Daily AI News — 20 Oct 2026

Nvidia on 19 October launched new open-weight Nemotron 3.5 models for always-on agents. Super VL is a 120-billion-parameter model built for enterprise agents that, Nvidia says, matches the accuracy of its far larger Nemotron 3 Ultra on instruction-following and tool use, and it claims it completes tasks faster than Anthropic and OpenAI models when post-trained for specific jobs like document intelligence, financial accounting and insurance claims. Two smaller Bolt models are made to run locally on new RTX-powered AI PCs and DGX Spark workstations, keeping work off the cloud. Nvidia named Skydance, EY, Ramp, CrowdStrike, Perplexity and Salesforce among early users.

Enterprises increasingly want to run capable agents on their own infrastructure rather than rent closed APIs. A strong open model from the dominant chipmaker, tuned on a customer's own data and owned outright, lowers the barrier to doing exactly that.

A finance or insurance team can post-train Super VL on its own documents and deploy it inside its own data centre, keeping sensitive records in-house while automating claims review and accounting.

Anthropic said on 18 October it will commit 5% of its shares, subject to stockholder approval, to a new Anthropic Public Benefit Fund structured as a wholly owned LLC. The board may add another 2% of shares for every 50% rise in the company's value above a baseline it has not yet disclosed, so the faster Anthropic grows, the more equity flows to the fund. Possible uses include large donations to nonprofits, direct payments to specific groups and 'bolder approaches' to AI-driven job disruption, building on an earlier $200 million labour-market commitment. None of the equity will fund lobbying or political giving. With co-founder pledges and an employee match, roughly 20% of shares are now earmarked for public purposes. CNBC reports the IPO is slated for November.

As AI companies approach trillion-dollar valuations, how they share the gains and cushion the disruption becomes a live political question. A structural pledge tied to valuation, days before a historic IPO, sets a template others will be measured against.

A nonprofit or workforce-retraining group focused on AI's labour-market effects now has a named, scaling funding pool to approach, rather than relying on one-off corporate philanthropy.

A Google DeepMind team placed four Gemini 4 Argon agents in a fresh Minecraft survival world on Easy difficulty with one goal, defeat the Ender Dragon, and the agents did it in 27 hours and 14 minutes, according to a paper on arXiv (2610.22319). Nobody assigned roles: within ten minutes the agents split the work in chat, one on tools, one on food, one on diamonds, one on the Nether portal, logged coordinates and lessons to shared notes, retrieved items from dead teammates and wrote a new program roughly every 35 seconds for over a day. Each controlled a player through the open-source Mineflayer library. The caveats matter: it was a single Easy-difficulty run, the agents read privileged game state rather than pixels, and a pathfinding plugin was used.

The boss fight is a stand-in for the real prize: a team of general-purpose agents that divides labour, recovers from hundreds of failures and sustains a multi-day task with no human-written playbook. That shape, long-horizon planning and coordination, is exactly what enterprise and software work demands.

An engineering team can take the same pattern, several agents coordinating through shared memory and chat, to tackle a large migration end to end, dividing subtasks among themselves and recovering from errors without a human scripting each step.

Chinese lab MiniMax released M3.1 Flash on 19 October, a model built for enterprise workflows, coding and computer-use tasks. It is a 428-billion-parameter mixture-of-experts design with about 23 billion active parameters and a one-million-token context, priced at $0.15 per million input and $0.60 per million output tokens, roughly 70% cheaper than its predecessor while running nearly three times faster. Independent tester Artificial Analysis scored it 42 on its Intelligence Index, up 13 points from MiniMax M3, at less than half the per-task cost, putting it near DeepSeek and Google's Flash models. The same tests flagged a high tendency to answer rather than admit uncertainty. MiniMax says model weights will follow on 23 October.

The cost of running capable agents is falling fast, led by Chinese labs undercutting Western frontier prices. Cheaper, faster models widen who can afford to deploy agents at scale, but the reliability trade-offs mean accuracy-critical work still needs guardrails.

A startup running high-volume agent workflows, like customer-service automation, can switch to M3.1 Flash to cut token spend sharply, while adding verification layers for any task where a confidently wrong answer would be costly.

doubleAI, the stealth venture of former Mobileye CEO Amnon Shashua, unveiled its first model, Maxwell, on 19 October and disclosed a $200 million Series A co-led by Lightspeed and Bessemer, with Nvidia and Dell Technologies Capital participating, valuing the company above $1 billion. doubleAI calls its technology 'Code Execution Models', which operate between software and hardware: Maxwell analyses a running program, its hardware and its data, then optimises the underlying machine instructions to run faster without changing the source code, formally verifying that each change preserves the program's behaviour. The company says Maxwell has made data pipelines over 38 times faster and AI inference kernels over 4.5 times faster on real workloads, figures it concedes are not independently verified. doubleAI plans to charge based on the compute savings it delivers.

Compute is the biggest line item in modern AI, and most effort goes into writing code, not running it efficiently. A verified layer that squeezes more out of existing hardware attacks cost from a different angle, and a pay-for-savings model aligns the vendor with the customer's bill.

A team running expensive inference or data-processing jobs could route them through Maxwell to cut GPU hours and cloud cost, paying only a share of the savings, without touching its existing codebase.

Tenstorrent said on 19 October it raised $1.5 billion in a Series E round valuing the Santa Clara chip startup at $6.2 billion post-money, more than double its $2.6 billion valuation from December 2024. The round was led by Trace Capital and Innovation X, with Cathie Wood's ARK Invest, Sandisk, Morgan Stanley, T. Rowe Price, Cadence Ventures and Tether Investments among the backers. Run by veteran chip designer Jim Keller, Tenstorrent will use the money to scale production of its Galaxy servers, which combine dozens of its AI chips to power chatbots, coding assistants and video generators, and to develop the AI and CPU designs it licenses to firms building their own chips, including Samsung and LG.

The economics of AI hinge on who supplies the chips, and a well-funded alternative to Nvidia, built on open-weight-friendly hardware and licensable designs, gives data-centre operators and chipmakers more leverage on price and supply.

A cloud or enterprise operator planning an AI build-out can weigh Tenstorrent's Galaxy servers as a second source alongside Nvidia, while a chipmaker can license its CPU and AI cores instead of designing from scratch.

Dream Labs, founded by former Nvidia robotics researchers, announced on 19 October a $90 million seed round led by B Capital, one of the year's largest seeds for a company whose entire product is intelligence other people's robots can use. It is not building a robot; it is building Dream-1, a foundation model trained on video of people doing everyday tasks, wrapped in a cloud service, Dream API, that turns a robot's camera feed plus a plain-English instruction into motor commands. Reporting confirms Dream-1 can transfer skills across robots from different makers, including gripper-equipped machines it has not seen before, without retraining from scratch, though the company's broader 'any robot' framing is its own claim. Pricing is set at $3 per hour for Dream-1 and $2 for a smaller variant.

The hardware for useful robots largely exists; the missing piece is a general brain that works across different bodies. An hourly cloud model, billed like AI inference, could do for robotics what cloud computing did for software, letting operators skip building their own control stack.

A warehouse or small manufacturer with off-the-shelf robot arms could connect them to Dream API and direct them with plain-English instructions, paying by the hour rather than funding a bespoke robotics-AI team.

Amazon Web Services released TorchNeuron on 19 October, a tool that lets AI startups easily move their applications onto Amazon's in-house Trainium chips, directly targeting the software advantage that has made Nvidia the default AI hardware. AWS argues porting to Trainium is not only easy but gives startups access to Amazon's entire cloud, pitching it as a better choice than Nvidia-tied neoclouds like CoreWeave and Nebius. Amazon is also contributing changes upstream to PyTorch itself, working with the community to remove code that ties the framework to specific chips, with the stated goal of making PyTorch more hardware-agnostic for all developers rather than just Trainium users.

Nvidia's moat is as much software as silicon. If a major cloud makes switching chips genuinely easy and loosens PyTorch's Nvidia defaults, it pressures GPU prices and gives AI builders real hardware choice for the first time in years.

A startup facing high GPU bills can test porting its training and inference to Trainium with TorchNeuron, comparing cost and performance against Nvidia without rewriting its PyTorch code from scratch.

Inception Labs released Mercury 3 on 19 October, calling it the most capable diffusion-based large language model to date and its proof that text diffusion scales. Unlike the autoregressive models behind most chatbots, which generate text one token at a time, diffusion LLMs refine many tokens in parallel, which Inception says lets Mercury 3 run four to eight times faster, around 1,068 tokens per second on Nvidia GPUs, while matching or beating GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks. The model has a one-million-token context and is available now through Inception's API and via OpenRouter, Baseten and Vercel.

Speed and cost increasingly decide which models get deployed in real products. A credible high-speed alternative to the dominant autoregressive approach keeps pressure on latency and price, and validates diffusion as a serious architecture for text, not just images.

A developer building a latency-sensitive feature, like live coding assistance or real-time chat, can route requests to Mercury 3 for much faster responses, benchmarking its quality against mainstream models before committing.

Microsoft has ended its agreement to be the main customer at Nscale's planned data centre in Loughton, Essex, The Telegraph reported on 18 October. Nscale had pledged roughly £2 billion to build the facility, billed as the UK's largest AI data centre and meant to house about 23,000 Nvidia GB300 units, but grid officials said they could not supply enough power until the 2030s. Nscale called its Microsoft relationship 'strong and long-term' and said it is exploring alternative power sources for the site. The loss lands as Nscale prepares a New York IPO in which Microsoft and Anthropic account for about 85% of its contract value, and as some investors push for a lower valuation than the up to $35 billion the company has sought.

The AI build-out is colliding with a hard physical limit: electricity. A flagship project stalling for lack of grid power, and losing its anchor tenant, is a concrete sign that power availability, not chips or capital, may be the real bottleneck on how fast AI can scale.

An enterprise planning major AI capacity should factor grid connection timelines into site selection and vendor contracts, since a data centre's announced size means little if the power to run it is years away.