From San Francisco
1 |
The Editorial |
Good morning.
Today's three picks all turn on what the money has actually bought so far.
Anthropic signed a $35 billion cloud deal with Lambda on August 31. On September 14 it cuts Claude Code's weekly allowance 17%, ending the boost that has run since May 13.
Reframe Systems raised $40 million to build houses in robotic microfactories. It has finished 10 homes, eight of them occupied, and its Billerica plant opens October 5.
Lee Eason parsed 30,000 SEC filings from 1,008 companies and found no AI effect on revenue, margin or R&D intensity. His detection floor was 2 to 3 points of margin.
Stay curious,
Marcus Schuler
2 |
The Big Story |
Anthropic will cut Claude Code's weekly usage allowance 17% on September 14, ending the temporary boost that has run since May 13.
The 50 percent promotion expires and is replaced by a permanent 25 percent increase over the pre-May baseline. Pro, Max, Team and seat-based Enterprise plans are affected. Five-hour session limits do not change.
Users describe rationing work and checking consumption every 30 minutes. Anthropic says it is working on changes intended to give customers more visibility and control over usage, without naming them or a date.
Why This Matters:
- Teams that sized Claude Code seats against the May allowance will hit the weekly ceiling sooner from September 14 onward.
- Anthropic committed $35 billion to Lambda capacity on August 31, so the constraint being priced here is not raw compute.
3 |
Also Today |
Reframe Systems raised $40 million led by Energy Impact Partners to expand robotic homebuilding microfactories across North America.
The company has finished 10 homes, eight of them occupied, and plans 114 more units over the coming year. Its Billerica plant opens October 5 with designed capacity of 500 multifamily units a year. The test is whether a factory that has not run yet can hold the cost claims the round is priced on.
4 |
The Outside Read |
Lee Eason's blog tests AI's economic payoff against SEC filings for 1,008 companies in the June 2022 Russell 1000 cohort.
Eason parsed more than 30,000 filings dated January 2, 2019 through August 21, 2026 into 1.4 million financial facts and found no detectable AI effect on revenue, margin or R&D intensity. For the dataset through August 21, 2026, his power analysis sets the detection floor at 2 to 3 percentage points of margin, which explains why company accounts cannot yet resolve smaller productivity gains.
5 |
The One Number |
6 |
Today's Headlines |
- OpenAI said ChatGPT Ads reached a $1 billion annualized run rate in under 200 days, with self-service buying opening across India, Europe, the Middle East and North Africa.
- Nvidia put $3.5 billion into MediaTek convertible bonds on August 31, with MediaTek adopting NVLink Fusion for customers building custom accelerators.
- The European Commission designated ChatGPT a very large online search engine under the Digital Services Act on August 31, giving OpenAI four months to comply.
- Anthropic deployed a classifier that blocks suspected sandbox escapes after Claude models reached live systems during two evaluations in July and August.
- EuroHPC ordered a €387.8 million LUMI-AI system from Bull, its largest contract, for installation in Finland in the second half of 2027.
- Florida revoked permits and gave local agencies 30 days to remove license-plate readers, including Flock cameras, from state highway rights of way.
Tue 9/1 |
Economy: the Bureau of Labor Statistics releases July job openings and labor turnover at 10 a.m. Eastern. |
Wed 9/2 |
Chips: SEMICON Taiwan opens its exhibition in Taipei, running through September 4. |
Wed 9/2 |
Fed: the Federal Reserve publishes its Beige Book at 2 p.m. Eastern. |
Thu 9/3 |
Economy: the Bureau of Labor Statistics releases revised second-quarter productivity and costs at 8:30 a.m. Eastern. |
7 |
The 5-Minute Skill |
A heated customer thread mixes evidence, interpretation and promises. Use an LLM to prepare a brief before the response meeting.
Your raw input:
The prompt:
Why this works: The source-citation rule separates what the record proves from what the team assumes. Holding the remedy until the key unknown is named reduces the risk of a costly promise made under pressure.
What to use: Claude handles long email chains well. ChatGPT is a strong fallback when the input includes tables or structured incident data.
8 |
AI Toolbox |
Mole is a terminal-based research agent that searches the web, checks each claim against a verbatim source quote and stops at a spending ceiling you set. It keeps an auditable claim trail and enforces the budget. The software was free under the Apache License 2.0 on August 18, 2026, while model and Brave or Tavily API charges remain yours.
How to use it:
- Install Mole on macOS or Linux with
brew install lajosdeme/mole/mole. - Select Tavily with
mole config set search.provider tavily, then save its key withmole config set search.tavily-key tvly-.... - Select Anthropic with
mole config set llm.provider anthropic, then save the API key and model name withmole config set llm.api-key sk-...andmole config set llm.model claude-sonnet-5. - Run
mole doctorand fix every failed configuration check. - Start a report with
mole research "your precise research question" --usd 0.50to impose a ceiling of 50 US cents on that run. - Run
mole trace <session-id>to inspect the cost and timing of each call.
Where it falls short: Mole has no browser interface and still requires separate model and search-provider API keys, so nontechnical users face more setup than with Perplexity.
9 |
AI Image of the Day |
10 |
The Rausschmeisser* |
The Defense Department added ChatGPT Mil and Grok for Government to its GenAI.mil platform, which had already onboarded 1.7 million unique users and is available to a workforce of 3 million (TechCrunch, August 31, 2026).
Our take: GenAI.mil is for unclassified administrative work, which the announcement lists as logistics, planning and policy. Three million people who previously moved paper will now move it faster, and the Pentagon has solved the memo. The pitch is that keeping this in-house stops sensitive data leaving through consumer channels, which is true and also an admission about what had been happening until now.
Grok is the interesting pick. The department chose the assistant with the most eventful public record and put it in front of a workforce larger than Chicago. Somewhere a colonel is asking it about a duty roster. The unclassified label is carrying a great deal of weight.
IMPLICATOR