vocal, and I’ll be listening and joining them as soon as my board arrives.

Samsung took the wraps off the Galaxy S26 FE, which is the company’s more affordable version of its flagship phone that it launched earlier this year. That translates to a $699 price, down from the S26’s $799 to $899, making it slightly more accessible while including some noticeable cost-reduction component choices. The CPU is now a Samsung Exynos 2500 in the U.S. market, while the display is a larger 6.7-inch OLED panel, offering more value than the standard S26. Memory and storage are also reduced, to 8GB of RAM and 128GB of storage, while the camera configurations remain largely unchanged. With the larger body, the battery capacity increases to 4,900 mAh, potentially offering longer battery life than the original S26. I believe Samsung has done a good job with the S26 FE, adding value in some areas (like screen size and battery) while balancing it with storage and memory reductions to reach a more accessible price point. I think we’ll see Samsung be more creative with its FE offerings like this for as long as the industry remains in a storage and memory crunch.

Qualcomm’s 6G Leadership Day in San Diego demonstrated the company’s strong position in 6G, both in terms of hard tech and in the R&D and IP that it has developed over the years. The company also discussed some expected issues with spectrum and where it expects 6G to be more successful than 5G. As with any new cellular generation, many of the things we are seeing in the later stages of 5G look a lot like what will be mainstream in 6G. The most notable thing is that AI will likely be everywhere in 6G, both as a workload and as an accelerating technology to enable new features, performance, and efficiency. There was also a healthy amount of shade thrown on GPU-accelerated approaches from NVIDIA and partners like Nokia. Qualcomm also gave live demonstrations of 6G features like integrated sensing and communication (ISAC), using an actual drone, cars, and a giga-MIMO antenna. This demo also used a Qualcomm add-in card with its own accelerator technology, which I believe will compete directly with NVIDIA’s GPUs in the radio access network (RAN).

HP’s earnings showed record revenue, even with PC units being down 16%. The company’s Personal Systems unit recorded $11.8 billion in revenue, with commercial sales growing 22% and consumer sales growing 10%. Revenue growth came almost entirely from increases in memory and storage prices, so the operating margin was only 4.6%. The stock dropped 10% after earnings but recovered in the days that followed, suggesting an unrealistic reaction to the realities of the PC market and how OEMs are handling the situation. I believe that right now PC OEMs are going through one of the toughest times in the history of the industry, but I also believe this is temporary and should end up as just another crisis they end up working through.

Apple’s launch of the M5 Ultra and M6 Apple Silicon chips as Mac Studio and Mac Mini PCs shows that the company is doing what it can to satisfy the demand for local AI compute. Even though Apple is launching these products — and even swapping pending M5 Mini orders for M6 chips — it will probably not be able to satisfy demand. Apple isn’t launching the 512GB M5 Ultra Mac Studio until October, likely because it won’t be able to supply many units, and 512GB is a lot of RAM to source for a single system. The M6 also serves as a new baseline for Apple’s Mac line of PCs across desktops, laptops, and even tablets. I will say that I was expecting Apple to go with the M6 in an iPad Pro first, as it has in previous years, but the Mac Mini approach also makes sense from a yield perspective.

Neutral-atom systems have emerged as a leading contender for achieving fault-tolerant quantum computing (FTQC). This modality uses lasers to trap and control thousands of identical, pristine qubits in flexible 2-D and 3-D grids. QuEra Computing is already demonstrating how neutral atoms can be dynamically rearranged mid-calculation to run complex quantum error-correcting codes, which could make reliable, large-scale quantum computation practical sooner than previously thought.

Quantum scientists use tightly focused laser beams (also called optical tweezers) to trap and hold individual neutral atoms in place. These function much like a grid of microscopic tractor beams. This technique has become one of the most promising tools for building large-scale quantum computers, simulating complex materials, and creating ultra-precise atomic clocks. Another advantage is that researchers have learned how to freeze atomic vibrations at their absolute lowest energy level and rearrange them on the fly.

Precise control of the positioning of atoms and their movements provides three major capabilities:

• Shuttling atoms during calculations: Instead of forcing atoms to remain in place and limiting their interactions to only immediate neighbors, lasers can move atoms across the processor even mid-computation. This allows any atom to interact with any other, making complex quantum programs much easier to run.
• Using the atom’s motion to store extra data: In addition to storing quantum information in the atom’s internal energy states, scientists can also encode data in how the atom oscillates back and forth inside its laser trap. That is an important capability because it shields quantum information from errors.
• Tunneling between traps: A distinctive characteristic of quantum mechanics is that particles can tunnel through barriers. By tunneling between nearby optical traps, the hardware mimics electron behavior in materials, which avoids cumbersome mathematical workarounds.

A Big Problem
Because atoms have mass, moving them takes time. If the move takes too long, then the quantum state will decay before the calculation finishes. Speeding up the process to improve reliability remains a challenge that researchers must solve. They are working on increasing the speed of moving atoms with several practical engineering tricks.

Optimal control shapes trajectories to minimize heating and residual vibrations for an atom by using smarter acceleration curves. This concept can be explained using an analogy: Everyone knows that if you move an open cup of water too quickly, it spills. And if you move it slowly, it takes longer. By calculating the proper acceleration and braking curve, scientists can figure out the top speed that will allow the atom to cross the chip and come to a dead stop without it vibrating out of place.

Raising the laser power increases the hold on an atom by creating a deeper, stiffer well that retains the atom more securely. That tighter hold enables researchers to accelerate and decelerate the atom much harder without losing control.

Since heavier atoms have greater inertia and move more slowly, switching to lighter elements (like lithium) instead of heavier ones (like rubidium or cesium) makes the atoms much nimbler and quicker to move.

Rather than moving atoms one by one with a single laser, optical conveyor belts use moving light waves to shift entire rows of atoms simultaneously.

Better chip layouts — shorter commutes — can be achieved by redesigning the processor layout so atoms don’t have to travel as far across the chip in the first place. This cuts transit time significantly.

Looking to the Future
As companies like QuEra continue to improve the neutral-atom modality, that modality is positioned to lead the transition from today’s noisy experimental devices to large-scale, fault-tolerant quantum computers. Here’s what I think could happen:

• Near-term (next one to three years): Acceleration curves and optimized zone layouts are expected to further reduce shuttling latency by an order of magnitude. Those changes should enable processors to run deeper error-correcting circuits without requiring further hardware overhauls.
• Longer-term (three to five years or more): Integrating lighter atomic elements with optical conveyor belts should increase atomic transit speeds, bringing them even closer to fundamental physical limits. When those changes are combined with 3-D array architectures, it should become possible to build quantum processors with tens of thousands of physical qubits.

The result should be sustained, high-speed simulations for materials science, chemistry, and complex optimization.

Ahead of its Snapdragon Summit in Hawaii, Qualcomm disclosed details on its next-generation Oryon CPU architecture for upcoming mobile platforms, headlined by a 5 GHz clock speed and a new dynamic cache architecture dubbed “Oryon FlexCache.” Moving past the previous generation’s 4.74 GHz ceiling, Qualcomm is bringing the 5 GHz capability proven on its PC silicon (Snapdragon X2 Elite Extreme) down to the mobile envelope, which should help it maintain a competitive edge against Apple and rival Android SoC vendors. To balance high clock speeds with power efficiency and memory constraints, Qualcomm is introducing a shared, dynamic cache pool (likely targeting approximately 24MB of L2 and 8MB of L3). This is designed to allow high-demand prime cores to dynamically access larger cache blocks, minimizing off-chip memory access and keeping intensive workloads on-die. The architectural combination of 5 GHz clocks and unified on-chip cache should directly benefit multi-agent AI execution (agentic handoffs), frame consistency in mobile gaming, rapid multitasking app switching, and local video rendering.

Beyond that, Oryon is now firmly established as Qualcomm’s unified CPU architecture across both PC and mobile. While hitting the 5 GHz benchmark sets a critical high-water mark against Apple’s A-series silicon, the real differentiator should be FlexCache’s ability to optimize daily thermal and power efficiency across complex multi-threaded workloads. Full platform disclosures and precise cache metrics are slated to follow at Snapdragon Summit.