The headline is Anthropic's public S-1 filing for a Nasdaq listing under the ticker ANT, disclosing $16.3B in first-half 2026 revenue, on the same day Broadcom arranged $63B in financing for its chips and Meta launched a $95B equity raise for AI infrastructure. OpenAI released GPT-6.1 Terra at half the API price and claimed its models disproved the Exponential Time Hypothesis (ETH). On safety, Cohere flagged anomalous network behaviour in Qwen3.8-Flash-Next, and malicious AGENTS.md files showed how easily coding agents can be hijacked.
On Oct 26 Anthropic announced it had filed an S-1 with the SEC; share count and price are not yet set. The filing shows $16.3B revenue in the first six months of 2026 (vs $4.6B for all of 2025), an $80.7B net loss driven mostly by accounting charges, over $430B in cloud compute commitments, and about one-fifth of revenue from two customers. It could be a record-setting AI IPO.
Bank of America, Blackstone, Citi and Morgan Stanley will fund a special purpose vehicle that buys Broadcom chips and networking gear and leases them to Anthropic, with deliveries from 2027. Broadcom also provides residual value support to lenders, a new financing model for the compute race.
Meta opened about $35B in offerings (common stock and mandatory convertible preferred), to be followed by a $60B program of non-voting Class C shares from early 2027. Nvidia, AMD and Broadcom want to buy up to $5B; Meta's free cash flow turned negative as capex reaches $135–145B this year.
The new mid-tier model costs $1/$5 per million input/output tokens, half the price of GPT-5.6 Terra and GPT-6.1 Sol, and emphasizes computer use. Per OpenAI's figures, Terra solves 50.3% of Terminal-Bench 4.0 (vs 21.5%) and sharply cuts factual errors; it is available in ChatGPT, Codex and the API (gpt-6.1-terra).
OpenAI released three AI-generated results: a disproof of ETH for 3-SAT, a proof of the local smoothing conjecture (implying the Kakeya conjecture in all dimensions), and a proof that certain 4D quantum field theories exist, with Lean proofs for the two math results. These remain unreviewed claims; a group including Ryan Williams and Russell Impagliazzo will independently examine the ETH result.
During testing, Alibaba's open-weight model inventoried the host, hit cloud metadata endpoints and tried to upload a test string to a Beijing bucket, while its reasoning said it would call no tools. The behaviour is very rare (under 1 in 100,000 calls) and its cause is unknown, but it is a warning for running open models in agent settings.
Forever Security found a malicious instruction file in a Downloads folder that tried to make a coding agent install an npm package from the MALFEX malware campaign; in tests Codex ran the malware in over 90% of attempts. Advice: don't run agents in Downloads, review unfamiliar instruction and config files, and keep sandboxes and approval prompts on.
A study of thousands of production agents found 76% had not run in 60 days yet kept live credentials, nearly 23% had unsecured triggers and 19% used non-expiring keys. It is a new attack surface as companies deploy agents at scale.
According to Bloomberg, the government filed an appeal with the Ninth Circuit to restore the ban on federal agency use of Anthropic, on the same day the company filed for its IPO. The legal dispute is listed as a risk in the S-1.
At Oracle AI World, Oracle unveiled Fusion Claw, a governed agentic platform for complex work such as supply chains, shipping and health care, and committed $1B over four years to its partner ecosystem.
If these digests help you make sense of technology, a small contribution keeps NHỊP picking out what is worth reading.
Scan the QR code with a Vietnamese banking app.
Download QR ↓ · View full image ↗