Algorithmic Progress Is Quietly Outpacing Your Data Strategy

The dominant assumption behind most enterprise AI roadmaps built over the last three years is that proprietary data is the primary defensible asset. Accumulate enough of it, clean it well, and you have a moat. That logic made sense when model architectures were relatively stable and the marginal gain from better algorithms was modest compared to the gain from more data. Neither condition holds cleanly in 2026, and the investment implications for engineering and data leaders are worth examining directly.

The Greenblatt Thought Experiment and What It Reveals

The Greenblatt framing asks a pointed question: if you hold data constant and allow only algorithmic improvements to vary, how much of observed model performance improvement do you recover? The answer, across multiple capability domains, is a substantial share. This is not a theoretical curiosity. It is a direct challenge to the capital allocation logic that sent billions into data lake infrastructure, labelling pipelines, and proprietary dataset accumulation.

The mechanism matters here. Algorithmic progress compounds through better training objectives, improved post-training methods such as reinforcement learning from human feedback and its successors, architectural changes that extract more signal from the same token budget, and inference-time compute strategies that shift capability from training to deployment. Each of these improvements applies to whatever data you already have. The data does not need to grow for the model to get meaningfully better.

The commercial implication is that teams who built their moat argument on data volume alone may have underestimated how much of the performance gap they were attributing to data was actually attributable to the model and training regime sitting on top of it.

Why Data Moats Are Weaker Than They Appeared

Data moats rest on two conditions: the data must be genuinely proprietary, and the model must not be able to substitute for it through other means. Both conditions are eroding.

Synthetic data generation has matured to the point where high-quality training signal can be produced at scale without proprietary corpora in many domains. Distillation from frontier models, combined with careful filtering, has allowed smaller organisations to close capability gaps that were previously assumed to require years of data accumulation. This does not eliminate the value of domain-specific data, but it does compress the lead time that proprietary data was supposed to guarantee.

The second erosion is more subtle. As foundation models improve, the marginal value of additional domain-specific fine-tuning data decreases for tasks where the model already has strong priors. The cases where proprietary data remains genuinely decisive are narrowing: they are concentrated in domains with highly idiosyncratic structure, regulatory constraints on data sharing, or where ground truth labels require expensive expert annotation that cannot be replicated synthetically.

Where Algorithmic Investment Actually Compounds

If data accumulation is not the primary driver of durable advantage, the question becomes what is. The answer is not simply "better models" in the abstract. It is the organisational capability to evaluate, adopt, and adapt algorithmic improvements faster than competitors.

Evaluation Infrastructure

The bottleneck for most enterprise teams is not access to better models. Frontier models are increasingly accessible via API or open weights. The bottleneck is the ability to evaluate whether a new model or training method actually improves performance on the specific tasks that matter commercially. Teams without rigorous internal evaluation infrastructure cannot tell whether a new release is genuinely better for their use case or simply better on public benchmarks that do not reflect their workload.

Post-Training Agility

Post-training methods, including instruction tuning, preference optimisation, and reinforcement learning with process-level reward signals, are where a significant share of practical capability gains are now concentrated. We have written previously about progress advantage as an annotation-free step-level scoring signal derived from RL post-training, and the degree to which it changes what is achievable in long-horizon agentic tasks. Teams that can run these methods internally, rather than waiting for a vendor to incorporate them into a hosted product, compress the adoption cycle from months to weeks.

Architecture Selection Cadence

Mixture-of-experts architectures, speculative decoding, and inference-time compute scaling have each materially changed the cost-performance frontier in the last eighteen months. The teams benefiting most are those with the internal fluency to evaluate architectural trade-offs and the infrastructure to test them without full retraining cycles. We covered the evaluation framework for this in our piece on training infrastructure as a competitive moat.

What This Means for Capital Allocation

None of this argues for abandoning data infrastructure investment. Domain-specific data remains valuable, particularly where it captures proprietary signals that synthetic generation cannot replicate, such as internal transaction histories, customer behaviour sequences, or annotated outcomes from proprietary workflows. The argument is about concentration and sequencing.

The reallocation question for a CTO is whether the next incremental dollar spent on data pipeline capacity would generate more return than the same dollar spent on model evaluation tooling, post-training capability, or internal research headcount with the depth to track and implement algorithmic advances. For most enterprise teams, the honest answer is that the marginal return on data infrastructure has declined relative to the marginal return on model agility.

The practical test is straightforward. If your team cannot currently evaluate whether a new open-weight model release outperforms your existing deployment on your own tasks within two weeks of release, you have an evaluation infrastructure gap. That gap costs more than the data pipeline ticket you are currently deprioritising.

Rebalancing Without Abandoning What Works

The rebalancing we are describing is not a pivot away from data discipline. Teams that have invested in clean, well-governed, domain-specific datasets retain a real advantage, particularly in regulated industries where data provenance and auditability are non-negotiable. The point is that data quality and governance are necessary but not sufficient conditions for AI advantage in 2026.

The teams that will compound their position over the next two to three years are those that pair strong data foundations with genuine algorithmic agility: the ability to evaluate new methods quickly, adopt post-training improvements without full retraining cycles, and build internal research capability that does not depend entirely on vendor roadmaps. That combination is harder to replicate than a data warehouse, which is precisely what makes it a more durable source of advantage.