Local AI Server at $18K, Browser Windows Revival Hits 200K Users, and Coatue's Decision-Function Playbook

Key Takeaways

1. A 10-person startup built a 128GB local AI server for $18K, running Qwen3.8 and DeepSeek V4 Flash at up to 538.5 tps aggregate decode across 16 concurrent sessions.
2. Pieter's browser-based Windows 3.1/XP preservation project crossed 200K users, with 80K playing in a single day.
3. Coatue's AI lead detailed a decision-function framework for finance, stressing that not every step should be automated with LLMs.
4. Nomads.com launched an API priced at $2,490–$9,990/year, targeting AI agents and data integration.
5. A no-code founder's trade-website component now earns AU$85K/month after door-to-door sales landed the first 10 customers.
6. Noah Kagan shared a repeatable X growth system and reported 68,200 new impressions in one week.
7. Acquire.com listed multiple SaaS businesses, including a real-estate rewards platform at $999K and an AI SVG tool at $250K.

1. AI Models and Infrastructure

1. A 10-person startup built a local AI server for $18K: a single Threadripper host with four Radeon AI PRO R9700 32GB GPUs, totaling 128GB VRAM, running Qwen3.8 and DeepSeek V4 Flash with support for 16 concurrent AI sessions. Qwen3.8-Flash-Next achieved 538.5 tps aggregate decode at 16 concurrent users with 48K context each, and 365 tps at 128K context each, with GPUs undervolted below 210W each. The author sees a trend toward companies eventually running all inference locally, while noting this setup may not satisfy many frontier developers and asking whether that future is desirable. — via 1
2. Coatue's AI lead described applying AI to financial services by treating a hedge fund as a decision function: external context (news, prices, filings via API) plus internal context (positions, liquidity, risk requirements, historical cases) produces investment targets, sizing, and timing. He framed it as a compression problem, said everything visible on an investor's screen should be API-accessible to agents, and credited Coatue's decades of rigorous internal-context documentation as an advantage, while acknowledging that stuffing all data into a context window will not yield three key numbers. The function is decomposed into data collection, analysis, and decision-making, each with dozens of substeps, and each step should use the best executor: deterministic code for rules like P/E below 15, classical ML for more complex relationships, LLMs for tasks like summarizing earnings calls, or humans. He stressed maximizing absolute quality per step rather than automating for its own sake, and estimated each node has at least 10 knobs with at least 20 valid values each, making the search space effectively infinite; presenting this in a technically understandable and culturally absorbable way is the biggest challenge. The same pattern applies to policy, nonprofits, drug discovery, and corporate strategy. — via 1

2. Products, Launches, and Data

1. Pieter's browser-based Windows preservation project surpassed 200K users, with 80K people playing Windows 3.1 or Windows XP in a single day, chatting with MSN/AOL Messenger AI bots, or playing 3D Pinball and Quake. The non-commercial project ports Windows 3.1/XP, Quake, and more to browser WASM with WebSockets for multiplayer, and took about two years to gain traction, which the author says is typical; it took off alongside AI-assisted reverse engineering and WASM porting. — via 1
2. Nomads.com launched an API after its MCP server, priced at $2,490–$9,990 per year, aimed primarily at AI agents and secondarily at companies integrating Nomads data into apps. The data includes cost of living for nomads/expats/locals/families, nomad-tested internet speeds, real-time weather and air quality, quality-of-life scores, short-stay hotel and coworking prices, tens of thousands of city reviews, best neighborhoods (from Hoodmaps), and demographics like population, currency, median age, gender ratio, and dating suitability. The author calls the data a moat built over years of manual collection and continuous updates from hundreds of sources, with subscriptions providing ongoing updates. — via 1
3. Acquire.com listed several SaaS businesses for sale: a real-estate media creator workflow platform at $129.6K asking price with $74.7K TTM revenue; an AI SVG conversion tool at $250K with $85.2K TTM revenue; a creative platform for animated music visuals at $55.6K with $12.3K TTM revenue; and a real-estate partner rewards SaaS at $999K with $371.8K TTM revenue. The account also shared a bootstrapped founder earning $30K/month ($360K/year) with no investors or board pressure, arguing startup culture has distorted success and more founders should take pride in profitable, owned businesses. Alex completed three acquisitions through the platform, all still operating, treating the first as a learning experience. — via 1 2 3 4 5 6

3. Founder Playbooks and Growth Tactics

1. Tom, a non-programmer, used a decade of savings from a marketing company to hire a lead designer, lead developer, and front-end developer, building an MVP with his brother and a friend in February of last year. With no customers before launch, he visited music studios, repair shops, and electricians door-to-door with an iPad, showing customized branded components and handing out printed brochures, landing the first 10 customers. The electrician/trades website component now earns AU$85K/month (over $50K) and is embedded in client sites and integrated into Google's "Book Online" button through direct partnerships. His advice: sell before you build; if people pay before the product exists, the direction is worth pursuing. — via 1
2. Noah Kagan shared a system from @codyschneider for building a seven-figure company in eight months on X: talk about only three topics and remix them in a dozen ways; post three times daily, eight hours apart; add links one hour after posting (Kagan says X's open-sourced ranking code has no link penalty); use AI to mine old podcasts and posts for reusable content; and conduct weekly AI-assisted self-interviews combining calendar, email, and Slack to turn answers into posts. Kagan reported 68,200 new impressions in the past week, up from 10,800 in the prior 10 days. He also shared his favorite X content prompt, used since September 21, which grew verified impressions from 333,600 to 448,600: every Friday, read the past 90 days of posts sorted by views, summarize the top three formats and why they worked, pull stories with real numbers, named people, or strong opinions from the past seven days of email, Slack, calendar, and chats, write four new posts in the best formats, score each 1–10 on whether it would be saved, shared, or replied to, delete anything below 8, and label the remaining posts with the past hit they mimic and its view count. — via 1 2
3. Max, a non-programmer, built a fitness app earning $13K/month by filming 5–10 videos daily, overcoming embarrassment from friends, analyzing retention and skip rates, and turning the best-performing content into paid ads. After seeing product-market fit, he gradually increased budget to about $250/day. A separate account advised that when asked how to ship products while working full-time, the answer is no product meetings, no calls, no standups, and every hour after work spent building or writing. — via 1 2

4. Market Signals and Observations

1. Tony Dinh said he is replacing a SaaS service costing over $600/month that was used only to send simple weekly reports. He also commented on a claim that humans are bad random number generators and LLMs are worse, noting that when asked to pick a number from 1 to 10, people usually choose 7 because 1 and 10 are often spoken aloud and people gravitate toward primes near 10, and that LLMs solidify such biases into behavioral patterns. — via 1 2
2. Noah Kagan reviewed $46M in software sales on AppSumo and found AI agents and app-building tools accounted for only 12% of sales, saying he will break down where founders actually spend by sales share. — via 1
3. Flippa listed a faceless home-tips YouTube channel registered in 2007 with templates and workflows, earning $44K/year with 63K subscribers and 8.97M total views, asking $100,758, now in First Access. A Shopify apparel and footwear dropshipping brand founded in September 2023 requires about 12–15 hours per week of owner time, earns $856K/year with $5,600 average monthly profit and $85 average order value, asking $210K, also in First Access. An AI narrative app for romance and fan-fiction fans, one year old, offers role-play on iOS, Android, and web, earning $389K/year with $3,520 monthly profit and 3,200 paid subscriptions, asking $495K. Sequoya Borgman of Borgman Capital said on The Exit episode 317 that most owners sell only once, markets are rarely perfect, and professionalization typically takes about two years before buyers stop pushing prices down. — via 1 2 3 4