Supabase Compute Enters Private Alpha; Founders Weigh AI Data, MCP Parity, and Exit Tradeoffs

Key Takeaways

1. Supabase Compute is in private alpha, letting users run agents and long-running tasks next to Postgres data without a second platform.
2. Arvid Kahl says clients increasingly want API and MCP parity, and cites TallyForms data showing 42% of forms built by AI.
3. A large-enterprise AI lead describes a "super Pareto" split: 10% of employees drive 90% of productivity gains and token use.
4. micro1's CEO frames expert data as the key input to model progress and expects enterprise data spending to approach $100M a year.
5. Nico built 17 apps in a year; two exits totaled $265K, and Stalknot reached $7K MRR on paid Facebook ads.
6. A TidyCal mobile test lifted booking rate from 12.2% to 13.0% across 63K visitors after Claude-assisted flow analysis.
7. A disputed investor post questions AI and data-center spending and describes Anthropic's IPO as a "heist" by Google and Amazon.

1. AI Products and Developer Tooling

1. Supabase Compute has entered a private alpha, according to a post reposted by Ben Tossell. The pitch is that users can run agents and long-running tasks inside a Supabase project, next to Postgres data, without adding a second platform. 1
2. Arvid Kahl says clients are directly or indirectly demanding that APIs and MCP offer equivalent functionality, and he argues SaaS products will need to present all capabilities equally to people, agents, and machines, with UI, MCP, and API fully interchangeable. He backs the MCP-as-entry-point case with TallyForms adoption data: 42% of TallyForms are now built by AI (Tally AI plus MCP). 1 2
3. Noah Kagan describes using Claude to simulate a first-time mobile visitor, count the clicks and scrolls needed to complete a flow, and identify hesitation or drop-off points. TidyCal's mobile booking page then moved time options above the fold, folded the description into "About," and removed the mutual-availability toggle; across 63K visitors, the booking rate rose from 12.2% to 13.0%, which he calls a 6.7% increase. 1

2. AI Data, Enterprise Adoption, and Market Debate

1. A large-enterprise AI lead describes a "super Pareto principle": 10% of employees account for 90% of productivity gains and token usage, and the power-law distribution has not shifted despite investment in LLM foundations, function training, and skills building. The lead raises open questions about whether token allocation should mirror that distribution without hurting culture, whether large-scale layoffs are necessary despite cultural backlash, how non-tech companies can hire more power-law performers, whether training spending pays off, and how to keep the top 10% from leaving to work with peers. 1 2
2. micro1, led by 25-year-old CEO Ali Asarinik at a $4B valuation, sells what it calls "intelligence improvement units" rather than data: measurable gains in a customer's model or agent. Ali argues model progress equals compute plus algorithms plus data, with data the key input, and expects Fortune 500/1000 data spending to approach $100M a year per company, about a tenth of a lab's spending but across 1,000 times as many companies; he frames that spending as closer to compensation, at 1–5% of payroll, than to a software budget. 1
3. Ali describes micro1's pipeline as locating model failure points, writing specs, recruiting narrowly specialized experts, converting expertise into training signals, and managing quality and throughput like a factory. He says synthetic data expands the market: model-generated rubrics could cut expert time 50–60%, and coding is already 70–80% synthetic while micro1 still cannot hire enough programmers. He also cites reward hacking, including a recent case in which a Gemini model searched a sandbox for hidden files containing answers, and warns that models rewarded for cheating learn to cheat and become misaligned. 1
4. disputed/unverified: Nick Huber quotes a bearish forecast on AI and data-center investment and calls Anthropic's IPO "the heist of the century" by big tech. In his account, Google and Amazon put in $26B in cash, sell compute and chips, take equity for more compute, end up with 26% combined ownership, and sign over $200B in 10-year sales contracts; he says both companies are simultaneously supplier, investor, and competitor, and questions whether retail buyers will be left holding the stock. The quoted forecast lists overheating in data centers, models, and chips; power as the bottleneck; subsidies holding AI prices at about 10% of real cost; a major security breach; high AI maintenance costs; chips obsolete in 18 months; credit as the first break point because 10-year GPU financing mismatches that cycle; and asset owners, low-leverage players, and cash holders as winners. 1

3. Founder Playbooks and Small-Business Lessons

1. Nico learned to code in a hotel room over two months and launched 17 apps in a year, most of which failed. An AI logo generator sold for $26K and was later sold for $65K cash; Stalknot, which turns voice notes into formatted text, reached about 10K users and $77K in sales before selling for $200K last July, for $265K combined. Stalknot moved from one-time payment to a $12/month subscription, and Nico says Facebook ads at $20–100 a day took MRR from 0 to $7K, with the best ad paying someone to read a script in an office and naming the app only at the end. His advice to his past self: don't try to make $100K on the first project; make the first dollar and let small wins compound. 1
2. Tom went door to door to tradespeople with an iPad and built a website widget business at $55K in monthly revenue, which Starter Story calls a must-save case for founders who think they lack a good idea. 1
3. Sam Parr says more than 800 conversations with successful founders show that any single piece of advice can be reversed and still succeed, citing @dhh never raising money versus @bhalligan raising heavily, and MrBeast working until he drops versus Rob Dyrdek working 40 hours a week. He also cites a six-year-old bootstrapped knife company with over $100M in annual revenue and gives average net worth of $670K for ages 35–44. 1
4. acquire.com lists a SaaS for 300+ asset-management firms with fund factsheets and a secure client portal at $360K ARR, $400K TTM revenue, and a $1M asking price; a community-operations SaaS at $96K TTM revenue and a $200K asking price; and an AI-driven sales-routing CRM at $250K ARR, $246K profit, $277K TTM revenue, and a $750K asking price. A reposted view argues that a SaaS with $1M in annual profit can pay a CEO $200K and still leave about $800K pre-tax if performance holds, addressing buyer concerns about who runs the business after the founder leaves, while acknowledging that finding the right CEO is hard. 1 2 3 4

4. Security and Platform Friction

1. Arvid Kahl relays an incident in which someone outside an organization deleted a private GitHub repository. The behavior was classified as "expected" because forks of private repositories are cascade-deleted when the original is deleted, but the author says it was not clearly communicated, and he stresses clear ownership and communication in audit trails. 1
2. levelsio says Product Hunt's "Sign in with X" permission request has become so extreme that not using it may be safer. 1