AI News Roundup: October Twenty-Six - The Funding Wall

Overview: Monday turned AI's capital problem into arithmetic. Anthropic filed to go public, Meta moved to raise $95 billion in equity, and SpaceX locked in a $40 billion GPU lease that never touches its balance sheet. Microsoft offered buyouts to roughly 60,000 people, OpenAI shipped a computer-use model, and the pull request itself took its first real hit from an agent-native Git forge.

Anthropic Files for Nasdaq IPO, Opening Its Books to the Public

Anthropic filed its Form S-1 with the SEC to list on Nasdaq under the ticker ANT, disclosing $16.3 billion in first-half revenue against an $80.7 billion net loss driven mostly by accounting charges. For the first time, a top-tier AI lab has been forced to show the shape of its unit economics to the public market.

The filing matters less for the loss than for the commitments. Anthropic has signed multi-year take-or-pay compute contracts, so from here the company manages pricing, rate limits and model access against quarterly expectations rather than a research roadmap. Builders should expect access terms to become a financial instrument, not a product decision.

---

Meta Moves to Raise $95 Billion in Stock to Fund AI

Meta plans to raise about $95 billion in three steps: $35 billion in underwritten public equity offerings, then up to $60 billion through an at-the-market program once its non-voting Class C shares trade. Nvidia, AMD and Broadcom were each interested in buying up to $5 billion of the stock, per CNBC - chip vendors financing their own demand.

Shares fell on the news. Meta guides full-year capital expenditure at $135-145 billion, is building four superclusters above one gigawatt each, and posted negative free cash flow last quarter. The money is earmarked for infrastructure, its Muse agent and the app family: capacity, not headcount, is what gets funded.

---

SpaceX Closes $40 Billion GPU Debt Deal, Warns Neoclouds

SpaceX reached a binding commitment for roughly $40 billion in financing, structured as a lease: an Apollo-managed entity buys Nvidia's next-generation Vera Rubin hardware and leases it back over about five years. The equipment stays off SpaceX's balance sheet, leases start in the first half of 2027, and the CFO claims sub-one-year payback on AI infrastructure.

Musk's framing was blunt - with debt markets closing to companies whose revenue does not exist outside AI data centers, most of the 200-plus neoclouds face what he called "guppy-level death and consolidation." Buyers with cash flow from something other than GPUs get preferential terms.

---

Microsoft Offers Buyouts to 60,000 Staff in AI Restructuring

Microsoft opened a voluntary exit program to employees at experience level 70 and below in the US, Canada, India and the UK, covering around 60,000 people, or about 27 percent of its 223,000-person workforce. Employees decide in early December and leave on December 31.

The exclusions are the signal. Data center and sales roles are largely shielded while general engineering and operations absorb the cut, six months after a smaller voluntary retirement round and a July layoff of 4,800. Capital expenditure for the September quarter was guided above $50 billion against $34.9 billion a year earlier.

---

Entire Replaces the Pull Request With Agent-Scored Evals

Thomas Dohmke, GitHub's former CEO, launched Entire, a Git forge built for agent-written code. It swaps pull requests for "Trails": customizable evals scored by agents on every push, with humans holding the gates instead of reading diffs. The pitch includes 89x faster pushes than GitHub.

A pull request assumes a human reads every line and that reading is the quality bar. Once agents generate most commits, the reviewable artifact has to become the eval history and the agent trace. Ars Technica's reporting that coding agents generate more code but not more software is the counterweight Entire is answering.

---

OpenAI Ships GPT-6.1 Terra With Sharper Computer Use

Late in the day OpenAI released GPT-6.1 Terra, available immediately to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex. OpenAI describes it as "a fast, efficient model that combines detailed visual reasoning and precise computer use alongside clearer, more natural writing," and says it excels at "the communication and computer-based tasks that fill the workday."

Terra is an upgrade to an existing line rather than a new frontier tier, and the emphasis on computer use is the tell: the competitive ground has moved from reasoning benchmarks to reliably driving real interfaces.

What to Watch Tomorrow

• Hyperscaler earnings: Microsoft, Alphabet and Amazon report this week, and funding - not demand - is expected to dominate the calls after Meta's raise. (CNBC Investing Club)
• Neocloud stress: CoreWeave hit a 52-week low of just above $51 on Monday before recovering. Watch whether smaller GPU clouds can roll debt as the window narrows.
• Agent-hostile retail: The RealReal tightened bot defenses after The Verge showed Meta's Muse could hold items in a cart for hours, blocking other shoppers. Agents can still surface listings but now struggle to log in, add to cart and check out. (The Verge)

Builder Impact

• Your model vendor now has a share price. Anthropic's S-1 makes pricing and rate-limit decisions legible as revenue management. Budget for terms changing quarterly and keep a second provider warm.
• Equity and leases are replacing cheap debt. Meta raising $95 billion in stock and SpaceX leasing $40 billion of GPUs through Apollo say the same thing: capital cost is the binding constraint, so compute prices will not fall as fast as model efficiency improves.
• Assume neocloud consolidation. If you rent inference from a small provider, contract length and egress terms matter more than list price this quarter.
• Review artifacts are moving from diffs to evals. Whether or not you adopt Entire, the pattern it encodes - agents score, humans gate, traces are the audit trail - is what CI should look like once agents write most commits.
• Test computer use, not benchmarks. Terra's launch is a reminder that the models competing on operating a screen are the ones that will run your automations.