MTS

Startup Trends · YC vocabulary 2025–2026

What 1,000 founders are talking about.

We read every one-liner and long description for every company in YC's last four batches (Winter 2025 through Winter 2026), then tracked how the language has shifted. What's surging, what's collapsing, and what's conspicuously absent.

Y CombinatorBrowse YC's job boardSpeedrun (a16z)Join the 1337 Talent Network
682Companies analyzed667 with usable text
4YC batches coveredW25 · S25 · F25 · W26
agenticFastest-rising word3.1% → 8.0% over the year
voiceFastest-falling word9.2% → 3.5% — voice peaked W25
zero"world model" mentionsacross all 4 batches

What's surging16 terms

Each card shows a word's trajectory across the last four YC batches. The bigger the swing up and to the right, the more this concept has taken over founder vocabulary.

infrastructure+9.4pp
0.0%5.0%10%15%20%W25S25F25W266.1%8.0%9.1%15.6%
ai-native+6.6pp
0.0%2.5%5.0%7.5%10%W25S25F25W262.5%5.6%8.4%9.0%
agentic+5.0pp
0.0%2.5%5.0%7.5%10%W25S25F25W263.1%5.6%3.5%8.0%
physical+4.2pp
0.0%1.9%3.8%5.6%7.5%W25S25F25W261.8%1.2%2.8%6.0%
autonomous+3.7pp
0.0%2.5%5.0%7.5%10%W25S25F25W264.3%3.7%9.8%8.0%
robotics+3.3pp
0.0%1.3%2.5%3.8%5.0%W25S25F25W261.2%3.1%3.5%4.5%
agent+3.2pp
0.0%3.8%7.5%11%15%W25S25F25W269.8%14.2%12.6%13.1%
hardware+3.0pp
0.0%1.9%3.8%5.6%7.5%W25S25F25W263.1%6.2%2.8%6.0%
context+2.9pp
0.0%2.5%5.0%7.5%10%W25S25F25W263.7%6.2%9.1%6.5%
coding agents+2.5pp
0.0%1.3%2.5%3.8%5.0%W25S25F25W260.0%0.6%4.2%2.5%
agents+2.5pp
0.0%10%20%30%40%W25S25F25W2625.2%32.7%37.8%27.6%
legal+2.4pp
0.0%1.3%2.5%3.8%5.0%W25S25F25W260.6%1.9%0.7%3.0%

What's collapsing6 terms

Words that defined an earlier batch and have since fallen out of favor. The voice-agent wave broke in Winter 2025 and hasn't recovered. Compliance and call-center pitches are gone too.

voice-5.7pp
0.0%2.5%5.0%7.5%10%W25S25F25W269.2%8.0%6.3%3.5%
model-3.6pp
0.0%2.5%5.0%7.5%10%W25S25F25W268.6%6.8%3.5%5.0%
compliance-3.1pp
0.0%1.9%3.8%5.6%7.5%W25S25F25W266.1%6.2%4.2%3.0%
copilot-2.6pp
0.0%1.3%2.5%3.8%5.0%W25S25F25W263.1%0.6%2.1%0.5%
open-source-2.1pp
0.0%1.3%2.5%3.8%5.0%W25S25F25W263.1%4.9%1.4%1.0%
voice agents-2.1pp
0.0%1.3%2.5%3.8%5.0%W25S25F25W263.1%1.9%1.4%1.0%

The full-year shiftW25 → W26

The biggest risers and fallers comparing the oldest batch (Winter 2025) to the newest (Winter 2026). All values are share of companies that use the word at least once in their description.

Biggest risers W25 → W26
scale+10.6pp2.5% → 13.1%
infrastructure+9.4pp6.1% → 15.6%
models+8.9pp9.2% → 18.1%
built+7.9pp9.2% → 17.1%
data+7.0pp19.6% → 26.6%
real+6.9pp6.1% → 13.1%
hours+6.9pp3.7% → 10.6%
ai-native+6.6pp2.5% → 9.0%
learn+5.8pp1.2% → 7.0%
starting+5.4pp3.7% → 9.0%
production+5.3pp1.2% → 6.5%
without+5.3pp4.3% → 9.5%
Biggest fallers W25 → W26
voice-5.7pp9.2% → 3.5%
integration-4.3pp4.3% → 0.0%
revenue-4.1pp8.6% → 4.5%
answers-3.7pp3.7% → 0.0%
eliminating-3.7pp3.7% → 0.0%
contact-3.7pp3.7% → 0.0%
model-3.6pp8.6% → 5.0%
knowledge-3.5pp5.5% → 2.0%
able-3.3pp4.3% → 1.0%
reports-3.2pp3.7% → 0.5%
reps-3.2pp3.7% → 0.5%
enterprise-3.1pp6.1% → 3.0%

Each batch's signaturedistinctive words per batch

The words most over-represented in this batch compared to the other three. Lift score = percentage-point gap between this batch and the rest. Think of it as 'what would surprise you about this cohort if you only read the other three.'

Winter 2025
168 cos · 163 w/ text
mission+4.2integration+3.7powered+3.5unique+3.5voice+3.4model+3.4able+3.3reps+3.3revenue+3.2trained+3.1eliminating+3.1quality+2.9
Summer 2025
167 cos · 162 w/ text
built+6.9problem+5.3test+4.6tasks+4.6fix+4.4production+3.9deploy+3.8coding+3.8past+3.7web+3.6order+3.6scalable+3.5
Fall 2025
148 cos · 143 w/ text
agents+9.3already+6.1live+5.8run+5.8directly+5.4ai-powered+5.0fully+4.4start+4.3autonomous+4.3engineers+4.2personalized+4.0power+3.9
Winter 2026
199 cos · 199 w/ text
models+8.5infrastructure+7.9scale+6.7data+6.5starting+5.0hours+4.4custom+4.3physical+4.1manual+4.1agentic+4.0real+3.9automate+3.7

Two-word combinationsdistinctive bigrams

Phrases that are over-represented in a single batch. These tend to be more concrete than single words — actual product framings or recurring claims.

Winter 2025
168 cos · 163 w/ text
voice agents+1.7
Summer 2025
167 cos · 162 w/ text
open source+3.2agents automate+2.3
Fall 2025
148 cos · 143 w/ text
coding agents+3.1
Winter 2026
199 cos · 199 w/ text
data sources+1.9foundation models+1.3coding agents+1.0

Conspicuous absences17 terms

Words that AI Twitter argues about constantly but show up two or fewer times across all four YC batches combined. The gap between online discourse and what founders actually pitch is striking.

agent loop
0 mentions total
anthropic
1 mention total
browser use
0 mentions total
call center
1 mention total
chrome extension
1 mention total
deepseek
1 mention total
diffusion
2 mentions total
embedding
1 mention total
fine-tune
1 mention total
mcp server
0 mentions total
physical world
0 mentions total
pretraining
0 mentions total
regulated industries
0 mentions total
transformer
2 mentions total
vector
1 mention total
world model
0 mentions total
world models
0 mentions total
"world model" — the concept dominates AI research discourse but appears zero times across ~680 YC companies. "MCP server" hasn't entered the pitch vocabulary yet either. "OpenClaw", a brand-new product term, has 3 mentions — small but rising.

Methodology

Corpus

YC's public company directory for the Winter 2025, Summer 2025, Fall 2025, and Winter 2026 batches. For each company: one_liner + long_description concatenated, lowercased, stripped of punctuation, English stopwords removed.

Counts

Document-level frequency: a word counts once per company, regardless of how many times it appears in that company's text. All percentages are share-of-companies, not share-of-tokens.

Distinctive words

For each batch, we compute the percentage-point lift versus the other three batches combined. Words with a high lift are over-represented relative to neighboring cohorts. Minimum 5 mentions to make the cut.

Watch list

~80 hand-picked terms (agent, voice, world model, claude code, openclaw, agentic, etc.) get a per-batch trajectory line regardless of frequency, so we can see emerging or fading signals before they hit the auto-discovery threshold.

Caveats

  • YC only. This is YC's vocabulary, not all-of-startups vocabulary. Speedrun and Neo cohorts use slightly different language and are excluded here for cleaner cohort-on-cohort comparison.
  • Founders write to YC's audience. Descriptions are written for the YC directory and partners. Words like "claude code" or "openclaw" show up because they're a shorthand YC partners recognize.
  • Small batches. Each batch is 150-200 companies. A 5pp swing is meaningful; a 1-2pp swing is noise.
  • This is what they pitch, not what they build. Vocabulary shifts before tech does. A company saying "agentic" in 2026 may be doing the same thing a 2025 company called "automation".
Y CombinatorBrowse YC's job boardSpeedrun (a16z)Join the 1337 Talent Network