The casual reader of Earth observation announcements in 2026 sees a normal year for the field. A new model from IBM here, another from Google DeepMind there, a few academic releases, an open platform launch. Each announcement reads as a steady-state technical maturation, the kind of incremental progress that any active research field produces.

The aggregate is not normal. The aggregate is that the release cadence for geospatial foundation models has gone from roughly one per year before 2024, to monthly across 2025, to weekly across academic and commercial labs in the first half of 2026. The wave's organisers know it is real. The downstream industry that depends on these models, the ESG verifiers and climate analytics firms and natural-capital data products, has not yet caught up.

This piece is a data read of the wave. Five findings, each anchored to public material. The first three describe what is happening; the last two describe what has not yet happened, and what that means for the firms that build on top of the wave.

Finding 1. The release cadence shifted from yearly to weekly in roughly 24 months.

Before 2024, a new geospatial foundation model was a once-a-year event. The field had a handful of models in active use, mostly academic, each with a narrow scope: a vision transformer fine-tuned on Sentinel-2, a self-supervised encoder for Landsat, a multi-sensor fusion network for crop classification. The output was useful but each model had to be paired with its own pipeline.

The cadence began to shift in late 2023 and accelerated through 2024. The Cloud-Native Geospatial Forum's taxonomy of Earth embedding products, published in early 2026, catalogues the inflection: roughly one foundation model release per year through 2022, jumping to a handful in 2024, then to a release cadence in 2025 measured in weeks rather than months.

A non-exhaustive list of named releases since late 2024, sequenced by ship date:

The list omits at least a dozen smaller models (Satlas, SatCLIP, SatMAE, Presto, ScaleMAE, SpectralGPT, Major-TOM, MOSAIKS, and others) that ship with similar regularity. Counting only the named flagship releases, the cadence in the first five months of 2026 was approximately one substantial release per fortnight; the pre-MTEB moment that the community started talking about in May 2026 captures this aggregate without saying it directly.

The casual reader sees one model. The data shows a wave.

Finding 2. Architectures differ by an order of magnitude. The output structure does not.

The wave's models do not agree on size, training data, or sensor strategy. The variance is large.

That is more than an order of magnitude in parameter count between Clay (26M) and Prithvi-EO-2.0 (600M). It is a doubling of embedding dimension between AlphaEarth (64) and TESSERA (128), and a factor-of-eight resolution difference between the 10m models and Prithvi's 30m alignment to the Harmonized Landsat Sentinel time series. The training datasets vary. The sensor combinations vary. The pre-training objectives vary.

What does not vary is the output structure. Every model in the table produces a dense per-pixel embedding of the Earth's surface, queryable for downstream tasks without retraining, retrievable through k-nearest-neighbour or linear-probe queries. The output convergence is the coordination point that did not exist three years ago.

A 2026 working group paper on Earth embedding products notes the same convergence with a different phrase: each product ships at a per-pixel grain, each works with the same downstream query patterns, each can be plugged into an analytical workflow in days rather than the years that traditional Earth observation pipelines required. The architectural divergence is real and visible. The analytical convergence on top of it is more striking.

For a climate data team or an ESG verifier choosing which model to bet on, this is the most important finding in the data. The choice of model determines the architecture, the cost, the licensing, the deployment story. The choice of model does not, in practice, determine what kinds of questions you can ask. Any of the eight named flagship models can answer the same downstream questions about an asset's land use, methane plume signature, vegetation index, or material flow. The convergence on output structure is what makes a market possible.

Finding 3. Distribution is fragmented across at least four mutually incompatible channels.

A foundation model's value to downstream users depends on how easily its outputs can be accessed. The wave has not converged on distribution. The Cloud-Native Geospatial Forum's 2026 audit lists the current distribution landscape with explicit frustration.

The audit uses the phrase "the snowflake problem": every new Earth embedding product ships like a snowflake. The underlying file formats (Cloud-Optimised GeoTIFF, GeoParquet, raw NumPy, native HDF5) are inconsistent. Coordinate reference systems are sometimes attached, sometimes missing. Licensing terms differ from research-only to permissive open-source to commercial-restricted. There is no equivalent of PyPI or DockerHub for Earth embeddings.

The lack of a common channel imposes real cost on downstream users. A climate analytics firm building a portfolio-screening tool that wants to combine AlphaEarth embeddings (for the global geographic baseline) with TESSERA embeddings (for the time-series detail) cannot simply import both. The firm has to write custom adapters for each, manage two distribution channels, reconcile coordinate systems, and rebuild the integration whenever either source updates its API.

The 2026 Earth Embeddings as Products paper proposes a TorchGeo integration as a partial fix, exposing the major embedding products through a single Python API. The effort is real, the API is shipping, but adoption across the wave is still partial. A common distribution channel is the field's most obvious missing piece in 2026.

Finding 4. Benchmarks fragment in the same way the models do. There is no MTEB moment yet.

The text embedding world reached a coordination point in 2022 with the release of MTEB (Massive Text Embedding Benchmark), which let buyers compare text embedding models head to head on a single leaderboard. Earth observation does not yet have an equivalent.

Each major foundation model currently claims state-of-the-art performance on its preferred benchmark suite. Prithvi-EO-2.0's release paper reports an 8-point GEO-Bench improvement over the 2023 version. TerraMind's release reports outperforming the existing twelve geospatial models on the PANGAEA benchmark by 8 per cent or more. OlmoEarth claims wins over Meta's DINOv3, Prithvi, and TerraMind on the Allen Institute's own evaluation. TESSERA outperforms baselines on its own downstream task selection.

All four claims can be true simultaneously. They use different evaluation suites, different downstream tasks, different metrics. The community has used the phrase "pre-MTEB moment" to capture the current state: each lab grading its own homework, with no neutral arbiter that lets a buyer rank the models on a common scale.

This matters because benchmark fragmentation slows commercial adoption. A natural-capital data firm that wants to commit to one foundation model for the next three years has no clean way to choose. The benchmarks the model's builders published do not transfer to the buyer's actual workflow. The buyer has to run an internal evaluation, which costs three to six months of engineering time, before committing.

The benchmark moment is coming. Earth observation has the same constituency push (researchers, standards bodies, downstream users) that text embeddings had in 2021-2022. A neutral leaderboard is a probability not a question. But it has not arrived yet, and the wave's commercial uptake will be capped at the limit of buyers' patience for running their own evaluations until it does.

Finding 5. Storage cost is the wave's silent ceiling.

The release cadence describes the supply side of the wave. The storage cost describes its ceiling.

A single global Earth embedding product, queried at 10-metre resolution across the planet's land surface, occupies a number of bytes large enough that distribution becomes a meaningful cost. The AlphaEarth migration from Google Cloud to Source Cooperative moved 465 terabytes. TESSERA's per-pixel 128-dim embeddings, generated globally, would occupy a similar magnitude. Prithvi's HLS-aligned outputs at 30m are coarser but cover more temporal slices.

The cost arithmetic is unfavourable for the downstream user who wants to run analytics across the global embedding. Cloud egress charges for moving the embeddings out of the originating provider's network can exceed the cost of running the model. Source Cooperative's role as a neutral distribution channel is partly a response to this: bulk-distribute once, let users compute against the embeddings without paying per-query egress.

For climate analytics firms and ESG verifiers, this means the choice of model is partly a choice of where the storage lives. AlphaEarth's migration to Source Cooperative made it cheaper to integrate; TESSERA's university-hosted API is fast for small queries but does not scale to portfolio-level analytics; Prithvi on Hugging Face is downloadable but every user pays the storage cost again. The model that gets adopted will not always be the model with the best benchmarks. It will often be the model whose distribution channel makes it cheapest to use at the scale the buyer needs.

The wave's release cadence is not the bottleneck. The wave's storage and distribution are.

What the data adds up to

Five findings stacked together produce an editorial picture.

The wave is real and accelerating, releasing more models in the first five months of 2026 than the field produced in 2022 and 2023 combined. The architectures vary by an order of magnitude in size but converge on a common output structure, which is the coordination point that makes a downstream market possible. The distribution channels and benchmarks do not converge, which holds commercial uptake below where the technical maturity would otherwise put it. Storage and egress cost is the silent ceiling on how usable the embeddings actually are at portfolio scale.

For the ESG verifier or climate analytics firm choosing where to invest engineering effort in 2026 and 2027, the practical implication is that the model selection question is less important than the distribution and benchmark questions. Picking AlphaEarth versus Prithvi versus TESSERA matters less than choosing which distribution channel and which evaluation framework the firm builds against. The first firm to standardise on a common integration pattern across several models will be in a stronger position than the firm that bets on one model in particular.

The wave is real. The convergence is real. The market is still missing.