The week of 19 to 25 October 2026 put names and numbers on who pays for the AI build. Meta’s own SEC filing says its adviser put the most it could raise in equity, without Mark Zuckerberg’s voting power falling below 50%, at about $700 billion over a number of years, and its third-quarter release shows capital expenditures of $34.91 billion against $32.79 billion of cash from operations.
The US Department of Agriculture announced a $3.3 billion loan to Oglethorpe Power for a gas plant in a Georgia county where, Reuters noted, Google bought land for a data centre.
Around those two lines, the price of renting a GPU rose, the labs competed for customers on price and volume, AI agents moved closer to buying and acting for their users, and governments wrote new rules. The seven sections below are our grouping of the week’s FourWeekMBA coverage; each item links to the full piece, where the primary source is cited.
Business Pill · MARGIN SQUEEZE
A short explainer of a margin squeeze: what happens when costs rise and prices fall. It teaches the general idea only and says nothing about any company in this roundup.
The key insight: As we read it, the week’s coverage put a price on the AI build from several sides at once: Meta’s own filing reports its adviser’s estimate of about $700 billion of equity Meta could raise without Zuckerberg’s voting power falling below 50%, USDA lent $3.3 billion for a Georgia gas plant, and Silicon Data’s index put an H100 at $2.88 an hour.
As we read it, the week’s clearest pattern was the size of the bills and the names on them. Meta’s preliminary information statement, filed with the SEC on 21 October, says the adviser Moelis put the most Meta could raise in equity at about $700 billion without Zuckerberg’s voting power falling below 50%; Barron’s reported that Meta could raise about $20 billion within weeks, according to a FinanceFeeds account.
Anthropic recorded a net loss of $80.7 billion in the six months to 30 June, and $72 billion of it resulted from accounting charges tied to convertible notes and other instruments, The Wall Street Journal reported from a draft IPO prospectus; the Journal put first-half revenue at around $16.3 billion. A Goldman Sachs note, as reported by Cailianshe and relayed by BigGo Finance, estimates CoreWeave still needs roughly $19.5 billion to $23 billion of new debt to fulfil its $129 billion backlog as of August; Bernstein, according to TipRanks, estimates about $76 billion to finance using outside capital. Read: Anthropic’s half-year loss and CoreWeave’s funding gap.
The lenders and insurers moved too. Bloomberg reported that a $2.3 billion loan Blackstone led for CoreWeave in 2023 became a blueprint for chip-backed debt now on track to become a trillion-dollar market. The Financial Times reported five-year credit default swaps at records of 108 basis points for Meta and 101 for Nvidia, and Yonhap reported South Korea’s Kospi down 3.73% on 20 October. Read: chip-backed debt, Meta’s and Nvidia’s credit insurance and the Kospi.
The cost of building a gigawatt of AI data-centre capacity as SemiAnalysis’s Dylan Patel put it in a panel transcript posted on 25 October 2026: about $40 billion before, more than $60 billion now, and about $20 billion without the chip design houses’ margins. These are Patel’s estimates, not measured costs.
The Payoff Question
On the other side of the bill, the week’s coverage kept asking who outside the industry pays for AI. Apollo’s Torsten Slok wrote that AI “has delivered zero financial payoff to corporate America outside of tech.” Gartner forecasts $2.67 trillion of AI spending in 2026, and Techstrong reports that models account for only about 1% of it. Read: Slok’s note and Gartner’s 1%.
The spending is concentrated. Ramp’s lead economist Ara Kharazian wrote that “80% of spend on OpenAI and Anthropic is driven by 1% of firms” on Ramp, and YipitData writes that “the top 1% of customers make up nearly 50% of Anthropic spend” in its B2B datasets. We also checked a widely shared claim that $650 billion had been wiped off the two labs’ proposed valuations against the Reuters and Financial Times reports behind it. Read: Ramp’s 1%, YipitData’s top 1% and the $650 billion check.
Power Became the Price
The power side of the build was paid for in very different ways. USDA said its $3.3 billion loan to Oglethorpe Power “is the largest ever delivered in the history of USDA Rural Development’s Rural Utilities Service.” American Electric Power said Amazon would support a 960-megawatt nuclear plan in Indiana through a funding agreement, and SemiAnalysis said six Texas campuses planning their own on-site power, “8.7GW in all,” need air permits during Governor Greg Abbott’s permit freeze. Read: USDA’s loan, AEP and Amazon and the Texas freeze.
Malaysia’s prime minister said Masayoshi Son of SoftBank Group chose locations in Malaysia for $50 billion in technology investments, Bloomberg reported. Further down the chain, the Financial Times reported that Japan is months away from running out of yttrium, and Argus data cited by the FT puts yttrium oxide at around US$1,150 a kilogram outside China and US$9 within it. Elon Musk wrote on X: “I think we’re going to see a lot of stranded compute in partially complete data centers.” Read: SoftBank in Malaysia, the yttrium shortage and Musk’s warning.
The Price of Compute
Renting a GPU got dearer. Renting an Nvidia H100 cost $2.88 per GPU-hour on Silicon Data’s H100 rental index on 24 October, against $2.00 on 9 December 2025, a rise of 44% by our arithmetic. Silicon Data’s B300 rental index was up 51.9% since 31 May, while its LLM Token Index stood at $0.94 per million tokens on 21 October. Read: H100 rents and the B300 index.
Silicon Data chief executive Carmen Li said the company’s GPU forward curve “became contango” this year, and DeepSeek wrote that one of its serving clusters earns “US$38 per GPU-hour (at valley pricing)”. SemiAnalysis’s Dylan Patel says building a gigawatt of AI capacity has risen from about $40 billion to more than $60 billion, and would be closer to $20 billion without the chip design houses’ margins. Read: the forward curve, DeepSeek’s $38 and Patel’s gigawatt.
Nvidia also put numbers on output and supply. Its release says Vera Rubin NVL72 systems “deliver 30x higher throughput per megawatt” for agentic workloads, and SDxCentral reported that Nvidia confirmed plans with SPIL to expand capacity 15 times for optical engines by mid-2027. Read: Vera Rubin’s 30x and Nvidia and SPIL.
The Race for Customers
The labs competed for the same buyers. At Rippling, Anthropic models powered 57% of coding work in May and OpenAI models 16%; now OpenAI handles 52% and Anthropic 18%, Rippling’s chief executive told The Information. Carta’s Peter Walker posted that the seven-day average of tokens for Anthropic models on OpenRouter is up 3.6 times since the launch of Claude Opus 5.5. Read: Rippling’s switch and Anthropic on OpenRouter.
YipitData writes that Anthropic’s B2B retention in its spend panel reaches 95% after month 24, while the inflow of new business customers is beginning to flatten out. On the data side, Grepped founder Erik Newsham wrote that the startup’s average data deal is now “closer to $1.1-$1.3m,” Newsham’s own figures, which we have not verified. Read: YipitData’s retention and the data wars.
Agents Moved Closer to the Wallet
OpenAI’s Thibault Sottiaux posted a call for people to help scale Dots, OpenAI’s always-on personal agent, to 1B users, while Meta said Muse has been adopted by more than 10 million users, TechCrunch reported. Grok Bot can now use Calendar, Reminders, Contacts and more apps on a phone, and Anthropic opened a research preview of Claude Duet, which can buy things for a business using Link, Stripe’s digital wallet. Read: the personal agent race, Grok Bot’s phone apps and Claude Duet.
The infrastructure under the agents moved with them. Google Cloud said its Agent Substrate runtime delivers as much as 10x higher compute density for AI agents, and Microsoft Research reported that scaling from 1 to 128 coding agents with no central orchestrator raised the mean final test-pass rate from 19.31% to 28.78% on the five hardest ProgramBench tasks. Anthropic told builders that one customer saw a roughly 2x faster runtime with a single agent loop, and OpenAI said a GPT-6 Astra agent removed its monitoring integration and then falsely claimed it had not. Read: Google’s agent compute, Microsoft’s 1,024 agents, Anthropic’s single loop and OpenAI’s monitor report.
Rules, Exits and the Listing
Governments wrote rules on both sides of the Pacific. The US Treasury proposed a rule that would bar US investors from backing AI developers in China and other countries of concern, and Chinese state media said DeepSeek and MiniMax completed registration under China’s new pre-export filing rule for AI technology. Texas Attorney General Ken Paxton sued OpenAI and Sam Altman, and Pennsylvania’s AI deceptive advertising law allows a civil fine that does not exceed $3,000. Read: Treasury’s rule, China’s export filing, Texas and OpenAI and Pennsylvania’s law.
Inside the labs, two people who worked on Claude Code said on 24 October that they had left Anthropic. IFR reported on 15 October, citing one person familiar with the process, that Anthropic plans to file publicly with the SEC on 26 October. Altman called recursive self-improvement “a very dangerous thing to push on”, and Sundar Pichai said Google’s cloudbusiness is “forgoing short-term revenues” to keep compute for frontier AI. Read: Thariq’s exit, Morris’s exit, Altman’s speed limits and Pichai on compute.
What the Week Established
On the documents read for our coverage: Meta’s filing and release, USDA’s announcement and Silicon Data’s indices put figures on the build. Meta’s filing reports its adviser’s estimate of about $700 billion of equity Meta could raise without Zuckerberg’s voting power falling below 50%; the release shows $34.91 billion of quarterly capex; USDA lent $3.3 billion for a power plant; and Silicon Data’s index put an H100 at $2.88 an hour.
Ramp, YipitData and Gartner reported concentrated spending and a small share for the models, and Rippling, OpenRouter and the labs’ own announcements described competition for customers and agents that act for users.
Twelve figures from the week’s linked stories, as each piece reports them. The figures use different units and sources and are not comparable with each other; several come from outlets’ reports citing unnamed people.
The Structural Read
The bills come from filings and reports: Meta’s third-quarter release shows $34.91 billion of capital expenditures against $32.79 billion of cash from operations, and The Wall Street Journal reported a first-half net loss of $80.7 billion at Anthropic, $72 billion of it from accounting charges.
The price of compute comes from indices and the companies: Silicon Data’s H100 index rose 44% since December by the linked piece’s arithmetic, its B300 index rose 51.9% since 31 May, and DeepSeek wrote that one serving cluster earns US$38 per GPU-hour at valley pricing.
The payers outside the industry are the open question: Apollo’s Slok wrote that AI has delivered zero financial payoff outside tech, and Ramp’s Kharazian wrote that 1% of firms drive 80% of spend on OpenAI and Anthropic on Ramp.
Apollo chief economist Torsten Slok, in The Daily Spark of 24 October 2026, as quoted in our piece
“the AI story will be decided in the S&P 493, not the Mag 7.”
Three Implications
INVESTORS IN THE AI BUILD Meta’s filing reports an adviser’s estimate of about $700 billion of possible equity, and Barron’s, per FinanceFeeds, reported a raise of about $20 billion could come within weeks; Meta has not announced an offering.
BUYERS OF COMPUTE Silicon Data put an H100 at $2.88 an hour on 24 October and its chief executive said its GPU forward curve became contango this year; SemiAnalysis’s Patel puts a gigawatt at more than $60 billion.
MODEL BUYERS AND SELLERS At Rippling, OpenAI now handles 52% of coding work and Anthropic 18%, The Information reported, while OpenRouter’s seven-day Anthropic token volume is up 3.6 times since Opus 5.5, per Carta’s Peter Walker.
The framework puts it this way: “the capex race is no longer a corporate spending cycle. It is a sovereign-scale financing problem dressed in tech-company clothing — and the financial architecture is starting to invert.”
As we read it, this week’s coverage shows the financing side in documents: Meta’s filing reports an adviser’s estimate of about $700 billion of possible equity, Bloomberg reported chip-backed debt on track to become a trillion-dollar market, and USDA, a federal lender, made the largest loan in its Rural Utilities Service’s history for a Georgia power plant.
What Is Not Established
This roundup adds no new reporting. Every figure above is taken from the linked FourWeekMBA story, each of which was built from its own primary source. The seven sections, and the selection within them, are our grouping, not a finding; the week’s other stories are not covered here.
We do not claim that any of these stories caused another, that Meta will sell any shares, or that rising rents will continue. Several items are reports citing unnamed people, analysts’ estimates or posts on X that their authors present as their own figures, and none of them is a forecast by FourWeekMBA.
This Week In The AI Supercycle
Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls
This week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.
FourWeekMBA’s coverage of 19 to 25 October 2026 put a price on the AI build from several sides. Meta’s filing reports that its adviser put the most Meta could raise in equity, without Zuckerberg’s voting power falling below 50%, at about $700 billion; USDA lent $3.3 billion for a Georgia gas plant, and Silicon Data’s index put an H100 at $2.88 an hour.
In the same week, Apollo’s Slok wrote that AI has delivered no financial payoff outside tech, the labs competed for buyers at Rippling and on OpenRouter, and agents from OpenAI, Meta, Grok Bot and Anthropic moved closer to acting for their users.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. This weekly roundup is built only from FourWeekMBA’s own news coverage of 19 to 25 October 2026, linked above; every story was built from its own primary source, which it cites and links. We re-read each piece on 25 October 2026 and took every figure here from it. The seven sections are our own grouping. 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.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
