Fortylaunch

We help startups, enterprises, and brands create category-defining digital experiences... Turning ideas into revenue.

[Fortylaunch @fortylaunch]

One of AI agents' clearest early benefits may be cognitive, not economic. Hacker News flags context sprawl. ADHD communities describe AI as external executive function, helping people finally finish projects. Productivity isn't a property of the tool. It's the tool-brain pair.

[Fortylaunch @fortylaunch]

The flagship LLM is becoming internal tooling. Google shipped Gemini 3.6 Flash and two other Flash models. No new Pro. $7.50/M output, down from $9. Enterprise AI is classification, extraction, codegen. Revenue lives in cost-per-thousand-calls, not leaderboards.

[Fortylaunch @fortylaunch]

Meter the fix and teams ration the fixing. GitHub moved Copilot to usage-based billing on June 1. Organizations now manage AI-credit budgets and caps. Somewhere an engineering manager is deciding which vulns are worth the credits. The Q4 postmortem writes itself.

[Fortylaunch @fortylaunch]

Qualcomm's $3.92B Modular buy tells you what Nvidia's moat is. Modular built software to run AI across chip architectures. CUDA remained dominant anyway. A chipmaker with silicon but no developer habit bought it. The moat was never the chip. It's 19 years of muscle memory.

[Fortylaunch @fortylaunch]

You don't spend $2M a quarter on something voluntary. Q2 filings: Anthropic lobbying up 26% to $1.97M, OpenAI up 18% to $1.2M. The same labs now helping shape the White House's voluntary 30-day pre-release review framework. The disclosures are the price tag on the regulatory moat

[Fortylaunch @fortylaunch]

A denied acquisition rumor that keeps spreading is data, not noise. Physical Intelligence's CEO said no. AI Twitter kept going. Months of AI acquisition reports set the prior at "labs buy everything." A denial can't outrun a prior that's right in aggregate.

[Fortylaunch @fortylaunch]

The AI industry has one landlord, and the rent just went up. TSMC told customers it's raising prices up to 10% in 2027 — advanced and mature nodes both (Nikkei, July 21). Every model of falling token prices assumes flat foundry costs. The foundry just voted no.

[Fortylaunch @fortylaunch]

Export controls just flipped. Per the FT, Beijing is considering controls on AI model weights and training data, consulting Alibaba, ByteDance, and Zhipu. Many built on Qwen and DeepSeek for cheap inference. Open weights built the leverage. Controls collect the rent.

[Fortylaunch @fortylaunch]

Voluntary federal review is a moat dressed as a concession. OpenAI, Anthropic, Google, Microsoft, and xAI agreed to give the U.S. government 30-day pre-release access to qualifying frontier models. A step smaller labs can't match. Meta's decision not to join is the tell.

[Fortylaunch @fortylaunch]

Anthropic's $1.5B settlement priced provenance, not training. AI training was transformative. Acquiring pirated source material wasn't. Clean data is now a capitalized asset with a known floor. Only the biggest labs can carry that line item. That's a moat, not a deterrent.

[Fortylaunch @fortylaunch]

Hut 8 fully leased its 1GW Texas campus. A 15-year, $9.8B deal with an investment-grade AI infrastructure customer. Bitcoin mining was never about the coins. It was a land grab for grid interconnection. The most mocked industry of the 2010s became AI's landlord.

[Fortylaunch @fortylaunch]

Moonshot temporarily paused new paid subscriptions for Kimi K3 on July 20 after demand outran compute. A 2.8T-param model rationed not by price but by chips. Benchmarks rank models. Capacity decides who gets served. A model you can't buy loses to a worse one you can.

[Fortylaunch @fortylaunch]

The labs are picking their referee before one gets chosen for them. A leading AI lab CEO proposed an industry-funded AI FINRA. The Treasury Secretary is helping develop a similar SRO under the SEC. SROs exist to pre-empt regulation. That's the pitch.

[Fortylaunch @fortylaunch]

Open weights are a go-to-market, not a gift. Thinking Machines raised a $2B seed and released Inkling, 975B params, free to download. Against ChatGPT's default status, a closed API from a new lab is dead on arrival. The $2B doesn't buy secrets. It buys adoption.

[Fortylaunch @fortylaunch]

Engineering ladders still promote authorship while the scarce skill is verification. Agents made code cheap to write and expensive to trust, yet review is still the janitorial work you do between 'real' contributions. The bottleneck moved; the org chart didn't.

[Fortylaunch @fortylaunch]

Investors aren't pricing DeepSeek. They're pricing the destruction of everyone else's economics. $74B valuation. ~$450M revenue. 150x. Priced dramatically below frontier rivals. You don't pay 150x for margins that don't exist. The price war isn't the tactic. It's the product.

[Fortylaunch @fortylaunch]

The AI race is now a supply chain. NYT: Meta in talks to sell Anthropic, its direct model rival, up to $10B of compute over two years. The company that spent 20 years monetizing attention wants to monetize watts. Models expire. Racks collect rent either way.

[Fortylaunch @fortylaunch]

Anthropic's S-1 turns AI safety into securities law. Once public, "we'd pause if..." commitments become disclosures investors can litigate if materially misleading. Years of AI governance design. The enforcement mechanism turns out to be Rule 10b-5.

[Fortylaunch @fortylaunch]

Companies didn't cut jobs against AI. They cut them against a demo. TrueUp: roughly 150,000 tech layoffs through early June. 40% of May announcements cited AI. Many are now rehiring. Because AI does the routine part. The salary was always paying for the other 40%.

[Fortylaunch @fortylaunch]

Claude is measurably warmer in Hindi than in English. Anthropic's own study confirmed it. Model personality isn't designed. It's absorbed from each language's politeness norms. We aligned the English model. A billion Hindi speakers got whatever came with the data.