Tessera
Series A · 2026
Series A · 2026

Better data
beats more data

Tessera is the curation engine for training sets — dedupe, score, and compose corpora that make the same model 4–9 points better on the same compute.

68
ML teams
$3.8M
ARR
$16M
raising
01 / 20
The problem

Every lab knows data quality is the highest-leverage knob — and still curates with grep and vibes.

Web-scale corpora are 40% near-duplicates and noise. Compute is priced to the penny while the thing it trains on goes unmeasured.

02 / 20
Why now
4–9 pts

benchmark improvement from curation alone at fixed compute, replicated across three model families in our published evals (2025) — cheaper than any scaling law says compute can buy.

03 / 20
Product

A refinery between your crawl and your cluster

01
Semantic dedupe
Near-duplicate detection at 100B-document scale.
02
Quality scoring
Model-based raters tuned to YOUR downstream evals.
03
Mixture optimizer
Composes domain ratios against a target eval suite.
[ SCREEN — corpus composition ]
04 / 20
Pipeline
STEP 01
Ingest
Crawls, licensed sets, synthetic pools.
→
STEP 02
Score
Every document rated on 14 axes.
→
STEP 03
Compose
Optimizer builds the training mixture.
→
STEP 04
Trace
Full lineage per training token.
05 / 20
Traction
01
68
ML teams
02
$3.8M
ARR
03
2.1T
tokens curated to date
04
131%
net revenue retention

Two frontier labs run Tessera inside their own VPCs — the deals that anchor the enterprise tier.

06 / 20
Growth

ARR, trailing 8 quarters

Q3 ’24Q4Q1 ’25Q2Q3Q4Q1 ’26Q2
07 / 20
Unit economics

Pricing scales with tokens, margin scales with reuse

TierPriceGross margin
Starter (100B tokens)$4k/mo68%
Scale (1T tokens)$22k/mo79%
Enterprise VPC$390k/yr84%
Blended—78%
08 / 20
Market · Epoch AI 2025

Data ops: the unbudgeted 20% of every training run

0100%
14%
38%
48%
SOM — AI-native labs + model teams 14%
SAM — enterprise fine-tuning data ops 38%
TAM — training data tooling + services 48%
09 / 20
Competition
Tessera
Labeling vendors
OSS dedupe stacks
In-house scripts
Scale (tokens) → ↑ Eval-linked quality
10 / 20
Moat
01
01

Rater ensembles

Quality models distilled from 400+ customer eval suites — accuracy compounds cross-customer.

02
02

Scale engineering

100B-doc dedupe in 9 hours; rivals quote weeks.

03
03

Lineage standard

Our per-token provenance format is entering two compliance frameworks.

11 / 20
Customers

Teams training on Tessera-curated corpora

10 partners
[ logo 01 ]
[ logo 02 ]
[ logo 03 ]
[ logo 04 ]
[ logo 05 ]
[ logo 06 ]
[ logo 07 ]
[ logo 08 ]
[ logo 09 ]
[ logo 10 ]
12 / 20
“
Customer proof

“Same cluster, same architecture, Tessera mixture: +6.2 on our eval suite. That is a result we had budgeted eight figures of compute to reach.”

Dr. Hannah Cole
Head of Pretraining, applied-AI lab (400 GPUs)
13 / 20
Business model

Revenue per customer-year

Curation SaaS
+ Mixture optimizer
+ Lineage/compliance
Total
sources uses ending
14 / 20
plan
The plan

Own the data layer of the training stack

Compute got its platforms. Serving is getting its runtimes. Data — the highest-leverage layer — is still up for grabs.

15 / 20
Roadmap
Now
Enterprise VPC GA
From 2 design partners to 12 accounts.
H1 2027
Synthetic pipeline
Generation + verification in the same loop.
H2 2027
Multimodal curation
Image-text and audio corpora, same raters.
2028
Compliance suite
Provenance audits as regulation lands.
16 / 20
Team
[ photo ]
Dr. Felix Braun
CEO — ex-DeepMind data research
[ photo ]
Ines Costa
CTO — ex-Databricks Spark core
[ photo ]
Kenji Mori
Research — dataset papers, 8k citations
[ photo ]
Sasha Ivanova
GTM — ex-Scale AI enterprise
17 / 20
Financial plan

ARR with this round

$24M
ARR exiting 2028
25
50
75
100
$2.8M
$3.8M
$6.7M
$10.6M
$16M
$24M
H1 ’26H2 ’26H1 ’27H2 ’27H1 ’28H2 ’28
18 / 20
Use of funds

Deploying $16M over 24 months

0100%
46%
26%
28%
Research + engineering 46%
Enterprise GTM 26%
Compute + buffer 28%
19 / 20
Tessera

Train on signal.

felix@tessera.dev · tessera.dev
[ QR ]
20 / 20