Since joining node.vc, I’ve been thinking about ways for us to become a more data-driven investor and have been building a few tools. One of the first tools I built was something to keep track of Nordic launches on ProductHunt, and thought the dataset was so interesting I wanted to share a few of my findings from 2023.

In this post I’ll showcase the top 25 Nordic launches on ProductHunt in 2023. I also clustered product descriptions to give some insights into what kind of products were launched during the year, and which categories were most and least popular. Maybe this could even give a hint about what the startup ecosystem looked like in a broader sense?

The top 25 Nordic launches

First, a little disclaimer: my geographical classifications are not perfect. I’ve checked the top 25 so there are no false positives, but it’s not really viable to find all the false negatives. If you see a Nordic launch missing from the top 25 — let me know and I’ll update my data.

Without further ado, here are the top 25 Nordic launches in 2023:

(Table: Top 25 Nordic Launches on ProductHunt in 2023, upvotes as of 1st of Jan 2024)

Out of these 25 launches, 20 are from startups, four from indie makers and one marketing agency (Martekings).

The startups

The best-performing Nordic launch for 2023 was Swedish-founded networking app Anyone, allowing users to find interesting people all over the world for a quick five-minute phone call. Anyone raised a pre-seed early 2020 from Antler and a €3.6m seed from Cavalry, Supernode and Futurology Ventures in mid-2021.

Very impressive work by the Klu team to bring home both the second and third place. Klu offers a unified search bar and chat interface to find and talk to your organization’s data across apps like Slack, Notion, Github, Gmail and Google Calendar, and recently became SOC2 compliant. To my knowledge, they only have raised a pre-seed from Antler after starting out in late 2022.

Hypertype (pos 4), with its AI email assistant drafting replies using your internal data, alongside sales prospecting and automated outreach platform Zaplify (pos 6) also did very well.

Looking at themes: LLM-powered features, collaboration tools and tools geared towards other tech companies are frequent, not surprising as I think many in the ProductHunt audience are working in tech.

Searching for commonalities in the products: Klu, Hypertype, and Question Base (pos 14) are all touching on surfacing company knowledge in some form. Bucket (pos 22) and Delibr (pos 25) are both building tools for product managers; while Butter (pos 9) and Wudpecker (pos 11) are building meeting enhancement tools, which is also a popular type of product in ProductHunt in general.

There are also a few Nordic companies in the top 25 that have raised larger venture rounds. These are feature evaluation platform Bucket (pos 22) with $5.7m from Creandum and Project A; customer-support platform Fullview (pos 17) with $9.3m Lightspeed and Cherry; spreadsheet platform GRID (pos 15) with $16.5m from NEA, AceCap and Slack’s (now Salesforce’s) CVC arm; UI design tool Uizard (pos 8) with $18.6m from Insight Partners and byFounders; and last but not least Spotify (pos 23 and 24) with $2.1bn from Northzone, Creandum, Accel, Kleiner Perkins and Founders Fund among others.

The indie makers

Among the Nordic indie markers, Dan Mindru and Sandra Djajic (CMO at Klu by day!) are the uncrowned rulers. Dan built node.js boilerplate generator Shipixen (pos 7), and API speed and load tester Clobbr (pos 18) alone. He also launched, together with Sandra and fellow builder Alex Szczurek, ProductHunt launch insights platform Hunted Space (pos 5) and ascii chart generator MRRArt Pro (pos 13). To learn something from Dan & Sandra and keep up to date with the indie maker space, check out their podcast Morning Makers Show!

What kind of products were launched?

To answer this question, I encoded the product taglines and descriptions, clustered them, and used GPT-4 to assign a label the clusters. I ended up with 90 clusters with ~29% of launches as outliers. I describe the methodology in more detail at the end of the post.

A disclaimer here as well: My dataset contains all featured posts with >25 upvotes in 2023.

(Table: My 90 clusters and the three top performing launches in each cluster)

2023 was a year of AI

During 2023, quick progress in three kinds of AI models really started pushing the boundaries of what we thought possible. Transformers models, diffusion models and speech models (tts and stt) were definitely around in 2022, but felt more like proof-of-concepts. In 2023, we saw ChatGPT demonstrating the power of transformers, reaching 100m users in January after two months. We saw diffusion models like Dall-E 3 and Midjourney v5 showing the power of image generation, and Whisper v2 (December 2022) unlocked multi-lingual speech transcription. It really felt like magic and entirely new features and products became possible.

Unsurprisingly, the cluster GPT-Apps and Chatbots is by far the largest. Many of these products are quite simple, e.g. ChatGPT UI improvements, prompt templates, and ChatGPT for your website or Discord server. The products also include AI assistants like Collato; Zapier-style AI-enabled automation platforms like n8n and Relay; customer support chatbots Botsonic & Algomo, and support automation platform Desku; as well as a few foundational models.

The second largest cluster, AI Art and Photos, contains all the amazing products built on top of diffusion models, e.g. product visual generation platforms Blend, Pic Copilot and Mokker AI; LinkedIn headshot generators like InstaHeadshots or sketch-to-image product Freepik Pikaso, alongside the foundational models themselves. DALLE-3 topped the chart and with Midjourney v6 and v5 also performing well.

Another big AI cluster was Video Tools, with products primarily falling into three categories: tools for creating product demo and how-to videos such as Guidde AI, Focusee and Zing; platforms allowing podcasters, streamers or other creators to generate short social media-ready clips from long videos like Spikes Studio, Opus Clip or vizard; video summarization tools like Eightify and TubeOnAI. Other products include video editing tools like Videobolt, LLM-powered video script generators such as Maekersuite and Google’s foundational prompt-to-video model VideoPoet.

Other AI-fueled clusters include AI Writing Tools, Coding Copilots, Sales Copilots, AI-Assisted Website Builders, PDF Chat & Tools, AI Quiz & Flashcard Generators and AI Fitness Coaches. They are also scattered in clusters like CRM Tools with conversational CRM interface CRM Chat, SEO Tools with tools for generating SEO content like RivalFlow, and Product Feedback Tools with LLM-powered insights extraction platforms Cycle and Enterpret.

For 2024, performance in these models will continue to improve — allowing even more new use-cases and a higher confidence in the output (maybe thanks to guardrail products). For me, I think prompt-to-video will have its POC-to-magic moment in 2024, perhaps pioneered by Runway ML?

The popularity of clusters

The cluster with the highest median upvotes was, perhaps unsurprisingly, resources for launching on ProductHunt and analytics led by Nordic Hunted Space mentioned above, followed by a cluster with marketing prompts and resources. Animation Tools was also a top-performing cluster, with the four top products relating to the Lottie animation file-format, such as Lottielab and Jitter, and Sales Copilots was led by digital sales room tools Flowla and trumpet with a tail of copilots like Winn.

A few devtool clusters also performed well: Analytics & Database Tools with Airbook for quickly building Jupyter Notebook-style reports from 150+ data sources, the super-simple database App Backend and a few “chat with your SQL database” products like BlazeSQL and AskYourDatabase; Simple Deployment Platforms with ambitious self-serve devops platforms such as Humalect and Codesphere, plural helping engineers manage Kubernetes clusters, and full prompt-to-backend & no-code editor Backender; and Webdev Platforms & Resources with e.g. UI library Pines and React framework Refine.

The link-in-bio platforms are another top performer, used to create link trees from social media bios, with Sequoia-backed German Bento performing best, and actually being acquired by Linktree a few months after its ProductHunt launch.

Shifting focus to the least popular clusters, it’s interesting that the only two EdTech related clusters, AI Quiz & Flashcard Generators and Language Learning Apps, perform poorly. The same seems to hold true for other direct-to-consumer clusters with products for personal use like Personal Finance Tools, Notetaking Apps, Productivity and Time Trackers, Productivity and Focus Tools, and Relaxation and Sleep Aids.

Reasons explaining this could be that these areas are more crowded or that there are fewer startups in the clusters, meaning less launch marketing. It’s also possible that the ProductHunt audience prefers tools for building and running a business over tools for personal growth and productivity.

Notable mentions

I saw a lot of buzz around the side hustle of creating children’s books using ChatGPT and Midjourney to sell on Amazon, and thought the AI Kids Story Generators cluster was interesting: check out top-rated Oscar Stories!

Another interesting, and less popular, cluster is Dating Apps & Copilots, with AI-powered products Meet Mille helping users craft the perfect response and CharmCheck providing feedback on your dating profile.

“This is not legal advice” is a disclaimer you see a lot, but maybe not for the products in the AI Legal Assistance Tools. One example is LegalNOW offering a range of chatbots, each specialized in providing advice in consumer law, taxation and corporate governance. These kinds of products, alongside the AI Therapists and Doctors cluster, really have to innovate with safeguards against hallucinations and output monitoring.

A few notes on methodology

To gather data on launches, I used the ProductHunt API which is free. Check it out!

To create the clusters I embedded product taglines and descriptions, and ran them through a pipeline of dimensionality reduction (tough to do high-dimensional clustering as distance metrics break down) and clustering. I finetuned the hyperparameters in the pipeline using the share of outliers as my loss function, while keeping a reasonable number of clusters. Finally, I fed the 10 most central products in each cluster into a few-shot prompt to GPT-4 to assign labels, and manually changed the labels that didn’t make sense.

I ended up using all-mpnet-base-v2 as my embedding model after trying a few different ones, UMAP for dimensionality reduction and HDBSCAN for clustering.

Closing remarks

If you are a founder building something big and disruptive in the Nordics, and want a pre-seed or seed from a VC run by ex-founders and operators with technical DNA, please drop us a line!

Feel free to ping me and the team on LinkedIn, send us your pitchdeck on our website, or follow us on LinkedIn :)