- AI Rundown
- Posts
- The AI Rundown by Lightscape Partners - 11/17/25
The AI Rundown by Lightscape Partners - 11/17/25
SoftBank doubles down on OpenAI, Anthropic ramps a $50B compute build-out, and Parallel arms agents with live web access
Good morning and welcome back to another edition of The AI Rundown by Lightscape Partners.
SoftBank is reshaping its AI bets, selling out of Nvidia and committing $22.5 billion to lift its OpenAI stake to ~11%, in a deal that implies a $500 billion valuation. It’s a clear signal that some large investors now see more upside in owning the application and platform layer than in riding GPU upside alone.
Anthropic, meanwhile, is planning $50 billion of U.S. data center investment with Fluidstack, building AI-optimized campuses in Texas and New York. The scale still trails Meta and OpenAI’s ambitions, but it underlines how access to power and cooling is becoming as strategic as model architecture.
Parag Agrawal’s Parallel raises $100M at a ~$740M valuation to wire agents directly into the live web. By turning real-time search and browsing into a clean API for coding, analytics, and risk agents, Parallel is aiming to become default “internet lungs” for the next generation of autonomous AI workflows.
Stay tuned as we explore these stories and their implications for the future of AI, technology, and innovation.
If you haven’t yet, please support the newsletter by subscribing!
Hardware + Software
Baidu unveils new M series AI chips and supernode systems to ease reliance on US hardware. Link.
The M100 inference chip arrives in early 2026, while the larger M300 for training follows in 2027.
Baidu is also launching supernode systems that connect hundreds of chips, enabling large scale clusters for AI workloads.
The company positions the designs as domestic alternatives after US restrictions tightened access to advanced Nvidia accelerators.
Baidu pairs the hardware with a new Ernie multimodal model that handles text, images, and video understanding.
Google introduces Agent Sandbox and GKE Pod Snapshots to harden and accelerate Kubernetes based AI workloads. Link.
Agent Sandbox extends Kubernetes so developers can run AI agents in isolated sandboxes without risking core production systems.
The tool supports agents that browse the web, access databases, or invoke tools which might otherwise threaten security.
GKE Pod Snapshots can capture a running container image with an AI model loaded, then restart it within seconds.
Google claims snapshots cut startup for some large models by about 80 percent, reducing cloud costs and latency.
Models
Baidu open sources multimodal Ernie 4.5 VL to rival larger Western vision language models. Link.
The 28 billion parameter Mixture of Experts model activates only about 3 billion parameters per query for efficiency gains.
Baidu reports strong scores on benchmarks for image, chart, and video understanding compared with models like GPT 5 and Gemini.
The company says Ernie 4.5 VL runs on a single 80 gigabyte GPU, lowering hardware requirements for multimodal workloads.
Released under an Apache 2.0 license, the model is free for commercial use, inviting adoption by enterprises and developers.
Product Launches
Google launches Gemini powered Ads Advisor and Analytics Advisor to make campaign management more conversational. Link.
Ads Advisor lives inside Google Ads, answering natural language questions and recommending bid, budget, and creative optimizations.
Analytics Advisor in Google Analytics can explain anomalies, surface trends, and generate charts or comparisons from plain English prompts.
Both assistants use Gemini models to interpret advertiser intent, lowering the learning curve for small teams managing complex accounts.
Google plans a global rollout for English accounts, positioning AI copilots as the default interface for advertising workflows.
Enterprise + Consumer AI Applications
Extreme Networks previews Extreme Exchange, a curated marketplace for enterprise ready AI agents and tools. Link.
Customers will access pre built agents for networking, security, and analytics inside ExtremeCloud, avoiding complex custom integrations.
Extreme plans to certify third party AI tools so they respect network policies, telemetry formats, and existing management workflows.
The marketplace is designed to reduce AI sprawl by centralizing discovery, procurement, and deployment of agents on one platform.
Extreme says every agent will plug into its fabric and management plane, automating network tasks with minimal scripting effort.
Google turns Gemini shopping tools into conversational assistants that search, call stores, and even auto checkout. Link.
Shoppers can describe desired products in natural language and get tailored results from more than 50 billion indexed listings.
A Let Google Call feature uses an AI agent to call local stores, confirm stock or hours, then text updates.
Users can set a price target, letting an agentic checkout tool automatically buy items with Google Pay when discounts appear.
Early partners include Wayfair and Shopify merchants, signaling how retailers may offload more shopping journeys to AI agents.
Content Creation
ElevenLabs opens a marketplace for licensed AI celebrity voices, backed by Matthew McConaughey. Link.
The platform sells official voice models for actors and public figures, covering use cases like audiobooks, ads, and trailers.
ElevenLabs says deals include consent and compensation arrangements, aiming to address concerns from Hollywood unions about AI cloning.
Investor partner Matthew McConaughey will offer his own model and use it to localize his newsletter audio in Spanish.
The launch suggests growing mainstream acceptance of synthetic voices as brands seek scalable production without losing recognizable talent.
Data Centers + Energy
Anthropic plans 50 billion dollars in US data center spending with cloud partner Fluidstack. Link.
The companies will build AI optimized facilities in Texas and New York by 2026 to host Anthropic workloads.
Anthropic frames the investment as securing long term access to compute for future Claude models amid capacity crunches.
The scale still trails rivals, with Meta reportedly targeting 600 billion dollars and OpenAI eyeing a 500 billion project.
Fluidstack says its liquid cooling designs will cut energy waste, making large language model clusters more efficient and sustainable.
Startup Funding & Valuations
AI sales startup 1mind raises 30 million dollars to build autonomous enterprise sales agents. Link.
The Series A round is led by Battery Ventures and supports a platform that handles inbound sales conversations automatically.
Co founder Amanda Kahlow previously built account based marketing unicorn 6sense, bringing deep go to market experience to 1mind.
1mind’s agent, Mindy, can qualify leads, run demos, and respond to objections while coordinating handoffs to human sellers.
The company pitches its AI as a way for sales teams to cover more inbound interest without adding headcount.
Parag Agrawal’s startup Parallel raises 100 million dollars to power AI agents with live web access. Link.
The Series A funding, co led by Kleiner Perkins and Index Ventures, values Parallel at about 740 million dollars.
Parallel builds web search APIs that let AI agents query current internet data rather than relying on static training snapshots.
Enterprises can use the platform to equip coding, analytics, or risk assessment agents with up to date information sources.
The funding will help Parallel scale infrastructure and partnerships as competition intensifies around agentic access to the open web.
WisdomAI secures 50 million dollars for hallucination resistant AI analytics on messy enterprise data. Link.
Kleiner Perkins led the Series A, with Nvidia’s NVentures participating, just six months after a 23 million dollar seed.
WisdomAI lets users query disparate business systems in natural language, returning answers backed by verifiable structured data.
The platform uses large language models only to generate database queries and text, avoiding free form model generated facts.
Customers include mid market and enterprise firms that want AI speed without the compliance risk of hallucinated analytics.
CloudX raises 30 million dollars to automate mobile ad pricing with Claude based agents. Link.
Founded by former MoPub and AppLovin executives, CloudX builds agents that continuously test ad pricing strategies for app publishers.
The AI runs on Anthropic’s Claude models and connects through an SDK to major ad networks and monetization partners.
By automating experiments that humans rarely revisit, CloudX claims publishers can improve revenue while also protecting user experience.
The funding supports go to market hiring and additional integrations as mobile games and utilities lean harder on ad income.
Alembic raises 145 million dollars to scale causal AI for enterprise decision making. Link.
New funding from Prysm Capital, Accenture, and others values Alembic at roughly thirteen times its previous round.
Alembic is deploying an Nvidia NVL72 SuperPOD supercomputer to run models that uncover cause and effect in business data.
Customers like Delta Air Lines and Mars use the system to understand which marketing or operational levers truly drive outcomes.
Alembic positions causal AI as a way to move beyond correlation heavy dashboards toward decisions grounded in experimental evidence.
Regulation + Legal
US judge lets X Corp antitrust lawsuit against Apple and OpenAI proceed to discovery. Link.
Judge Mark Pittman denied motions to dismiss from Apple and OpenAI, refusing to toss X Corp’s claims at this stage.
The complaint alleges the companies conspired to monopolize app distribution and AI chatbot markets, harming Musk’s rival xAI products.
Pittman emphasized the ruling is not a judgment on the case’s merits, only a decision to keep litigation moving.
The case now heads into evidence gathering, which could surface internal documents about Apple’s and OpenAI’s competitive strategies.
Safety + Ethics
OpenAI backs biosecurity startup Red Queen Bio to curb AI enabled bioweapons risks. Link.
OpenAI led a 15 million dollar seed round for Red Queen Bio, which builds tools to detect dangerous designs.
The company spun out of an mRNA therapeutics venture and focuses on screening DNA orders and lab workflows for misuse.
OpenAI frames the deal as part of a broader effort to raise the AI ecosystem’s resilience against catastrophic biological threats.
It follows a recent investment in Valthos, another biosecurity startup, signaling sustained attention to dual use AI risks.
OpenAI
OpenAI rolls out GPT 5.1 with new style controls and deeper reasoning options for ChatGPT. Link.
GPT 5.1 comes in Instant and Thinking variants, balancing faster responses against more deliberate reasoning for complex tasks.
Users can pick from tone presets such as Professional, Friendly, or Quirky and control response length directly in settings.
ChatGPT can now suggest style changes mid conversation, for example asking whether to be more concise, technical, or casual.
The release is rolling out first to paid subscribers on web and mobile, with broader availability expected after initial feedback.
SoftBank commits 22.5 billion dollars to lift its OpenAI stake while exiting Nvidia completely. Link.
The Japanese group will own roughly 11 percent of OpenAI after the deal, up from about 6 percent previously.
SoftBank raised cash by selling its remaining Nvidia shares for 5.83 billion dollars and trimming its T Mobile US stake.
The investment follows a restructuring that valued OpenAI around 500 billion dollars in a new deal with Microsoft.
SoftBank’s CFO said the firm is reallocating capital toward OpenAI and other strategic tech bets while preserving balance sheet strength.
Thank you for reading the AI Rundown by Lightscape Partners. Please send any questions, comments, or suggestions to [email protected].
