The pitch has changed: from what models can do to what it takes to run them
Read the announcements of the last three weeks together and a pattern emerges. The headline is no longer a model benchmark. It is a firewall service, a partner programme, a data pipeline and a chip. Fortinet launched the FortiGuard Inline AI Security Service, which brings AI-specific inspection and guardrails to FortiGate firewalls through a cloud-delivered service.1 Cisco made Cisco IQ generally available to partners and expanded its Cisco Cloud Control offerings so partners can build, secure and manage AI-ready infrastructure.2 The common message is that AI is now something a company has to secure, plumb in and pay for.
A caveat on method first. Most of the material below comes from company press releases carried by NewsEOD. Our measured reliability check found that 54% of 6,613 claims checked from that feed held up.1 The Crunchbase News piece fares worse on a much smaller sample: 49% of 41 checked claims.4 Press releases state what companies want believed. We separate what is quoted, what is a reported figure, and what we cannot confirm.
Security: AI is being sold as an add-on to infrastructure you already have
Fortinet's pitch is about where AI security should sit. Michael Xie said: “Organizations need AI security that can scale with how quickly AI is being adopted across the enterprise... By extending advanced AI inspection and guardrails through FortiGate, we are making AI security available through the infrastructure customers already use.”1 The release also says FortiAIGate is the foundation of the strategy. It lets customers choose how to deploy and consume AI security based on cost, performance, privacy and sovereignty requirements.1 For practitioners, the point is that guardrails are being built into network equipment instead of into each model or application. The dossier gives no adoption figures or independent test results for the service, so how well it works is unproven here.
Cisco's version is commercial. Tim Coogan said: “Our partners are where Cisco meets the customer, and the AI era is asking more of all of us.”2 The update ties the Cisco 360 Partner Program to Cisco IQ and Cisco Cloud Control.2 This is a channel story: vendors are reorganising how AI infrastructure gets sold and supported.
T-Mobile supplies the one result in this set that reads like an outcome. Its published report, Advancing Cybersecurity at T-Mobile, covers how it rebuilt its security programme, what held up when tested and where the work goes next, including Bug Bounty programme results.3 Jeff Simon said: “Today, phishing-related credential compromises among our people are down more than 99%.”3 The accompanying claim says teams skipped pilots and multi-year rollouts and moved faster than most companies would try because of the stakes.3 This is T-Mobile's own figure in its own report. The dossier does not say what the baseline was or who audited it, and the report is not specifically about AI. It does show the scale of change large organisations are now attempting under security pressure.
Physical AI: the shortage is data, and it is expensive
The most candid technical commentary comes from Ulrik Stig Hansen of Encord. His argument is that physical AI, meaning robots and embodied systems, cannot lean on text-scale data. “Physical AI is very different because there is no internet of physical AI data, and so the data first has to be collected, and you have to convert it from sensor tokens to action tokens.”4 He also says the industry is still in the pre-training phase of physical AI and will soon enter post-training.4 On cost, he said: “Storing all that data, curating all that data, enriching all that data, and validating that all the data works for these systems is a very expensive operation.”4 That is a data-infrastructure bottleneck, not a model-capability one, and it matches the framing executives are using elsewhere.
Money is arriving. Crunchbase data cited in the report put global venture funding to robotics startups at $9.9 billion in Q3 2026, more than double the $4.3 billion in Q3 2025, with the bulk going to companies based in China and the US.4 The article's headline cites a 185% jump. The excerpt we hold does not say what that figure is measured against, and $9.9 billion versus $4.3 billion is more than double, not a 185% rise. Treat the headline percentage as unreconciled until the comparison is clear. The dollar totals are the safer anchor.
On the hardware side, Harvest Praxis launched its “Hyper AI Brain” at the 9th China Robot Summit in Hangzhou. It is an on-device computing platform built on MediaTek's 3nm Genio Pro 5100 SoC. The company says it is the world's first fusion of a robot's “brain” (perception and decision-making) and “cerebellum” (motion control) on a single chip.5 That “world's first” is the company's claim and we have not verified it. Founder Philip Cai Zhenhui frames the problem this way: “The industrialization of embodied intelligence relies on three critical pillars: large models, real-world data, and on-device computing.”5 The claim summary adds that he calls on-device computing the most fundamental yet weakest link.5 Two sources, one from a data company and one from a chip designer, point at different bottlenecks (data and compute). Both are about the cost of running AI, not about model intelligence.
Consolidation and the edge
Salesforce signed a definitive agreement to acquire Braze, a customer engagement platform, for about $5.66 billion in enterprise value, paying $47.50 in cash per share to holders of Braze Class A common stock.6 Braze founder Bill Magnuson said: “This past year we shipped at the fastest pace in our history, and BrazeAI is helping marketers get their best ideas live faster.”6 The claim summary adds that Salesforce's platform will help Braze innovate faster.6 The deal pairs a CRM giant that calls itself “the world's #1 AI CRM” with a company that already ships AI features. The dossier does not give a closing date or any regulatory detail, so the timeline is unknown.
At the consumer edge, Apple unveiled a new iPad mini with an 8.5-inch OLED display and the A20 Pro chip.9 Pre-orders opened October 28 and the device reaches customers on November 6.9 Our brief also lists NVIDIA RTX Spark laptops as part of this theme. The dossier we hold contains no source document or figures on them, so we do not describe them.
The small-cap test: what an “AI-driven” restructuring looks like in numbers
TEN Holdings (Nasdaq: XHLD) is a useful contrast to the large vendors. For the quarter ended September 30, 2026, revenue rose 17.9% year over year to $640,000.7 That is a very small business. The results also include a $1.8 million write-off of capitalised service contracts and about $0.5 million of legal expenses, both larger than the quarter's revenue.7 CEO Virgilio Torres said the company has cut its annual operating cost structure by about $2.5 million over 12 months “through strategic cost reductions, process improvements, and the integration of AI-driven tools to streamline engineering and accelerate product development.”7 The balance sheet moved sharply too: short-term related-party borrowings fell from $4.6 million to about $0.1 million, and total liabilities fell about 74%, from $7.7 million at year-end 2025 to $2.0 million.7 The dossier does not say how those liabilities were reduced, so we cannot say whether the cause is repayment, conversion or write-offs. It also does not break down how much of the $2.5 million saving AI tools produced. “AI-driven” here is management's label, not a measured result.
A bright spot that is not really about AI
Impinj, an RFID and Internet of Things company, reported record third-quarter revenue of $112.1 million, with record endpoint IC revenue and unit volumes.8 CEO Chris Diorio credited M800 market strength, resilient retail demand, lighthouse accounts and Gen2X momentum, and said trends should continue.8 Its Q4 guidance is $104.0 to $107.0 million in revenue,8 below the record Q3, so the company expects a step down next quarter. We include it because it sits in the same earnings season and the same hardware-led demand, but nothing in the dossier ties its results to AI specifically.
What to watch
- Evidence behind the security claims. T-Mobile's 99% figure and Fortinet's service both need baselines, test results or customer numbers before they count as proof.
- Physical-AI data costs. If Stig Hansen is right that the industry is about to enter post-training, expect demand for curation and validation tooling to show up in funding rounds. The next Crunchbase quarter will show whether the $9.9 billion level holds.
- Braze closing. The $47.50 per share cash price is set; terms and timing beyond what is stated are not in our material.
- TEN Holdings' next quarter. Whether revenue grows without more one-off charges will show if the restructuring is working.
- Impinj's Q4 against its $104.0 to $107.0 million guidance.