Google accused of poisoned search results, Salesforce explains Koa model-choice pitch, Netflix, IBM, Nvidia - Big Tech Report

Nozzle, a search engine results page monitoring firm, says it has found evidence that Google is deliberately serving manipulated results to rank-tracking tools and other automated services. Analysing 27.5 million SERPs, the firm found more than 660,000 keywords associated with poisoned results pages, with roughly 8 percent of SERPs affected across both desktop and mobile. Rather than blocking automated requests outright, as it has in the past, Google appears to have shifted toward swapping out normally high-ranking links for less relevant results, often pulled from social media platforms.

Google has long sought to disrupt AI scrapers that harvest its search data to train competing models or reverse-engineer its ranking algorithm, and Nozzle's findings suggest a tactical shift from blocking to data poisoning over the past month. The side effect, according to the report, is that legitimate SEO services and rank-tracking tools are being caught in the same net, receiving misleading data about where client sites actually rank. Google shares fell 1.58% to $348.18 in regular trading.

Salesforce has published an interview with Chief Data Officer Michael Andrew laying out why its CRM software lets customers switch between AI models, including its own newly built Koa, rather than committing to a single supplier. Koa is described as Salesforce's "first CRM reasoning model," built with NVIDIA and trained on simulations drawn from 27 years of Salesforce CRM workflow data, running inside what the company calls its "trust boundary" alongside frontier models from partners such as Anthropic and OpenAI rather than replacing them. The piece gives no parameter count, benchmark, pricing or availability date for Koa.

Andrew's broader argument is that deterministic CRM systems, which always return the same result for the same action, now have to coexist with probabilistic AI agents that behave more like a new hire than a software button. His pitch is that customers should pick whichever model suits a task's cost and speed requirements and keep changing that choice as the model market shifts. Salesforce shares were down 0.83% at $236.05 in regular trading.

Netflix is cutting roughly 1,000 jobs, about 6% of its global workforce, the Wall Street Journal reported Wednesday, citing an email sent to employees. It marks the company's biggest round of job cuts to date, exceeding reductions made during its 2022 subscriber slowdown.

Netflix shares were up 0.28% at $70.66 in regular trading.

IBM and Siemens are collaborating on research to advance nanostack, the three-dimensional nanosheet transistor architecture IBM unveiled earlier this year. The design vertically stacks and staggers transistors, packing nearly 100 billion of them into a fingernail-sized chip area at a 7 angstrom node, while allowing different materials within each layer to tune performance and power independently. Siemens will bring its electronic design automation, process simulation, lithography and manufacturing software to the effort, with the two companies building virtual models of nanostack to test performance and flag fabrication issues before chips are built.

IBM demonstrated nanostack's viability earlier this year; the new work targets the design, simulation and lithography advances needed to move the architecture beyond 7 angstroms and toward manufacturing at scale. IBM shares traded up 0.62% at $232.24 in the regular session.

Microsoft's cloud division has discussed significant cost reductions in its Azure server-rental business, with the unit's president asking leaders to find savings of around 20%, according to a person with direct knowledge of the matter cited by The Information.

The push for leaner spending comes as Microsoft continues to pour capital into data centres and AI infrastructure to meet demand for Azure's AI services. Microsoft shares were roughly flat, up 0.09% at $543.95 in regular trading.

American Electric Power plans to invest about $4 billion in a 702-megawatt fuel cell project near Cheyenne, Wyoming, to supply power for Google, the utility said on Wednesday.

The project is the latest sign of how surging electricity demand from data centres is reshaping utility investment plans, with power companies striking direct supply arrangements to meet AI-driven load growth. Google shares were up 1.01% at $357.36 in regular trading.

ASML Holding has raised its full-year sales outlook for the third time in 2025, citing demand for chipmaking equipment driven by the AI buildout. The company said it will be unable to meet customer demand for its machines in 2027 and 2028, pointing to a supply gap stretching years out rather than a near-term bottleneck.

ASML shares fell 1.86% to $1,717.76 in the regular session despite the upgraded forecast. The outlook also lifted European tech stocks more broadly, with shares in chip-equipment and semiconductor names tracking the demand signal; TSMC, the industry's largest contract chipmaker and a major buyer of ASML's lithography systems, edged up 0.17% to $445.12.

Nvidia CEO Jensen Huang on Tuesday proposed investments in four French startups — Doctolib, Genesis AI, Pigment and Entalpic — as part of a $2 billion funding push across Europe. The move extends Nvidia's strategy of taking stakes in AI-focused companies that build on its hardware and software stack, deepening its ties to the continent's startup ecosystem.

The French investments come alongside a separate $250 million compute donation Nvidia is making in the UK. Nvidia shares were up 0.97% at $228.52 in regular trading.

French cloud provider Scaleway will build an AI supercomputer in Sweden, its first deployment outside France and Europe's first commercial AI supercomputer built on AMD chips. Announced at the company's annual AI conference in Paris, the system will run on AMD Instinct MI355X GPUs, which compete with Nvidia's Blackwell line, and will be open to startups, AI companies and businesses across Europe. Scaleway picked Sweden for its largely low-carbon electricity and developing AI sector, building on a partnership with telecom operator Tele2 that brought its cloud and AI services into the Swedish market in May.

French firm Bull will assemble and maintain the machine, with servers from Giga Computing, networking from HPE, storage from VAST Data and cooling liquid from Motul. AMD's EMEA senior vice president, Stephanie Dismore, said the deployment widens AI infrastructure options in Europe through Instinct GPUs and the open ROCm software ecosystem. AMD shares were up 1.90% at $605.53 in regular trading.

Google DeepMind has demonstrated Gemini 4 Argon converting ordinary video into training material for humanoid and quadruped robots, compressing a process that normally takes weeks of manual simulation work into a few hours. In one demonstration, the model watched a video of a person throwing beanbags at a cornhole board, mapped the motion onto the joints of Apptronik's Apollo humanoid, built a physics simulation to refine the throw, and transferred the result to the physical robot, which then sank seven consecutive shots. In another, Argon directed a Boston Dynamics Spot quadruped to build a digital replica of a pedal bin in the MuJoCo simulator, test its balance on three legs, then step on the pedal and drop a soda can inside after checking its wrist camera.

Google said Argon scored 87% on its Embodied Reasoning benchmark of 3D spatial and physical problems, outperforming other frontier models tested on the same questions despite receiving no robot-specific tuning. From a single photo of a table, the model can generate a simulated version in MuJoCo, devise more than 4,500 task variations, and train a vision-language-action model that lifted a robot's real-world success rate on the task from 56% to 75% while shaving 11 seconds off execution time. Argon can also retro-label footage of a robot in motion, down to finger-level contact with objects, and reproduce published robotics control papers in a browser-based MuJoCo environment from a text prompt. Access is available now through the paid Gemini API and to Google AI Ultra subscribers.

London Stock Exchange Group has partnered with Nvidia to let financial institutions analyse historical market data using GPU acceleration, cutting processing times from hours to seconds. Benchmark tests show years of LSEG's Tick History data — 110 trillion rows spanning 100 million instruments, more than 30 years and over 580 exchanges and venues — can be processed in seconds using Nvidia's cuDF software library, rather than the hours required by conventional methods.

The collaboration brings the data closer to the computing environments used for quantitative research and AI without requiring firms to duplicate the full dataset, preserving existing licences and access controls. LSEG said traffic on its real-time network rose 29% in the first half of 2026 versus a year earlier, peaking at 26 million messages per second in June, while API requests for Tick History data climbed 48%. Expected uses include quantitative strategy testing, order book reconstruction, liquidity analysis and historical market replay. Nvidia shares were up 0.97% at $228.52 in regular trading.