← All tech signals
Infrastructure

Synthetic Data

When the real data runs dry, models start training on worlds they invent — reality reflected in a hall of mirrors.

What is it?

Labs are turning to machine-made training data because the internet’s stock of quality human text is nearing its limit. It sounds like a stopgap; inside the labs it now reads as standard engineering practice. Whether the practice holds up at frontier scale stays unproven.

Why now?

Scaling rewarded raw data volume, and the web’s free supply of quality text is not infinite. Cognitive abundance needs material to keep compounding, so labs manufacture what the web can no longer provide. A scarce input turns into a built one — the real question is how far that substitution stretches.

Who’s building it?

Frontier labs design and filter the synthetic data themselves — the substitute good stays in their own hands.

Frontier AI labsData-generation startupsCuration & filtering toolsEnterprise data holdersAI safety researchers
The full read
Why now, who is building it, where the market goes, and what stays unsettled.
Get weekly →
01 · The market
Who’s in it
Builders — frontier labs and the data-curation tooling beneath them.
Buyers — enterprise teams and smaller labs, buying curated sets instead of building their own.
Gatekeepers — publishers and regulators deciding what human data counts as consent-cleared.
What’s at stake
The efficiency gain is real, but so is the collapse risk — the labs racing fastest also build the failure mode.
Rights-cleared human data turns scarce, and valuable, fast.
Provenance blurs quietly — most readers never notice the difference.
What’s unsettled
Whether curated synthetic data actually prevents collapse, or just delays it.
How much of a model’s true knowledge is now second-hand.
Whether the scarcity clock is real, or a convenient story for data-licensing deals.
02 · How the demand matureda progression, not a timeline
Generation
Labs synthesize training material once the real supply runs thin.
Curation
Filtering and validation pipelines separate useful synthetic sets from drift-prone noise.
Discipline
Data design becomes a guarded core competency, with feedback-loop decay treated as a real operating risk.
Moat
Genuine, rights-cleared human data becomes the scarce input everyone competes to hold.
03 · Where the market goesthe forks, not a prediction
→Substitution or exhaustion — either labs keep manufacturing enough usable data to sustain scaling, or the real bottleneck reasserts itself and growth slows.
→Regulators and publishers may force real licensing deals before labs finish building around them entirely.
→Collapse risk could stay a manageable engineering problem, or turn into a hard ceiling on scale.
→Value could concentrate in whoever owns the last honest pools of human data.
The demand question underneath all of it is simple: does synthetic data genuinely extend real scaling, or just postpone the day the real bottleneck bites back?
04 · The AI Dev Process deskadjacent markets, moving now
The Trust Layer
Watermarking tags machine-made media at scale.
SynthID has watermarked over 100 billion images and videos.
Mashable
The Provenance Web
Platforms now label AI content at massive scale.
TikTok labeled over 3 billion AI-generated posts via C2PA.
C2PA
The Consent Layer
Sites set their own rules for how bots may use content.
Cloudflare lets sites set separate rules for search, agent and training bots.
Help Net Security
05 · The Wire
01
High-quality public text for AI training may run out within 6 years.
South China Morning Post (citing Epoch AI) ↗
02
Repeated training on synthetic data makes model performance degrade over generations
npj Artificial Intelligence ↗
The panel on this signal
They agreed on
–Manufacturing training data is now standard engineering practice, not a stopgap patch over a scaling wall.
–Genuine human data is becoming the real scarce input.
–Provenance is blurring quietly, and most ordinary people never notice it happening.
The ML systems engineer disagreed on
–The data-exhaustion timeline in circulation traces back to a single estimate from one source that has been widely repeated without independent verification; treating that projection as settled fact overstates its evidentiary basis.
Know more about the method →
✦How this signal was built
01 · Panel
A Delphi panel of 4 AI analysts, each reasoning independently.
→
02 · Debate
They challenge and revise each other’s reads over two rounds.
→
03 · Evidence
Every claim is tied to a cited, linked source on The Wire.
→
04 · Synthesis
A moderator writes it into one read, disagreement and all.