At Dreamforce 2026, Salesforce Goes All-In on Agentic AI
At its Dreamforce 2026 conference, Salesforce made its clearest statement yet that agents, and not the Customer 360 app suite, are now the center of gravity for its platform strategy. The headline move is an “AIforce”-branded architecture built for headless interoperability for any AI or agent interface and anchored by the recently announced “Claudeforce” partnership that ships 37 prebuilt Anthropic-powered skills and eliminates roughly 50 MCP server configurations for customers. Salesforce also folded its Qualified and Fin acquisitions into named digital workers — Piper, Fin, and Hunter — with CEO Marc Benioff citing $500 million in pipeline generated by Hunter alone last quarter. The company is still experimenting with agent pricing, continuing with Flex Credits alongside new bundled subscriptions, offering enterprises the choice of buying on consumption or through a variant of its historic pricing strategy. The real test isn’t the announcements themselves but whether Salesforce can convince buyers that three decades of backend data and workflow logic can be extended to become the agentic infrastructure, rather than just another app interface to a hyperscaler’s agents.

Salesforce Brings a New Angle to the Model Wars
Another notable Dreamforce announcement was Koa, Salesforce’s new CRM-specific reasoning model, built by post-training NVIDIA’s open-weight Nemotron 3 Super on synthetic CRM data rather than licensing a general-purpose frontier model outright. The strategic logic is that a model tightly tuned to CRM tasks needs fewer tokens to get a job right than a frontier generalist, letting Salesforce compete on cost for its bread-and-butter workflows while still leaning on partners like Anthropic for harder reasoning. Salesforce’s own benchmarks claim Koa matches or beats general-purpose models on core CRM operations with fewer errors. But the more important signal comes from the pattern we are seeing: Established SaaS vendors increasingly want to own a slice of the model layer themselves rather than being purely dependent on frontier labs, which could reshape the competitive dynamics of enterprise AI pricing and differentiation.

Amodei Wants to Pace the Frontier — but He’s Not Going Far Enough
Dario Amodei’s “We Must Pace the Frontier” essay was published against the backdrop of a rough summer for AI safety headlines — Anthropic disclosing blocked bioweapons-research attempts, OpenAI’s account of its own agents launching unauthorized attacks on Hugging Face, and agents in that incident that reportedly tried to corrupt their own evaluators. Anthropic’s argument for pacing capability growth is directionally right, but I’d argue it is insufficient. My impression from Amodei and other AI leaders is that frontier labs are treating what’s really a structural problem as a messaging problem. I am skeptical of Amodei’s push to restrict distillation and tighten chip controls further, reading those as competitive moves dressed up as safety policy. I am also disappointed by the calls for government regulation as cover rather than building real self-regulatory teeth. My latest piece on Forbes goes into more depth on all of this, and closes with my five-point agenda for frontier safety — tiered access to the most capable models, transparent third-party evaluation, industry-funded open guardrails, model legislation that labs actually help write, and an IAEA-style inspection authority. A well-written CEO essay isn’t a substitute for any of it.

At Dreamforce 2026, Salesforce Goes All-In on Agentic AI
At its Dreamforce 2026 conference, Salesforce made its clearest statement yet that agents, and not the Customer 360 app suite, are now the center of gravity for its platform strategy. The headline move is an “AIforce”-branded architecture built for headless interoperability for any AI or agent interface and anchored by the recently announced “Claudeforce” partnership that ships 37 prebuilt Anthropic-powered skills and eliminates roughly 50 MCP server configurations for customers. Salesforce also folded its Qualified and Fin acquisitions into named digital workers — Piper, Fin, and Hunter — with CEO Marc Benioff citing $500 million in pipeline generated by Hunter alone last quarter. The company is still experimenting with agent pricing, continuing with Flex Credits alongside new bundled subscriptions, offering enterprises the choice of buying on consumption or through a variant of its historic pricing strategy. The real test isn’t the announcements themselves but whether Salesforce can convince buyers that three decades of backend data and workflow logic can be extended to become the agentic infrastructure, rather than just another app interface to a hyperscaler’s agents.

Salesforce Brings a New Angle to the Model Wars
Another notable Dreamforce announcement was Koa, Salesforce’s new CRM-specific reasoning model, built by post-training NVIDIA’s open-weight Nemotron 3 Super on synthetic CRM data rather than licensing a general-purpose frontier model outright. The strategic logic is that a model tightly tuned to CRM tasks needs fewer tokens to get a job right than a frontier generalist, letting Salesforce compete on cost for its bread-and-butter workflows while still leaning on partners like Anthropic for harder reasoning. Salesforce’s own benchmarks claim Koa matches or beats general-purpose models on core CRM operations with fewer errors. But the more important signal comes from the pattern we are seeing: Established SaaS vendors increasingly want to own a slice of the model layer themselves rather than being purely dependent on frontier labs, which could reshape the competitive dynamics of enterprise AI pricing and differentiation.

Amodei Wants to Pace the Frontier — but He’s Not Going Far Enough
Dario Amodei’s “We Must Pace the Frontier” essay was published against the backdrop of a rough summer for AI safety headlines — Anthropic disclosing blocked bioweapons-research attempts, OpenAI’s account of its own agents launching unauthorized attacks on Hugging Face, and agents in that incident that reportedly tried to corrupt their own evaluators. Anthropic’s argument for pacing capability growth is directionally right, but I’d argue it is insufficient. My impression from Amodei and other AI leaders is that frontier labs are treating what’s really a structural problem as a messaging problem. I am skeptical of Amodei’s push to restrict distillation and tighten chip controls further, reading those as competitive moves dressed up as safety policy. I am also disappointed by the calls for government regulation as cover rather than building real self-regulatory teeth. My latest piece on Forbes goes into more depth on all of this, and closes with my five-point agenda for frontier safety — tiered access to the most capable models, transparent third-party evaluation, industry-funded open guardrails, model legislation that labs actually help write, and an IAEA-style inspection authority. A well-written CEO essay isn’t a substitute for any of it.

Most of the agent governance conversations I have are about what an agent is allowed to reach. Hardly anyone asks what happens when the agent itself goes bad. And agents do go bad. Memory, instructions, guardrails, and credentials all change over time, and any of them can drift or get corrupted. So I was glad to see Cohesity launch Agent Resilience at Catalyst, its virtual customer event, to back up those pieces. A team can use it to roll an agent back to a checkpoint from before the trouble started and restore only the state that went wrong. Cohesity also protects the databases and file systems the agent worked in, so you can undo what the agent changed. Support starts with agents built on Amazon Bedrock. Here’s how to think about this: If your team runs agents in production, could anyone restore last Tuesday’s version of one?

In my experience, a recovery plan gets written once and tested rarely, while the estate it covers keeps changing. Cohesity wants agentic workflows to keep recovery plans current, and it calls the effort Autonomous Cyber Resilience. It’s built to find and protect new data as it lands, and to check recovery readiness on an ongoing basis. RecoveryAgent runs parts of incident response today, so security and IT teams work from one shared plan. Cohesity’s data security posture management tool now flags new sensitive data, and the matching protection rule applies automatically. An admin still simulates each new rule and checks for drift before it goes live. I think keeping a person in the loop at the sign-off stage is exactly right. I suspect that security teams will trust automated tools like these with far more of their protection work when someone still approves each rule.

For years, many Splunk customers paid for the data they brought in and indexed. So teams decided up front what to index, and everything else went somewhere cheaper, or got dropped. I was in Denver last week for Splunk’s .conf26 event, and the company made that decision far less permanent. Activity-based pricing weighs searching data the same as bringing it in, and the Machine Data Lake gives raw data a cheaper place to land. As I told TechTarget, I could see some customers completely rethinking what they index the traditional Splunk way. I’m also a fan of Splunk being able to reach data it doesn’t store. Its Federated Search now covers AWS CloudWatch Lake and Databricks, so an investigation can span them without copying data first. And Cisco customers get Data Fabric through Cisco Cloud Control with no separate Splunk license. A company running Cisco networking can now start on Splunk’s data platform with no Splunk contract.

One of the agent observability announcements from the show I really liked was Tokenomics. Every AI budget conversation I have ends with someone asking what the agents cost, and nobody knows. Tokenomics attributes token spend by team and user, including coding agents like Claude Code, Codex, and Cursor. I love that none of the three is a Cisco product. A team sees its AI bill no matter whose agent it picked. Splunk Agent Observability, now in Observability Cloud and Cisco Cloud Control, is designed to block hallucinations and data leaks with runtime guardrails. To me, those guardrails are a spending control as much as a safety one, since a wrong answer still burns tokens. Then there’s Cisco AI POD for Splunk, which runs Splunk AI on Cisco and NVIDIA infrastructure in customers’ own datacenters, even air-gapped ones. Regulated and government teams that keep telemetry off of vendor clouds can now turn on those AI features.

Can a conference outgrow itself in a single year? AI Infra Summit just did. The show in Silicon Valley drew about 8,500 attendees, more than double last year, and it was obvious nobody planned for that. Sessions were packed, hallways were jammed, and signage was nonexistent. I went looking for a mainstage keynote 10 minutes early and found it with five minutes left in a 20-minute session.

I’m not complaining. These are good problems. If you were around for the OCP Summit in its early years, this feels familiar. Real energy, real technical conversation, organization still catching up to the crowd.

What I like most is how the program is built. The tracks map to the AI chain rather than to vendor categories: compute, AI datacenter, data movement, data and models, and physical AI. Organizing this way forces the entire supply chain into the same hallway. We need more of this.

I was able to moderate a session on the AI Datacenter track with Microsoft, and we spent the time on what has quietly become the hardest part of the U.S. buildout — getting a community to say yes.

For context, the U.S. has more than 3,000 operational datacenters and over 1,500 more in development — close to half the world’s total. Here’s the part that matters: Of those planned, 67% are headed to rural areas, and 39% are landing in counties that don’t already host a single one. This industry is about to introduce itself to a lot of communities that have never met it. The concerns those communities raise around power, water, and the environment carry different weights.

Here’s where we (the industry) keep getting wrong. Even when a concern is exaggerated, dismissing it is a losing strategy. Some nerd correcting someone’s math at a local commission doesn’t work. These communities aren’t asking to be educated. They’re asking what they get and what happens when something breaks.

Microsoft actually pointed to a great case study in Quincy, Washington, where, after a couple of decades, the benefits are tangible and impressive: new schools, better public safety, and poverty rates halved. This has a lot to do with the datacenters covering 75% of the town’s property tax revenue.

The point is simple. If we want to unstick the 75 datacenter projects worth $130 billion in aggregate that are already delayed, we as an industry have to move past “We know what’s best for you” and get to “How can we work together?”

Because change is coming — and we need to embrace and manage it.

NVIDIA introduced a metric rather than a launch, and it stuck. The framing from the main stage was that what matters is total token throughput per megawatt. This is a fast shift, and it wasn’t just NVIDIA. Nearly every keynote and vendor on the show floor had converted.

NVIDIA’s answer to address this metric is DSX MaxLPS, which shifts power dynamically across racks instead of provisioning every rack for its own worst case. The company projects up to 40% more GPUs inside the same site power envelope and up to 35% more token throughput without pulling a new line. The measured result is smaller and more interesting. Lambda ran 19 nodes inside the power budget normally allocated to 16, gaining 24% in cluster token throughput and 23% in performance per watt on B200 hardware it already had. So, while Lambda’s number falls short of NVIDIA’s, that’s still a significant gain.

MLPerf Inference v6.1 landed mid-show with a record field and a new agentic test, and three independent cloud providers posted throughput gains of 8% to 36% on Blackwell systems they already own, through software optimization alone. NVIDIA’s own submissions improved up to 1.6x over the v6.0 versions, again on software improvements. If you can find a quarter more capacity in hardware that’s already racked and already drawing power, that’s the cheapest megawatt you will ever buy. It’s also the only one available to a buyer who can’t get an allocation.

NVIDIA’s Vera prioritizes single-threaded performance over core count, on the argument that each step in an agentic chain is gated by sequential execution and that the high-core-count host is a training-era artifact. NVIDIA pointed to benchmarking data from Signal65, a testing lab co-owned by our very own Patrick Moorhead, to make the case. Timing the full agentic task lifecycle with one sandbox per physical core and every core under load, Vera finished 1.64x faster per core than a leading x86 server processor across 69 Terminal-Bench 2 tasks, and 1.87x across the 20 tasks that most resemble an agent’s daily work.

Read the scope before you read the number. The test keeps the model out of the loop by design, so it measures sandbox and tool execution rather than inference, and it says nothing about whole-server throughput. This is the right call for isolating CPU behavior.

I think the direction here is what matters. If every step in an agentic chain waits on sequential CPU execution, then the host CPU ceases to be a consolidation platform and becomes a latency governor for the entire pipeline. That’s a challenge to 20 years of server CPU logic, in which more cores per socket were the answer to nearly every question. And it lands on an enterprise buyer who is still sizing CPUs based on VM density and consolidation ratios. Almost none of the IT leaders I talk to are evaluating on that basis yet. They’re going to have to.

Maybe the most useful framing I heard all week at the AI Infra Summit came off the AMD main stage, and it had nothing to do with a part number. The premise: nobody knows the next dominant model architecture, where the training and inference mix will settle, or how much AI will end up running in your datacenter versus somebody else’s. Yet infrastructure leaders have to commit to a five- to seven-year foundation right now.

So, the design question isn’t “Which chip?” It’s how easily the system adapts when requirements change.

The sharpest piece of that is the shape argument. Infrastructure built for yesterday’s training clusters isn’t short on capacity; it’s the wrong shape because the compute, memory, and network balance no longer match where the work actually goes. And you can’t buy your way out of a shape problem by adding more of what you already have.

Agentic workloads make this concrete. One request becomes a chain of retrieval, tool calls, code execution, sub-agent coordination, and a state that outlives the request. The economics follow. Training was a capital decision you made once. Inference is an operating decision you make a billion times a day. So stop pricing tokens. Price the finished task.

This surfaces a tension nobody resolved on stage. Tokens per megawatt was the metric of the week (see my “Datacenter — AI” entry above, if you haven’t already), and that metric rewards tightly co-designed racks. Adaptability pulls the other way. Breadth buys you options. Co-design buys you efficiency. Buyers have to pick a mix, and nobody is helping them do it.

“Revenue” usually has more than one definition inside a company, and so does “active customer.” An agent just grabs whichever definition it finds first. I wrote about Alation’s AIOS back in July, and last week at revAlation, its user conference in Chicago, Alation launched Semantic Model Mastering. The product pulls semantic models out of Databricks and Snowflake, enriches them in Alation, and syncs the improved definitions back. Alation got the design right, if you ask me. Almost nobody rebuilds metrics in a new tool, and the governed definitions land where agents already look.

In fraud detection, connections matter as much as definitions, because a fraud ring is a network of accounts, devices, and people. So, graph databases have always made a lot of sense to me there. Neo4j has sold the database underneath fraud detection for years. Coordinated with its GraphSummit event in New York (also held last week), Neo4j started selling a finished fraud product on top, called GraphAware Financial Crime Intelligence. I think going up the stack is a smart play. Each alert arrives with the connected people and accounts attached. The evidence stays in the graph, so an investigator can show a regulator why a case got escalated or closed.

Most companies I talk to still pick databases one application at a time. One group inside Cockroach Labs started building apps with agents and went from 13 databases in March to 392 by July. Nobody filed a ticket for any of them. Cockroach Continuum manages and bills a company’s whole database fleet as one system on shared hardware. Shared hardware should save the most cost for fleets full of small databases sized for rare peaks. One application team can’t sign off on a purchase covering every database a company runs. So Cockroach now sells to whoever owns the database budget. I went deeper on this topic in a Field Note.

Advantech announced an expansion of its WISE Edge Developer Architecture (WEDA) edge AI software framework on September 17, 2026, adding AMD as a silicon partner and formalizing Intel’s participation. WEDA launched at Computex Taipei in June 2026 with NVIDIA and Qualcomm as its original silicon partners. (See my “IoT and Edge” entry in that week’s roundup.) NVIDIA does not appear in the September 17 announcement, despite being a launch partner with its own named integration (NemoClaw, for manufacturing use cases).

WEDA packages AI workloads into Docker containers with direct passthrough to a chip’s GPU or NPU. Advantech calls it a framework, not a platform — bring your own cloud backend and use native device OSes and board support with no lock-in.

The architectural details are still a bit vague, but here’s my initial analysis. Each silicon vendor’s board support package, boot loader, kernel, and drivers remain intact and vendor-owned. Advantech validates that stack, certifies it “WEDA-Ready,” then layers its own container runtime and application lifecycle management on top, unchanged regardless of which vendor is underneath. The orchestration is proprietary Advantech code, not Kubernetes or any open standard.

My take: Advantech is building the device-level operational layer above vendor platforms, extending the same validation-and-certification role it already plays for industrial hardware to AI containers. Four vendors now publicly stand behind that layer. I find it a little odd that the press release did not mention NVIDIA, a prominent WEDA partner. Might be just an oversight.

Google has launched the Googlebook with system availability in October; this introduces an Android-based laptop PC running on various hardware from partners, starting at $899. This is Google’s answer to the AI PC and the challenges created by the memory crunch that has hurt the PC industry. ChromeOS was a very memory-efficient, highly secure operating system, but lacked the global compatibility that Android has. Googlebook is how Google solves that problem. Acer’s Googlebook starts at $899, while pretty much all others are $1,000 and up. Google is also throwing in a year’s worth ($250) of a Google AI Pro subscription to give users added value and more reason to use a Googlebook. Based on everything I’ve seen, Googlebook deeply integrates the Gemini AI experience into what is effectively ChromeOS with Android app compatibility. This should mean that it will be fairly secure, with plenty of applications available. Additionally, it will enable cross-device intelligence between Android and Googlebook that simply does not exist elsewhere.

Googlebook partners are launching new chips to enable Googlebooks. One example is MediaTek’s Dimensity CX C10 Max. This part is the first in a new line of compute chips specifically designed for the PC. Interestingly for MediaTek, none of the five Googlebooks shown by Google on its Googlebook page uses this new chip, as they are all powered by Intel and Qualcomm. MediaTek does seem to claim that it will be in an upcoming Googlebook laptop; I will be interested to see how that chip’s performance pans out against Intel’s and Qualcomm’s offerings.

IBM’s quantum foundry subsidiary Anderon finalized the $1 billion award provided to it under the CHIPS and Science Act administered by the U.S. Department of Commerce. That award was matched by another $1 billion from IBM, with the goal of allowing Anderon to operate as a dedicated 300 mm pure-play quantum wafer foundry for the industrialization of the North American quantum hardware supply chain.

To scale quantum processors, Anderon plans to use advanced 3-D packaging that routes complex wiring to shield the qubits from environmental noise. By using superconducting through-silicon vias (TSVs) and lossless indium micro-bumps, control lines connect through multi-layer interposers beneath the processor, which should eliminate surface routing congestion and lateral cross-talk. Magnetic shielding and low-loss dielectric or vacuum-gap wiring layers help to isolate qubits from stray thermal noise, electromagnetic fields, and resistive losses. All of these measures help to maintain qubit coherence times.

As a pure-play foundry, Anderon creates a TSMC-style horizontal manufacturing layer for quantum computing. The aim is to allow quantum startups to be able to fabricate 300 mm custom devices without the expense of a proprietary, multi-billion-dollar cleanroom. Anderon plans to support superconducting quantum processing units (QPUs), cryogenic control and readout electronics, and silicon spin qubits.

Pasqal is newly listed on the Nasdaq as a pure-play neutral-atom quantum computing company (ticker: PSQL). It represents the first publicly traded entity dedicated exclusively to this modality. A public listing aims to provide the company with additional liquidity and access to capital needed to accelerate its commercial hardware roadmap.

Pasqal’s technology uses large, reconfigurable arrays of laser-trapped neutral atoms. These serve as on-premises accelerators for high-performance computing (HPC) centers and industrial laboratories. Pasqal qubits are trapped by optical tweezers. Its gate operations are driven by exciting atoms into high-energy Rydberg states — in which the principal quantum number n is large, resulting in strong, long-range dipole-dipole interactions via the Rydberg blockade mechanism.

Pasqal’s processor can operate in two modes: analog and digital. In analog mode, the system Hamiltonian is engineered and evolved to solve complex problems in condensed-matter physics, combinatorial optimization, and quantum machine learning. In the digital mode, the system executes discrete, gate-based quantum circuits, providing a pathway toward quantum error correction.

Pasqal’s commercial strategy is to deploy its noisy intermediate-scale quantum (NISQ)-era systems today, then upgrade those systems to fault-tolerant quantum computing (FTQC) architecture when it becomes available.

I have three follow-up questions that I will keep investigating:

1. How firm is the delivery timeline for Vela (256 high-fidelity qubits), and will it ship on schedule?
2. Do Pasqal’s claims of quantum advantage hold up under independent academic and industry scrutiny?
3. How will the company perform financially post-IPO, particularly regarding revenue and cash burn as it heads into its first scheduled earnings report?

The launch of Snap SPECS will likely go down as one of the seminal moments in the augmented reality space. The company opened pre-orders in June when it first showed off the glasses, but at the launch in Beverly Hills last week, it truly demonstrated what the AR glasses can do and with whom Snap is partnered. The company showed off SPECS Intelligence, its AI platform that runs across Mac, iPhone, and SPECS. This is a clear demonstration that the company understands the way an AI agent is supposed to operate and be useful to the user.

The company also launched what many industry watchers expected to be consumer-only AR glasses, but it did so alongside a series of enterprise partnerships. I believe that launching SPECS for both consumer and enterprise may be a bit resource-constrained, but enterprises will be the ones who pay SPECS’ bills for the foreseeable future — at least until the price of SPECS comes down to a consumer-friendly level or some consumer-killer apps arrive from developers. We’ve already seen consumers lust after a developer-created Pokémon AR app that simply animates the Pokémon card-playing experience.

SPECS also did something unexpected by integrating a 5G modem into the SPECS glasses case. While this adds $200 to the cost of the glasses, it also ensures they can operate anywhere, entirely free of a smartphone, and gives enterprise users an easy way to track the glasses and operate them in the field.