AI Daily Digest · November 3, 2026
Seven stories from the past week, curated by KD Agentic · AI Daily Digest.
1. OpenAI and ServiceNow take their partnership into the enterprise workflow
On October 28, OpenAI and ServiceNow announced a major expansion of the partnership they first struck in January. The core of the deal is a new integration that brings ServiceNow business context and workflow actions directly into ChatGPT. An IT engineer on call will be able to find similar incidents, assess the risk of a change, and update a ServiceNow record without ever leaving the chat window. ServiceNow's Action Fabric applies the customer's existing business rules and permissions, so the actions an agent can take still reflect the organization's own access configuration.
Beyond the ChatGPT integration, the two companies laid out three product lines. ServiceNow's AI Workflow Factory plans to integrate Codex so developers can build applications with ServiceNow's platform context, and OpenAI models will also power the Build Agent and Autonomous Engineer. A joint security solution will combine OpenAI's cybersecurity models with ServiceNow's asset, identity, and business context to investigate vulnerabilities and coordinate remediation. A voice experience in ServiceNow's Super Intelligence Control Tower, where people describe an AI use case aloud, is expected in the first quarter of 2027.
The rollout is phased. The ServiceNow MCP plugin is available in the OpenAI marketplace now, security capabilities enter early access in January 2027 with general availability planned for Q2 2027, and Codex inside the Workflow Factory arrives during 2027. ServiceNow is also giving all of its employees ChatGPT Work and Codex, extending access beyond engineering, and OpenAI itself is using ServiceNow workflows in its own operations. After Uber moved 30,000 employees onto ChatGPT Work and Oracle embedded OpenAI products in its apps, this deal is another sign that enterprise workflow vendors, not just cloud providers, are becoming distribution channels for frontier labs.
— OpenAI
🔗 OpenAI
2. Sharon AI and NVIDIA plan a 72MW Australian AI factory with up to 40,000 GB300s
On October 31, Australian AI cloud provider Sharon AI announced a six-year strategic compute collaboration with NVIDIA, unveiled through a Business Wire press release. The agreement covers 72 megawatts of new data center capacity in Australia, deploying NVIDIA's DSX AI factory design and scaling to as many as 40,000 Grace Blackwell GB300 GPUs to serve AI startups, enterprises, and university researchers in the region.
The commercial structure is the interesting part. Sharon AI will sell NVIDIA-powered cloud services while NVIDIA earns both standard product revenue and a share of the cloud revenue on the supported capacity, backed by a credit-support model. NVIDIA has been quietly building this kind of usage-linked earnings stream across the neocloud ecosystem, and this is one of the clearest formalizations of it: the vendor helps finance the buyer's capacity and then participates in the buyer's upside.
Following the deal, Sharon AI's total AI factory capacity expands to 132MW, of which 102MW is already contracted to end customers, and the company expects more than 55,000 NVIDIA GPUs deployed by mid-2027. Sovereign AI remains the framing: an Australian-certified NVIDIA Cloud Partner building local capacity so regional customers are not entirely dependent on US-based clouds. Whether the revenue-share model proves durable across a full six-year cycle is the question worth tracking, but the capacity itself is real and contracted.
— Sharon AI · Business Wire
3. Anthropic files to sell up to $3.5B of new shares to Japanese investors
On November 2, filings submitted in Tokyo showed Anthropic plans to offer as much as 552.9 billion yen, roughly $3.5 billion, of new stock to Japanese investors as part of its US IPO process. The filing states the proceeds from the newly issued shares will fund working capital, operating expenses, and capital expenditures. Specific IPO size and valuation remain undisclosed in the document.
The Japan allocation is notable in its own right. When SpaceX went public in June, Japanese investors ultimately received about $2.2 billion of allocations. Anthropic's planned offering to Japanese institutions exceeds that figure, which makes it one of the largest Japan-side allocations ever prepared for a US tech listing and a deliberate signal to a market that has been hungry for growth assets.
The broader IPO picture is still moving. Media reports have suggested the listing could come as soon as mid-November, with investor discussions centering on a valuation around $1.5 trillion, down from the roughly $2 trillion floated earlier in the autumn. Anthropic's second quarter revenue of $11.6 billion and its first quarterly operating profit are on the table, but so is a heavy infrastructure commitment profile. How the books get priced will set the reference point for every AI financing that follows, including OpenAI's own reported fundraising.
— SEC filing
🔗 Nikkei coverage · Anthropic S-1 announcement
4. Hugging Face launches Project ProtoStar to make public science AI-ready
On October 29, Hugging Face announced Project ProtoStar, an initiative to make publicly funded scientific research discoverable and reusable in AI-ready, machine-actionable formats. The founding partners include Carnegie Mellon, Purdue, Georgia Tech, Arizona State, and George Washington universities through the Open Forum for AI, plus the Astera Institute and the Open Research Community Accelerator, with Hugging Face as infrastructure partner. Participating organizations formalize commitments through memoranda of understanding.
The first open challenge, hosted by Carnegie Mellon, opens November 2 and runs through November 13. Participants get historical weather data from 210 US cities and must forecast four indicators, including average annual temperature and cooling degree days, for the already-observed 2019 to 2026 period. Because the targets are historical records, forecasts can be scored against actual weather rather than projections, and a public leaderboard compares AI approaches against simple statistical baselines. Teams also propose a short climate adaptation policy idea informed by their analysis.
The problem ProtoStar targets is real. The US federal government invests roughly $190 billion a year in research and development, and agencies like NIH and NSF now require public sharing of funded research, but datasets sit in legacy formats with thin documentation, which makes them hard for both humans and AI systems to reuse. Hugging Face hosts more than 17,000 public science datasets, researchers shared three times as many new science datasets in the first nine months of 2026 as in the same period of 2025, and 90 percent of dataset downloads come from fully documented datasets. Documentation, it turns out, is the product.
— Hugging Face
5. IBM ships Granite Switch 4.1.1, updating a model like a software patch
On October 29, the IBM Granite team published Granite Switch 4.1.1 on Hugging Face, a patch release built on the modular Switch architecture introduced on October 15. The version number means what it means in software: fully backwards compatible with 4.1, no migration required, incremental improvements under the hood. The 30B-parameter, 128K-context base model is untouched.
Two components changed. The Granite Speech ASR model now handles multilingual speech-to-text with native punctuation and capitalization, so transcripts are ready for downstream models without a formatting pass. The Decision adapter in granitelib-core was retrained and posts improved performance on the Jev Decision Index test set. Everything else, including the guardian-core safety adapter, stayed the same. IBM's point is the release process: apart from training the new ASR model and the decision adapter, nothing was recomputed. You compose the adapters you want and deploy.
The team is working on tutorials and an Adapter SDK so developers and third parties can build, version, and ship their own adapters the way they ship software libraries. That is a quiet but consequential shift in how open models get maintained. If models become base weights plus swappable adapters with semantic versioning, the release cadence of AI stops looking like foundation-model launches and starts looking like point releases, with all the pinning, changelogs, and compatibility guarantees that developers already know how to consume.
— IBM Granite · Hugging Face
6. GXO adds two more warehouses to Agility's Digit rollout
On October 26, GXO Logistics and Agility Robotics announced an expansion of their humanoid deployment to two additional US sites, in Statesville, North Carolina, and Lebanon, Tennessee, with installations beginning in early 2027. The sites will start on Digit 4 hardware and migrate to the newer Digit 5 platform once it is commercially available. The announcement came through a formal GlobeNewswire press release from both companies.
The expansion rests on an operating record rather than a demo. At GXO's Flowery Branch, Georgia facility, where the industry's first commercial humanoid deployment and first robots-as-a-service agreement went live in mid-2024, Digit has moved more than 100,000 totes working alongside existing conveyor automation, and Agility counts over 65,000 hours of real-world operation across its customer programs. Agility's RoboFab factory in Salem, Oregon is designed for up to 10,000 robots per year to support the scale-up.
The honest complication is the economics. GXO's own CEO said on an August earnings call that the company had run 45 humanoid pilots, had not yet achieved ROI on humanoids, and estimated being a couple of years away. Signing two more sites anyway is a bet on the learning curve and on Digit 5's cooperative-safety design, not a claim that the machines pay for themselves today. Agility is also preparing to go public through a business combination with Churchill Capital Corp XI expected to close in the fourth quarter, trading on Nasdaq as AGLT, so the commercial evidence base is about to face public-market scrutiny.
— GXO · Agility Robotics
🔗 GXO press release · Container News
7. Hugging Face ships on-device speech and decision packages for Apple platforms
On October 29, Hugging Face released two Apache-2.0 Swift packages aimed at Apple developers. AnySpeechRecognizer mirrors Apple's SpeechAnalyzer API from macOS 26 and iOS 26 but turns the recognition engine into a pluggable choice: NVIDIA's Parakeet TDT 0.6B v3 covering 25 European languages, Moondream's post-trained Parakeet Ultra, English-only Parakeet Unified, and Alibaba's Qwen3-ASR in 0.6B and 1.7B sizes, which covers 30 languages and 22 Chinese dialects that Parakeet misses. All run locally on Core ML. Switching engines changes one import line.
The benchmarks are pointed. On LibriSpeech test-clean files transcribed on an M4 Max Mac Studio, Parakeet Unified scored a 1.43 percent word error rate against 2.12 percent for Apple's built-in transcriber, and an hour of file audio takes roughly 20 to 40 seconds. Apple's API has no speaker diarization, so the package adds one based on NVIDIA's Streaming Sortformer that tracks up to four overlapping speakers.
The second package, AnyDecisionModel, answers yes-or-no, multiple-choice, and rating-scale questions by reading the probabilities of allowed answer tokens instead of generating text, returning typed results with per-option probabilities and entropy-based confidence. It supports six backends including MLX, Core AI, llama.cpp, and decision-head models like Cloudflare's Clef, and adds temperature scaling calibration plus rotation debiasing to counteract position bias. Together the packages push a familiar argument: task-specific small models, run on device, beat general-purpose LLM calls on both cost and accuracy for narrow jobs.
— Hugging Face
KD Agentic · AI Daily Digest
