Today’s through-line is compute expansion meeting grid politics, pre-IPO financial transparency, and the shift from models that “chat well” to systems that “get things done”: OpenAI launches a nationwide Grid Protectors Program and pilots up to about 1,000 megawatts of demand response at its Effingham, Georgia campus; Anthropic’s amended S-1 fills in third-quarter results and nearly $900 billion in compute commitments; OpenAI releases LifeVal for everyday tasks; and Trajectory closes a $45 million Series A betting on continual post-training for domain experts.

1. OpenAI launches Grid Protectors Program; Effingham pilot may cut up to about 1,000 MW

OpenAI published a blog on November 2, 2026 formally launching the Grid Protectors Program, saying it will work with utilities, regulators and policymakers so future compute campuses can lower electricity demand when the grid is under stress, and stressing that households and small businesses should not subsidize the power costs of compute campuses. The company said U.S. data-center capacity is expected to exceed 100 gigawatts by 2030 and wants “flexible load” written into the core capabilities of subsequent campus construction and operations; how much can be cut, how fast, and for how long will vary by location, with details to follow as agreements are finalized. Depending on local needs and equipment, on-site generation or storage, batteries and power controls may also be used to smooth load changes.

The pilot sits at the planned Effingham County, Georgia campus. OpenAI said it has committed to paying the full infrastructure and power-service costs required to serve the project; the agreement with Georgia Power includes up to about 1,000 megawatts of flexible demand response, which the company said could free electricity equivalent to the average use of about 700,000 homes in extreme-heat scenarios. The Verge (Emma Roth, November 2) and Axios (Ben Geman) noted the move comes as backlash over data-center power and community impact is rising in many parts of the United States; OpenAI has also been shifting its narrative from traditional “data centers” to “AI compute campuses.” Axios quoted OpenAI energy-policy lead John McCarrick saying the industry is less insistent that data centers must maintain “five nines” availability. OpenAI also stressed that flexibility cannot replace new generation and transmission investment.

Takeaway: Packaging curtailable load as a national program tied to a local utility agreement shows data-center conflict has moved beyond land use and noise into electricity rates, peak capacity and political acceptability.

Sources: OpenAI; The Verge; Axios

2. Anthropic amends S-1: about $17B Q3 revenue, roughly $895.4B in compute commitments

According to CNBC (Ashley Capoot, about Eastern Time November 2) and The Verge (Jess Weatherbed), Anthropic filed an amended Form S-1 with the U.S. Securities and Exchange Commission, filling in third-quarter financials through September 30, 2026. Coverage said the company’s first-nine-months revenue was about $33.24 billion, up roughly 1,200% year over year; net loss over the same period was about $104 billion, largely from accounting fair-value changes on convertible notes, derivative liabilities and SAFEs. On a quarterly basis, third-quarter revenue was about $17 billion, up about 1,127% year over year and about 47% quarter over quarter (growth slowed from roughly 144% in Q2 and 132% in Q1); third-quarter net loss was about $23.29 billion.

The amendment disclosed purchase commitments and lease agreements totaling about $895.4 billion in committed spend, “almost all” related to compute and infrastructure. CNBC said Anthropic draws substantial compute from Amazon, Google and SpaceX and also relies on some neoclouds; market concerns about data-center financing costs and capacity delivery are entering the pre-roadshow window alongside this filing. The SEC amendment also shows nine-month revenue through September 30 of $33.242 billion and reiterates Claude’s rapid enterprise adoption since commercialization in the first quarter of 2023.

Takeaway: Putting revenue scale in the same frame as nearly trillion-dollar compute commitments hands both the growth story and capacity-and-financing risk to public-market pricing.

Sources: CNBC; The Verge; SEC S-1/A

3. OpenAI releases LifeVal: measuring personal-assistant progress on everyday tasks

OpenAI released LifeVal on November 2, saying its goal is to measure model progress on “tasks that actually matter in everyday life,” rather than stopping at knowledge or academic benchmarks. The post lists parenting, accessing government services, budgeting, school prep, adapting recipes, ordering groceries, designing fitness plans and similar scenarios, and stresses that a truly useful personal assistant must take action and finish work—not merely offer advice or chat replies.

On task sourcing, OpenAI said it drew on the American Time Use Survey and the Multinational Time Use Study, and added a survey of 3,995 respondents described as demographically representative to identify where people most want personal-assistant help; coverage spans caregiving, health, education, household administration, personal finance, shopping and government services. The page gives sample workflow scores such as about 93.1% for choosing and enrolling in a school after a move, about 98.0% for organizing volunteers before a flood, about 100% for finding a home and handling title-related steps for a family, about 96.3% for watching a game remotely with grandparents, and about 90.4% for preparing a meal plan for a parent newly diagnosed with diabetes; a table also shows GPT-6 Astra “best recorded scores” on selected tasks, including 100% entries for booking a sensory class for a toddler, scheduling a car inspection and ordering halal groceries. OpenAI said it spent the past year training models on browser use, computer use and connector tasks to push ChatGPT from a chat tool toward a more action-capable personal assistant, but the page does not provide a fully reproducible public task set or scoring rubric for independent download.

Takeaway: The company is shifting the evaluation narrative from leaderboard contests toward “getting things done,” while still keeping task design and result release inside the lab, so external verifiability remains limited.

Sources: OpenAI

4. Trajectory closes $45M Series A at about a $300M valuation

According to TechCrunch (Marina Temkin, about 3:13 p.m. Pacific Time on November 2, 2026), roughly year-old startup Trajectory announced a $45 million Series A at about a $300 million valuation; Sequoia led, with participation from Nvidia, Bessemer Venture Partners and Bain Capital Ventures. The company was founded by three Stanford classmates with backgrounds spanning Google DeepMind and Apple and currently has about 18 people; co-founders include Ronak Malde, Michael Elabd and Arjun Karanam.

The product thesis is continual learning and post-training automation: enterprises can start from open-source models, connect the platform to their own applications to collect performance feedback, then retrain so models keep improving on specialized work such as law and finance, reducing reliance on in-house research teams or expensive “forward deployed engineers” (FDEs). Coverage said Harvey and Rogo have used the platform to post-train open-source models; other customers include Decagon, Mercor and Nvidia; the company is opening the post-training platform to the public. Co-founders emphasize that humans define “what good looks like” and write that into the improvement loop, which they see as more controllable than fully autonomous recursive self-improvement (RSI). On the competitive map, coverage named Thinking Machines Lab (Tinker), Applied Compute and Prime Intellect among others.

Takeaway: The round bets on a post-training interface operable by domain experts, productizing continual improvement from a research workflow and tying it directly to the delivery cost of vertical AI applications.

Sources: TechCrunch

---

One-line takeaway: OpenAI answers compute-campus power disputes with the Grid Protectors Program, Anthropic’s amended S-1 shows about $17 billion in quarterly revenue and nearly $900 billion in compute commitments, LifeVal pivots evaluation toward everyday task completion, and Trajectory raises $45 million Series A for continual post-training.