Several layers of the stack moved at once
The most useful way to read the week of 12-14 October 2026 is as a build-out reaching past the GPU. Cisco described new networking silicon and optics for AI data centers.1 Hesai launched a data engine meant to feed robots with real-world training material.2 Zeta and Parloa, with TELUS Digital, shipped software aimed at running business processes with AI agents.3,4 No single announcement is decisive. Together they show where the industry thinks the next bottlenecks are: moving data between chips, getting real-world data to train on, and turning models into products that complete tasks.
One caution applies throughout. Nearly every item below is a company press release distributed through a newswire. Our measurement of that source class found that 56% of 4,960 checked claims held up.1 That means roughly four in ten claims did not pass. Treat the announcements as what companies say they are doing, not as confirmed results.
Layer one: open networking at OCP 2026
At the OCP Global Summit, Cisco introduced new silicon, systems and optics for open AI infrastructure. The company said the first G300-powered systems are shipping to customers and that it has new 1.6T G302 silicon. It also said P200-powered systems are deployed in production.1 Cisco and NVIDIA also said they plan to develop an open industry specification for a next-generation 400G-per-lane pluggable optics design, aiming to raise bandwidth density while keeping serviceability and customer choice.1
Cisco's Kevin Wollenweber framed the shift this way: "AI is redefining what it means to scale infrastructure. It's no longer just about individual data centers or clusters, but how we connect compute at massive scale without compromising performance or resilience."5 The practical point for practitioners is that the constraint he describes is the network between accelerators, not only the accelerators. The word "plan" in the optics item matters. A planned specification is not a shipped one.
What Nvidia's verified numbers show
Our verified data comes from Nvidia's SEC filings. It tells a clear story about the scale of spending on delivering AI hardware, with one limit: the dossier has cost figures but no revenue figures, so we cannot say anything about margins.
Nvidia's annual cost of revenue was $16.621 billion in fiscal 2024,6 $32.639 billion in fiscal 20256 and $62.475 billion in fiscal 2026.6 That is roughly double each year, and more than 3.7 times the fiscal 2024 level over two years. Quarterly figures follow the same direction: $15.157 billion in Q3 fiscal 20266 and $20.458 billion in Q1 fiscal 2027.6 Cost of revenue is what it costs a company to make and deliver what it sells, so a rise this steep means a much larger volume of product is moving.
Cash is less dramatic and moves around. It was $7.28 billion at the end of fiscal 2024,7 $8.589 billion in fiscal 20257 and $10.605 billion in fiscal 2026.7 The latest quarter in the dossier, Q1 fiscal 2027, shows $13.237 billion,7 below the $15.234 billion recorded in Q1 fiscal 2026.7 Cash can fall for many reasons, and the dossier does not say why, so we do not draw a conclusion.
Earnings per share needs care. The filings show $11.93 for fiscal 2024,8 $2.94 for fiscal 20258 and $4.90 for fiscal 2026.8 Read naively, that is a collapse and then a recovery. Given cost of revenue quadrupled over the same span, a collapse looks implausible. The dossier does not explain the pattern, and we have not checked whether share counts or reporting basis changed. Do not compare these three figures directly until that is confirmed.
For readers wondering whether they hold this indirectly: the dossier gives no index weightings, so we cannot say how much of any fund this represents. Our knowledge graph also links Nvidia to a software layer around its chips, including NeMo Guardrails, the NeMo Agent Toolkit, Nemotron 3 Super and DGX Spark, and lists AMD as a competitor.9 That is a map of relationships, not a performance claim.
Layer two: data for physical AI
On 14 October Hesai launched Kosmo, a 3D world data engine. It pairs the Kosmo One AI spatial camera with the Kosmo World cloud platform to turn physical environments into 3D digital spaces for robot training, world-model development and immersive content.2 Hesai also announced partnerships with embodied-AI companies Galbot and Dexmal, and showed applications built with Pixar's RenderMan and Amap.2 Hesai's Shaoqing Xiang stated the problem: "Physical AI needs real-world data that is accurate, affordable and available at scale. Meeting all three requirements has been a persistent problem."2 This is a data-supply claim, not a capability claim. The press release says Kosmo is designed to solve the problem. It does not show that it has.
A sober counterweight sits in the same dossier. A 8 October MIT Technology Review newsletter item was headlined "AI breakthroughs in robotics won't change your life any time soon."10 We have only the headline and excerpt, and that source has a poor measured record in our checks (0% of 11 claims held up), so we cite it only to show that the optimism around physical AI is contested, not to support any particular claim.
Layer three: agents as software
Zeta introduced AthenaOS, which it calls an enterprise intelligence operating system. CEO David A. Steinberg said: "For decades, software forced people to work the way applications were designed, but AthenaOS changes that. AthenaOS marks an expansion of Zeta's intelligent AI infrastructure beyond marketing."3 The significant detail is the scope: a company known for marketing technology is presenting a general layer that starts from the outcome a user wants.
TELUS Digital announced a partnership with Parloa and becomes an implementation partner. It will introduce enterprise clients to Parloa's AI agent platform and provide integration, change management and managed services.4 Chris Silver's quote argues the market case: "consumer expectations are high, and brand loyalty is fragile."4 The structure matters for adoption. Selling agent software through a services firm suggests the hard part is integration and change management, not the model.
On the demand side, a Snipp survey of 1,000 U.S. adults who regularly use AI for search and shopping found that 61.8% had purchased after an AI recommendation.11 The sample was people who already use AI regularly, so it cannot be read as a measure of the general population.
Safety and governance remain unresolved
Our topic summary for this week also includes OpenAI's capability gains, a safety-lead resignation, leadership pledges on monitoring and open-source timelines, and hardware-cost stories involving Acer, Micron and Samsung. The dossier contains no source document for the resignation, the pledges or the memory-price items, so we do not state specifics about them.
What we can quote is OpenAI chief research officer Mark Chen, speaking about the company's response to a hacking incident: "I do kind of reject the premise that OpenAI is a company with visible impacts in the world and therefore OpenAI is not training safe and aligned models."12 In another interview he said: "If you disappeared OpenAI, that would be bad for the world."13 Both are undated in our records, and both are self-description by an executive of the company under scrutiny. They show the company's stance, not independent evidence of safety.
What to watch
- Whether Cisco's G300 shipments and the 400G-per-lane optics specification turn from "plan" into published, multi-vendor standards.
- Whether Kosmo shows customer results from Galbot, Dexmal or others, rather than launch-day partnerships.
- Nvidia's next filing: whether cost of revenue keeps roughly doubling, and a clear explanation of the earnings-per-share pattern.
- Primary-source reporting on the OpenAI safety departure and the memory-price warnings, which this dossier cannot yet support.