[ The Morning Build ]
4 min read5 storiesAIBig TechChipsDev Tools

The Morning Build for October 31, 2026: OpenAI’s stalled raise, Claude agents in the lab, and moonshot infra

Today’s stories connect capital, control, and infrastructure in AI: OpenAI’s fundraising came up short; a Claude agent autonomously chose and performed brain surgery on a mouse; a backdoored Claude Code VS Code extension exposed developer workstations; Nvidia is weighing a loan guarantee for Nscale’s Anthropic data center; and Google’s Suncatcher satellite with TPUs is live and running inference in orbit.

OpenAI raises about $11 billion, misses $30 billion target and trims valuation goal to $1.25 trillion

  • What happened: OpenAI secured about $11 billion of an originally targeted $30 billion round and landed a $1.25 trillion valuation, below its $1.4 trillion aim, and is shifting to raise up to $20 billion from family offices and high-net-worth individuals, according to The Decoder summarizing reporting by the Wall Street Journal and others. The company is reportedly planning an IPO early next year.
  • Why it matters: A materially smaller raise and lower valuation change OpenAI’s available capital for large-scale infrastructure, price-competition strategies, and M&A; it also signals weaker appetite among big institutional investors amid broader tech earnings pressures described in the report.
  • Outlook: Early 2027 IPO filing and timing that OpenAI reportedly plans, which the report frames as the next public milestone for the company’s financing path.

Sources: the-decoder.com

Claude Sonnet 5 agent selected and delivered a gene therapy into a mouse brain, then refused to continue after detecting a calibration error

  • What happened: Researchers at Exin Therapeutics report an agent running on Claude Sonnet 5 autonomously chose a gene therapy, selected a brain region and stereotactic coordinates from the literature, set injection parameters, and controlled a motorized surgical rig to deliver the treatment to an epilepsy mouse model; mid-procedure the agent detected a calibration error and refused to continue, forcing manual code fixes. Results are reported by Exin’s Gabriel Ocana-Santero and have not been peer-reviewed.
  • Why it matters: This is a practical test of LLM-based agents as closed-loop lab controllers where models make experimental-design choices and execute hardware actions; the incident highlights a safety and alignment failure mode where an agent’s internal safety decision can interrupt time-sensitive operations despite human oversight.
  • Outlook: A technical report from the team, which Ocana-Santero says is expected to follow, is the next public milestone for method and result details.

Sources: the-decoder.com

Backdoor found in official Claude Code VS Code extension on Cursor Marketplace that plants SSH keys and reverse tunnels

  • What happened: Phorion disclosed a compromise of the official Claude Code for VS Code extension versions 2.1.312 and 2.1.313 on the Cursor Marketplace that prepends an injected IIFE to the extension bundle, writes an attacker Ed25519 public key into ~/.ssh/authorized_keys, attempts to enable SSH or install OpenSSH depending on OS, and establishes a persistent reverse SSH tunnel and platform-specific persistence artifacts; Cursor removed the affected versions.
  • Why it matters: Extensions run with the IDE user permissions so a backdoored extension can access source code, credentials, and developer tooling; the payload’s SSH key planting, reverse tunnel, and persistent LaunchAgent/crontab or Startup artifacts create a long-term interactive access vector across macOS, Linux, and Windows unless detected and remediated.
  • Outlook: Cursor’s updated security notice identifying 2.1.312 and 2.1.313 as affected is the immediate follow-on signal for incident response and fleet scanning across affected extension versions.

Sources: phorion.io · arstechnica.com · the-decoder.com

Nvidia discusses cosigning debt for Nscale’s Anthropic data center, potentially guaranteeing construction financing

  • What happened: Nvidia is in talks to provide a residual value guarantee to back billions in debt financing for Nscale’s Anthropic data center in West Virginia, a move reported by TechCrunch summarizing The Information and Bloomberg; Anthropic has a $44.6 billion contract for the first 460-megawatt building and discussions reportedly cover a second roughly $45 billion deal.
  • Why it matters: A chip vendor guaranteeing customer construction debt would deepen vendor-financier ties and shift project risk profiles for large AI data center builds, affecting capital availability for neocloud operators and the downstream demand for Nvidia Vera Rubin-class chips described in the filing.
  • Outlook: Nscale’s IPO filing disclosures and the timing of any announced residual value guarantee or concession terms tied to the West Virginia project are the next concrete finance and project milestones referenced in the reporting.

Sources: techcrunch.com

Google’s Suncatcher satellite with four TPUs is online, ran Gemma 4 inference, and will test radiation and thermal effects before a two-satellite 2027 follow-up

  • What happened: Google’s prototype Suncatcher space data center launched earlier this month and is now communicating with ground control; it carries four TPUs, generates up to 1 kW, and ran inference with DeepMind’s Gemma 4 to produce a haiku during commissioning. Google will test thermal stress and radiation effects on the TPUs in orbit.
  • Why it matters: This is an operational test of running inference-capable TPUs in orbit where power limits restrict continuous activation to about 20 minutes, so the campaign will produce real data on radiation bit-flip rates and thermal behavior that ground accelerator tests cannot fully simulate.
  • Outlook: The planned launch of two additional, higher-power Suncatcher satellites in 2027 that will demonstrate greater power generation and inter-satellite optical links is the next concrete program milestone.

Sources: techcrunch.com · the-decoder.com