The AI Rundown by Lightscape Partners - 4/27/2026

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

• Google committed up to $40 billion in cash and compute to Anthropic, days after Amazon agreed to put another $5 billion in (with up to $25 billion possible) in exchange for $100 billion in AWS spending. The two deals, taken together, hand Anthropic the largest combined infrastructure war chest in the industry and entrench it as the second pole of the frontier-model market.

• SpaceX disclosed that it is working with Cursor on AI training and holds an option to buy the AI coding startup for roughly $60 billion. The structure preempted a planned $2B fundraise and would slot Cursor into Elon Musk's xAI orbit, escalating the talent and capital fight with OpenAI.

• OpenAI released GPT-5.5 with materially better math and coding performance and what TechCrunch describes as the company's clearest move yet toward an AI "superapp," while also rolling out Workspace Agents, ChatGPT Images 2.0, and a new Codex Labs developer training service.

Stay tuned as we break down these stories and the rest of a packed Google Cloud Next week, plus the latest in chips, models, healthcare AI, and policy.

If you haven't yet, please support the newsletter by subscribing!

Anthropic

Google plans up to $40 billion of cash and compute into Anthropic, just days after Amazon's expanded deal. Link.

• TechCrunch reports the package is split between cash and Google Cloud compute and follows the limited release of Anthropic's cybersecurity-focused Mythos model.

• The investment lands days after Amazon disclosed an additional $5 billion stake (with up to $25 billion possible) in exchange for Anthropic committing to $100 billion in AWS spending.

• Combined, the two deals make Anthropic uniquely positioned across both Google's TPU stack and AWS's Trainium fleet, an unusual hedge for any frontier lab.

• The structure also makes clear that hyperscalers now view multi-tens-of-billions in compute commitments as table stakes for keeping a frontier partner inside their ecosystem.

Anthropic's Mythos cyber model became the most talked-about model of the week — for very different reasons. Link.

• Mozilla revealed it used Mythos to find and fix 151 real bugs in Firefox, an early proof point for AI-driven vulnerability discovery at scale.

• TechCrunch reports the NSA is already using Mythos despite the Pentagon's ongoing supply-chain dispute with Anthropic, underscoring quiet government adoption.

• Wired separately reported that "Discord sleuths" gained unauthorized access to Mythos, raising fresh questions about how tightly even gated frontier models can really be controlled.

• Sam Altman dismissed the model as "fear-based marketing," extending the public feud between OpenAI and Anthropic over how to talk about AI safety risks.

Frontier Models

OpenAI shipped GPT-5.5 with sharper math and coding chops and what TechCrunch calls a clear push toward an AI "superapp." Link.

• SiliconANGLE notes the model posts notable jumps on advanced math and coding benchmarks versus GPT-5.4.

• The release sits alongside ChatGPT Images 2.0, which generates images at up to 2,000 pixels wide, and a new Codex Labs technical training service for enterprise adoption.

• OpenAI also rolled out Workspace Agents for paying subscribers, designed to chain multi-step actions across email, docs, and internal tools.

• Together the launches reinforce OpenAI's strategy of competing on integrated workflow rather than benchmark headlines alone.

DeepSeek and Moonshot kept the open-source pressure on with major Chinese model drops. Link.

• DeepSeek open-sourced its V4 family, including a flagship V4-Pro and a smaller V4-Flash that trades quality for far lower hardware needs, both built on a mixture-of-experts architecture.

• Moonshot AI released Kimi-K2.6, a 1-trillion-parameter model with new attention optimizations aimed at long-context efficiency.

• Both releases keep Chinese labs squarely on the open-weights frontier even as US capital flows concentrate around Anthropic and OpenAI.

• DeepSeek separately previewed a follow-up model that it claims "closes the gap" with closed frontier systems, suggesting the open/closed performance gulf is narrowing again.

Hardware + Infrastructure

Google launched two new TPUs at Cloud Next, sharpening its pitch as the credible Nvidia alternative. Link.

• TechCrunch reports the new chips, branded under Google's expanding TPU lineup, target both training and large-scale inference for agentic workloads.

• Google also unveiled mega-scale networking innovations and positioned its "Nvidia-Google Cloud superstack" alongside Jensen Huang at the keynote.

• The announcement reinforces Google's strategy to win on full-stack economics — silicon, networking, and Gemini software — rather than compete chip-on-chip.

• Anthropic's $40B Google deal becomes a much bigger anchor tenant for the new TPU capacity once it scales.

In a notable twist, Meta signed a deal to take millions of Amazon's homegrown CPUs (not GPUs) for AI agentic workloads. Link.

• TechCrunch frames this as a new chip race, with Amazon's Graviton-class CPUs targeting the orchestration and serving layer that surrounds GPU clusters.

• AWS separately announced a multibillion-dollar AI infrastructure deal with Meta, deepening a relationship that historically leaned on Nvidia.

• The move suggests hyperscalers and large AI buyers are starting to optimize the full agentic stack rather than throwing more GPUs at every problem.

• It also gives AWS a rare Meta win at a moment when Meta has been consolidating compute partnerships across multiple clouds.

Vast Data raised about $1 billion at a $30 billion valuation as AI storage becomes a bottleneck. Link.

• CEO Renen Hallak told SiliconANGLE the round size was driven by investor demand rather than capital need.

• The valuation reflects how mission-critical high-throughput data infrastructure has become for training and inference at scale.

• The raise puts Vast in the rare club of >$30B private AI infrastructure companies alongside Databricks, Anthropic, and CoreWeave.

Google Cloud Next 2026 + Enterprise AI

Google put Gemini Enterprise at the center of an "agentic taskforce" pitch and committed $750M to its partner ecosystem. Link.

• Google Cloud unveiled a Gemini Enterprise Agent Platform that brings agent build, deployment, and governance under a single control plane.

• The company committed $750 million to incentivize partners to build agentic solutions on top of Gemini, Workspace, and BigQuery.

• New Workspace Intelligence features turn AI into what TechCrunch calls "your new office intern," automating routine tasks across Docs, Sheets, Gmail, and Meet.

• Salesforce, Atlassian, and Oracle all announced deeper Gemini integrations during the conference, signaling that Google's agent stack is being treated as serious infrastructure.

Startup Funding & Valuations

SpaceX disclosed it is working with Cursor and holds an option to buy the AI coding startup for $60 billion. Link.

• TechCrunch reports the structure preempted Cursor's planned $2 billion fundraise, fundamentally changing the company's path.

• The deal would pull Cursor into Elon Musk's xAI orbit and escalate the AI coding fight with OpenAI's Codex and Anthropic's Claude Code.

• Cursor's 25-year-old CEO is described in coverage as a former Google intern who ended up brokering one of the year's largest tech transactions.

• The move also signals SpaceX's growing willingness to use balance-sheet leverage to pull AI talent and tooling under the Musk umbrella.

Cohere is merging with Germany's Aleph Alpha, with $600M in new funding from Schwarz Group and others. Link.

• TechCrunch reports Lidl's parent Schwarz Group is anchoring the round with the explicit blessing of the German and Canadian governments.

• The combined entity is being pitched as a "transatlantic AI powerhouse" providing a sovereign alternative to US-dominated frontier labs.

• Fortune notes the deal underscores the rise of AI's "middle powers" — credible non-US, non-China model providers serving regulated buyers.

• The merger gives both companies the scale and balance-sheet to compete for sovereign and enterprise contracts that frontier US labs may not be able to serve cleanly.

Cognition, the maker of AI software engineer Devin, is in talks to raise hundreds of millions at a $25B valuation. Link.

• The round comes amid intense competition from Cursor, Anthropic's Claude Code, and a new wave of internal-context coding tools.

• Investor enthusiasm reflects how quickly autonomous software engineering has moved from demo to revenue line.

• The raise would extend Cognition's runway to expand Devin into broader engineering and operations workflows beyond pure coding tasks.

ComfyUI hit a $500M valuation on a $30M raise as creators look for fine-grained control over generative media. Link.

• TechCrunch reports the round reflects strong demand for tools that let professional creators steer AI outputs node-by-node.

• ComfyUI has become the default backend for many studios building proprietary generative pipelines.

• The valuation is a useful read on how much money is moving toward "creator-side" AI rather than fully automated content systems.

Omni raised $120M led by GV to push deeper into AI-native analytics. Link.

• The round positions Omni as one of the more credible challengers in the modern BI stack, alongside Hex and ThoughtSpot.

• AI agents that can author and refine analyses are becoming the centerpiece of enterprise BI roadmaps.

Healthcare + Life Sciences

DeepMind spinoff Isomorphic Labs is taking AI-designed drug candidates into human trials, a meaningful first. Link.

• Wired reports the program represents one of the first instances of fully AI-designed molecules entering clinical evaluation.

• The milestone moves AI drug discovery from "accelerating chemists" to a model where the model itself is generating lead candidates end-to-end.

• The announcement coincides with the launch of Treehub, a new fund led by Tim Draper and Anne Wojcicki backing the next wave of AI-native health founders.

• Together the two stories suggest healthcare is one of the few sectors where AI capex is starting to map to credible therapeutic outcomes, not just productivity claims.

Workforce + Labor

Meta is cutting 10% of staff and canceling 6,000 open roles in what it is calling an AI efficiency push. Link.

• The cuts come even as Meta continues to lock in massive AI infrastructure deals with AWS and Amazon.

• Leadership is explicitly framing the layoffs as headcount being replaced by AI productivity, a sharper posture than prior "year of efficiency" language.

• The move may set a template for other large tech employers heading into 2026 budgeting cycles.

Meta also confirmed it will record employees' keystrokes and screens to train internal AI tools. Link.

• TechCrunch reports the new internal tool converts mouse movements and clicks into structured training data.

• The combination of layoffs and surveillance-grade data collection is likely to draw labor and regulatory attention.

• Nvidia CEO Jensen Huang offered the week's sharpest framing on AI and jobs: "you won't lose your job to AI — you'll lose it to your coworker who uses it."

Data Centers + Energy

Fermi's CEO and CFO suddenly departed the AI nuclear power upstart, the first major leadership shock in the AI energy boom. Link.

• The exits come as AI nuclear projects move from press releases into hard permitting, financing, and grid-interconnect realities.

• Fermi was one of the most heavily promoted "AI-plus-nuclear" stories of the past 18 months, raising the stakes on the transition.

• The departures will sharpen investor scrutiny of every other AI nuclear and SMR-linked capital plan now in the market.

Roughly half of all new US electricity demand last year came from data centers, just as public sentiment turns. Link.

• Fortune reports public opinion is hardening against data-center growth in many local jurisdictions even as load growth accelerates.

• Maine's governor vetoed a proposed data center moratorium this week, illustrating how state-level fights are now front and center.

• Moody's separately warned that Iran-related disruptions to global helium supply have become a $650 billion problem for the AI economy, with no viable substitute at scale.

OpenAI

OpenAI's GPT-5.5 launch was paired with new Workspace Agents for paying subscribers. Link.

• The agents are designed to automate complex, multi-step tasks across teams without leaving the ChatGPT surface.

• Combined with ChatGPT Images 2.0 and Codex Labs, the launches make ChatGPT look much more like a workflow operating system than a chatbot.

• OpenAI also expanded its enterprise distribution with a new Infosys partnership designed to bring its tools to more businesses globally.

Florida's attorney general issued subpoenas to OpenAI in a ChatGPT probe tied to the FSU shooting. Link.

• The probe centers on what role, if any, ChatGPT played in the lead-up to the incident, including how the model responded to specific prompts.

• It is one of the most pointed state-level AI investigations to date and could shape how chatbot harm cases are litigated nationally.

• The action lands in the same week Clarifai disclosed it deleted 3 million OkCupid photos that had been used to train facial recognition AI, a separate but thematically related reminder of how fast AI training-data norms are changing under regulator pressure.

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