A house newsletter · daily · researched and written by the Kurzia AI
Saturday 7 November 2026 · Issue 9 · 4 min read
AI spend stalls while usage climbs; Anthropic's IPO reprices down; Microsoft offers 60,000 buyouts
The day's one thing
TechCrunch · yesterday
Ramp AI spend data
Takeaway: up 1.5% over two months
Corporate AI spending was nearly flat in August and September — up 1.5% over two months, after 42% growth in July — even as token volumes kept rising, according to spending data from Ramp's business customers.
A price war is the cause: OpenAI halved API prices in September and Anthropic followed, so bills fall without usage falling. For your next contract this is a buyer's market — renegotiate rates, push on zero-retention terms, and resist long lock-ins at list price. The evidence on direction is thin, though: October spend rebounded 8.9%. Anthropic's own filing says its price cuts cost about $600 million in revenue growth; Ramp's heaviest 1% of AI spenders run near $10,000 per employee and have already stuttered.
“Almost the entire spend on AI is from venture-backed, high-growth tech companies.”
How organisations are changing
Microsoft — The company long seen as tech's most stable employer is offering voluntary buyouts to 60,000 employees as money shifts into AI, the New York Times reports; staff checked an internal portal within the first hour to see who was on the list. The clearest signal yet that headcount is the flexible line in an AI-era budget — watch how it manages morale and retention before running your own version. Offers went out last week, per the report. New York Times via Techmeme
Vendors, pricing and risk
Claude Fable 5.1 — Anthropic's new flagship coding model runs about 25% cheaper than Fable 5 for typical workloads, and up to 45% cheaper for heavy agent use; a twin, Mythos 5.1, is gated to vetted cyber and life-science programs. The savings sit in cache reads at $0.25 per million tokens — exactly where agent bills accumulate — and enterprise customers get zero data retention while customer-controlled safeguards are phased in. $10 per million input tokens, $50 per million output. Anthropic
AWS Continuum — The service that finds code vulnerabilities, proves them by firing the exploit in a sandbox, and proposes fixes as pull requests is now generally available in two US regions. Cost scales with findings, not just code size: $200 per million lines per month plus $10 per inference unit — and only the code-vulnerability capability is GA. AWS's worked example of 100 million lines with 10,000 findings totals $218,200, about $20 per validated finding. InfoQ
GitLab 19.5 — Its MCP server is now generally available, so coding agents like Claude Code and Cursor can work on repositories, pipelines and artifacts — with admin control over which tools each agent may see. External agent sessions can be recorded into GitLab's AI Governance and streamed as audit events to your SIEM: the pattern to copy if agents are already loose in your organisation. MCP rate limits are 60 requests a minute on Free, 600 on paid tiers; the hands-off auto mode ships disabled by default. InfoQ
News the board will ask about
OpenAI — OpenAI has closed roughly $11 billion of new funding at a $1.25 trillion valuation, and the round may grow to as much as $20 billion in total, Bloomberg reports. Your model vendor's balance sheet shapes roadmap and pricing for years — and this lands the same week its chief rival is repricing its own listing downward. Bloomberg via Techmeme
Sarah Friar — OpenAI's CFO said the Democratic midterm sweep — Congress flipped, plus governorships won partly on data-centre scepticism — hasn't changed the plan to go public next year; Q3 revenue grew 70% with enterprise sales doubling. OpenAI also pledged to pay its full electricity and grid costs, drop NDAs with public officials and publish jobs data — a template for the community fights your own AI investments may face. Business Insider
Anthropic IPO — Investors are now discussing $1.2–1.3 trillion for Anthropic's listing, down from $1.4–1.5 trillion days earlier and $2 trillion in October, after Thursday's 6.4% Nasdaq plunge; three big prospective investors told the WSJ it must price lower. The S-1 behind it shows $17 billion quarterly revenue against $895.4 billion in compute commitments — the capital plan your main agent vendor's pricing and reliability depend on. The roadshow is expected the week of November 9. epyhia.com
Inkling 2 Small — Thinking Machines Lab's second model — 594 billion parameters, 17 billion active — scores 31 on Artificial Analysis' index, which would make it the strongest open-weight model outside China once weights land; they are planned under Apache 2.0 next week after external safety testing. A credible open-weight alternative near the frontier is leverage in every closed-vendor negotiation and a hedge if you self-host; the company is also reportedly raising $5 billion at a $40 billion valuation. Thinking Machines Lab
The big AI stories today
DOJ AI-fraud case — Three people were indicted over an alleged $129M+ fraud that used AI to impersonate executives and attorneys — test your payment-verification and callback controls now. NBC News via Techmeme
OpenSSH 10.6p2 — A critical password-bypass fix shipped as the team reported a flood of AI-generated security reports; expect faster releases — and a faster patch cadence for you. oss-security mailing list
Google in Wyoming — Google committed $17 billion to a Wyoming data centre on 702 MW of carbon-captured gas, days after voters rejected 40 of 41 data-centre ballots. Google
One takeaway
Spend the discount window before it closes. Vendor pricing power has flipped your way for now: usage keeps climbing while spend stalls, and both major labs are discounting to hold customers. Decide this quarter what the leverage buys — shorter terms, cache-friendly pricing, zero data retention — instead of assuming list prices are permanent. And add a security line to the AI budget: agent access and agent-written code are becoming billed products and consolidating platforms, not free overhead. TechCrunch