Daily AI News — 23 Oct 2026

At Nvidia's GTC conference in Berlin on 22 October, Saudi AI company HUMAIN announced it is expanding its Nvidia partnership to build up to 84 megawatts of AI computing capacity, about 27,000 GPUs, in two phases. Phase one is 14MW of Blackwell HGX B300 systems, roughly 6,000 GPUs, deploying from the fourth quarter of 2026 for HUMAIN Cloud customers. Phase two adds 70MW of next-generation Vera Rubin infrastructure, about 21,000 GPUs, from mid-2027, with 50MW reserved for an unnamed US customer. HUMAIN, owned by the Public Investment Fund, plans roughly 3GW of data-center capacity by 2030.

Compute is the scarce input of the AI era, and the Gulf is now buying it at national scale. HUMAIN's expansion shows Saudi Arabia intends to host foreign AI workloads, not just run its own, turning cheap energy and US chip access into a globally relevant cloud business.

An AI startup without the capital to secure Blackwell racks in the US could rent HUMAIN Cloud capacity to train or serve models, tapping sovereign Gulf compute instead of waiting in a multi-month GPU queue elsewhere.

On 22 October, OpenAI said Uber is rolling out ChatGPT Work company-wide, making the agentic platform available to more than 30,000 employees after an initial pilot of 1,000. ChatGPT Work takes on complex, multi-step knowledge tasks and connects to tools Uber uses daily, including Google Drive, Gmail, Google Calendar and Slack, letting staff research, prepare briefs, build presentations and analyse data across apps. GPT-6.1 Sol runs as the default model, balancing performance and cost, with voice and image generation available too. Uber's engineers continue using Codex for software work. It is Uber's first company-wide deployment of an agentic knowledge-work platform.

This is what enterprise AI adoption looks like when it stops being a pilot. A Fortune 500 operation standardising an agent for every employee signals that agentic tools are now core infrastructure, not experiments, and sets a template rivals will be pressured to match.

An Uber operations analyst can ask ChatGPT Work to pull data from Drive, cross-check a calendar, draft a stakeholder brief and assemble a slide deck in one workflow, instead of stitching together four separate tools by hand.

In research published on 16 October, Anthropic said an internal Claude model traced the origins of the mysterious seeds behind five NIST elliptic curves, P-192, P-224, P-256, P-384 and P-521, in under 30 minutes, settling a question open for roughly 25 years. Rather than chasing a long-rumoured secret-phrase theory, Claude spotted statistical patterns pointing to lrand48, an old C random-number generator, then used the LLL lattice-reduction algorithm to recover its internal state and reproduce the published seeds exactly. The starting values corresponded to Unix timestamps from 1997 and 1998. Cryptographer Filippo Valsorda independently reproduced the result; Anthropic found no evidence of a deliberate backdoor and says no deployments need changing.

The unexplained seeds had fuelled decades of suspicion that the NSA might have weakened these curves. Resolving it with ordinary timestamps eases that fear, but the bigger story is method: an AI independently chose a sophisticated attack strategy and executed it fast, a concrete sign of research-grade reasoning.

Security researchers can point Claude-style models at long-standing provenance and forensic puzzles, scanning for statistical fingerprints in legacy systems and reconstructing how old parameters or artifacts were generated far faster than by hand.

Sony AI published research in Nature, 'Outplaying Elite Table Tennis Players with an Autonomous Robot', describing Ace, the first robot to achieve expert-level play against elite and professional human players under International Table Tennis Federation rules. Ace pairs a high-speed perception system, nine active-pixel-sensor cameras plus event-based vision sensors for measuring spin, with a model-free reinforcement-learning control system and high-speed robotic hardware. It returned a wide range of spins with over a 75% return rate up to 450 rad/s, reacting to rare shots like net bounces. Sony says the work extends its earlier Gran Turismo Sophy research from the virtual domain into the physical world.

Table tennis demands perception, planning and action at the edge of human reaction time. Cracking it under official rules shows physical AI can now operate reliably in fast, unpredictable real-world settings, a milestone with implications well beyond sport.

The same low-latency sense-plan-act loop could drive robots that work safely alongside humans on a moving assembly line or in other dynamic environments where split-second reactions to unexpected motion matter.

Microsoft said it will roll out its in-house Maia AI chips at massive scale starting in 2027, making them the main hardware for running AI across its products and reducing reliance on Nvidia. The chips are already deployed in the US, Europe and Australia and deliver up to 1.7x more output per watt than leading GPU-based systems for AI features in Teams and Copilot, the company said. New server racks will hold 96 Maia chips each, twice as many as before, handling up to seven times the workload, with engineers focused on cutting the power wasted moving data inside machines, a growing bottleneck as AI tasks get more complex.

Nvidia dominates AI training, but inference, running models in production, is now the bigger prize, and every hyperscaler wants its own silicon there. Microsoft scaling Maia pressures GPU margins and signals custom accelerators are becoming table stakes for cloud giants.

An enterprise using Copilot in Microsoft 365 could get faster, cheaper AI responses as Maia silicon handles the inference behind the scenes, with the efficiency savings potentially passed through in cloud pricing.

On 22 October, OpenAI made its Decisions API generally available. Built on GPT-6 Luna, it focuses the model on questions with bounded answers, classify this, route that, pick the next action, and returns a probability, a choice, or a score up to 10x faster than calling Luna through the Responses API. Developers can pass native JSON, include images in instructions, and ask up to 400 questions per request. Pricing is $0.10 per million input tokens, $0.01 cached, with no output-token charges, plus US and EU regional processing. It integrates natively with GPT-Live for real-time voice apps.

Agents make countless small decisions, and at scale their latency and cost dominate. A dedicated, dirt-cheap decision endpoint lets builders route and classify without burning frontier-model budgets, a practical building block for production agent systems.

A customer-support bot can use the Decisions API mid-conversation to instantly decide whether to answer, ask a clarifying question or escalate to a human, without the lag of a full chat completion on every turn.

On 21 October, OpenAI and Google Cloud said they are partnering to bring OpenAI's frontier models and Codex workflows into Google Cloud's Model Garden, letting customers buy GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna through the Google Cloud Marketplace and draw on existing cloud spend commitments. It puts OpenAI's models on all three major US clouds, following Amazon in April after Microsoft's exclusivity ended. Google Cloud said it is now the only hyperscaler offering all three leading closed-source LLMs, OpenAI, Anthropic and its own Gemini, on one platform, with the OpenAI models also coming to its Antigravity coding tool and Gemini Agent. General availability is planned in the coming weeks.

AI's supposedly clean rivalries are really a web of dependencies, compute, distribution and customers cut across competitors. OpenAI models in Google's Model Garden show that meeting enterprises where they already run workloads matters more than brand purity in the race for adoption.

An enterprise standardised on Google Cloud can now evaluate and call OpenAI models from within its existing environment, on a single invoice with unified governance, instead of signing a separate vendor contract and wiring up another provider.

Zurich-based physical-AI startup Enact Intelligence came out of stealth on 22 October with $30 million in pre-seed funding to build robots that feel force and touch through their own motors, a deliberate break from the camera-heavy approach that dominates humanoid and manipulation demos. Enact says it is rethinking the entire robotics stack, from model architectures and data pipelines down to the actuators and sensors, starting with high-precision force data from bespoke actuators. Investor redalpine's Harald Nieder described it as tackling dexterity with a new data modality, building from the hardware up.

Vision alone struggles with delicate, contact-rich tasks, inserting a plug, handling fragile objects. Teaching robots to sense force could unlock manipulation skills cameras can't, and a large pre-seed signals investors see touch as a missing piece of general-purpose robotics.

A warehouse or manufacturing robot with force sensing could pick up and assemble fragile or irregular items, gauging grip pressure the way a human hand does, rather than crushing or dropping objects it can only see.

Boston-based AI startup ASI said on 20 October it raised $325 million in a Series C valuing it at $3 billion, amid growing appetite for AI used in real-world operations. The round was led by Spark Capital with affiliates of Citadel Securities, Meritech Capital and Craft Ventures. ASI builds simulation and monitoring software that helps operators forecast disruptions hours, days or weeks ahead across complex systems such as air traffic and power grids. Founder and CEO Phillip Buckendorf contrasted the approach with text-generating AI, saying ASI is focused on systems that manage physical assets. It recently won a 12-year, $875 million Federal Aviation Administration air-traffic-management contract.

Most AI capital has gone into language. A $3B valuation for software that predicts and steers physical infrastructure shows investors betting the next frontier is operational AI, systems that keep grids stable and skies safe, where reliability, not fluency, is the product.

A grid operator can use ASI's simulations to anticipate where demand spikes or equipment failures might cause outages days ahead, rerouting power or scheduling maintenance before a disruption cascades into a blackout.

In a roundtable about Google's new Gemini 4 model, Argon, which Google shared mid-October, co-founder Sergey Brin said AI development is accelerating internally and that models are now helping build their next versions. 'Our researchers and the model work hand in hand to make the next version of the model,' he said, describing a recursive-self-improvement (RSI) flywheel that speeds up each release cycle. Brin said that by the time people watched the Argon video, Google would already be 'a couple of models ahead' internally, and that he saw 'no slowing down' in the cadence teams were seeing.

RSI, AI improving AI, is one of the field's most consequential and debated ideas. A Google founder stating it is already happening in production, however carefully, is a notable data point on how fast capability may compound, and it lands amid industry debate over whether to deliberately pace the frontier.

Teams planning AI roadmaps should assume shorter gaps between frontier releases: budgeting, evaluation and migration processes built around annual model updates may need to handle a faster, rolling cadence of capability jumps.