What Is Dark Energy?

On the night of 14 April 2026, five thousand robotic fibres on a telescope at Kitt Peak swung towards a patch of sky near the Little Dipper and took their last scheduled observation. The Dark Energy Spectroscopic Instrument had finished the map it set out to make: 47 million galaxies and quasars, against a target of 34 million, covering 11 billion years. The survey is done. The question that built it is not. And in the eighteen months before that final exposure, the evidence for the answer everyone wanted got weaker, twice, for reasons worth understanding.

What dark energy is, and what the name hides

Two teams found in 1998 that distant supernovae were fainter than an expanding, decelerating universe allowed. The expansion is speeding up. Something is pushing.

Dark energy is the name for that something. It is a label attached to an observation, not an explanation of it. What physicists actually measure is one number: the equation of state, written w, the ratio of the pressure of this component to its energy density. Ordinary matter has w = 0. Radiation has w = 1/3. To push space apart you need w below -1/3.

The simplest candidate is Einstein’s cosmological constant, Λ, an intrinsic energy of empty space. A constant has w = -1 exactly, everywhere, for all time. That is what makes it simple and what makes it testable. Measure w to be anything other than -1, or catch it changing, and the cosmological constant dies. It sits inside the standard model of cosmology, ΛCDM, which currently accounts for about 70 percent of the universe this way.

Simple does not mean comfortable. Quantum field theory says the vacuum carries energy, and the obvious calculation of how much overshoots the observed value by something between 60 and 120 orders of magnitude, depending on where you stop the sum. Nobody has explained that gap. So the cosmological constant is at once the best-fitting description of the data and the worst-understood number in physics.

How you measure something that emits nothing

You measure geometry. Dark energy leaves no light, so every constraint on it comes from comparing distance against redshift and asking which expansion history fits.

Three rulers do the work. Baryon acoustic oscillations are a preferred separation of roughly 150 megaparsecs between galaxies, frozen into the matter distribution by sound waves in the early plasma. Theory fixes that scale, so its apparent size gives the distance. Type Ia supernovae are standard candles. Astronomers can standardise their intrinsic brightness, so how bright they look gives a distance. The cosmic microwave background anchors the whole ladder at one end, fixing the early universe with high precision.

DESI measures the first of these. It takes spectra of millions of galaxies and quasars, converts redshifts into a three-dimensional map, and looks for the acoustic scale at seven epochs. No single probe settles anything. The conclusions everyone argues about come from combinations, and the combination you choose is most of the argument.

What DESI actually found

DESI’s second data release covered three years and more than 14 million galaxies and quasars. Taken by itself, it fits a flat ΛCDM universe comfortably. The trouble starts on combination.

Under ΛCDM, the parameters DESI prefers sit in mild tension with the parameters the microwave background prefers, at 2.3 standard deviations. Allow w to vary with time and that tension eases. The favoured solution has dark energy stronger than a constant in the past and weaker now, crossing w = -1 somewhere around redshift 0.4. For DESI plus the microwave background alone, the evolving model beats ΛCDM at 3.1σ. Add supernovae and the preference lands somewhere between 2.8σ and 4.2σ.

Then George Efstathiou checked the supernovae

If the answer depends on which supernovae you use, the supernovae deserve scrutiny. George Efstathiou, of Cambridge, gave them some.

Roughly a fifth of the supernovae appear in both the Pantheon+ and DES-SN5YR compilations, so two independent pipelines have measured the same objects. Writing in Monthly Notices in 2025, Efstathiou found the two disagreed by about 0.04 magnitudes between low and high redshift. That is a tiny number. It is also enough. His argument ran as follows. A claim of this weight needs low-redshift and high-redshift photometry consistent to well under 0.04 magnitudes. An offset this size could manufacture a preference for evolving dark energy out of nothing.

The Dark Energy Survey answered directly. They reproduced the offset, then traced 43 percent of it to genuine improvements in DES-SN5YR: better treatment of intrinsic scatter, host-galaxy properties, and selection bias. On their reading the difference belongs there. It is not an error. The exchange ran sharp and public. That is how this should work.

The recalibration that moved the number

Rather than argue further, DES rebuilt the measurement. Brodie Popovic led the reanalysis. It went up as a preprint in late 2025 under the name DES-Dovekie.

Four changes mattered. They cross-calibrated the surveys using observations of white dwarfs, whose spectra are well modelled. They retrained the SALT3 light-curve model on the new calibration. They replaced a polynomial approximation to a dust colour law that had been sitting in the analysis software since around 2013. And they fixed an input file in which nine calibration weights had been rounded to 0.3 instead of 0.33, which had quietly understated the photometric uncertainty by about 20 percent.

Combined with DESI DR2 and the microwave background, the recalibrated sample gives w₀ = -0.803 ± 0.054 and w_a = -0.72 ± 0.21. The preference for evolving dark energy falls from 4.2σ to 3.2σ. A colour-law approximation and a rounding error were worth a full standard deviation.

3.2 sigma and five-to-one odds are the same statement

The Dovekie paper quotes two numbers for the strength of its own evidence. Frequentist: 3.2σ. Bayesian: odds of roughly 5 to 1, which the standard scale calls weak. Those look contradictory. They are not.

The frequentist number asks a narrow question. Given that ΛCDM is true, how improbable is a fit this much better? It penalises the extra model parameters lightly, just by counting them.

The Bayesian number asks which model the data support, and it charges a heavier fee for complexity. A model with two spare parameters can fit many datasets it was never asked to fit, and that flexibility costs it evidence. Adding freedom to a model always improves the fit, so the real question is whether it improved enough to justify the freedom. Here the answer is: a bit, not much.

Both readings agree on the thing that matters. Physics treats 5σ as the threshold for a discovery. Nothing in this story is close.

DESI’s own high-redshift probe moved towards the constant

In July 2026 DESI published a second 2026 development, and it also pointed away from evolution.

At redshifts above about two, galaxies are hard to see, so DESI switches tracer. Hydrogen gas between galaxies absorbs quasar light at a fixed wavelength. Each cloud sits at a different distance, so one quasar spectrum carries a whole forest of absorption lines. That forest maps the matter along the sightline. Earlier analyses used only the acoustic peak in those correlations. The new work fits the entire shape of the correlation function, which carries geometric information the peak alone discards. The result is a one percent measurement at an effective redshift of 2.33, twice as tight as the acoustic-peak analysis of the same data.

Its central value sits closer to ΛCDM than the earlier measurement did. DESI’s own summary stays careful. The new numbers agree with the standard model. That could mean the hints of evolving dark energy fade away, or that fitting everything at once will take a more complicated model. That is the collaboration talking about its own headline result, and it is more measured than most coverage of it.

Why a changing w would break more than a model

Here is the part that usually goes missing. The favoured solution does not merely have w drifting. It has w crossing -1, from below to above. That crossing is very hard to build.

Take the obvious model: a single scalar field rolling down a potential, minimally coupled, with a normal kinetic term. For such a field the energy density and pressure satisfy ρ + p = (dφ/dt)², the square of the field’s rate of change. A square cannot be negative. So ρ + p is never negative, which means w is never below -1. Quintessence of this kind can approach the cosmological constant. It cannot go past it.

Getting across the line takes something more: two fields with opposite-sign kinetic terms, a field coupled non-minimally to gravity, an interaction between dark energy and dark matter, or modified gravity outright. Each route has to dodge its own instabilities, from ghosts to a sound speed that blows up at the crossing. The claim on the table is therefore not a small adjustment to one parameter. It is a claim that the dark sector has structure.

The neutrino mass bound rides on the answer

This is not a self-contained argument about cosmology. It propagates into particle physics.

Massive neutrinos suppress the growth of structure, so cosmological data bound the sum of the neutrino masses. From DESI plus the microwave background, that 95 percent upper limit is 0.064 eV assuming ΛCDM. Allow dark energy to evolve, and the same data give 0.16 eV. The bound loosens by a factor of two and a half purely because of what you assumed about dark energy.

That matters because the tighter version is in uncomfortable territory against laboratory oscillation results, which set a floor on the sum. Cosmology’s neutrino bounds are model-dependent in a way that is easy to forget when a single number gets quoted. Any cosmological limit on neutrino mass should be read with the dark energy model it assumed printed alongside it.

What will actually settle it

More data, and better supernovae.

DESI is now processing its full five-year dataset, with first dark energy results expected in 2027. The instrument keeps observing into 2028, expanding from 14,000 to 17,000 square degrees and aiming at 63 million extragalactic redshifts. Euclid’s first major data release lands in October 2026, with cosmological constraints following rather than arriving with it. The Vera Rubin Observatory will eventually deliver supernova samples two orders of magnitude larger than DES-SN5YR. At that point the limit stops being statistics and becomes systematics, which is precisely the problem Dovekie set out to attack.

Note what does not settle it. Another paper reanalysing the same supernovae with a new parameterisation will not. The literature has produced a great many of those since 2024, and the significance has drifted up and down inside the same window throughout.

Established, contested, unproven

Established. The expansion of the universe is accelerating. Something with strongly negative pressure dominates the energy budget. DESI has measured the acoustic scale across 11 billion years at unprecedented precision, and those measurements are consistent with DESI’s earlier data and with SDSS. Under ΛCDM, the combination of current datasets fits less well than it used to.

Contested. Whether dark energy evolves. The preference is real in the data and sits between roughly 3σ and 4σ depending on choices that are defensible either way: which supernova sample, which microwave background likelihood, frequentist or Bayesian. Spatial curvature is a live alternative explanation for the same tension. Supernova systematics remain the leading suspect for part of the signal, and the community does not agree on how much.

Unproven. That w crosses -1. That the cosmological constant is dead. Nothing here reaches the 5σ discovery threshold, two separate 2026 results moved the evidence towards the constant rather than away, and the decisive dataset has been taken but not yet analysed.