CRO fee-for-service and robotic wet labs.One brief. The right lab.
Spark Lab Network maps who can run which experiment, at what price, how fast, and how it can be ordered. It covers contract research providers that work per sample or per study, and the automated and robotic labs that are starting to sell results instead of machine time. We use the map to route a single experimental brief to vetted providers and return comparable quotes and structured data.
Order an experiment. We find the lab, normalise the quote and check the data.
Send one brief. We match it against the map below, shortlist vetted providers in the US, China or Europe, and come back with one normalised quote. Scoping and quotes are free. When you go ahead, you pay the provider price plus a 5% Spark service fee, shown as its own line on the quote; nothing is purchased until you approve it in writing.
Provider counts come from the directory on this page. Turnaround ranges are typical for standard formats and are confirmed on each quote.
Three ways to order
Reference
We reply from [email protected] within one business day. Keep the reference for every follow-up.
Prefer a conversation first? Send a direct message to @jimmy_build2026 on X with the short brief below. We reply in the thread, and move to email when there is something to quote.
Give your AI agent a skill that interviews you, builds a schema-valid request and, once you approve it, submits it to our order API (POST /api/labs/order). It works with any agent that reads Markdown skills or instructions, such as Claude Code, Codex, OpenClaw or a custom agent.
Install the skill at https://sparkscience.xyz/labs/agent/SKILL.md and use it to order my experiment from Spark Lab Network.Or paste the full skill into your agent's skills folder as spark-lab-order/SKILL.md. It validates against the order schema (spark-lab-order/v1).
Loading skill…
What happens after you send it
- ReceivedAcknowledgement quoting your reference.Instant reference · reply within 1 business day
- TriageFeasibility check, follow-up questions, NDA if you asked for one.1–2 business days
- MatchTwo or three vetted providers per stage, with a second source for the key readout.Target 3 business days
- QuoteOne normalised quote: price per sample, turnaround, controls, deliverables and data format.You approve in writing
- RunMaterials logistics, provider order, weekly milestone updates against the plan.Per quote
- Check & deliverControls and reference samples checked before release; raw data and report with provenance.Before sign-off
What we ask for, and why
- Objective and decision: decides the assay format, and whether a cheaper screen is enough
- Samples, materials and modality: decide which providers can take the work, and the price band
- Budget, deadline and regions: decide the shortlist and the logistics route
- Confidentiality and flags: decide the NDA, biosafety, animal-welfare and export checks before anything is shared
Ground rules
- Requests and quotes are free; approved work is billed at the provider price plus a 5% Spark service fee, itemised on the quote
- Confidential details move only after an NDA; we never forward a brief marked strict
- We do not take GLP studies, clinical or human-subject work, select agents, or anything we cannot verify as lawful in every jurisdiction involved
- Animal work goes only to providers with recognised welfare oversight
Six things the map tells us
A summary of what we found across — fee-for-service provider profiles and — robotic and automated labs. Hover any chart for detail; the full data sits in the directory below.
- Supply is deep, ordering is shallow. Every preclinical stage has six or more providers in both the US and China, yet only — of mapped providers take orders online, from a catalogue or by API.
- Software ordering barely exists yet. Only a handful of providers, led by Adaptyv and Ginkgo ADME-One, accept orders by API, and marketplaces such as Scientist.com are just adding agent ordering. China is strong in catalogue e-commerce (— of its providers, against — in the US) but has no API ordering for services.
- Transparency ends where molecules end. Genes, building blocks and standard mice have list prices. Cell function, safety panels, in-vivo efficacy and PK are almost entirely quote- or project-based.
- Small buyers pay up to 2.2× for speed. Fast per-order work costs a one-person company about 2.2× the standard rate, falling to 1.2× for large pharma. Pooling demand onto a standard menu is the fix.
- A standard in-vitro cascade takes about 8–10 weeks. From molecules through binding, developability, ADME and panels, it can be assembled today from at least two providers per stage. Target-specific cell assays and orchestration are the gaps.
- Robotic labs win by selling results. — of the tracked labs sit at the reconfigurable L2.5 layer. A US work cell beats manual labour at about 42% utilisation; in China, automation pays off only around the clock or per sample.
Providers mapped per stage
Count of provider profiles by headquarters region
How each market takes orders
Share of providers by ordering mode, from most to least programmable
The small-buyer premium
Fast per-order price as a multiple of standard per-order price, by client size (Spark cost model)
When a robot beats a technician
Cost per 96-well plate vs work-cell utilisation (Spark unit-economics model)
A standard in-vitro cascade
Indicative weeks from order, assembled from fee-for-service providers
What you can buy from robotic labs
Tracked labs by access model, split by automation layer
Supply is not scarce. Ordering it is.
Every stage of preclinical outsourcing has at least six providers in both the US and China. What a researcher, a small biotech or an AI agent actually lacks is a way to find the right one, compare like with like, and place the order without weeks of email.
Prices are mostly hidden
Public per-sample prices exist for gene synthesis, building blocks, standard mice and a few new ADME and binding packages. Cell function, safety panels, in-vivo efficacy and PK are almost all quote-only. In China, almost no provider publishes a per-sample price.
Small buyers pay the most
Our cost model puts the premium for fast per-order work over standard per-order work at about 2.2× for a one-person company, 1.8× at Series A, 1.4× mid-size and 1.2× for large pharma. A one-person company that buys an FTE contract uses only about 19% of it, which makes FTE dearer still.
Robots sold time, not results
Three generations of cloud labs asked customers to program the lab themselves. The models growing fastest today sell a fixed result per sample, in days, at a list price.
A standard in-vitro cascade can already be assembled from fee-for-service providers
From molecule supply through binding, developability, ADME and selectivity panels, every stage has at least two providers you can order from per sample or per panel. A clean run takes about 8–10 weeks end to end. The missing pieces are target-specific cell-function assays and the orchestration layer that stitches providers together.
Price transparency by stage
Share of mapped providers in each stage that take orders online, from a catalogue or by API rather than by quote.
What stitching providers together really costs
- Time stacks. Materials take 2–4 weeks; binding, developability and ADME can run in parallel; one late provider delays the whole cascade.
- Paperwork multiplies. Every extra stage adds a shipment and often a CDA or MTA.
- Project management is per order. About three hours of coordination per order in our model, so small, fragmented orders carry the highest overhead.
- Data does not line up by default. Merging results from a catalogue provider and a traditional CRO needs shared reference samples and an agreed report format, or rankings from screening will not hold on re-test.
Fee-for-service providers
Who does what, how it is ordered, and the public price or turnaround where one exists. Figures are as reported or listed in October 2026. Inclusion is not an endorsement, and we publish no contact details: introductions go through Spark Science.
API programmatic submissionOnline cart or instant quoteCatalogue list priceQuote request for quoteProject scoped study or FTE° background profile, not re-checked this round
| Provider | Region | Stage | What they do | Ordering | Public price · turnaround | Scale |
|---|
Automated and robotic labs, by what you can actually buy
We grade labs by changeover cost, not by robot count: how expensive it is to switch the same hardware to a different experiment. Almost every "autonomous lab" in 2026 sits at L2.5. No provider has yet published unit costs or FTE replacement for general-purpose work.
| Lab | Region | Layer | How you access it | Public evidence | Spark read |
|---|
What sells
Fixed price per sample, days to result, machine-readable output. Adaptyv grew revenue about tenfold in 18 months on that model; Ginkgo's $199 ADME panel signed 17 customers in six weeks.
What does not, yet
Renting robot time to customers who must write the protocol. General-purpose arms are still far from the ~99.9% per-step reliability a 50-step protocol needs; imitation learning reports 65–75% in simulation.
Where China stands
Large integrators and captive robot fleets, but no external service desk, no lab-OS layer independent of hardware, and no public API or MCP ordering endpoint. Labour at about a third of US cost means single-shift automation never breaks even; only 24/7 or per-sample models work.
When robots beat people, and when they do not
From our in-house model of a 96-well plate assay, with every assumption exposed and stress-tested across five scenarios.
| Manual · US | Manual · China | Automated work cell · US | Automated work cell · China | |
|---|---|---|---|---|
| Annual fixed cost | 1 RA ≈ $140k | 1 technician ≈ $37k | ≈ $675k (depreciation, service, software, 0.5 engineer, space) | ≈ $575k |
| Cost per plate at 60% utilisation | $64 | $17 | $45 | $38 |
| Cost per plate at 20% utilisation | — | — | $135 | $115 |
| Break-even utilisation vs manual | — | — | ≈ 42% | ≈ 135%: not reachable on one shift |
The lever is first-pass success
A master's-level lab technician in China costs about RMB 350k a year fully loaded, for about 1,050 hours of bench execution. A robot replacing one FTE today costs RMB 1.4–2.0M. Success rate per full run, and the engineer-to-robot ratio, move the answer more than capex does.
Where robots enter first
Plate handling, doors and lids, and structured capping already run at 88–100% in operator-level tests. Small tubes, VLA-driven capping, powder dosing and buttons sit at 4–55%. Start with loading and unloading, then feed liquid-handling stations, then add vision monitoring.
Why aggregation matters
The only viable path for fast per-order work is pooling many small clients onto a standard menu, adding capacity subscriptions, and taking orders through an API or agent. That is the job this network is built to do.
From one brief to comparable results
The network runs as a managed service. People order through a form or on X; AI agents order through a published skill and schema. Automatic pricing follows once the menu is standard enough.
Built for
- AI-native biotechs and protein-design teams that need fast, standard readouts
- Academic groups without in-house ADME, panels or vivarium access
- AI agents that need a lab they can call; see Research Agent
- Providers who want standard-format demand instead of one-off RFQs
What we do not do
- Run GLP studies or sign off regulatory packages
- Publish provider contacts or confidential price sheets
- Rank providers on undisclosed criteria; every number on this page carries its source tier
Research that feeds the network
The directory is the visible layer of a larger, continuously updated body of first-party research on outsourced and automated wet-lab work.
US–China fee-for-service census
About 150 providers across four segments, with order format, public price, turnaround and scale, compiled from filings, price lists and provider materials.
Who outsources, and how much
US venture-backed biotechs are the main buyers of China-based research services; for a preclinical small-molecule company, 15–30% of a round can go to Chinese CROs. We model what onshoring does to that budget.
In-vitro labour, globally
An estimated 260–380k people do industrial in-vitro work, 60–75k of them at CROs. Graduates are plentiful; DMPK method developers, automation engineers and AI-literate scientists are not.
Three-stage lab framework
A five-layer stack (action, instrument, orchestration, intent, closed loop) to place every robotic lab, with an eight-condition test for when intelligent automation beats people.
Robot-versus-technician model
A fully linked model with five scenarios, closed-form break-even utilisation and operator-level success rates checked over three review rounds.
Machine-orderable windows
Which assays are already orderable by software, which are productisable next (in-vitro safety panels, structure determination, fixed-protocol rodent PK), and which should stay human.
Method notes. Provider data was compiled in October 2026 from annual reports, exchange filings, public price pages, provider brochures and press releases. Entries marked ° come from background knowledge and were not re-checked in the latest round. Third-party database figures often conflict; where they did, we kept the primary source or left the field blank. Public prices are list or promotional prices and change often. Names and trademarks belong to their owners; inclusion does not imply partnership or endorsement.
Have an experiment you would rather not manage?
The service is in early access. Send a brief through the order form, on X or from your agent; providers who want to join the network can write to us at [email protected].