The AI build-out’s cash and credit was the day’s biggest thread on 31 October 2026, by our grouping of the news. Amazon, Alphabet, Microsoft, Meta and Oracle spent about $31.8 billion more on property and equipment in their latest quarter than their operations brought in, by our arithmetic from their 10-Q filings, against about $61.4 billion left over a year earlier; the commitments eight companies keep off their balance sheets reached about $3.94 trillion, also by our arithmetic.
The threads below are cash and credit; OpenAI’s round; memory; compute at scale; agents and pricing; models and research; and power, policy and pushback. Each item links to the full piece, which names its sources.
AI Daily Briefing · video
The day in 84 seconds: six stories, read from the published pieces.
Business Pill · FINANCIAL CONDITIONS
An evergreen Business Engineer explainer on financial conditions, how rates and market mood change the price of capital. It explains the general idea and does not discuss the companies in this roundup.
The key insight: As we read it, the day’s coverage put the cost of the AI build next to the price of what it sells: by our arithmetic from their 10-Qs the five biggest AI builders spent about $31.8 billion more than their operations brought in, while on OpenRouter Anthropic drew about 2.1 times OpenAI’s spend per token, also by our arithmetic.
The five biggest AI builders spent about $31.8 billion more on property and equipment in their latest quarter than their operations brought in, by our arithmetic from their 10-Qs; a year earlier the same five had about $61.4 billion left over, so the swing is about $93.2 billion. A Goldman Sachs chart shared on X puts the five at minus $31 billion for the third quarter. Read: the free-cash-flow swing.
Eight companies now list about $3.94 trillion of commitments that are not on their balance sheets, up from $3.56 trillion nine days earlier, by our arithmetic; Alphabet’s lines reached $1.06 trillion at 30 September, the first company in our tracker above $1 trillion. Read: the off-book tracker, Alphabet’s support pledges and Meta’s data-centre ventures.
Microsoft sold $25 billion of notes rated AAA by S&P, according to its pricing term sheet, and the Financial Times reported that Blackstone told investors Anthropic is likely to be rated in the double-B range after its IPO; on S&P’s scale that range sits 10 to 12 notches below AAA, by our arithmetic. Read: the rating gap.
Oracle’s Investor Day release says it “committed to limiting additional funding to be raised in the capital markets for its FY27-FY30 investment cycle to $20 billion”, and the Wall Street Journal reported that Arini Capital Management sold about $800 million of debt holdings after margin calls. Read: Oracle’s cap and Arini.
The filings also showed Meta’s construction in progress at $89,499 million on 30 September, up from $50,521 million at the end of 2025; Amazon carrying its two Anthropic holdings at about $238.7 billion, by our arithmetic from its 10-Q; and Intel recording a $12.6 billion loss on shares held for Washington. Read: Meta’s construction, Amazon’s Anthropic stake and Intel’s loss.
Digital Realty’s backlog of signed leases due to start paying rent in 2027 rose to $831 million at the end of September from $480 million at the end of June, according to its investor presentations. Read: Digital Realty.
A mainstream PC’s parts bill rose from $435 to $699 between the first quarter of 2025 and the third quarter of 2026, and memory and storage went from about $70 to about $294 of it, according to Omdia’s figures as our piece reports them; the split of the remainder is our arithmetic.
OpenAI’s Round
OpenAI “is now seeking about $20bn at a $1.25tn valuation as it tempers its ambitions”, the Financial Times reported, and the Wall Street Journal reported that it raised less than half of its $30 billion target at a $1.25 trillion value, “less than its $1.4 trillion target”; by our arithmetic that is $150 billion, or about 11%, below the target. Read: the FT report, the WSJ report and the infrastructure marks.
Bloomberg reported that Baseten is set to raise at about $30 billion, 2.3 times its June valuation, and that Fluidstack is being pitched to investors at $40 billion, 5.3 times its January mark by our arithmetic. Read: Baseten.
OpenAI says that “Starting in December, Plus, Pro, and Business plans’ usage allowance will be shared across ChatGPT and Codex”, counted in tokens, with Instant chat not drawing from it. Read: the shared allowance.
Memory
Memory and storage made up 42% of the parts cost of a mainstream PC in the third quarter of 2026, up from 16% in the first quarter of 2025, according to Omdia; they account for about $224 of the $264 rise in the parts bill, or about 85%, by our arithmetic. Read: the PC parts bill.
Samsung’s chip division earned about 72% operating margins, by our arithmetic, in the quarter that Counterpoint counted 21% fewer smartphone chips shipped, a fall Counterpoint said was driven by rising memory costs, longer replacement cycles and cautious inventorymanagement; Bloomberg reported that Sam Altman met Samsung Group’s top executives in South Korea on Saturday. Read: Samsung’s margin, Counterpoint’s count and Altman in Seoul.
SK hynix told analysts that some customers are moving to 8-high instead of 12-high HBM stacks because of a supply shortage, and that requested volumes exceed more than twice what it can supply, according to Korean outlets. Read: SK hynix.
UBS estimates that an “unconstrained/aspirational” agentic CPU build-out by 2030 would need 282 exabytes of DRAM, which it puts at 3.5 times current industry estimates of 77 exabytes, and TechInsights expects memory to take about 62% of a chip market above $2.6 trillion in 2027, The Elec reported. Read: the UBS estimate and TechInsights.
Compute at Scale
Elon Musk wrote on X that “There will be a total of around 1.2M Blackwells at Colossus 2 by end of year”; the four buildings he named are to hold about 860,000 GB300s between them, by our arithmetic. In another post he wrote that orbital compute deployed per month “probably exceeds terrestrial around the end of 2028”. Read: Colossus 2 and orbital compute.
Starlink satellites make up 11,184 of the 16,848 active satellites in CelesTrak’s public list, or 66.4%, by our count, and two DayOne sites in Johor rank sixth and eighth on Epoch AI’s list of AI data centres by compute. Read: Starlink and Malaysia’s data centres.
Serving Kimi K3 on Nvidia’s B200 cost $0.179 per million tokens against $0.077 on AMD’s MI355X at 100 tokens a second per user, according to SemiAnalysis’s InferenceX benchmark, about 2.3 times as much by our arithmetic. Read: B200 and MI355X.
China’s demand for AI computing power rose 417% in the first quarter of 2026 while supply grew 128%, according to the China Academy of Information and Communications Technology as reported by Caixin on 31 August; Kuaishou said the fundraising for its Kling AI unit has been completed. Read: China’s compute gap and Kling AI.
Agents and Pricing
Peter Walker, Head of Insights at OpenRouter, posted its chart of Anthropic and OpenAI shares on 31 October: in the most recent week Anthropic had 48% of the two labs’ tokens and 66% of their spend, which works out to about 2.1 times OpenAI’s spend per token, by our arithmetic. Read: OpenRouter’s chart.
OpenAI lists gpt-6-rosalind at $15.00 per million input tokens and $75.00 per million output tokens, against $10.00 and $50.00 for GPT-6 Astra, which uses the same weights; researchers from UC Berkeley, Microsoft Research and Microsoft Security released TokenEconomizer, which their paper says cut coding agents’ average token cost by 18.4% to 30.9%. Read: Rosalind-6 and TokenEconomizer.
Agentcard launched opencommerce, one endpoint that lets an agent buy from 5,026 online stores, with its standard plan charging 1% of each completed order; Google released Gemini API Computer Use to general availability; and Android Studio now lets developers pick Claude Agent, Codex or other agents. Read: opencommerce, Computer Use and Android Studio.
Microsoft made the software that runs its own AI agents available to developers, with billing from 1 December, and Reducto released a free plugin that brings Word, PowerPoint and Excel editing to ChatGPT and Codex. Read: Microsoft’s harness and Reducto Office.
A single stolen cloud key made approximately 1.7 million calls to Amazon Bedrock in thirteen hours, spread across a dozen Claude models, according to a report Wiz Research published on 30 October, and Walmart committed to an initial additional spend of about $6 billion on Symbotic’s warehouse automation, according to Symbotic’s 8-K. Read: the stolen key and Walmart and Symbotic.
Models and Research
A paper from researchers at FAIR at Meta reports that splitting an agent’s memory into routed clusters cut estimated API cost by 65–72%, and a Google DeepMind paper reports that training on simulated psychiatric patients cut the share of acute risks a model under-triaged. Read: Meta’s memory paper and DeepMind’s intake model.
Rock Matrix released EdgeRock-4B, an open-weight model whose cache stays at 72 MiB per request whether the context is 32K or 256K tokens, according to the model card, and developer Divyansh Agrawal compiled a vector unit into the weights of Alibaba’s open Qwen3.5-4B model to run Karpathy’s stories260K inside the model’s own forward passes. Read: EdgeRock-4B and the LLM inside Qwen.
Anthropic’s Alignment Science Blog says it instructed a Claude model to write training data that would make a small research model misaligned while passing a monitor; Apple said the first Apple Intelligence features are coming to users in China. Read: Anthropic’s monitor test and Apple in China.
Power, Policy and Pushback
A federal judge in Mississippi on 30 October denied a motion by X.AI Corp. and MZX Tech LLC to dismiss an NAACP lawsuit over the gas-turbine plant behind xAI’s Colossus 2; the order lists at least 27 turbines with a combined 495 MW, by our arithmetic from its figures. Read: the NAACP suit.
Campaigns and outside groups have spent $173 million on almost entirely negative ads about data centers this election, according to AdImpact data reported by Axios; 61% of US adults now oppose new data centres in their area, the Annenberg Public Policy Center found, while GE Vernova’s orders rose 96% organically. Read: the ad war, power orders and opposition and the utility races.
Michael Kratsios, director of the White House Office of Science and Technology Policy, told Colossus magazine that Anthropic could “Just announce tomorrow that you’re spending a billion dollars on safety”, and SpaceXSI will watermark text from seven Grok models on its API from 7 November, according to its documentation. Read: Kratsios and the watermark.
Three price multiples for comparable AI work, as our pieces report them (each our arithmetic): Nvidia B200 against AMD MI355X per Kimi K3 token on SemiAnalysis’s InferenceX benchmark, Anthropic against OpenAI spend per token on OpenRouter’s chart, and GPT-Rosalind-6 against GPT-6 Astra list prices on the same weights.
The Structural Read
By our arithmetic from company filings, eight companies list about $3.94 trillion of commitments that are not on their balance sheets, and Alphabet’s lines reached $1.06 trillion at 30 September.
Omdia’s figures put memory and storage at 42% of a mainstream PC’s parts cost in the third quarter of 2026, up from 16% in the first quarter of 2025, and SK hynix told analysts that requested HBM volumes exceed more than twice what it can supply, according to Korean outlets.
As we read it, the pricing coverage sat beside the cost coverage: SemiAnalysis’s benchmark put B200 at about 2.3 times MI355X’s cost per Kimi K3 token by our arithmetic, and OpenAI lists gpt-6-rosalind at 1.5 times GPT-6 Astra’s price on the same weights.
Elon Musk, on X, 31 October 2026, as quoted in our piece
“There will be a total of around 1.2M Blackwells at Colossus 2 by end of year”
Three Implications
THE BIGGEST AI BUILDERS By our arithmetic from their 10-Qs, Amazon, Alphabet, Microsoft, Meta and Oracle spent about $31.8 billion more on property and equipment than their operations brought in during their latest quarter, against about $61.4 billion left over a year earlier.
LENDERS AND RATING AGENCIES Microsoft sold $25 billion of notes rated AAA by S&P, according to its pricing term sheet, and the Financial Times reported that Blackstone told investors Anthropic is likely to be rated in the double-B range after its IPO.
BUYERS OF MEMORY AND TOKENS Omdia’s figures put memory and storage at 42% of a mainstream PC’s parts cost, and OpenAI says ChatGPT and Codex will share one usage allowance on Plus, Pro and Business plans starting in December.
The analysis says: “The $1T capex cycle is being underwritten by a $1T concentrated demand commitment from two counterparties whose combined free cash flow is deeply negative.”
As we read it, the day’s coverage shows the cash side of that cycle, in the builders’ spending above their operating cash, the commitments kept off balance sheets and a smaller OpenAI round. We do not say one caused the other. The analysis was published in May 2026 and does not discuss these events.
What Is Not Established
This roundup adds no new reporting and selects from the day’s coverage. Every figure above is taken from the linked FourWeekMBA piece, each built from its own primary source. The seven threads are our grouping, not a finding.
We do not claim that these stories are connected or that one caused another. The OpenAI round figures rest on reports by the FT and the WSJ, the Blackstone rating view on summaries of an FT report, the Baseten and Fluidstack marks on Bloomberg, and the Colossus 2 counts on Elon Musk’s posts; the off-book total and the free-cash-flow figures are our arithmetic from company filings.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
By our arithmetic from their 10-Qs, the five biggest AI builders spent about $31.8 billion more than their operations brought in during their latest quarter, against about $61.4 billion left over a year earlier. The same day, the FT and the WSJ reported a smaller OpenAI round at a $1.25 trillion value, and Omdia’s figures put memory and storage at 42% of a mainstream PC’s parts cost.
This week’s AI Supercycle issue on The Business Engineer reads the build-out as a credit story: read the issue.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. This roundup draws on FourWeekMBA’s news coverage of 30 and 31 October 2026, linked above; every story was built from its own primary source, which it cites and links. We re-read each piece on 31 October 2026 and took every figure here from it. The seven threads and the figures in our charts are our own grouping and selection. Nothing here is a forecast, and nothing here is financial or investment advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
