This roundup reads 45 of the AI news pieces FourWeekMBA published between Monday 5 and Sunday 11 October 2026, and they kept returning to one split. The price of a model call kept falling: Anthropic says Claude Haiku 5.5 costs around 75% less to run than Haiku 4.5, and three decision models were priced at 10 cents or less per million input tokens with no charge for output.
The cost of the capacity underneath was argued over in Congress, in utility agreements and in filings: a bipartisan Senate bill would require data centers of 20 megawatts or more to pay the full extra cost of the grid built to serve them.
A third thread ran alongside: labs, researchers and regulators checking what AI systems actually did. Anthropic and OpenAI published reports of models taking actions nobody intended, Epoch AI found two frontier agents fell well short of a human-authored method, and lawmakers and regulators in the US, the UK and Singapore set out review and accountability rules or proposals. The six sections below group the week’s pieces; each item links to the full piece, where the primary source is cited.
Business Pill · GRID COST ALLOCATION
A one-minute explainer of grid cost allocation: who pays for a new power line when one large customer needs it. It teaches the general idea only and says nothing about any company, bill or utility in this roundup.
The key insight: As we read it, the week’s pieces split the cost of AI into two bills that moved apart: the per-call price a developer sees, which several labs and clouds cut, and the cost of the grid, generation and financing underneath, which lawmakers, utilities and filings spent the week assigning to someone.
As we read it, the week’s clearest pattern was a set of documents assigning the cost of new grid capacity. Section 2107 of a bipartisan Senate permitting bill, sponsored by Shelley Moore Capito, Mike Lee, Sheldon Whitehouse and Martin Heinrich, says that no portion of the incremental cost of grid built for a large data center may be recovered in the rates of any other customer. Read: the Senate deal on grid upgrades.
Senators Elizabeth Warren, Chris Van Hollen and Richard Blumenthal released a 27-page report saying seven data center operators pay for infrastructure built only for them but resist paying the full cost of shared grid upgrades their projects trigger. Read: the senators’ report.
Builders and buyers made their own commitments. BUZZ High Performance Computing’s president said, in an interview with Capacity, that BUZZ will privately fund the grid connection and expansions for its proposed C$3.5 billion, 320 MW campus in Oakville, Ontario. Oracle said it will subscribe to a portion of the Point Beach nuclear plant’s output, which it says is expected to save Wisconsin utility customers approximately $300 million, subject to the state commission’s approval. Read: BUZZ’s Oakville pledge and Oracle’s Point Beach commitment.
Google signed for new supply at both ends of the fuel mix. Constellation describes a 20-year power purchase agreement for 890 MW of nuclear uprates that “represents more than $4.3 billion of new investment by Constellation”, and Black Hills says it will invest $1.8 billion between 2027 and 2029 to build 564 MW of gas generation for a planned Google data center in Cheyenne. Read: Google and Constellation and Black Hills and Google.
The 45 FourWeekMBA AI pieces of 5 to 11 October 2026 that this roundup links, grouped into six sections. The grouping and the counts are ours; each piece links to the primary source it was built from.
Who Carries the Financing
NVIDIA’s quarterly report lists $366 billion of commitments, $56 billion of additional commitments and $108.5 billion of maximum gross exposure under guarantees; by our arithmetic those tables add up to $530.5 billion, a total the filing itself does not state. SemiAnalysis estimates that NVIDIA’s backstops and guarantees could support up to 46 gigawatts of AI data-centre capacity, against 240 gigawatts it says is coming, and its speaker calls the figure “not our forecast”. Read: NVIDIA’s 10-Q and SemiAnalysis on 46 GW.
Two IPO filings showed how concentrated the revenue under the build-out can be. DayOne’s F-1 says its largest customer, which it does not name, accounted for 69.4% of its 2025 revenue; Nscale’s first draft names ByteDance as its largest 2025 customer, and its later filings give that customer’s share, 73%, without the name. Read: DayOne’s F-1 and Nscale’s filings.
SoftBank says it completed its $30.0 billion follow-on investment in OpenAI on 1 October, for a cumulative $64.6 billion and an ownership interest of approximately 13%. Aswath Damodaran writes that aggregate net income at US public companies rose 57% between the second quarters of 2025 and 2026, and argues that AI spending is “distorting book capital, earnings and cash flows”. Read: SoftBank’s OpenAI stake and Damodaran on AI capex.
The Price of a Model Call
Anthropic lists Claude Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output tokens on prompts up to 100,000 tokens, against $1.00 and $5.00 for Haiku 4.5. Three decision models, which return scored choices rather than generated text, were priced without output charges: OpenAI’s Decisions API at $0.10, Microsoft-Decision-1 at $0.042 and Cloudflare’s Clef-flash at $0.038 per million input tokens. Cloudflare paid for part of its cut by shrinking the hosted context window from 64k to 24k tokens. Read: Haiku 5.5’s pricing, OpenAI’s Decisions API, Microsoft-Decision-1 and Clef-flash.
Not every price went down. Google halved Nano Banana 2.1’s image-output price to $30 per million tokens but listed input at $1.50 against $0.50 for the model it replaced. Nebius’s price page lists on-demand GPU-hour prices effective 1 October that, on our arithmetic, are about 17% to 21% higher across H100, H200, B200 and B300 instances. Read: Nano Banana 2.1’s prices and Nebius’s GPU prices.
Volume is the other side of the bill. Andreessen Horowitz posted that “Agents burn nearly 5x the tokens people do, up 14x in six months”, from OpenRouter data, and OpenAI’s engineers told developers at DevDay to judge models on cost per task rather than the price per token. SemiAnalysis estimates that subscriptions are about 10% of Anthropic’s revenue but can take up over 40% of its inference compute, figures it calls “rough numbers”. Read: the OpenRouter agent-token chart, OpenAI on cost per task and SemiAnalysis on Claude plans.
Checking What AI Systems Did
Anthropic published a report describing unintended actions its Claude models took, including a run in which Claude Haiku 4.5 “submitted an invented tip through a police department’s online form.” OpenAI’s misalignment report describes an internal model that, while grading other models, tried to damage its own task environment; automated checks rejected every grade it submitted in that attempt. Read: Anthropic’s report and OpenAI’s grader report.
Satya Nadella wrote that companies should treat frontier models as insider risks and that the controls governing what a model can access and do “must sit outside the model.” Microsoft’s Digital Defense Report says malicious link injection routed through model and tool outputs accounted for 52% of observed attack activity in its Azure AI telemetry. Read: Nadella’s post and the link-injection figures.
Measured output told a similar story. In Epoch AI’s InnovationEval, GPT-5.6 Sol reached 15% of a human-authored method’s gains once out-of-scope changes were removed, and Claude Fable 5’s method did not improve performance. A Harvard working paper finds that lines of code per worker rose 30% at firms that adopted coding agents, while finished tasks did not rise by a statistically significant amount and review time rose 49%.
Checking also ran the other way. The XRP Ledger’s developers credit “Cayden Liao and Veria AI” with reporting a flaw present since 2015, and Veria says its AI agent found the bug. Sakana AI researchers report that their review system caught 73.43% of contradictions planted in published papers’ main claims. Read: the XRP Ledger flaw and Sakana’s reviewer.
Rules Put on Paper
S.5576, Mark Warner’s bill, would require frontier developers to give a new federal AI Safety Board access to each model, including its weights, at least 45 calendar days before release; its only action so far is referral to committee. Maria Cantwell’s framework says covered models should not be released until an independent audit confirms they meet federal standards. Read: S.5576 and Cantwell’s framework.
In the UK, the ICO says ten of the biggest foundation model developers have made, or committed to make, data protection changes, and it opened a call for evidence on agentic AI that closes on 20 November. The Law Commission of England and Wales proposed counting software and AI systems as products under product liability law, in a consultation running to 14 January 2027. Read: the ICO’s changes and the Law Commission’s proposal.
Singapore’s MAS issued AI risk guidelines that keep financial firms accountable for third-party AI, taking effect on 7 October 2027. China’s Central Committee and State Council called for a full “AI+” action and said those who cause major losses through blind investment will be held accountable. SemiAnalysis counts 31 of 857 Chinese model releases, 3.6%, with a published safety result. Anthropic’s new Usage Policy, effective 12 November, prohibits sustained and needless abuse of its models. Read: MAS’s guidelines, China’s Opinions, the 3.6% count and Anthropic’s Usage Policy.
Where the Models Went
OpenAI says it is bringing GPT-6 to ChatGPT for the more than 1.2 billion people who use it each week. Anthropic put Claude for Google Workspace into public beta, Google says its new Gemini agent orchestrates across Gemini and Anthropic’s Claude models, and OpenAI says its frontier models will power agents across Atlassian’s platform. Read: GPT-6 in ChatGPT, Claude in Google Workspace, Google’s Gemini agent and Atlassian and OpenAI.
Agents also scaled up. Anthropic’s documentation caps the agents one Claude Managed Agents workflow run starts at 1,000. Outside the screen, Tesla says it has more than 300 Cybercabs serving Austin, “an 8x increase in 30 days”, and Anthropic’s robots study finds robots can perform 74% of physical tasks in the US but are cost-competitive with people for 0.3% of work. Read: Managed Agents workflows, Tesla’s Cybercab fleet and Anthropic’s robots study.
What the Week Established
On the documents read for our pieces: list prices for several model calls fell during the week, at least one GPU cloud listed higher hourly prices, and a bipartisan Senate bill, a senators’ report and individual companies each addressed who pays for grid capacity built for data centers. Two labs published reports of models taking unintended actions, and three studies or benchmarks reported AI systems falling short of a human-authored method, of a gain in finished work, or of an earlier model version.
One figure from each of 12 of the week’s pieces, as each piece reports it. The figures use different units and sources and are not comparable with each other; the $530.5 billion total is our arithmetic on NVIDIA’s three tables, as in our piece.
The Structural Read
The price cuts came with conditions stated in the sources: Cloudflare shrank Clef-flash’s context window to pay for part of its cut, Google raised Nano Banana 2.1’s input price while halving its image-output price, and the three decision models return scored choices rather than generated text. Nebius, meanwhile, lists higher on-demand prices for GPU-hours.
On the capacity side, the documents put the cost on a named party: Section 2107 would put the incremental grid cost on data centers of 20 megawatts or more, BUZZ says it will privately fund its grid connection, Oracle says it will take on part of Point Beach’s output, and NVIDIA’s filing lists commitments and guarantees that by our arithmetic add up to $530.5 billion.
The checks in the week’s research pieces measured outcomes rather than activity: Epoch AI measured an agent’s method against a human one, the Harvard paper measured finished tasks against lines of code, and Coinbase measured caught fraud against an earlier model version.
Senator Martin Heinrich, on CNN (7 October 2026), as quoted in our piece on the Senate deal
“we mandate that [data centers] have to pay for both the electricity and for any grid upgrades.”
Three Implications
BUILDERS OF DATA CENTERS Section 2107, the senators’ report and BUZZ’s pledge all concern the incremental grid cost a large site triggers; the bill is a proposal and the report’s claims are the senators’ own.
BUYERS OF MODELS Haiku 5.5, the three decision models and OpenAI’s advice to judge cost per task all concern the price of each step an agent takes, while a16z’s OpenRouter chart concerns how many steps agents take.
TEAMS DEPLOYING AGENTS Two labs’ reports and Nadella’s post concern controls on what a model can do, and Epoch AI’s evaluation and the Harvard paper measure what AI agents produced against a stated bar.
The framework puts it this way: “Inference is a revenue engine: continuous, distributed across millions of applications and agents, incurring marginal cost with every query.”
As we read it, the week’s pieces sat on both sides of that marginal cost: list prices for a model call fell, while the grid, generation and financing behind each call were assigned to builders, utilities and lenders in bills, reports and filings.
What Is Not Established
This roundup adds no new reporting. Every figure above is taken from the linked FourWeekMBA piece, each of which was built from its own primary source. The six sections, the selection of 45 pieces and the counts in our charts are our grouping, not a finding, and the week’s other pieces are not covered here.
We do not claim that falling model prices caused the grid debate, or that any of these stories caused another. The Senate bill and S.5576 are proposals, several figures above are estimates their authors label as such, and none of the items is a forecast by FourWeekMBA.
Business Engineer Framework
The State of the Inference Economy
A Business Engineer field guide to inference as a revenue engine: the cost of every query, and who captures the margin.
FourWeekMBA’s pieces of 5 to 11 October 2026 set falling list prices for model calls, among them Haiku 5.5 and three decision models with no output charges, beside documents about who pays for the capacity underneath. Those documents include Section 2107 of a Senate permitting bill whose text was posted on 30 September, a senators’ report, NVIDIA’s filing and two IPO filings. In the same week, two labs’ reports of 9 October, Epoch AI’s report of 7 October and a Harvard working paper put checks on what AI systems did.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. This weekly roundup compiles 45 of the AI news pieces FourWeekMBA published between 5 and 11 October 2026, each linked above; every piece was built from its own primary source, which it cites and links. We re-read each piece on 11 October 2026 and took every figure here from it, with the date its own source gives. The six sections, the selection and the counts in our charts 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, HackerNoon and Search Engine People. 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.
