At 9:30 in the morning on February 27, 2025, the National Weather Service sent out a two-page notice about a single Arctic town. Effective immediately and until further notice, the agency was suspending weather-balloon launches at Kotzebue, Alaska — upper-air site identifier PAOT, World Meteorological Organization number 70133 — "due to a lack of WFO staffing." The notice described no storm and no broken instrument. It described an absence: a station that would no longer send a radiosonde into the sky twice a day, indefinitely, because there were not enough people left to do the launching.

The agency's own mitigation plan was modest and specific. Kotzebue's gap would be covered by balloons from neighboring sites, by polar and geostationary satellite soundings, and by sensors on aircraft using nearby airports. Nothing in the notice claimed the loss was harmless. Nothing claimed it was catastrophic. It simply registered, in the flat administrative language the agency uses for all such notices, that one node had gone dark in a network its own upper-air program describes as 69 sites in the continental United States, 13 in Alaska, nine in the Pacific and one in Puerto Rico, plus ten more it supports in the Caribbean.

It did not stay isolated. On March 20, a second notice from the same division director reduced six more offices — Aberdeen, South Dakota; Grand Junction, Colorado; Green Bay, Wisconsin; Gaylord, Michigan; North Platte, Nebraska; and Riverton, Wyoming — from two flights a day to one. Reporting from Inside Climate News that same week found NOAA had by then suspended all launches indefinitely at three stations — Kotzebue, Omaha, and Rapid City — and paused two more, Albany and Gray, Maine, promising to resume "when staffing permits." By mid-April, a third notice turned the pattern into standing policy: the agency would, "until further notice," reserve the right to "temporarily reduce or suspend scheduled radiosonde launches at selected NWS upper air sites due to staffing limitations or operational priorities," as conditions required, with no list of sites and no end date attached.

The staffing behind these notices had a documented cause. NOAA had lost roughly 650 employees to job cuts driven by the Department of Government Efficiency, and colleagues at the agency said they had been told to identify another 1,029 positions for elimination, disproportionately among probationary staff in their first two years. Those early-career hires, one former NOAA official said, had the machine-learning and data-science skills the agency needed to build its next generation of models. That official, Michael Morgan, described the mechanism plainly: "We're going to lose data because of this staffing... That loss of data then translates into less precise forecasts, more uncertainties in the forecast." He added his own caveat: "Does it mean every single forecast is going to be poor? No, but it does mean that the uncertainties in our forecast will grow over time."

An Associated Press review of internal staffing data showed the scale behind the individual notices. Fifty-five of 122 weather field offices carried vacancy rates of 20 percent or higher; the systemwide rate had risen from 9.3 percent a decade earlier to 19 percent by March 2025. Rapid City stood at 41.7 percent vacant and Omaha at 34.8 percent — the same two offices that had just gone to zero balloon flights a day. That week, staffing shortages also kept Louisville meteorologists from immediately surveying tornado damage after an outbreak, a step normally used to sharpen future warnings. Brad Colman, a past president of the American Meteorological Society, called the vacancy numbers "a crisis situation." Wyoming meteorologist Don Day, whose regional forecasts depend on the Riverton and Rapid City launches, was less measured: "The absolute starting point for weather forecasts is the balloon data. There is nothing more important than the balloon data." Losing it, he said, was "like not changing the oil in an engine."

A radiosonde is a small thing to build a forecasting system around — sixty to eighty grams of sensor, rising for about two hours and reporting once a second as it climbs. The National Weather Service's own upper-air program lists several uses for that ascent: input for prediction models, local severe-storm and aviation forecasts, climate research. And one that is easy to overlook. The balloon provides "ground truth for satellite data," a physical check against which the far larger volume of orbital measurement is calibrated. Satellites see more of the atmosphere than any balloon network could. They still need something independent to tell them when they are wrong.

The ocean side of the observing system is thinning for a slower, less visible reason: chronic underfunding rather than a staffing order. Many U.S. wave and weather buoys run through NOAA's National Data Buoy Center, alongside the Integrated Ocean Observing System, the federal-regional partnership whose regional associations fund much of the network's day-to-day maintenance. An independent study estimated IOOS needs roughly $715 million to fully realize its mission; the most it has ever received is $42.5 million, a level it has been effectively stuck at for years.

In 2022, the Northeastern Regional Association of Coastal Ocean Observing Systems pulled a buoy from the Gulf of Maine's Northeast Channel, where warm Gulf Stream water meets Arctic meltwater, because stagnant funding made routine servicing impossible; forced to choose, the group kept the buoys closer to shore that mattered more for marine safety and let go of the one that primarily supported research. Buoys that should be serviced five times a year now sometimes get one visit. "If one of our buoys goes offline, I hear about it from fishermen before our data guys or sensor technicians," said the association's executive director, Jake Kritzer. IOOS Association head Kristen Yarincik was blunt about what a further funding lapse would cost: reduced data would hit "navigational safety for commercial shipping, fishermen and recreational boaters; local flood monitoring for coastal communities; and weather forecasting, especially for hurricane intensity forecasts."

None of this data disappears into a void. Along with millions of other daily measurements, it feeds the reanalysis datasets used to train and grade the current generation of AI weather models — and reanalysis is a specific, limited thing, not a synonym for the real atmosphere. Three categories are easy to blur. Observations are direct physical measurements. A forecast model, physics-based or machine-learned, turns a starting snapshot of the atmosphere into a prediction. Reanalysis is neither: it is a fixed-model reconstruction of the past, built after the fact by blending both. The European Centre for Medium-Range Weather Forecasts defines the process behind its widely used ERA5 reanalysis precisely: "data assimilation is a process whereby a model forecast is blended with observations to obtain the best fit to both the forecast and the observations, given the known uncertainties of both. The result is called an analysis."

ERA5 runs that blend twice daily using a single forecast-model version that was operational in 2016, held constant so the same yardstick applies across decades. Between 1979 and 2019, the volume of observations it assimilated grew from about 750,000 a day to 24 million, drawn from radiosondes, aircraft, ships, and, increasingly, satellites. ECMWF's own documentation is candid about the limits: reanalysis uncertainty "becomes larger... in data sparse locations," and "spurious changes will still be included in the reanalysis, due to changes in the observing system." A reanalysis is not independent of the observing network beneath it. It is a sophisticated best estimate built from that network — and the paper documenting ERA5's construction notes that independent buoy data gave "a much improved fit for ocean wave height" over the prior reanalysis, one of the ways its own builders check whether the reconstruction can be trusted. The record checks itself against the buoys now being pulled from the water.

That distinction matters because reanalysis has become the near-universal scoreboard for machine-learning weather models. A widely cited 2023 preprint, from researchers at Excarta, a commercial weather-forecasting company, found FourCastNet's advantage over a physics-based forecast vanished when the two were compared against real ground observations rather than ERA5 — though the same experiment also used a coarser starting state, so the yardstick was not the only thing that changed. The authors' conclusion was the modest one: models validated in simulation need testing against measurement before anyone trusts the ranking. A 2026 benchmark called RealBench was built on a related but narrower complaint: reanalysis products, its authors write, are assembled through delayed processing that does not reflect the constraints of real-time operational forecasting, so grading against them risks rewarding skills operational forecasting will never use. RealBench replaces reanalysis with a live network of more than 10,000 weather stations — and its authors are careful to credit what reanalysis still does well, calling ERA5 "a high-quality retrospective reference."

Two further peer-reviewed studies show where the mismatch hides. Physics-based forecasts still consistently beat AI models on record-breaking heat, cold and wind events, even in cases where the same AI models win on ordinary accuracy scores. And in a detailed look at Storm Ciarán, a 2023 windstorm that killed at least sixteen people across northern Europe, four machine-learning models captured the storm's broad shape but underestimated its peak winds and missed its sharpest frontal structure. That last case cuts both ways. The study's authors ruled out one explanation cleanly: forecasts made using the ERA5 system itself did not share the low wind bias, so the underestimation was, in their words, "unlikely to be the result of training the ML models on the ERA5 data." A closer dynamical analysis found the likelier culprit was what the authors called "inadequacies in the geopotential field," an unphysical roughness built into the models' own output, rather than the coarseness of their grids.

A companion essay on this site raised the possibility that thinning observations could let forecasting systems "congratulate themselves against a deteriorating yardstick," and declined to chase it further. The chase turns up a documented mechanism and nobody visibly minding the seam. ECMWF's own documentation names observing-system change as a source of spurious signal in the reanalysis record — that is not an inference, it is the dataset's builders naming the failure mode themselves. The National Weather Service has spent 2025 executing exactly such a change: unplanned, staffing-driven, with no end date attached. No published study joins the two, and nothing in either institution's public materials suggests anyone is tracking whether it should.

Radiosondes are one contributor among hundreds of millions of daily inputs to ERA5, most of them now satellite-based; losing a handful of U.S. sites does not, by itself, compromise a global dataset built from four decades of data. The notices preserve the ability to launch special observations ahead of a specific dangerous storm. The routine baseline is what stopped; the emergency response held. Retiring an aging balloon network in favor of satellites and aircraft sensors is a legitimate scientific choice, not automatically a loss — provided the substitution is measured rather than assumed.

What is harder to dispute is the shape of the arithmetic. A network that thins one staffing notice or one missed grant deadline at a time loses two things at once: an input to the forecast, and a fragment of the record used to check whether the forecast — or the model being praised for beating it — is actually right. In Kotzebue, the balloon that used to leave the ground twice a day now does not leave it at all. The gap will not announce itself in the next forecast cycle, and it will not announce itself in the next reanalysis run either. It is simply absent, on both sides of the ledger, at the same time.

Source note

  • National Weather Service, Public Information Statement 25-08, "Suspension of Radiosonde Observations at Kotzebue, Alaska, Effective Immediately," issued by Mike Hopkins, Director, Surface and Upper Air Division, 9:30 AM EST Thu Feb 27, 2025 (weather.gov/media/notification/pdf_2025/pns25-08_kotzebue_ak_upper_air_suspension.pdf); Public Information Statement 25-18, 4:35 PM EDT Thu March 20, 2025 (weather.gov/media/notification/pdf_2025/pns25-18_Reduction_RAOB_Launches_ABR_GJT_GRB_APX_LBF_RIW.pdf); Service Change Notice 25-36, "Temporary Reduction of Radiosonde Observations, from Selected Sites," 10:30 AM EDT Thu Apr 17, 2025 (weather.gov/media/notification/pdf_2025/scn25-36_suspension_of_raob_launches.pdf); and the NWS "Radiosonde Observation" factsheet (weather.gov/upperair/factsheet), for site identifiers, dates, station counts, radiosonde specifications, the listed uses of a sounding, and the "ground truth for satellite data" phrase.
  • Dennis Pillion, "NOAA Cuts Weather Balloon Launches Due to Staff Shortages After DOGE Layoffs," Inside Climate News, March 25, 2025, for the three-station indefinite-suspension list, the Albany and Gray pause, the DOGE staffing figures, and the Morgan quotations. (insideclimatenews.org/news/25032025/noaa-cuts-weather-balloon-launches-due-to-staff-shortages-after-doge-layoffs/)
  • Associated Press, "Nearly half of National Weather Service offices have 20% vacancy rates, and experts say it's a risk," April 2025, for vacancy statistics, the Louisville tornado-survey detail, and the Colman quotation. (apnews.com/article/doge-weather-cuts-tornado-dangerous-staff-warnings-aa7db3e0d0009d99c143742ab722c40a)
  • Andrew Rossi, "Wyoming Meteorologist Says Cutting Weather Balloon Launches 'Dumb And Unacceptable,'" Cowboy State Daily, March 21, 2025, for the Don Day quotations. (cowboystatedaily.com/2025/03/21/weather-balloon-launches-cut-just-in-time-for-wyomings-spring-storm-season/)
  • Ryan Krugman, "As NOAA Funding Lags, a Critical Ocean Weather System Nears a Breaking Point," Inside Climate News, Dec. 8, 2025, for IOOS funding figures, the Northeast Channel triage decision, and the Kritzer and Yarincik quotations (insideclimatenews.org/news/08122025/noaa-integrated-ocean-observing-system-funding/); NERACOOS (neracoos.org), for the association's own name; National Data Buoy Center (ndbc.noaa.gov) for the buoy network's operating role.
  • ECMWF/Copernicus Climate Change Service, "ERA5: data documentation" (confluence.ecmwf.int/display/CKB/ERA5%3A+data+documentation), for the definition of data assimilation and the discussion of reanalysis uncertainty and observing-system artifacts; Hersbach et al., "The ERA5 global reanalysis," Quarterly Journal of the Royal Meteorological Society 146 (2020): 1999–2049, for observation volumes, input types, and the buoy wave-height validation finding. (rmets.onlinelibrary.wiley.com/doi/10.1002/qj.3803)
  • "Verification against in-situ observations for Data-Driven Weather Prediction," arXiv:2305.00048 (2023, preprint, authors affiliated with Excarta), for the FourCastNet comparison; "RealBench: Benchmarking Data-Driven Numerical Weather Forecasting Under Operational Conditions and Extreme Event Challenges," arXiv:2605.24945 (2026), for the benchmark rationale and station count; Zhang, Fischer, Zscheischler and Engelke, "Physics-based models outperform AI weather forecasts of record-breaking extremes," Science Advances, May 1, 2026, doi:10.1126/sciadv.aec1433, for the record-event comparison; Charlton-Perez et al., "Do AI models produce better weather forecasts than physics-based models? A quantitative evaluation case study of Storm Ciarán," npj Climate and Atmospheric Science 7, 93 (2024), for the Storm Ciarán findings, the direct quotation on ERA5-system forecasts not sharing the low wind bias, and the geopotential-field finding from the paper's dynamical-balance analysis. (nature.com/articles/s41612-024-00638-w)
  • "The Instruments That Disagree," Large Language (large-language.ai/read/the-instruments-that-disagree), as the predecessor essay and the source of one attributed callback phrase. Its opening case (the March 20 six-site reduction) and its Canadian bias-correction study are deliberately not repeated here.