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

• OpenAI closed the largest private fundraise in history at $122 billion, valuing the company at $852 billion with anchor commitments from Amazon, Nvidia, and SoftBank. Within 48 hours, roughly $600 million in secondary shares sat without buyers at a 10 percent discount, signaling a growing gap between headline valuations and real investor demand for OpenAI equity.

• Microsoft launched three in-house foundation models under the MAI brand, with CEO Mustafa Suleyman declaring the company a top three AI lab alongside OpenAI and Google. The release challenges Microsoft's existing partnership with OpenAI by demonstrating independent capability across transcription, voice, and image generation.

• Microsoft, Google, and Meta each locked in major natural gas power deals this week for AI data center infrastructure, collectively representing over 10 gigawatts of new fossil fuel capacity. The speed of these commitments confirms that natural gas, not renewables, will underpin the next phase of AI compute buildout.

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

Nvidia Invests $2B in Marvell to Co-Develop Custom AI Chip Interconnects.  Link.

• Reuters reports Nvidia invested $2 billion in Marvell Technology on April 1 as part of a deal to co-develop NVLink Fusion interconnects for custom AI chips.

• Marvell shares surged 12 percent on the announcement as the partnership positions both companies against Broadcom in the custom silicon market.

• NVLink Fusion allows custom AI accelerators from companies like Amazon and Google to communicate natively with Nvidia GPUs on the same fabric.

• The investment signals Nvidia is shifting from competing against custom chips to integrating with them, potentially expanding its addressable market rather than defending it.

Models

Google Releases Gemma 4 Open Model Family With Native Agentic Support.  Link.

• Google released Gemma 4 on April 2 with four model sizes under the Apache 2.0 license, including a 31 billion parameter version that ranked third on Arena AI.

• The models include native function calling, structured output, and multi-turn agentic capabilities designed for enterprise workflow integration.

• The 31B model ranks as the third best open model globally on Arena AI, outcompeting models 20 times its size on reasoning and coding tasks.

• The release positions Google to compete with Meta and Mistral for developer ecosystem loyalty in the open weights model market.

Netflix Releases VOID, Its First Public AI Model for Physics-Aware Video Editing.  Link.

• The Register reports Netflix released VOID on April 3, its first publicly available AI model, designed for physics aware video inpainting and editing.

• In blind evaluations VOID was preferred 64.8 percent of the time versus Runway at 18.4 percent for maintaining physical consistency across edited frames.

• The model handles complex scenes including reflections, shadows, and fluid dynamics that typically cause artifacts in competing video generation tools.

• Netflix releasing a production grade model under open license signals that media companies are transitioning from AI consumers to AI infrastructure providers.

Microsoft Launches Three MAI Foundation Models, Claims Top Three Lab Status.  Link.

• TechCrunch reports Microsoft launched three MAI foundation models on April 2 covering transcription, voice synthesis, and image generation.

• CEO Mustafa Suleyman declared "we are now a top three lab" positioning Microsoft as a direct competitor to OpenAI and Google in foundational model research.

• Transcribe-1 reportedly outperforms Whisper on multi-accent and noisy audio benchmarks while running at lower compute cost for enterprise deployments.

• The launch challenges Microsoft's existing partnership with OpenAI by demonstrating independent model development capability across multiple modalities.

PrismML Ships 1-Bit Bonsai Model That Runs 8B Parameters on an iPhone.  Link.

• The Register reports PrismML released Bonsai on March 31, a 1 bit quantized 8 billion parameter model that fits in 1.15 gigabytes of memory.

• The model is 14 times smaller than its FP16 equivalent and runs at 44 tokens per second on an iPhone 17 Pro Max without cloud connectivity.

• PrismML claims over 10 times the intelligence density of full precision counterparts while eliminating inference API costs for on device applications.

• The release could accelerate edge AI deployment by proving that capable language models no longer require cloud infrastructure or GPU acceleration.

Enterprise AI Applications

Oracle cuts up to 30,000 jobs in its largest layoff ever to fund a $156B AI data center buildout.  Link.

• CNBC reports Oracle began terminating staff on March 31, eliminating roughly 18% of its 162,000 person global workforce with no prior warning.

• The company took a $2.1B restructuring charge and added $50B in new debt within two months to finance its AI infrastructure pivot.

• Oracle posted a 95% jump in net income last quarter to $6.13B, but its stock has lost more than half its value since September 2025.

• The restructuring signals that legacy enterprise vendors view compute capital as existential, even at the cost of massive human capital reduction.

Microsoft launches three MAI foundation models for transcription, voice, and image generation.  Link.

• TechCrunch reports Microsoft released MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2 on April 2, available through its Foundry platform.

• MAI-Transcribe-1 logs a 3.9% error rate across 25 languages, outperforming Gemini 3.1 Flash and OpenAI GPT-Transcribe on speed benchmarks.

• MAI-Image-2 debuted at number three on the Arena.ai leaderboard, using 10B to 50B non-embedding parameters for up to 1024x1024 output.

• The release underscores Microsoft's push to reduce dependency on OpenAI by building a competitive in-house model stack under Mustafa Suleyman.

Consumer AI Applications

Google launches Veo 3.1 Lite, cutting AI video generation costs to $0.05 per second at 720p.  Link.

• Google announced Veo 3.1 Lite on March 31, a developer-focused video model priced below half the cost of its existing Fast tier.

• The model supports text-to-video and image-to-video at up to 1080p with native audio generation, matching Fast tier generation speed.

• Free users get 10 generations per month via Google AI Studio, while Ultra subscribers receive up to 1,000 video generations monthly.

• The launch fills the vacuum left by OpenAI shutting down Sora and gives Google a three-tier video API stack with no direct competitor at scale.

MediaAlpha launches the first carrier-approved conversational AI app for auto insurance on ChatGPT.  Link.

• GlobeNewsWire reports MediaAlpha launched the app on April 2, delivering real-time auto insurance quotes through a ChatGPT conversational interface.

• The app draws listings exclusively from MediaAlpha's existing marketplace of over 1,150 carrier partners and sends users to official carrier sites.

• MediaAlpha's programmatic technology powered $2.2B in advertising spend in 2025, giving the app an immediate distribution and trust advantage.

• The launch signals that conversational AI commerce is moving from experimental to carrier-grade, with compliance and accuracy built into the stack.

Content Creation

ElevenLabs releases ElevenMusic, a free iOS app for AI music generation taking on Suno and Udio.  Link.

• TechCrunch reports ElevenLabs quietly released the iOS app on April 1, offering up to seven free AI-generated songs per day via text prompts.

• A Pro tier at $9.99 per month unlocks 500 tracks monthly, with features including song remixing, curated stations, and mood-based daily mixes.

• ElevenLabs raised a $500M Series C at an $11B valuation in February, positioning the music push as a hedge against voice AI commoditization.

• The move expands the competitive field beyond Suno and Udio and signals that voice AI companies see music as a natural adjacency for growth.

Google Vids adds AI avatars, custom music via Lyria 3, and free video generation powered by Veo 3.1.  Link.

• TechCrunch reports Google updated its Vids productivity app on April 2 with prompt-directed AI avatars in realistic and cartoon styles.

• AI Pro and Ultra subscribers can generate custom royalty-free soundtracks up to three minutes long using Google's Lyria 3 and Lyria 3 Pro models.

• Free accounts receive 10 monthly video generations via Veo 3.1, with direct YouTube publishing and a Chrome screen recorder extension.

• The update bundles video generation, music creation, and avatars into one productivity tool, challenging standalone players like Synthesia and HeyGen.

Data Centers + Energy

Microsoft and Chevron Enter Exclusivity Deal on a 2.5 GW Gas Power Hub in West Texas.  Link.

• Fortune reports the companies confirmed the exclusivity agreement on April 1 for a multibillion dollar project, with Engine No. 1 also involved.

• The plant in the Permian Basin would initially generate 2.5 GW, scalable to 5 GW, with seven GE Vernova turbines already ordered.

• Chevron has historically stayed out of power generation, making the deal its largest collaboration to date between a U.S. oil giant and Big Tech.

• The agreement signals a structural shift in how AI infrastructure gets powered, with dedicated fossil fuel generation built specifically for compute demand.

Google partners with Crusoe on a 933 MW behind-the-meter gas plant in Texas with no carbon capture.  Link.

• Axios reports the plant would be built on-site at Google's Armstrong County campus, codenamed Goodnight, and would not connect to the grid.

• Cleanview uncovered the project through Crusoe's January 2026 permit filings; Google confirmed it has not yet signed a formal offtake agreement.

• Google invested roughly $90B in capex in 2025 and plans to nearly double that to as much as $185B in 2026, mostly for data center buildout.

• The project's lack of carbon capture marks a departure from Google's prior clean energy commitments and reflects the urgency of the AI power race.

Startup Funding & Valuations

Sarvam AI Nears $350M Round at $1.5B Valuation as India Bets on Sovereign AI.  Link.

• Bloomberg reports Sarvam AI is close to raising $300M to $350M at a valuation of $1.5B to $1.55B, with a close expected within days.

• Bessemer Venture Partners is expected to lead, with Nvidia, Amazon, and Prosperity7 Ventures also participating in the round.

• Founded in 2023, Sarvam builds voice-first AI systems supporting 22 Indian languages and has launched open-source LLMs trained on Indian language data.

• The round would be the largest single infusion into a pure-play Indian AI company, reflecting growing sovereign AI investment globally.

Cognichip Raises $60M Series A for Physics-Informed AI Chip Design.  Link.

• TechCrunch reports Cognichip closed a $60 million Series A on April 1 to scale its physics informed approach to AI chip design automation.

• Intel CEO Lip-Bu Tan joined the board alongside Seligman Ventures' Umesh Padval, signaling major foundry interest in AI driven chip design tools.

• The platform uses simulation models trained on semiconductor physics to predict chip behavior before fabrication, reducing costly respins.

• The round suggests EDA tooling is becoming a critical AI infrastructure layer as custom chip demand outpaces traditional design capacity.

Rebellions Closes $400M Pre-IPO at $2.3B to Challenge Nvidia in AI Inference.  Link.

• TechCrunch reports Rebellions closed a $400M pre-IPO round led by Mirae Asset Financial Group and Korea's National Growth Fund, its first direct startup investment.

• The deal brings total funding to $850M and more than doubles its valuation from the $1.4B Series C in September 2025, signaling strong momentum.

• Rebellions designs inference-focused AI chips and launched RebelRack and RebelPOD, modular AI infrastructure products, alongside the funding announcement.

• CEO Sunghyun Park named Meta and xAI as target customers as the company expands into the U.S. ahead of a planned late-2026 IPO.

depthfirst Raises $80M Series B for AI-Native Security, Less Than 90 Days After Stealth Launch.  Link.

• SiliconAngle reports depthfirst announced $80M in Series B funding on March 31, led by Meritech Capital with participation from Forerunner Ventures, The House Fund, and existing investors Accel and Box Group.

• The round brings total capital raised to $120M after a $40M Series A just three months prior, reflecting urgent investor demand for AI-native cybersecurity platforms.

• Founded by former DeepMind, Databricks, and Faire leaders, depthfirst introduced dfs-mini1, a domain-specific security model that outperforms frontier models at 10-30x lower cost.

• The rapid back-to-back raises suggest VCs see AI-native security as a category where speed to market matters more than capital efficiency.

Linx Security Closes $50M Series B for AI-Powered Identity Governance.  Link.

• SecurityWeek reports Linx Security closed $50M in Series B funding on March 31, led by Insight Partners with continued backing from Cyberstarts and Index Ventures.

• The 100-person Israeli startup governs millions of identities for Fortune 500 firms, addressing a market where machine and AI agent identities outnumber humans roughly 80 to 1.

• Alongside the round, Linx launched Autopilot, an autonomous AI agent that continuously monitors, detects, and resolves identity governance issues in real time.

• The deal brings total funding to $83M and reflects growing VC interest in securing AI agent infrastructure as enterprise AI adoption accelerates.

Valar Atomics Lands $450M at $2B Valuation to Power AI Data Centers With Nuclear.  Link.

• Bloomberg reports Valar Atomics raised $450M ($340M equity, $110M debt) on March 31, five months after a $130M Series A at a far lower valuation.

• Backers include Palmer Luckey (Anduril founder) and Shyam Sankar (Palantir CTO), connecting defense tech capital to AI infrastructure energy needs.

• The company builds clusters of small high-temperature gas-cooled reactors for AI data centers, with its reactor achieving criticality at Los Alamos in November 2025.

• The round signals that AI infrastructure investment is expanding well beyond chips and software into purpose-built energy generation for compute demand.

Regulation + Legal

DOD Appeals Anthropic Injunction as Ninth Circuit Battle Begins.  Link.

• The Trump administration filed a Ninth Circuit appeal on April 2 challenging Judge Lin's injunction that blocked Anthropic's supply chain risk designation.

• Briefing deadlines are set for April 30, with multiple amicus briefs supporting Anthropic from Microsoft, military leaders, and theologians.

• The original ruling found the designation process lacked sufficient due process protections for AI companies flagged under national security authorities.

• The appeal signals the administration's commitment to asserting broad supply chain oversight powers over frontier AI developers.

Perplexity Faces Data Sharing Class Action Over Hidden Tracking Scripts.  Link.

• A 135 page complaint filed in Northern District of California alleges Perplexity embedded hidden tracking scripts that sent user prompts to Meta and Google.

• The scripts reportedly operated even in incognito mode, raising questions about whether California wiretapping law applies to AI prompt data.

• The lawsuit names Meta Pixel and Google Analytics integrations as the specific tracking mechanisms embedded without user consent.

• The case could set precedent for how privacy law treats conversational AI data, particularly prompt content shared with third party ad networks.

OpenAI

OpenAI Closes $122B Round at $852B Valuation, the Largest Private Raise Ever.  Link.

• CNBC reports OpenAI closed a $122 billion round anchored by Amazon at $50 billion, Nvidia at $30 billion, and SoftBank at $30 billion, plus $3 billion from retail investors.

• The round values OpenAI at $852 billion pre money, making it the most valuable private company in history by a wide margin.

• Roughly $35 billion of Amazon's commitment is contingent on OpenAI completing an IPO or achieving specific AGI related milestones within defined timelines.

• The contingency structure signals that even the largest backers are hedging against OpenAI's ability to convert research dominance into durable commercial returns.

OpenAI Secondary Market Demand Collapses With $600M in Shares Finding No Buyers.  Link.

• Bloomberg reports roughly $600 million in OpenAI secondary shares are sitting without buyers, with Goldman Sachs and Morgan Stanley waiving carry fees to attract interest.

• Shares are trading at an implied $765 billion valuation, a 10 percent discount to the primary round that closed just days earlier at $852 billion.

• Anthropic's API market share reportedly jumped from 12 percent to 32 percent over the same period, shifting competitive dynamics in the foundation model market.

• The divergence between primary fundraising success and secondary market weakness suggests institutional investors are repricing OpenAI's risk profile even as headline valuations climb.

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