AI news for Tuesday, November 3, 2026: Anthropic shows an $895B compute bill and an AI R&D warning

Anthropic's IPO filing had two numbers worth your time: what it owes for compute, and how close it thinks AI is to speeding up AI.

Business Hardware3 min
Anthropic's S-1/A shows $895B in compute commitments
Anthropic's amended IPO filing shows $17.0 billion in third-quarter revenue and about $895.4 billion in spending commitments, substantially all for compute. The revenue is real. The bill is bigger, and it is what will shape the price of your tokens.
Why it matters If you build on Claude, your supplier just showed you its cost structure: compute ate 58% of revenue last quarter, up from 48%, and about $69.8 billion of commitments come due within a year. Pricing and rate limits will follow that math. Keep a second model tested on your core work.
SRCSEC (Anthropic S-1/A)
Safety Policy3 min
Anthropic says it may "imminently" hit its AI R&D threshold
Anthropic's amended S-1 says internal estimates about future models indicate it "may imminently meet" its Responsible Scaling Policy's bar for dramatically accelerated progress driven by AI. Said in a securities filing, it deserves more weight than the usual lab forecast.
Why it matters The lab behind Claude says AI is speeding up its own model development, that it has spotted such speedups only after the fact, and that it is weighing rule changes to keep humans in control. If you build on frontier models, plan for faster model turnover and releases with more conditions attached.
SRCSEC (Anthropic S-1/A)
Agents Business3 min
Salesforce to buy Cresta, maker of AI voice agents
Salesforce signed a definitive agreement to acquire Cresta, whose AI voice agents handle customer calls end to end, without disclosing a price. The Information reports $2.7 billion. Voice is where support agents meet real customers, and Salesforce wants to own that line.
Why it matters If you run customer support on Salesforce, voice agents are heading into the stack you already pay for. If you build or buy voice agents elsewhere, the biggest support platform just picked a side. Put this on your vendor review before your next renewal.
SRCSalesforce
Research Models3 min
MIT team says an LLM can edit model weights directly
MIT CSAIL researchers say a frontier LLM, used as is, read and rewrote neural network weights without gradient descent, lifting a small Qwen model's MATH-500 score from 7.9% to 48.1%. It is an early blog post, not a peer-reviewed paper, and still the strangest result of the week.
Why it matters If it holds up, fixing one specific failure in a model could become an edit an AI makes in an afternoon, not a fine-tuning run. Nothing here is a product you can use today. The edited Qwen model and the full logs are public, so you can check the claim yourself.
SRCMIT CSAIL (Takehi et al.)
Hardware Business2 min
Marvell to buy Lumentum for $115B in an AI optics bet
Marvell agreed to buy Lumentum in an all-stock deal that values the maker of lasers and optical transceivers at about $115 billion, the companies said. Copper is running out of room inside AI data centers, and the chip companies are buying the light that replaces it.
Why it matters The links between GPUs are becoming as important to AI capacity as the GPUs themselves. This deal puts more of that supply under one company, years before it closes. It will not change your bill this quarter, but it shapes who sets prices for the networks your models run on.
SRCLumentum (SEC Form 8-K, Exhibit 99.1)

ShareX LinkedIn Hacker News Email

Prompt of the day

Stress-test how much your work depends on one AI provider

› I rely on AI tools for my work and want to know how exposed I am if one provider raises prices, tightens usage limits or changes its models. Here is how I use AI today:

[List each task you use AI for, the tool or model you use, roughly how often, and what breaks if it is unavailable for a week]

Do the following:
1. Build a table of my tasks with these columns: task, current provider, how critical it is (high, medium, low), how hard it would be to switch (easy, moderate, hard) and why.
2. Flag every high-criticality task that depends on a single provider.
3. For each flagged task, describe a fallback: what kind of alternative model or tool could do it, and what I would need to change (prompts, output formats, integrations).
4. Write a test plan I can run in one hour to check that each fallback gives acceptable results, with three sample inputs to try per task.
5. End with the three changes that would cut my risk the most this month, in order.

If my list is too vague to judge, ask me up to three clarifying questions first.

Open in Magai

Works in any AI chat. In Magai, run it against several models at once.

Get every edition

This one arrived in inboxes at 6am ET.

Anthropic's S-1/A shows $895B in compute commitments

Anthropic says it may "imminently" hit its AI R&D threshold

Salesforce to buy Cresta, maker of AI voice agents

MIT team says an LLM can edit model weights directly

Marvell to buy Lumentum for $115B in an AI optics bet