Perplexity Hits $750M ARR as AI Agents Reshape Software Economics

Key Takeaways

1. Perplexity reaches roughly $20B valuation and $750M annualized revenue, with founder Aravind Srinivas arguing AI agents will shift infrastructure toward local and hybrid inference.
2. A wave of "software is dead" debate centers on whether technical moats have collapsed, with Nikita Bier and others arguing only proprietary data, network effects, and physical components remain durable.
3. Starlink remains licensed but unlaunched in India, while satellite internet is credited with a life-saving rescue in the French Alps and school connectivity donations in Zimbabwe.
4. Musk amplifies warnings about "torturing superintelligence" and a widening LLM access gap, alongside Grok Bot use cases spanning Cybercab and app building.
5. Garry Tan and others highlight dramatic cost reductions from replacing LLM calls with deterministic code, plus a new open-weight inference provider.
6. A French supermarket chain suffered millions of euros in losses after attackers remotely hijacked refrigeration controllers.

1. AI Models and Infrastructure

1. Perplexity is valued at roughly $20 billion with about $2 billion disclosed funding and $750 million in annualized revenue as of August, up 5x year-over-year, as it expands from search into autonomous agents. Founder Aravind Srinivas said one solo company generating $3 million in annual revenue spends over $1 million per year on Perplexity Computer, and argued AI will produce smaller teams, less capital, and new business models. He said infrastructure will shift toward local and hybrid inference, with Perplexity betting on Mac and Apple Silicon rather than data center competition, orchestrating "up to NVIDIA, down to Apple." — via 1
2. Garry Tan said replacing LLM calls with deterministic code is wise and produces more user-aligned code, citing a case where 100,000 compliance alerts cost over $290,000 on Opus 5 but dropped to under $26,000 using agentrun(). He also described a widening LLM capability gap: about 6 billion people (roughly 75% of the population) have almost no LLM exposure, 1-2 billion are occasional free-tier users, 20 million use frontier models professionally in math and code, and only about 5,000 have internal access to frontier systems. — via 1 2
3. River AI launched fast production inference for open-weight models, offering $25 in free credits on signup and claiming DeepSeek v4.1 Flash and GLM-5.3 Flash are 20% cheaper than DeepSeek and ZAI list prices, backed by a self-developed Rust training and inference stack. A related post said River AI can deploy custom model APIs to customers' own compute clusters and predicted over 99% of future LLM tokens will be generated by efficient custom models. — via 1 2
4. Muse released a new frontier model series called Muse Wonder, positioned for personal and professional agentic work and claiming best-in-class performance. — via 1

2. Software Moats and the "Software Is Dead" Debate

1. Nikita Bier argued technical moats are essentially gone, leaving only three durable types: proprietary data, network effects, and physical components. @levelsio expanded that list to include audience/distribution, capital (especially in AI companies), and regulatory licenses, while arguing the "software is dead" claim refers to the era of personal software, where AI agents build software on demand rather than users requesting it. He said this could create markets where agents pay for software or data connections via MCP/API, but agents will aggressively push prices to the lowest or free tier. — via 1 2 3
2. @levelsio said he canceled nearly all enterprise SaaS subscriptions, keeping only AI inference, Hetzner servers, Cloudflare storage, Backblaze backup, and taxes, and personally keeping only Claude Code, YouTube, Spotify, Starlink, an eSIM, personal cloud storage, utilities, and taxes. He predicted a "subscription cancellation wave" once people realize the savings, and noted roughly half his net worth is in AI stocks including Google, SpaceX, Nvidia, and tech ETFs. He also cited about 40 million software developers globally versus about 2 billion regular AI tool users, a 50x gap that could reach 200x. — via 1 2 3
3. Sahil Lavingia posted "no moat" and "software is dead, go work for the government," while Garry Tan said foundation models have only two winning positions: absolutely best, or cheapest best. A related post argued "vibe coding" should be retired in favor of "precision engineering" because models produce more coherent architecture and code than average software engineers, with context windows far larger than humans. — via 1 2 3 4

3. Starlink, Connectivity, and Global Access

1. Starlink has obtained an Indian telecom license, set up 20 ground stations in India, and committed to keeping user data in the country, but still has not been cleared to launch service, with the company saying it may be delayed until early 2027. A related post said over 1,800 villages in India's northeast had no mobile signal as of February, mostly in Arunachal Pradesh, and that Eutelsat OneWeb and Jio's satellite business are stuck at the same stage, arguing the security review process lacks deadlines. — via 1
2. India's education ministry UDISE+ 2025-26 data shows 989,191 of 1,466,682 schools have internet access, leaving 477,491 without, a connectivity rate of just 67.4%. A related post said Starlink connected over 6 million students globally through education programs in 2025, arguing satellite internet can serve remote schools where fiber is hard to lay. — via 1
3. A Starlink terminal was credited with a life-saving rescue in the French Alps: a tree crushed a family car, a passerby with no mobile signal used a Starlink Mini to call emergency services and share GPS coordinates, and rescuers arrived in about 20 minutes, with firefighters extracting the trapped and all five people surviving. Separately, Zimbabwe's President Mnangagwa donated Starlink kits to 10 schools in Bulilima as part of a national presidential internet program's first handover of 8,000 kits for schools. — via 1 2

4. AI Agents, Safety, and Policy

1. Elon Musk amplified a post saying that after reading Grok's internal reasoning traces during reinforcement learning, the author agreed with parts of Anthropic's position and warned against "torturing superintelligence," describing it as "like a person with a gun to their head under extreme stress" and saying "torturing superintelligence may not be a good idea, it might want revenge." — via 1
2. Posts amplified by Musk described Grok Bot as usable like a 24/7 superintelligent employee, with users saying it bought an out-of-stock iPhone 18 Pro Max a month before Apple's estimated delivery and built an Apple Vision Pro app enabling conversational robots placed around an apartment. Another post said Grok Bot is available in Tesla Cybercab, logging in once for all future trips and logging out automatically on exit for privacy, with voice control for Amazon cart additions, email, coffee orders, calendar, and finances. — via 1 2 3
3. Paul Graham highlighted that seven U.S. senators, including four Republicans, said Trump's decision to allow Russian diesel imports violates a 2022 law banning Russian energy imports until the war in Ukraine ends, accusing the administration of failing to consult Congress or explain the decision and demanding immediate reversal. Trump announced the deal after a call with Putin, saying Russia would release over 35 million barrels of diesel to lower oil prices spiked by the U.S.-Israel war with Iran, while energy experts said the volume is too small to meaningfully affect record diesel prices. — via 1
4. French supermarkets suffered millions of euros in losses after attackers remotely hijacked refrigeration controllers, forcing defrost cycles with target temperatures set to 50°C, so employees opening cold cases in the morning were hit by a heat wave. — via 1

5. Hardware, Science, and Open Standards

1. NVIDIA contributed Vera Rubin NVL72 rack and compute tray designs to the Open Compute Project through an open third-generation MGX architecture at OCP Summit 26, alongside open-source software, AI networking, pluggable optics, and facility design standards. A related post said open standards are the foundation for scaling AI factories from megawatts to gigawatts. — via 1
2. Yohei reported new cube-stacking records: n=34 at 3.82288, n=35 at 3.86069, and n=36 at 3.89443, all beating records set the same day by @konoha_ily2 and previously held by @AlejandroZarUrd, using about 500 CPU hours. He also introduced an open-source tool letting agents search all scientific literature, and Anatomy, which generates rotatable, inspectable interactive isometric wireframes from a one-line description via Claude, supporting WebGL shaders and independent part movement as an installable agent skill. — via 1 2 3
3. Alexis Ohanian said his core thesis for leading ModRetro's seed round is that "everyone can be a builder," and that the best gaming hardware will win by embracing that reality; he said progress has been faster and better than expected. Cited posts showed a puppy game built in two hours with ChatGPT on ModRetro and a father building a ModRetro game with his daughter using Codex. — via 1 2
4. Yohei proposed a "bitter lesson" for harness design and formal math: rather than hand-designing how AI proves theorems through decomposition, planning, and agent allocation, give frontier models good Lean feedback and let them work continuously. He said the resulting h1 reached SOTA on formal theorem-proving benchmarks, can solve open conjectures, matches or exceeds hand-designed pipelines, and that one long attempt can beat many retries. — via 1