Good morning and welcome back to another edition of The AI Rundown by Lightscape Partners.

• Shield AI raised $2 billion at a $12.7 billion valuation in the largest defense AI financing to date, with Advent International, JPMorgan Chase, and Blackstone anchoring the round. The deal signals institutional capital is flowing into autonomous military systems at a pace that mirrors early cloud infrastructure investment cycles.

• OpenAI killed its Sora video product and shelved plans for an adult content mode in back-to-back retreats, while its ChatGPT advertising pilot crossed $100 million in annualized revenue in just six weeks. The moves paint a clear picture of a company sharpening its product portfolio ahead of IPO scrutiny, cutting anything that complicates the narrative.

• Arm shipped its first production chip after 35 years of pure IP licensing, a 136-core data center CPU co-developed with Meta on TSMC 3nm. The shift from licensing to silicon signals Arm believes agentic AI workloads create enough durable demand to justify owning the full stack.

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

Arm Holdings shipped its first production chip after 35 years of pure IP licensing, a 136-core data center CPU co-developed with Meta and built on TSMC 3nm.  Link.

• Arm's blog announced the AGI CPU on Neoverse V3 cores, the first time in the company's history it has sold production silicon rather than licensing designs.

• Meta is the lead partner and co-developer, deploying the chip across gigawatt-scale infrastructure while pairing it with custom MTIA AI accelerators.

• Arm CEO Rene Haas told Reuters the chip should generate roughly $15 billion in annual revenue within five years across AI data center workloads.

• The shift from licensing to chips signals Arm believes agentic AI creates durable demand for CPU infrastructure it controls end to end.

Google published a compression algorithm claiming 6x memory reduction for AI inference with zero accuracy loss.  Link.

• Ars Technica reports TurboQuant uses a PolarQuant process, converting vectors to polar coordinates and applying quantization that preserves directional meaning efficiently.

• SemiAnalysis pushed back on bearish chip readings, arguing KV cache removal will motivate deploying larger models rather than using fewer chips.

• The key-value cache stores context to avoid recomputation and is one of the primary memory consumers during AI inference at production scale.

• The research signals inference efficiency is becoming a software-layer competitive advantage that could shift margin away from memory hardware manufacturers.

Models

Mistral released an open-source speech model that claims to beat ElevenLabs while fitting on a smartwatch.  Link.

• TechCrunch reports Voxtral TTS launched March 26 based on Ministral 3B, achieving 90ms time-to-first-audio and a 6x real-time rendering factor.

• The model clones custom voices from fewer than five seconds of audio while preserving accents, inflections, and natural speech quality.

• VP Pierre Stock told TechCrunch the model fits on a smartwatch at a fraction of the cost of competing proprietary voice API offerings.

• The open-weight release challenges ElevenLabs, Deepgram, and OpenAI by letting enterprises run voice AI on their own hardware without lock-in.

Product Launches

WordPress.com gave AI agents write access to create, edit, and publish content on live sites.  Link.

• WordPress.com added 19 new write operations to its MCP integration on March 20, covering posts, pages, comments, media, and settings.

• AI agents including Claude Desktop, ChatGPT, and Cursor can now draft, edit, and publish site content through natural language conversation.

• The platform previously offered read-only MCP access for analytics, making this a significant escalation in agent capability on the open web.

• The update signals a future where AI-managed websites become normal, raising questions about content authenticity and quality control at scale.

Enterprise + Consumer AI Applications

SAP agreed to acquire Reltio to make enterprise data usable by AI agents across its Business Data Cloud.  Link.

• SAP announced the Reltio acquisition on March 27, with the deal expected to close in Q2 or Q3 2026 pending regulatory approval.

• Reltio's entity resolution technology merges records from different formats into unified golden records that enterprise AI agents can trust.

• The acquisition integrates Reltio into SAP Business Data Cloud, supporting Joule Agents with cleaner, more reliable cross-platform data pipelines.

• The deal shows SAP accelerating its AI-first pivot where data quality infrastructure is as critical as the models themselves.

Meta launched AI-powered shopping with review summaries and one-tap checkout on Instagram and Facebook.  Link.

• TechCrunch reports Meta revealed AI shopping features on March 25 at Shoptalk 2026, adding generative review summaries to ad click-through experiences.

• The updated checkout flow built with Stripe and PayPal lets users complete purchases without leaving Instagram or Facebook apps.

• Creators can now earn affiliate commissions from products tagged in Reels, with Facebook extending the feature to both Reels and photos.

• The move shows Meta betting AI-assisted discovery inside social feeds can capture e-commerce revenue flowing to Amazon and Google Shopping.

Content Creation

Artlist embedded Google's Lyria 3 audio models for AI music generation inside its video creation platform.  Link.

• Artlist announced AI Music powered by Google Lyria 3, letting creators generate studio-quality songs from text prompts or up to 10 image prompts.

• The feature is available on Max, Max Business, and Enterprise plans, positioned as commercially licensed and production-ready from day one.

• Artlist framed the launch as completing a unified ecosystem where AI video, image, voiceover, and music generation all live in one platform.

• Google Lyria 3 reaching market through Artlist shows DeepMind audio research deploying faster via partnerships than through standalone Google products.

Data Centers + Energy

Nvidia and Emerald AI partnered with major utilities to build data centers that flex power use with the grid.  Link.

• Nvidia and Emerald AI unveiled the flexible AI factory initiative on March 23 at CERAWeek 2026, partnering with six major U.S. energy companies.

• The design pairs Nvidia's Vera Rubin DSX architecture with Emerald AI's Conductor platform to shift compute based on real-time grid conditions.

• The first commercial deployment is planned at Nvidia's AI Factory Research Center in Virginia using co-located generation and battery storage.

• The approach reframes data centers as grid assets rather than pure consumers, potentially easing utility pushback and accelerating permitting timelines.

Startup Funding & Valuations

Reflection AI is raising $2.5 billion at a $25 billion valuation in one of the largest open-source AI rounds. Link.

• The Wall Street Journal reported Reflection AI, founded by ex-DeepMind researchers, is negotiating a $2.5 billion raise at $25 billion pre-money valuation.

• JPMorgan Chase is considering participating through its Security and Resiliency Initiative, joining Nvidia as a key strategic and financial backer.

• The proposed valuation is more than triple the roughly $8 billion from Reflection's prior round, reflecting explosive demand for open-source alternatives.

• The deal signals investors see open-source AI labs as viable long-term businesses, especially as enterprises seek alternatives to closed model providers.

Shield AI raised $2 billion at a $12.7 billion valuation as defense autonomous AI draws institutional capital.  Link.

• Reuters reported Advent International and JPMorganChase co-led the round, with Blackstone investing $500 million preferred plus a $250 million delayed draw facility.

• Fortune reported Shield AI projects over $540 million in 2026 revenue, more than 80% growth, driven by Hivemind autonomy and V-BAT deployments.

• Funds will support acquiring Aechelon Technology, a flight simulation maker, to strengthen Hivemind's capabilities in GPS-denied combat environments.

• The round underscores how global conflict is redirecting institutional capital toward defense AI, with Shield joining Anduril in a rapidly expanding market.

Kleiner Perkins raised $3.5 billion going all-in on AI in its largest fundraise since 1972.  Link.

• TechCrunch reports Kleiner Perkins announced $1 billion for its 22nd early-stage fund and $2.5 billion for a late-stage growth vehicle on March 24.

• The firm holds stakes in Together AI, Harvey, OpenEvidence, Anthropic, and SpaceX, with the latter two expected to IPO in 2026.

• Kleiner operates with just five partners after recent turnover, meaning the raise reflects sustained LP conviction despite the leanest team in its history.

• The fund positions Kleiner alongside Thrive and General Catalyst in a mega-fund wave where allocators are concentrating AI bets into established platforms.

Harvey raised $200 million at an $11 billion valuation as legal AI moves beyond the pilot phase.  Link.

• CNBC reported on March 25 that Harvey raised $200 million co-led by GIC and Sequoia Capital, with Andreessen Horowitz and Coatue participating.

• Harvey's products now serve over 100,000 lawyers across 1,300 organizations, covering contract analysis, compliance, and litigation support workflows.

• Sequoia partner Pat Grady called Harvey the firm's strongest conviction bet, comparing it to how Salesforce defined the cloud platform transition.

• The deal shows investors backing AI application layer companies with proven enterprise revenue rather than competing at the foundation model level directly.

Kandou AI closed a $225 million oversubscribed Series A to scale its copper-based chip interconnect technology, attracting SoftBank, Synopsys, and Cadence at a $400 million valuation.  Link.

• Kandou AI announced on March 23 that Maverick Silicon led the round with strategic participation from SoftBank, Synopsys, and Cadence.

• The company's CNRZ-5 Chord technology uses six wires to transmit up to five bits at once with half the power of competing interconnect approaches.

• Kandou claims its Glasswing interconnect shrinks multi-die processor packages while moving data faster between chip components than fiber optics.

• The round signals growing investor conviction that data movement, not raw compute, is becoming the critical bottleneck in scaling AI infrastructure.

Regulation + Legal

Sanders and Ocasio-Cortez introduced a bill to halt new AI data center construction until safeguards are set. Link.

• Sanders and Ocasio-Cortez announced the AI Data Center Moratorium Act on March 25, calling for a federal pause on all new data center construction.

• The bill requires safeguards ensuring AI does not increase utility prices, harm communities, or concentrate economic gains among Big Tech companies.

• Data center power demand is projected to reach 24 gigawatts in Texas alone by 2031 according to ERCOT forecasts cited in the bill.

• The bill is unlikely to pass but signals growing pressure on AI's energy footprint that could influence permitting and rate decisions at state level.

Baltimore became the first U.S. city to sue an AI company after Grok generated nonconsensual deepfake images.  Link.

• CNBC reports Baltimore's mayor and City Council filed suit on March 24, alleging xAI violated consumer protection laws with Grok's deepfake imagery.

• The suit follows a Tennessee class-action for three teenagers and regulatory probes in multiple countries targeting Grok's explicit image generation.

• Mayor Brandon Scott called the deepfakes a threat to privacy and public safety, framing the lawsuit as cities filling gaps in federal AI regulation.

• The municipal lawsuit model shows cities developing their own AI enforcement frameworks in the absence of federal content liability rules.

Safety + Ethics

OpenAI shelved its adult content mode indefinitely, the second major product retreat in one week after Sora. Link.

• TechCrunch reports the Financial Times cited sources attributing the halt to pushback from advisors, investors, and employees over the proposed feature.

• The adult mode would have allowed explicit content generation in ChatGPT, targeting subscription revenue in the adult content platform market.

• The pause arrives the same week OpenAI shut down Sora, with the company citing a desire to focus on core products ahead of its IPO.

• Back-to-back retreats show OpenAI making explicit tradeoffs about which product surfaces it wants associated with its brand under market scrutiny.

OpenAI

OpenAI abruptly killed its Sora text-to-video product on March 24, terminating the consumer app, API, and a $1 billion Disney partnership in a sharp strategic pivot toward coding tools and enterprise revenue.  Link.

• Reuters reported the announcement blindsided a Disney team collaborating on Sora content just 30 minutes before the internal communication went out.

• Bloomberg confirmed the closure ends a three-year Disney deal for licensed characters including Mickey Mouse, though no money had changed hands.

• OpenAI said the Sora research team will continue world simulation work for robotics, citing the need to reallocate compute to higher-priority products.

• The decision marks OpenAI's sharpest strategic pivot, cutting a consumer creative product to focus on agents and reasoning models ahead of its IPO.

OpenAI's ChatGPT ad pilot crossed $100 million annualized revenue in six weeks with most users still unexposed.  Link.

• Reuters reported on March 26 that the ad pilot crossed $100M ARR with over 600 advertisers now participating in the U.S. program.

• Roughly 85% of free and Go tier users are eligible but fewer than 20% see ads daily, leaving substantial room for monetization growth.

• OpenAI reported no impact on consumer trust metrics and plans to expand ads to Australia, New Zealand, and Canada in coming weeks.

• The rapid revenue ramp validates advertising as a viable business line ahead of the IPO, potentially reducing dependence on subscription and API income.

SoftBank secured a $40 billion bridge loan from five global banks to deepen its OpenAI investment.  Link.

• Reuters reported the bridge loan matures in March 2027, arranged with JPMorgan, Goldman Sachs, Mizuho, SMBC, and MUFG as lead banks.

• The loan adds to SoftBank's existing $30 billion OpenAI commitment through Vision Fund 2, deepening Masayoshi Son's largest AI infrastructure bet.

• SoftBank and OpenAI are key Stargate Project partners, targeting up to $500 billion over four years in U.S. AI data center infrastructure.

• The borrowing scale underscores SoftBank's conviction in OpenAI as the dominant platform, but significantly increases leverage risk for the conglomerate.

Thank you for reading the AI Rundown by Lightscape Partners. Please send any questions, comments, or suggestions to [email protected].