The headlines today are saying AI is killing SaaS. AI isn’t killing SaaS. It’s exposing it. It’s exposing who has built a real MOAT. It’s separating must-haves from nice-to-haves.
And here’s the good news: SaaS companies have huge advantages — deep workflow ownership, trusted customer relationships, and rich domain data. AI without that context is a demo. AI with workflow and trust becomes a durable business.
For twenty years, SaaS built systems of record. Massive value was created. But those systems were reactive. Humans clicked. Humans configured. Humans interpreted.
AI changes the center of gravity. Software no longer waits. It understands context. It recommends. Increasingly, it acts. We’re moving from systems of record to systems of work.
For legacy SaaS companies, five shifts matter:
Move from features to decisions. Treat data as a strategic moat. Re-architect teams, not just products. Progress from assistive AI to partial autonomy, then to full autonomy. And evolve pricing from seats to outcomes.
For founders, the message is clear: SaaS isn’t dead; complacency is. Model advantages will compress. Workflow ownership will compound. Shipping AI features isn’t enough. You must own outcomes. And often, the bottleneck isn’t the model; it’s organizational design.
This isn’t about becoming an AI company. The SaaS companies that endure will move to systems of work that orchestrate work semi-autonomously between humans and agents or autonomously amongst agents.
I feel we are entering an era of Work as a Service (what I am calling WaaS).
The Mayfield View: We believe the next generation of enduring software companies will be those that move beyond recording work and begin orchestrating it. Durable advantage will come from workflow ownership, data feedback loops, and the ability to act responsibly on behalf of the customer.
The Bottom Line for Founders: SaaS is not going away. But the bar is higher. The companies that win will be those that redesign around outcomes and gradually earn the right to do the work, not just document it.
Signals Shaping the Future of AI:
NVIDIA signs a $50 billion deal to supply Meta with millions of AI chips. The agreement covers Blackwell and next-generation Rubin GPUs, along with Grace and Vera CPUs, and could be worth an estimated $50 billion as Meta expands its AI infrastructure. Click here
SoftBank commits $33 billion to build a major U.S. natural gas power plant to support AI infrastructure. The multiyear investment will fund one of the largest gas generation projects in the country, aimed at meeting rising electricity demand from AI data centers and securing long-term power capacity for future compute expansion. Click here
Reliance announces $110 billion AI infrastructure investment in India. Mukesh Ambani outlined a seven-year plan to build gigawatt-scale data centers, expand edge computing through Jio, and deploy AI services nationwide, with initial capacity coming online in 2026. Click here
Taalas unveils model-specific “Hardcore” AI chips for ultra-fast inference. The startup hardwires individual AI models into custom silicon, claiming up to 10x faster performance and significantly lower cost and power use, with new model versions produced in about two months. Click here
AMD to backstop $300 million Crusoe loan for AI chip deployment. AMD will guarantee a $300 million loan to AI cloud startup Crusoe, enabling it to finance and deploy AMD chips in a new Ohio data center, with Goldman Sachs providing the underlying debt. Click here
Enterprise
Anthropic launches Claude Code Security to detect and patch software vulnerabilities. The new capability, now in limited preview for Enterprise and Team customers, uses AI reasoning to identify complex security flaws in codebases and suggest fixes for human review. Click here
Google launches Gemini 3.1 Pro with upgraded reasoning for complex tasks. The new model delivers a 77.1% ARC-AGI-2 score, more than doubling Gemini 3 Pro’s reasoning performance, and is rolling out in preview across the Gemini app, API, Vertex AI, and developer tools. Click here
Anthropic and Infosys partner on agentic AI for regulated sectors. The firms will integrate Claude with Infosys Topaz to build AI agents for telecom, financial services, manufacturing, and enterprise software, focused on compliance and legacy system modernization. Click here
OpenClaw creator Peter Steinberger joins OpenAI as agent tools move in-house. Sam Altman said Steinberger will lead the development of next-generation personal agents at OpenAI, with OpenClaw to operate as an open-source project supported within the company. Click here
Perplexity pivots from ads to subscriptions and enterprise sales. The AI search startup said it is deprioritizing advertising and focusing on paid subscriptions and business customers, expanding its enterprise sales team while reporting roughly $200 million in annual recurring revenue. Click here
Capital Flows
Fei-Fei Li’s World Labs raises $1 billion to advance a new AI development approach. Autodesk invested $200 million in the round, with participation from Andreessen Horowitz, NVIDIA, and AMD, as the startup scales its next-generation AI research efforts. Click here
Ricursive Intelligence raises $335 million at a $4 billion valuation. The AI chip design startup closed a $300 million Series A to build automation tools for semiconductor layout, targeting major chipmakers as customers. Click here
Ineffable Intelligence reportedly raises $1 billion at a $4 billion valuation. The London AI startup founded by former DeepMind researcher David Silver is said to be backed by Sequoia, with Alphabet, Nvidia, and Microsoft potentially participating, as it pursues reinforcement learning research. Click here
Code Metal raises $125 million to modernize defense software with AI. The Boston startup secured a $125 million Series B to expand its AI platform that translates and verifies legacy code for defense contractors, following a $36 million Series A and bringing its valuation to $1.25 billion. Click here
Research
Google DeepMind launches Lyria 3, its most advanced AI music generation model. The model creates high-fidelity tracks from text or image prompts, supports multilingual vocals and genre control, and includes SynthID watermarking and safety guardrails, expanding Gemini’s generative media capabilities. Click here
AI models reach research-level performance in advanced mathematics. Systems from OpenAI and Google DeepMind achieved gold-level results at the International Mathematical Olympiad and generated a new solution to an Erdős problem, marking mathematics as a leading benchmark for frontier AI progress. Click here
AI system DeepRare advances rare disease diagnosis. DeepRare integrates clinical records, genetic data, and medical literature to generate ranked diagnostic hypotheses for rare diseases, with transparent reasoning tied to verifiable evidence. Click here
Claude C Compiler shows AI building a full C compiler. Anthropic’s Claude generated a multi-architecture C compiler spanning frontend, IR, optimization, and backend components, marking progress in large-system AI coding while exposing limits in generalization and production readiness. Click here
Policy
DHS signs a $1 billion contract with Palantir for department-wide software. The deal streamlines purchases across agencies, including ICE and CBP, expanding Palantir’s role in immigration and homeland security operations. Click here
UAE bans generative AI tools for students under 13 and restricts classroom use. The new education framework prohibits AI use in exams, mandates teacher oversight, and imposes strict privacy and content controls across schools. Click here
Treasury launches AI cybersecurity initiative for the financial sector. The U.S. Treasury announced a public-private effort to release six resources aimed at strengthening cybersecurity and risk management for AI use across financial institutions. Click here
Global AI Strategy
NVIDIA deepens early-stage push into India’s AI ecosystem. NVIDIA unveiled new partnerships with local venture firms to back early-stage AI startups, expanding its footprint in one of the world’s fastest-growing developer markets. Click here
Global tech leaders call for coordinated AI action at India summit. Executives from leading technology companies urged governments and industry to collaborate on AI governance, infrastructure, and talent development during the India AI Impact Summit in New Delhi. Click here
Talent Signals
Each week, we spotlight key roles tied to the themes shaping this week’s AI headlines, connecting talent to the companies driving the news.
Skild AI builds general-purpose AI systems designed to enable robots to learn physical world interaction through large-scale reinforcement learning rather than hand-coded rules. As investment accelerates in Physical AI and autonomous systems that can do work beyond chat, SkildAI is hiring across robotics software, perception, and systems engineering. Open roles are listed on its careers page. Click here
Ramp integrates AI into finance automation, using models to streamline accounting, expense management, and procurement workflows. As AI becomes a built-in layer inside core SaaS platforms, Ramp is hiring across technical and operations teams. Open roles are listed on its careers page. Click here
WRITER builds enterprise AI software focused on secure content generation and workflow automation inside large organizations. As enterprises demand governance and control in AI SaaS deployments, Writer continues to grow its team. Open roles are listed on its careers page. Click here
You can see all the opportunities at Mayfield-backed AI companies here, and across the broader ecosystem here.
Social Signals
The most important conversations in AI are unfolding across social media, where top voices are shaping the next wave of signals and strategy. Here are some of the top social signals and their takes from the past week.
Hasan Toor (Click here) — “Microsoft Research + Salesforce just dropped a paper. They tested 15 top LLMs across 200,000+ simulated conversations. Single-turn prompt: 90% performance. Multi-turn conversation: 65% performance. Same model. Same task. Just talking normally.” Toor highlights results showing that performance drops sharply in multi-turn settings, while “aptitude only dropped 15%” and “unreliability exploded by 112%,” with issues like wrong early assumptions compounding, mid-conversation forgetting, and longer outputs introducing more errors. He notes that even reasoning models did not hold up and that tactics like more thinking tokens or temperature 0 did not solve it, arguing that real back-and-forth conversation remains a major fragility relative to single-turn benchmark results.
Lex Fridman (Click here) — “The power of AI agents comes from: 1. intelligence of the underlying model, 2. how much access you give it to all your data, 3. how much freedom & power you give it to act on your behalf. The more data & control you give to the AI agent: (A) the more it can help you AND (B) the more it can hurt you.” Fridman argues that as model intelligence rapidly scales, security is becoming the primary bottleneck for effective agent deployment. He suggests broad adoption will hinge less on raw capability and more on solving the risks tied to data access, autonomy, and cyber exposure, even as developers push forward in “full explore/experiment mode” with tools like Claude Code, Codex, Cursor, and OpenClaw.
Jeff Dean (Click here) —“Last week, we released our Gemini 3.1 Pro Deep Think model. Here are a succession of examples of that model doing heat transfer analysis. No tools, just Deep Think and Image Generation. Step 1: generate a CAD file from a technical drawing. Step 2: Do heat transfer analysis based on the CAD file and material properties. Step 3: Turn the heat transfer analysis at different times in the heating process into visual representations.” Dean showcases Gemini 3.1 Pro performing multi-step technical workflows end to end, from CAD generation to physics-based simulation and visualization, and notes that benchmark scores are “quite a bit better” versus the prior version. The demos highlight increasingly integrated reasoning and multimodal capabilities inside a single model, without relying on external tools.
