# She Was Stuck at Newark While a Million Trips Fell Apart

*A routine upgrade to the wires that carry air traffic messages overloaded the old system on Friday, and more than a million passengers paid for the failure with their weekend.*

By Jay Cadmus | Published November 15, 2026
Canonical: https://theainextwave.com/editions/2026-11-15/

## The Short Version

Louise Maher, 47, spent Friday night stuck at Newark Liberty International Airport, waiting on a British Airways flight to London that the meltdown had swallowed. She told Reuters she might cry at some point, but not yet. She had company. A nationwide ground stop ordered Friday afternoon held nearly every departing flight in the United States for about five hours, and the count kept growing after the stop lifted: 9,200 flights canceled on Friday and another 6,200 on Saturday, according to FlightAware figures reported by CNN, with more than a million passengers caught in the disruption. The trigger, on the account officials gave this weekend, was maintenance work meant to make the system newer. A contractor replacing copper wire with fiber optics installed a router that flooded the network, and the old equipment buckled.

The rest of this edition follows the same shape at different scales. A London committee heard this week that the data centres planned around the capital could add an average of £88 to a British household energy bill and push about 10,000 London households into fuel poverty, figures a campaign group presented as projections and a United States study published this month supports with observed bills. Workers who use AI heavily are logging workdays up to 90 minutes longer than colleagues who avoid the tools, in an October analysis of more than 9,000 employees that measures association and names its own limits. And the Federal Reserve reported Friday that 87 percent of the market contacts it surveyed now name AI as a salient risk to the financial system, alongside a tally of more than $680 billion in debt issued for advanced computing by October. Each story ends at a household: a traveler on a terminal floor, a family choosing between heat and other bills, an employee whose day grew longer, a saver whose retirement account rides on the AI trade heading into earnings week.

## She was stuck at Newark while a million trips fell apart

### A five hour ground stop, a million disrupted trips, and a router at the center of the investigation

The trouble began at 3:36 p.m. Eastern on Friday, November 13, when air traffic control centers started seeing instability in radar and communications systems, Transportation Secretary Sean Duffy said. Switches and routers were overloaded with data and began returning inconsistencies, and the Federal Aviation Administration ordered a nationwide ground stop at about 4 p.m. Nearly all departing flights sat on the ground for almost five hours, the longest nationwide ground stop in 25 years, until the FAA lifted the order shortly before 8:30 p.m. Airlines canceled nearly 10,000 flights on Friday and delayed more than 7,000 others, Reuters reported, in one of the worst travel days in United States history. The only other nationwide ground stop since the September 11 attacks came in January 2023 and lasted about 90 minutes.

Saturday brought the hangover. Major airlines canceled more than 3,700 flights and delayed 2,000 more by the time Reuters filed its Saturday report, as carriers tried to move aircraft and crews back into position. United Airlines ordered a ground stop for its own flights at Chicago O'Hare and Delta Air Lines did the same at New York LaGuardia, both to handle congestion and gate pressure. CNN, citing FlightAware, put the full Saturday cancellation count at 6,200, on top of Friday's 9,200, and counted more than one million passengers whose travel plans were upended across the two days. Duffy said Saturday that the system was back at full operations and that the worst was behind the country, with a fair amount of recovery still ahead. American Airlines said late Friday it was prioritizing international flights and smaller cities outside its hubs to position planes and crews for a faster recovery.

One detail in the Reuters account lands hard on the passengers sleeping in those terminals. Because the FAA caused the delays, airlines are under no requirement to compensate stranded passengers. That rule covers compensation for costs like hotels and meals. It does not cancel the fare itself: a passenger whose flight was canceled, or significantly changed, and who chooses not to travel is entitled to a refund of the ticket price under the Transportation Department's refund rule. A family that lost a hotel night, a rental car day, or a shift at work on Saturday carries that cost itself, and the bill arrived two weeks before Thanksgiving, one of the busiest travel periods of the year.

![Alt text: A new router with a yellow fiber cable beside a coil of old copper wire and a paper flight strip.](https://theainextwave.com/assets/inline-2026-11-15-1.jpg)
*The upgrade was the failure. The old network could not take what the new part sent it.*

### The fix that broke the system was an upgrade

The investigation now centers on the maintenance itself. Duffy said on Saturday that L3Harris, the contractor that manages the FAA Telecommunications Infrastructure, was doing routine work to switch telecommunication lines from old copper wire to fiber at a remote location when the failure started. An updated router began flooding other parts of the system, he said, with more than 1,000 connection attempts every four to five seconds, which disrupted communications with airplanes and some radar feeds. Officials told Reuters there may have been problems with routers installed in Arkansas and Texas. Whether the fault was an installation error or a defective router is still undetermined, Duffy said, and the FAA is running additional testing with L3Harris to look for further risk. L3Harris said Saturday that it has completed more than 70 percent of the planned conversion from copper to fiber across the airspace system, that it is supporting the FAA investigation, and that it takes the ground stop and its impact on travelers seriously. The network it manages supports about 45,000 flights a day and connects more than 4,400 facilities.

Homeland Security Secretary Markwayne Mullin said there was zero evidence so far that a cyberattack or a rogue AI system caused the outage, and investigators have ruled out neither possibility, with the National Security Agency working alongside the Department of Homeland Security. Duffy also said the failure had no link to the FAA's AI supported SMART airspace planning system, and Air Space Intelligence, the company behind that system, denied its technology had any effect on the ground stop. Readers who saw the two stories side by side on Friday, a global Microsoft outage the day before and a nationwide ground stop the day after, got an official answer on that pairing too. Mullin told reporters the two incidents were unrelated.

The deeper problem is the age of the system underneath the upgrade. CNN's account of the current network describes controllers handing off paper strips printed with flight names, floppy computer discs still in use, and copper wire carrying communications for nearly a third of the system. Duffy put the stack in one sentence on Friday night: 1980s technology, 1990s software, 1970s radar that the country no longer makes parts for. Reuters has reported that the system runs in places on decades old Compaq computers with replacement parts bought on eBay, and the controller workforce sits several thousand people short. FAA Administrator Bryan Bedford has set September 2027 as the target for replacing the outdated copper telecom wire completely. Duffy was blunt about the bind: building a modern system while keeping the antiquated one running is hard, and doing nothing injects more risk into the system.

Money is where the argument goes next, and the price moved during the crisis itself. Congress approved $12.5 billion for air traffic control upgrades last year. Earlier on Friday, before the outage, the White House and the Transportation Department had asked Congress for $30.1 billion to upgrade airports and air traffic control infrastructure. By Friday night Duffy had raised the ask to $50 billion and called the situation an emergency. A traveler planning a Thanksgiving trip has no lever on that vote. What she has is the record of this weekend: a system that handles 45,000 flights a day, fails rarely, and failed this time because a part meant to modernize it sent more data than the old network could carry.

## A British household could pay £88 more so a data centre can run

### A committee heard the bill, the water count, and the housing trade in one sitting

London has 99 data centres already built, holding more than 80 percent of the United Kingdom's data centre capacity, and another 26 are planned or moving through the development system. On Thursday, November 12, the London Assembly Environment Committee heard what that growth could do to the households living around it. Oliver Hayes, head of campaigns at Global Action Plan, a climate campaign group, told the committee that the group's research puts the added cost at an average of £88 on a British household energy bill, and that high growth projections for data centres could push 77,000 households across Britain into fuel poverty, including about 10,000 in London. Fuel poverty is the condition the name describes: a household that cannot afford to heat its home adequately on its income. Hayes told the committee the growth would further impoverish households already in fuel poverty. Two labels belong on those numbers before they travel. They are projections from a campaign group, presented as early analysis, and they describe households at risk under a high growth path rather than a count of bills already charged.

The water evidence arrived with its own label. Jon Chappell, deputy director of Water UK, the industry body for water companies, told the committee that water systems are built to meet the highest day of demand in the year, so a large new summer load raises the peak the whole system must be sized for. If every new data centre were built like the least efficient one now running, he said, the country would need an extra 77 million litres of water a day, about the daily use of three London boroughs. That figure is a worst case built on the least efficient example, and he paired it with a plainer fact: parts of the east and southeast of England already have places where houses cannot be built and businesses cannot expand for lack of water. Rory Murphy, a research associate at the University of Exeter, added the indirect count: in the United States, data centres used 66 billion litres of water directly to cool servers in a given year, against 800 billion litres used to generate the electricity that kept them running.

London's own planning rules already contain the test this evidence points at. The draft London Plan supports data centres in suitable locations in principle and allows approval only where water and electricity capacity is available. The mayor has commissioned a study on what new data centres mean for the capital's water supply and has told assembly members the city is examining the grid impact as well. Zack Polanski, the Green Party leader who sits on the committee, told the Local Democracy Reporting Service the city needs to slam the brakes on rampant data centre building while residents face high bills and low wages. That is a politician's position in a live argument, and the committee took no decision this week.

The American record gives a London household a measured version of the same argument. A Cornell study published November 9 in Nature Energy examined four United States power grids from 2020 to 2025 and found that data centres paid 38 percent of an $8.25 billion infrastructure bill tied to their growth in 2025, with residents covering most of the rest, and the lowest income households seeing extra energy costs reaching up to 28 percent of their incomes. The Cornell figures are observed, drawn from bills and filings that already happened. The London figures are modeled, drawn from projects that may never be built. A planning committee that holds both kinds of evidence at once has the honest version of the choice: the American count shows what the cost split looked like when nobody set terms in advance, and the London projection shows the size of the bill the terms could still change.

![Alt text: A row of brick terraced houses with lit windows at dusk, a large windowless data centre behind the rooftops.](https://theainextwave.com/assets/inline-2026-11-15-2.jpg)
*The houses were there first. The planning decision decides who pays for what arrived second.*

## The tool that was sold as time back is measuring out as a longer day

### AI users logged days up to 90 minutes longer, and the study comparing them is careful about what it proves

The pitch for AI at work usually ends with time handed back. Public advocates have talked about shorter workweeks arriving on the back of these tools. An analysis released October 29 by ActivTrak, a company that sells workplace monitoring software, measured the opposite pattern in the data its own product collects. Its Productivity Lab examined digital work activity from 9,140 employees at 467 companies between April 1 and September 25, 2026. Employees who used AI had workday spans averaging 8.4 to 9.1 hours. Employees who avoided AI averaged 7.6 hours. The gap reaches about 90 minutes a day at the top of the range and about 50 minutes at the bottom. Heavy users, defined as employees on AI tools for more than 45 minutes across at least five sessions a day, also logged an extra 18 minutes of work after hours compared with non users, switched between applications more often, and held shorter stretches of uninterrupted focus than light users. Gabriela Mauch, the company's chief customer officer and head of its Productivity Lab, said AI may help employees complete tasks faster, and the company's research shows it contributes to longer workdays, more application switching, and shorter focus sessions.

The limits on that finding matter as much as the headline. ActivTrak compared groups of different people during the same months. It did not follow the same employees from before adoption to after, so the data cannot show that AI added the hours. People carrying heavier workloads may reach for AI more often, and the sample comes from the customers of a monitoring vendor, which is a commercial interest a reader should hold alongside the numbers. Gizmodo's coverage of the release carried the same caution, offering the readings a careful editor would: longer hour employees may simply adopt more, AI may encourage people to take on more work, or sorting through weak AI output may eat the time the tool saved.

A second study, built on observation rather than software logs, points in the same direction. Researchers at the Haas School of Business at the University of California, Berkeley spent eight months inside one technology company with 200 employees and published their findings in February. They found that AI made work more intense: employees worked at a faster pace, took on a broader scope of tasks, and extended work into more hours of the day, often without being asked. Xingqi Maggie Ye, one of the authors, warned that the expansion can feel self driven and even exciting at first, and that expectations then reset until the extra effort reads as standard performance. One company and 200 people is a narrow base, and an ethnographic study describes a workplace rather than measuring an economy. Two different methods landing on the same pattern is still worth a worker's attention, because the cost in both accounts is paid in the same currency: the hours at the edge of the day, where family, sleep, and rest sit.

For the employee living inside those numbers, the practical point is about who captures the gain. A tool that helps a person finish a report in half the time creates value exactly once. What happens next decides who keeps it. If the saved hour fills with two more tasks and a later logout, the employer captured the gain and the employee paid for it with a longer day. That transfer happens one accepted extra task at a time, and it rarely appears in any announcement about productivity.

![Alt text: An empty office at night with one desk lit by a lamp, a dark monitor and a cold cup of coffee.](https://theainextwave.com/assets/inline-2026-11-15-3.jpg)
*The day ended on the schedule. The work kept the chair warm past it.*

## The central bank counted the debt, and the count reached the household saver

### Eighty seven percent of the Fed's market contacts named AI their top risk, before the Huang resignation and before the midterms

The Federal Reserve released its semiannual Financial Stability Report on Friday, November 13, and the survey inside it recorded a sharp turn. Staff at the Federal Reserve Bank of New York polled 23 market contacts at broker dealers, banks, investment funds, and advisory firms in September and October. Eighty seven percent of them named artificial intelligence a salient risk to the financial system, up from 50 percent in the spring survey. The next concerns trailed well behind: a correction in risk assets at 52 percent, higher long term interest rates at 48 percent, cyberattacks at 43 percent, and fiscal debt sustainability at 43 percent. The Fed states plainly that the survey reports the contacts' views, which are not the views of the Board or the New York Fed. A survey of 23 professionals is a measure of sentiment, and sentiment is a record of what the people closest to the money are worried about, rather than a forecast of what happens next.

The report's own tally gives that worry a balance sheet. Debt issuance tied to advanced computing infrastructure passed $680 billion by October 2026, about five times the $130 billion issued in all of 2023, and that borrowing now accounts for about 31 percent of nonfinancial investment grade corporate debt issuance, up from under 6 percent in 2023. Outstanding public debt tied to advanced computing stands above $1.1 trillion on the Fed staff's count. A second pile sits off the balance sheet: for seven large companies, leases signed but not yet started carry undiscounted values above $1 trillion, and other commitments such as take or pay contracts and resale guarantees add more than $2 trillion, figures the Fed cautions are not directly comparable with ordinary debt. Reuters summarized the combined picture as debt, guarantees, and leases that could exceed $3 trillion. The largest bank holding companies hold close to $1 trillion of exposure to large technology firms, data centre developers, related financing vehicles, and utilities, and those exposures sit inside the annual stress test. Fed staff compared the cycle with the broadband telecom build out that began in 1995 and said AI investment may surpass that cycle relatively soon, measured against the size of the economy.

The mechanism the report worries about has a household at the end of it. The contacts named three AI concerns: high equity valuations that could fall if earnings disappoint, cyberattacks by state actors or damage caused by frontier AI models themselves, and complex AI related financing. The Fed noted that asset valuations across many classes remain high, fed in part by enthusiasm for AI, and that an unexpected earnings downturn could bring a market decline, tighter credit, and reduced investment, each of which lands on ordinary borrowers and savers. The financing web cuts both ways in the report's telling: the largest firms may absorb losses in a downturn and shield the wider system, and their true exposure may be substantially larger than their balance sheet debt suggests. Hedge fund leverage, concentrated in a small number of large funds, reached the highest level since the data series began in 2013. Against that, the report describes the banking sector as sound and resilient overall, with capital near historic highs, and household debt relative to the economy still falling.

Timing sharpens the reading. The survey closed before Jensen Huang resigned from Nvidia on November 12 and before the November 3 midterm results, and the report's data run to October 28. The worry was already at the top of the list before the week's two largest AI stories broke. This week supplies the first earnings test of that sentiment: Nvidia reports on Tuesday, November 17 at about 4:20 p.m. Eastern, with a call at 5 p.m., and Anthropic's IPO site, filed with the SEC on November 12, expects pricing for its share sale on Wednesday, November 18 with trading expected Thursday, November 19. A household saver holding index funds meets both events through the same account statement, which is the channel the Fed's contacts are describing: concentration at the top of the market, financed with debt the report has now counted in public.

## Opportunity Radar

Cutover risk reviews for live infrastructure carry a dated lesson and a clear buyer. Friday's failure happened during an upgrade, on a system that cannot be switched off for repairs, and the investigation is still deciding between an installation error and a faulty part. A small specialist service can run pre cutover load testing, staged rollout plans, and rehearsed rollback steps for the contractors and operators who modernize systems that must keep running: air navigation, hospitals, utilities, transit networks, and emergency communications. The buyer is the operator or the prime contractor holding the modernization contract. Validation means reproducing a flood condition like Friday's 1,000 connection attempts every few seconds against a test copy of the network and proving the rollback works inside a set time. Success is measured in cutovers completed with no service loss and rollback times on rehearsal. The service fails if buyers treat it as paperwork after the headlines fade, which is a sales problem a fixed price rehearsal package can answer while this weekend's record is fresh.

Workday span audits give employers a second opening from the ActivTrak pattern. Many firms adopted AI tools with no baseline, so they cannot say whether the tools shortened work or stretched it. A consultant can measure workday span, after hours activity, and focus length for the same employees before and after an adoption, using data the firm already holds, and report the change in hours alongside output. The buyer is a mid size firm whose leaders promised staff that AI would ease workload. Validation is a baseline the firm's own records confirm for one department. Success is measured by whether span and output move in the direction the firm promised its people. The work fails if a firm wants the audit as a discipline tool against individual employees, because a measurement sold as workload design and used as surveillance will poison the data and the trust it depends on.

## What You Can Do With This

### If you are flying for Thanksgiving

Check your flight in the airline's app before leaving home and keep checking on travel day, because recovery from a system failure runs for days after the official all clear. Hold one fact from this weekend in mind while you plan: because the FAA caused Friday's delays, airlines are under no requirement to compensate stranded passengers, so a missed connection can leave the hotel and meal costs with you. If your flight was canceled or significantly changed and you choose not to travel at all, you are entitled to a refund of the ticket price, and the airline must tell you so. Keep receipts anyway, ask the airline what it will cover as a goodwill matter, and build a longer buffer into any trip with a deadline at the far end. The ground stop was lifted Friday night and officials said Saturday the worst was behind the country, with recovery still ahead.

### If you pay a power bill where data centres are proposed

Ask the planning body deciding the case for two numbers in writing: the expected effect on household bills and the expected daily water use at peak, with the assumptions named. London's draft plan already contains the test to cite, approval only where water and electricity capacity is available, and the Cornell study gives you an observed American record to bring with it: data centres paid 38 cents of every dollar of grid infrastructure tied to their growth in 2025, and residents paid most of the rest. A projection can be argued with. A cost split that already happened sets the starting point for the argument.

### If you use AI at work

Track your own workday span for two weeks, first login to last activity, and compare it with the span you kept before the tools arrived, if you can reconstruct it from your calendar. The ActivTrak gap was up to 90 minutes a day at the top of the range, and the Berkeley researchers found expectations resetting until extra effort read as standard. A personal record turns that pattern into a workload conversation with dates and hours in it. Bring output with the hours, because the pair together shows whether the gain from the tool stayed with you or moved to your employer.

### If your savings sit in index funds

Read Tuesday and Wednesday before acting on either. Nvidia reports Tuesday, November 17 at about 4:20 p.m. Eastern, and Anthropic's pricing is expected Wednesday, November 18 with trading expected Thursday, November 19, per the company's IPO site filed November 12. The Fed report is a sentiment survey of 23 contacts plus a debt tally, and it forecasts no correction. What it does give a saver is a concentration check: the report counts more than $680 billion of advanced computing debt issued by October and warns that high valuations rest partly on AI earnings arriving as expected. Check the share of your funds held in the largest technology names, a figure your fund's fact sheet publishes, and weigh it against the risk you thought you were carrying.

## The Bigger Picture

Four systems showed the same construction this weekend. Each one layers something new on top of something old, and each one sends the cost of the join to the person at the end. A fiber router met a copper era network and a million passengers sat on terminal floors. A data centre application met a water system sized for yesterday's peak and a British household faces a projected £88 on its bill. A fast drafting tool met an unchanged workload and an employee who uses AI logged a day up to 90 minutes longer than a colleague who does not. A historic build out met public credit markets and a saver's retirement account now rides on a market backed by $680 billion in fresh debt, counted in public by the central bank.

None of those joins was hidden. The FAA's copper wire had a replacement date of September 2027. The London plan carries its capacity test in writing. The ActivTrak data sat in software employers already run. The Fed's tally came from filings companies already publish. What changed this week is that each join produced a document with numbers on it: a cancellation count, a committee transcript, a workday measurement, a stability report. Documents like those give the person at the end of the chain something to act on before the next failure, hearing, adoption, or earnings report arrives. This week hands that person four of them at once, starting with an earnings report on Tuesday afternoon.

## References

[Reuters: Thousands of US flights canceled, delayed as air travel system recovers from shutdown, November 14, 2026](https://www.reuters.com/business/aerospace-defense/thousands-us-flights-canceled-delayed-air-travel-system-recovers-shutdown-2026-11-14/)

[CNN: The systems used to keep America flying are ancient, November 15, 2026](https://www.cnn.com/2026/11/15/business/faa-tech-systems-us-government?cid=external-feeds_iluminar_meta)

[USA TODAY: FAA contractor, router focus of investigation into air traffic halt, November 14, 2026](https://www.usatoday.com/story/news/nation/2026/11/14/faa-ground-stop-probe-l3harris/92780406007/)

[Salamander News, reporting the Local Democracy Reporting Service: London's new AI data centres will increase fuel poverty and water stress, November 14, 2026](https://www.salamandernews.org/2026-11-londons-new-ai-data-centres-will-increase-fuel-poverty-and-water-stress/)

[Europe Says, reporting the London Assembly Environment Committee hearing of November 12, 2026: London data centre boom could push 10,000 households into fuel poverty, November 14, 2026](https://www.europesays.com/britain/166382/)

[Tech Xplore: Data centers need power, lower income rural residents pay the price, Cornell study in Nature Energy, November 9, 2026](https://techxplore.com/news/2026-11-centers-power-income-rural-residents.html)

[Gizmodo: AI users are working up to 90 minutes longer every day, study finds, ActivTrak Productivity Lab analysis, November 3, 2026](https://gizmodo.com/ai-users-are-working-up-to-90-minutes-longer-every-day-study-finds-2000833982)

[IT Brief Australia: AI use lengthens workdays by up to 90 minutes, ActivTrak research on 9,140 employees at 467 companies, November 14, 2026](https://itbrief.com.au/story/ai-use-lengthens-workdays-by-up-to-90-minutes-study)

[Reuters via Kitco: AI rockets to top of worries about financial system in latest Fed survey, November 13, 2026](https://www.kitco.com/news/off-the-wire/2026-11-13/ai-rockets-top-worries-about-financial-system-latest-fed-survey)

[DailyMech: AI becomes the top risk in the Fed's Financial Stability Survey as data center debt passes $680 billion, November 14, 2026](https://dailymech.com/articles/fed-financial-stability-report-ai-top-risk-data-center-debt.html)

[US Department of Transportation: Refunds](https://www.transportation.gov/individuals/aviation-consumer-protection/refunds)

[Enjoy, reporting Anthropic's IPO site filed with the SEC on November 12, 2026: Anthropic's IPO is expected to price November 18 and trade November 19](https://enjoy.dev/blog/anthropic-ipo-filing/)

[World Money Roadmap: Nvidia Earnings and Anthropic IPO Test the AI Trade Next Week, November 15, 2026](https://worldmoneyroadmap.com/en/reports/2026-11-15-preview/)
