Skip to content

Deal economics

Cheaper compute and data licensing economics: which datasets gain, which lose

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

Cheaper compute raises demand for data others cannot get and lowers the value of data anyone can get or generate. The compute needed for a given language model result halved roughly every eight months (Ho et al., 2024), Epoch AI estimates most of OpenAI's 2024 compute went to experiments, and open models train on far more tokens per parameter. Generic, easily imitated data loses differentiation; unique, outcome-linked and evaluation records gain.

Value drivers

Cheaper compute data

This page covers cheaper compute and data licensing economics: which datasets gain, which lose.Illustrative

Key takeaways

  • Cheaper compute means cheaper capability per dollar, not smaller budgets: Epoch AI estimates frontier final-run costs grew about 2.4x a year since 2016 even as unit costs fell.
  • More experiments consume more distinct data: Epoch AI reports open-weight models' tokens per active parameter grew 3.1x a year since 2022.
  • Repetition has limits: up to about four epochs of repeated data is nearly as good as fresh data, after which returns fade, so unique records become the binding input for longer runs.
  • Cheap inference makes synthetic data cheaper too, which erodes the value of any dataset a capable model can imitate.
  • The exposure matrix scores six dataset classes: evaluation sets and outcome-linked operating records gain most, generic text loses most.

Does cheaper compute help or hurt data licensing?

Both, and the split runs by dataset class. Cheaper compute lets developers run more training and evaluation experiments, which consumes more distinct data, and it also makes synthetic data cheaper to produce, which erodes the value of any data a model can imitate. The investor decision is which licensed datasets land on the winning side of those two forces.

A useful answer needs evidence on three questions: how fast the cost of capability is falling, where the extra compute goes and how far synthetic or repeated data can stand in for new records. Each has dated public evidence. Turning that evidence into a forecast needs a framework, signposts with dates and a clear statement of what would prove the forecast wrong, which is what the rest of this article supplies.

What a lower-compute-cost world means, in numbers

It means each unit of capability costs less while total spending at the frontier keeps rising. Reading cheaper compute as smaller budgets misreads the evidence.

Unit costs are falling fast. Ho and colleagues estimated in 2024 that the compute needed to reach a fixed language model performance halved about every eight months between 2012 and 2023, with a 95 percent confidence interval of roughly 5 to 14 months. Epoch AI found in March 2025 that the price of reaching a given score on six benchmarks fell by between 9x and 900x per year depending on the threshold, with a median of 50x, while cautioning that the fastest declines were recent and less certain to persist. DeepSeek's V3 technical report (December 2024) put its full official training run at 2.788 million H800 GPU hours, which it converted to $5.576 million at an assumed $2 per GPU hour, excluding earlier research and ablation runs.

Total spending is still climbing. Epoch AI reported in May 2024 that training compute for frontier models had grown about 4-5x per year, and its cost study estimated that the amortized hardware and energy cost of frontier final training runs grew about 2.4x per year since 2016. Meanwhile cheaper adaptation spreads training down the market: the LoRA paper reported that, compared with full fine-tuning of GPT-3 175B, its method cut trainable parameters by 10,000 times and GPU memory requirements by three times.

Force one: more experiments burn through more distinct data

When compute gets cheaper, developers run more experiments, and every experiment needs data the model has not already absorbed. That is the demand side of the thesis.

Most compute already goes to experiments rather than final models. Epoch AI estimated that OpenAI spent about $5 billion on R&D compute in 2024 and that the majority went to research, experimental training runs and unreleased models rather than the final training runs of released models. It is an estimate built from reported spending, not a company disclosure, but it shows where a falling compute price lands first.

Models are also consuming more data per parameter. Epoch AI reported that the ratio of training tokens to active parameters in open-weight language models grew about 3.1x per year since 2022, from roughly 10 in 2022 to about 300 in 2025, around 20 times the ratio implied by the 2022 Chinchilla scaling results. Epoch suggests one reason is economic: smaller models trained on more data can match performance with fewer parameters and cost less to serve.

Repeating data stops helping after a few passes. Muennighoff and colleagues (2023) trained more than 400 models and found that up to four epochs of repeated data changed loss negligibly compared with unique data, but that with more repetition the value of added compute eventually decays to zero. NVIDIA's Nemotron training FAQ cites the same constraint: more unique tokens let its Nemotron-H models train over long horizons without exceeding 4 to 8 epochs, and higher epoch counts brought diminishing returns. Cheaper compute pushes developers toward that ceiling faster, which makes new, distinct records the binding input.

Evaluation scales with experiments too. Anthropic's January 2026 engineering guide to agent evaluations says 20 to 50 simple tasks drawn from real failures are a good start, and warns that an eval at 100 percent tracks regressions but gives no signal for improvement. More experiments mean more evaluation runs and faster saturation, so fresh tasks drawn from real work are used up sooner.

Force two: cheap compute also makes data easier to imitate

The same falling costs make synthetic data cheaper, so any dataset a capable model can imitate loses differentiation. This is the supply-side risk, and it is already visible in disclosed training recipes.

NVIDIA's Nemotron FAQ says synthetic data goes into both pre-training and post-training: its web corpus includes synthetic counterparts of the higher-quality crawl tiers made by rephrasing filtered documents, raw Python source rephrased by a model and specialized synthetic sets for areas such as STEM reasoning. As inference prices fall, generating such data gets cheaper, and the public stock matters less as a ceiling. Epoch AI's 2024 projection that models would fully use the roughly 300 trillion tokens of public human text between 2026 and 2032 is a constraint that cheap generation partly relaxes.

Solved problems lose value fastest. The same FAQ says reinforcement learning prompts that the starting checkpoint already solves with a 100 percent pass rate are dropped. Data that teaches a model what it can already do is worth little, whatever it cost to collect.

Yet the scarce input is still data, not compute. Asked what matters most among dataset, architecture and methodology, the Nemotron teams answer the dataset, adding that a perfect methodology and architecture cannot overcome poor data. Cheap compute amplifies the advantage of whoever holds data a teacher model cannot imitate: records of real decisions, outcomes and failures in workflows that are not on the public web.

The compute-cost exposure matrix for six classes of licensed data

Score each dataset class on both forces and the net effect becomes visible. Scores run from -2 to +2 and are editorial judgments made on October 26, 2026, not measurements; the last column names the evidence that would change each score.

  • The pattern is consistent: cheaper compute rewards data that is real, unseen and tied to outcomes, and punishes data that is abundant or easy to imitate.
  • A licensing business's exposure is roughly the revenue-weighted average of its rows, so ask for revenue by dataset class before accepting any thesis about compute.
The compute-cost exposure matrix for six classes of licensed data
Dataset classExperiment demand effectImitation risk from cheap generationNet scoreSignpost that would change it
Generic public-style textSmall: supply is already abundantHigh: rephrasing and filtering substitute-2Licensors of text disclose higher renewal terms
Domain documents such as manuals and policiesModerate: useful for adaptationHigh once a model has read similar documents-1Fine-tuning recipes stop relying on generated domain text
Expert judgments and preference labelsHigh: every post-training run needs themMedium: model graders replace some labels+1Technical reports shift preference data from people to model graders
Operational records with outcomes, such as tickets, claims and dispatch logsHigh: agents need real decisions and resultsLow: outcomes are not on the public web+2Synthetic workflow data matches real records on held-out tests
Evaluation and holdout task setsHigh: more runs and faster saturationLow: tests must stay unseen and real+2Synthetic tests become trusted for release decisions
Refreshing records of new periodsHigh: each run wants recent dataLow: new periods cannot be generated in advance+1Published deals stop paying for recency
A diverging bar chart of editorial net scores from -2 to +2 for six licensed dataset classes under cheaper compute: generic text -2, domain documents -1, expert labels +1, operating records +2, evaluation sets +2 and fresh records +1.
Compute-cost exposure by dataset class. Illustrative: editorial scores as of October 26, 2026, not measurements.

Six signposts to score this forecast by the end of 2027

The forecast is that cheaper compute widens the value gap between imitable and non-imitable data. These signposts would confirm or refute it. Leans are editorial and all six were open on October 26, 2026; each carries the end of 2027 as its date, the evidence that would confirm or refute it and where to check.

  • Signposts 3 and 5 pull against each other: rising synthetic shares alongside slowing price declines would mean imitation is driven by quality rather than cost, which would hurt generic data further while leaving outcome-linked records intact.
  • Signpost 4 is the one most likely to overturn the matrix: if models can reuse data usefully far beyond 8 epochs, unique records lose much of their premium.
Six signposts to score this forecast by the end of 2027
SignpostLean on 2026-10-26Would confirmWould refuteWhere to check
Tokens per parameter keep rising in open-weight modelsLikelyEpoch AI's tracking shows the ratio still growing in 2027 modelsThe ratio flattens or fallsEpoch AI data insights
Experiments stay the majority of lab computeLikelyNew estimates still show final runs as a minority of R&D computeFinal runs take most R&D computeEpoch AI estimates and company disclosures
Synthetic shares grow in disclosed training mixesLikelyTechnical reports show larger synthetic or rephrased sharesShares stay flat or fallOpen model technical reports
Useful repetition stays near 4 to 8 epochsToss-upNew scaling studies confirm returns fading after a few epochsStudies show strong returns far beyond 8 epochsPeer-reviewed and preprint scaling studies
Inference price declines slowToss-upMeasured declines fall well below the 2025 ratesDeclines stay at or above the 2025 ratesEpoch AI price trend updates
Evaluation built from real work becomes a purchase lineLikelyDevelopers publish evaluations built from professional work or buy private test setsEvaluations stay public and syntheticDeveloper engineering posts and procurement pages

Where the cheaper-compute thesis could fail

Five developments could break the argument, and each deserves a discount in position sizing.

  • Synthetic data could learn operational behavior. If teacher models trained on public and first-party data generate workflow records that pass held-out tests built from real records, the +2 rows fall.
  • Cost declines may not persist. Epoch AI notes that the fastest inference price declines were recent and less certain to continue; if compute stops getting cheaper, the experiment multiplier shrinks with it.
  • Frontier buyers may concentrate. With frontier final-run costs growing about 2.4x a year, per Epoch AI, fewer organizations can train at the frontier, so rising data needs could flow through a handful of buyers with strong bargaining power.
  • Data efficiency may improve. Research that lets models learn more from each token, or reuse data well beyond a few epochs, would cut demand for unique records even as compute gets cheaper.
  • Rights may bind first. More demand does not create permission; if owners will not license, cheaper compute raises scarcity without raising licensed volume.

Questions for a data licensing business when training gets cheap

Ask where revenue sits in the exposure matrix and how each row holds as experiments multiply. The same questions apply to SourceX.

  • What share of licensed revenue comes from datasets a current model could generate at comparable quality?
  • Which datasets include outcomes, decisions or failures that are absent from public sources?
  • How often do buyers return for refreshed periods or new test sets, and on what terms?
  • Do buyers test the data on held-out real tasks before accepting it, and what share passes?
  • How would revenue change if a buyer could run ten times as many experiments on the same budget?

What the two forces mean for SourceX

SourceX is a labeled example of the question, not evidence of an outcome. For businesses that license selected operational records to AI developers, it does the work around the data: finding and qualifying it, reviewing rights, contracting, preparing it, delivering it and handling payment. It holds no ready-made stock; datasets are sourced from U.S. businesses against a buyer's request, and each company approves the scope and terms and signs before data is delivered.

That puts its exposure in the operational-records rows of the matrix, which is a thesis to test, not a result. SourceX has not published demand, renewal or pricing metrics; it provides company and commercial metrics to investors one to one. Under Rule 506(c), it is raising from accredited investors; investor materials can be requested through its investor page.

Sources, scoring rules and what was not measured

The two-force framework and the exposure matrix are editorial. Each dataset class was scored from -2 to +2 on the net of experiment demand and imitation risk, using the evidence cited in this article, on October 26, 2026.

Sources checked that day: Epoch AI's publications on frontier compute growth, training costs, inference prices, OpenAI's compute allocation, tokens per parameter and the public data stock; the algorithmic progress, DeepSeek-V3, data-constrained scaling and LoRA papers on arXiv; NVIDIA's Nemotron training FAQ; and Anthropic's guide to agent evaluations. Epoch's OpenAI figures are estimates, not company disclosures.

Not measured: prices of licensed data, buyer budgets, how strongly experiment counts respond to compute prices and any SourceX metric. Each signpost is dated to December 31, 2027, and its lean is stated so a reader can set it beside the outcome.

Frequently asked questions

Does cheaper compute reduce the value of training data?

It reduces the value of data that is abundant or easy to generate and raises the value of data that is unique and tied to real outcomes. Cheaper compute means more experiments, and Epoch AI reports open-weight models training on far more tokens per parameter, which increases demand for distinct records.

How many times can a model reuse the same training data?

Muennighoff and colleagues found that up to about four epochs of repeated data performs almost as well as unique data, after which returns fade toward zero. NVIDIA's Nemotron documentation describes a similar 4 to 8 epoch range. Beyond that, longer training needs new data.

Will cheaper fine-tuning create more buyers of licensed data?

It lowers the cost of entry: LoRA cut trainable parameters by 10,000 times compared with full fine-tuning of GPT-3 175B. More organizations can then adapt models to their own domains, and that adaptation needs domain records. Whether they license outside data or use their own is still an open question.

Why does most AI compute go to experiments rather than final models?

Developers test architectures, data mixes and recipes before committing to a final run. Epoch AI estimates that most of OpenAI's 2024 R&D compute went to research, experimental and unreleased runs rather than final runs of released models. Cheaper compute makes those experiments easier to run, and each one needs data.

Which licensed datasets hold value as synthetic data gets cheaper?

Records of real decisions and outcomes that are not on the public web, held-out evaluation sets that must stay unseen, and refreshing records of new periods. Generic text and documents similar to public material are the most exposed to cheap generation.

Sources

  • Ho et al. (2024) estimate the compute needed for a fixed language model performance halved about every eight months from 2012 to 2023. Source
  • Epoch AI (March 2025) measured yearly price declines of 9x to 900x for reaching given benchmark scores, median 50x, with the fastest declines recent. Source
  • DeepSeek-V3's report puts full training at 2.788M H800 GPU hours, or $5.576M at an assumed $2 per GPU hour, excluding prior research. Source
  • Epoch AI reported in May 2024 that frontier training compute grew about 4-5x per year. Source
  • Epoch AI estimates the amortized hardware and energy cost of frontier final training runs grew about 2.4x per year since 2016. Source
  • LoRA cut trainable parameters by 10,000 times and GPU memory by three times versus full fine-tuning of GPT-3 175B. Source
  • Epoch AI estimates most of OpenAI's roughly $5 billion 2024 R&D compute went to research, experimental and unreleased runs, not final runs of released models. Source
  • Epoch AI reports tokens per active parameter in open-weight LLMs grew 3.1x per year since 2022, from about 10 to about 300 by 2025. Source
  • Muennighoff et al. found up to four epochs of repeated data nearly as good as unique data, with the value of added compute eventually decaying to zero. Source
  • NVIDIA's Nemotron training FAQ documents the 4 to 8 epoch limit, synthetic data in pre- and post-training, dropping solved RL prompts and that data matters most. Source
  • Anthropic's January 2026 eval guide recommends starting with 20 to 50 tasks from real failures and notes a 100 percent eval gives no signal for improvement. Source
  • Epoch AI projected models would fully use about 300 trillion tokens of public human text between 2026 and 2032 if trends continue. Source

Request SourceX investor materials

SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.

See if your company qualifies