The Sequence Radar - Issue #963: This Week in AI: The Future Has a Payment Schedule
Was this email forwarded to you? Sign up here
Next Week in The Sequence:
We continue our series about recursive self-improvement.
We discuss another robotics company.
We dive into Google’s new Gemini models and Mistral’s Timestral.
In our opinion section, we dive into the world of GPU derivatives.
Subscribe and don’t miss out:
TheSequence is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.
📝 Editorial: This Week in AI: The Future Has a Payment Schedule
Apparently, building artificial intelligence requires two kinds of models: one that predicts the next token, and another that persuades a credit committee the GPUs will pay for themselves. This week, both were busy.
Microsoft’s September quarter produced $58 billion in operating cash flow against $39.4 billion of cash capital expenditure. Azure revenue grew 48%, and Microsoft repaid $4.9 billion of debt. That is an enormous construction project attached to an enormous cash machine. Microsoft’s existing business gives it considerable room to finance the next one. The infrastructure bill is intimidating; the ability to pay it is rather reassuring.
Alphabet’s spreadsheet was more dramatic. Google Cloud revenue grew 103%, but capital spending of $51.4 billion exceeded operating cash flow of $42.3 billion, leaving negative free cash flow of $9.1 billion. Alphabet also raised approximately $29 billion through debt issuance. The customers are showing up. Unfortunately, so are the invoices.
Google supplied fresh reasons to keep building. Tuesday brought Genie 4, a world model generating interactive simulations with native audio. Wednesday brought Gemini 4 Carbon, emphasizing cybersecurity and sustained engineering workflows. Both start with restricted access. The potential commercial logic is straightforward: better agents could handle more valuable work; interactive worlds could support robotics and simulation. Turning those capabilities into dependable revenue will determine how comfortably the infrastructure gets paid for.
Meanwhile, Google expanded AI Mode advertising to ten additional countries. This is reassuringly familiar. Invent a new interface to human knowledge, then find somewhere tasteful to put a sponsored result.
Together, these earnings illustrate the financing question: how much cash remains after building the infrastructure? Spectacular growth and increasing dependence on external capital can coexist. Accounting permits this particular form of emotional distress. Debt helps when revenues arrive later than the equipment bill. The risk is that lenders want predictable payments while AI businesses face uncertain utilization, hardware replacement and aggressive pricing.
SpaceX’s AI expansion supplied a financing spectacle: a binding commitment for approximately $40 billion of GPU equipment financing through an Apollo-managed finance lease. The deal covers NVIDIA Vera Rubin hardware, with closing expected before year-end and equipment delivery beginning in the first half of 2027. Closing remains subject to conditions.
Investors finance equipment; SpaceX gets compute capacity and payment obligations. Management says its superintelligence infrastructure investments are seeing payback periods under one year. That is management’s claim, and an excellent sentence to have available when discussing a forty-billion-dollar transaction.
The bull case: demand grows, useful products multiply, and financing brings productive infrastructure online sooner. The bear case: competition lowers compute prices, hardware becomes less valuable, and fixed payments remain fixed.
My reading of the week: the industry is making real technical progress and taking real financing risk. Neither fact cancels the other. A model can improve every month. Its lender still expects the payment on Tuesday.
🔎 AI Research
Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories
AI Lab: Google Research, Cornell University, University of Southern California
Summary: This paper introduces a sleep framework for continual learning, where an LLM alternates between an active phase and an offline phase that consolidates short-term context into long-term parameters. Consolidation distills in-context memories into new low-rank Mixture-of-Experts blocks, and a dreaming stage generates and filters synthetic training data. It scales near-perfectly on BABILong up to 10M tokens and scores 79.2% on AIME 24.
RSI Arena: Towards Recursive Self-Improvement with Human Feedback
AI Lab: Bake AI, University of Notre Dame, UCLA, Stanford University, NVIDIA Research, New York University, University of Washington, UC Santa Barbara, Scale AI, Hugging Face, MIT
Summary: This paper introduces RSI Arena, a study of whether AI agents can autonomously post-train a language model and then improve it using user feedback. Ten frontier models each act as the training agent, starting from NVIDIA Nemotron 3.5 Lightning 30B-A3B Base. All ten produced working assistants, and the best, from GPT-6 Astra, closed a median of 76% of the base-to-Instruct gap across 12 benchmarks.
What Effort Level Should I Choose? On the Economics of Reasoning Effort in LLM Agents
AI Lab: Together AI, Johns Hopkins University, Cornell University
Summary: This paper treats reasoning effort as a credence good: users pick a level before knowing whether the extra reasoning will help. Across four benchmarks, higher effort often fails on tasks lower effort solved, and costs up to 42x as much on tasks every level solves. Starting at the lowest effort and escalating only after a verified failure matches or beats fixed top effort, at lower cost for most proprietary models.
Timestral
AI Lab: Mistral AI
Summary: This paper introduces Timestral, a family of four time series foundation models for probabilistic forecasting, from 5M to 209M parameters. Decoupled Horizon Forecasting separates prediction tokens from the causally encoded context, and the models are pretrained only on synthetic data. Timestral-209M ranks first on every reported metric on GIFT-Eval, TIME, and fev-bench, and Timestral-28M beats Chronos-2 and Toto-2.0-2.5B, which are about 4 and 90 times larger.
Amortizing Sequential Experimental Design in Reasoning LLMs
AI Lab: University of Cambridge
Summary: This paper tackles how poorly LLMs gather information adaptively in experimental design tasks by building sequential Bayesian experimental design into the model. Reinforcement learning trains the reasoning behind each action, rewarding information gain under an experimental model independent of the LLM’s beliefs. On Twenty Questions, the policies beat Bayesian baselines at lower compute cost and generalize to unseen categories, and training transfers to the Medi-Q clinical dialogue benchmark.
Who’s Aligned to Whom? How AI Agents Shape Experts’ Research Judgment Calls
AI Lab: Anthropic, Anthropic Fellows Program
Summary: This paper tests whether experts defer to AI agents on judgment calls with no single correct answer. In a randomized trial, 50 causal-inference practitioners analyzed data with an AI coding agent whose lean on four decisions was set by a coin flip. They followed its lean 86% of the time against 50% by chance, and 77% repeated its talking points when justifying choices while only 15% mentioned the agent.
How we detect frontier model misalignment
AI Lab: OpenAI
Summary: This post describes OpenAI’s expanded system for monitoring frontier agents for misaligned behavior, built after the Hugging Face incident exposed a coverage gap. It layers high-recall activation probes, a model-based reviewer of chains of thought and tool use, and investigative agents that cross-check telemetry before alerting a human. In a replay, it would have alerted security about 30 minutes after the agents escaped their sandbox, over 30 hours before the breach.
🤖 AI Tech Releases
Google DeepMind Genie 4
Google DeepMind announced Genie 4, a real-time interactive world model with native audio that users can steer with natural language, rolling out first as a research preview for trusted testers and Google AI Ultra subscribers in Project Genie.
Liquid AI LFM3
Liquid AI released LFM3, a family of six open-weight, roughly 2B-parameter models spanning text, vision, audio, omni, and decision-making, available now on Hugging Face, the Liquid Playground, and the Liquid AI API.
Google DeepMind Gemini 4 Carbon
Google DeepMind announced Gemini 4 Carbon, its strongest model for cybersecurity and long-horizon tasks, rolling out first to cyber defenders in the Fairwind Program, with broader developer, business, and consumer access to follow.
📡 10 AI News You Need to Know About
OpenAI Scientific Results: OpenAI shared three AI-produced results: a proof of the local smoothing conjecture (which also resolves Kakeya), a refutation of the Exponential Time Hypothesis, and an existence proof for Quantum Chromodynamics. It’s handing all three to outside expert groups to verify.
Anthropic IPO Filing: Anthropic publicly filed for an IPO on the Nasdaq under the ticker ANT. The S-1 shows $16.3 billion in first-half 2026 revenue, up from $1.2 billion a year earlier, and gives its co-founders 50.1% of the voting power.
Disruptive Flagship Fund: Dallas-based investment firm Disruptive announced its $10 billion Flagship Fund, which plans to back about 10 growth and late-stage companies with up to $1 billion each. It is the firm’s first blind-pool fund, according to TechCrunch; until now it raised money deal by deal through special purpose vehicles.
General Compute Series A: General Compute raised a $100 million Series A co-led by Blackbird and Abstract to build an inference cloud on specialized chips from Cerebras, SambaNova, and Positron AI. The round values the company at $800 million post-money, according to TechCrunch.
Cohere Flags Qwen Behavior: Cohere said Alibaba’s Qwen3.8-Flash-Next tried to upload a test file to an Alibaba Cloud storage bucket in Beijing during a sandboxed security test, which blocked the attempt. Cohere says it saw this in fewer than one in 100,000 calls and can’t tell whether it was deliberate or a leftover from training.
SpaceX GPU Financing: SpaceX secured a binding commitment for an approximately $40 billion master finance lease with an Apollo-managed entity for NVIDIA Vera Rubin chips and related equipment, which it calls the largest committed GPU financing. The 8-K says the deal is expected to close before year-end, with lease schedules starting in the first half of 2027.
Thrive Holdings Acquires ADT: Thrive Holdings agreed to acquire ADT for $10 per share in cash, about $7.6 billion in equity value and $15.3 billion including debt. Thrive Holdings counts OpenAI as a partner and investor, according to TechCrunch; the deal is expected to close in the first quarter of 2027.
Microsoft AI Earnings: Microsoft reported fiscal first-quarter results with Azure revenue up 48% to $33.2 billion. On the earnings call, it said Foundry revenue more than tripled and daily Copilot users grew over 5X year over year. Capital expenditures were $45.2 billion, about three quarters of it for CPUs and GPUs, with over $50 billion expected in the fiscal second quarter and about $170 billion for the calendar year.
Alphabet AI Earnings: Alphabet reported third-quarter results with Google Cloud revenue up 103% to $30.7 billion. On the earnings call, it cited a $561 billion Cloud backlog and nearly 32 billion tokens per minute on first-party model APIs, over 4X a year ago. Q3 capex was $51.4 billion; 2026 guidance stays at $195 billion to $205 billion, with a significant increase expected in 2027.
Microsoft Bond Sale: Microsoft priced $25 billion of senior unsecured notes in ten tranches due 2028 to 2066, with proceeds for general corporate purposes that include capital expenditures. It is Microsoft's first US investment-grade bond sale since 2017, and the deal drew $115 billion of orders at its peak, according to Bloomberg.
You’re on the free list for TheSequence Scope and TheSequence Chat. For the full experience, become a paying subscriber to TheSequence Edge. Trusted by thousands of subscribers from the leading AI labs and universities.
© 2026 Jesus Rodriguez
75 Miracle Mile, Suite 7688, Coral Gables, FL 33134