Hey — It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 5 most important things:

• Solid, the AWS of Fintech, has shut down — learn why below.

• A guide to disrupting incumbents.

• Startup funding hit records in Q1. But the outlook for 2025 is still awful.

• Nexad raised $6 million to put ads in your AI chats.

• OpenAI released two new reasoning models: o3 and o4-mini — learn more below.

Let’s get into it.

This Week In Startups

🔗 Resources

A guide to disrupting incumbents.

AI startups will thrive In adversity.

Data and Defensibility.

AI is making big tech even bigger.

📰 News

OpenAI is reportedly in talks to buy Windsurf for $3B.

Startup funding hit records in Q1. But the outlook for 2025 is still awful.

Anthropic’s Claude can now read your Gmail.

The most interesting startups showcased at Google Cloud Next.

💸 Fundraising

Nexad raised $6 million to put ads in your AI chats.

Marshmallow, the UK insurance startup for migrants, raised $90M.

Hammerspace, an unstructured data wrangler used by Meta, raises $100M.

Battery startup Nyobolt raises $30M.

Fail(St)ory

AWS of Fintech

This week, Solid, a fintech infrastructure startup once valued at $330 million, filed for bankruptcy.

It had promised to be the “AWS of fintech,” offering banking and payments via APIs so startups could launch their own financial products in weeks, not years. Now, it’s in court trying to sell or restructure what’s left of the business.

Let’s break down what went wrong.

What Was Solid:

Solid launched in 2018 with a compelling pitch: make fintech infrastructure as easy to plug in as cloud storage. 

Just like AWS helped developers spin up servers instantly, Solid wanted to let startups launch embedded financial services — think bank accounts, payment rails, debit cards — all through a few lines of code.

The idea wasn’t new, but Solid moved fast and made noise. By 2022, they said revenue had grown 10x, customers had doubled to 100, and the company was profitable. They’d raised $81M, landed FTV Capital as a backer, and hit a $330M valuation.

From the outside, things looked solid.

A year later, it was falling apart.

The Numbers:

• 📅 Founded in 2018

• 💰 Raised ~$81M in total

• 💸 Valued at $330M in 2022

• 🧾 100+ customers

Reasons for Failure: 

• Legal Battles With Investors: FTV Capital, their lead investor, sued the founders for fraud. They said the numbers were fake, churn was hidden, and the whole thing was misrepresented. They wanted their $61M back and the founders gone. Solid fought back, accusing FTV of using threats and pressure tactics once the investment stopped looking good. The case was settled in 2024. But by then, the damage was done. No one wants to invest in a company tangled in lawsuits.

• Fundraising Freeze: Solid hadn’t raised new capital since that 2022 round. With the lawsuit dragging on and investors losing confidence, fresh funding became nearly impossible. Meanwhile, legal costs piled up, eating into runway. Despite the bold talk of profitability, the company couldn’t sustain itself.

• Platform Risk and Partner Trouble: Solid’s infrastructure relied heavily on Evolve Bank & Trust — the same partner that worked with Synapse, another BaaS startup that also filed for bankruptcy. When your core partner is unstable, everything you’ve built is at risk. Mercury, another fintech, saw the writing on the wall and dropped Evolve. Solid didn’t.

• No Real Moat: The space is crowded. Stripe, Treasury Prime, Unit — they all do similar things, often with better funding or stronger reputations. Solid didn’t have a clear edge. Once the hype wore off, they blended into the background.

Why It Matters: 

• Litigation, even when settled, can paralyze a company by freezing future funding.

• Platform risk is real. Relying too heavily on a single partner can take you down with them.

• “We’re profitable” isn’t the same as “we’re safe.”

• Infrastructure startups need more than APIs. They need trust — from developers and from capital.

Trend

GPT o3 and o4-mini

This week, OpenAI introduced two new models — o3 and o4-mini — and gave us a glimpse of where reasoning-based AI is heading.

These are the latest releases in OpenAI’s “o” series, a family of models designed to think before they speak.  Instead of rushing to finish your sentence, these models pause, reason, and figure out the best way to respond. 

Now, for the first time, they can use every ChatGPT tool on their own and think with images.

Let’s break it down.

Why It Matters:

• Reasoning performance is improving fast: Even the smaller, cheaper o4-mini shows strong results in math, data science, and non-STEM tasks — at a much higher throughput.

• Visual reasoning is here: For founders building products that involve photos, charts, diagrams, or even messy whiteboards, this unlocks new possibilities.

• You can now run lightweight agents locally with Codex CLI. It’s open-source and built for serious devs, not just prompt dabblers.

Smarter, Faster and more Independent

o3 is now OpenAI’s most capable reasoning model to date. It improves on o1 and o3-mini with better depth, accuracy, and stability when the task isn’t simple. It’s great at math, code, science, and any question that takes a few steps to unpack. It doesn’t just guess, it works through the logic.

o4-mini is the efficient one. It’s cheaper, faster, and still surprisingly strong. It performs well in data science, language tasks, and coding — and it supports much higher usage. If you’re building something that needs a smart, low-latency engine on repeat, this one makes sense.

But the biggest change isn’t raw performance — it’s autonomy.

Both models can now use all ChatGPT tools — search, Python, image understanding, image generation — and they can decide on their own when to use them. They can also call your custom functions through the API.

Lets say you ask:

“How much did global smartphone sales drop last quarter compared to the same period last year?”

The model might:

• Search for the latest quarterly reports from major manufacturers,

• Extract and organize shipment numbers,

• Use Python to calculate the year-over-year change,

• Generate a chart comparing both periods,

• And summarize the key trends behind the drop.

In one example, the user asked it to create a video — something it can’t technically do. But instead of giving up, the model drew each frame as an SVG and stitched them together.

That’s exactly what makes this feel like an agent — it identified the gap, came up with a workaround, and carried out each step to deliver the final result. It didn’t just answer the question — it figured out how to answer it.

Thinking with Images

One of the biggest upgrades? These models can now “think with images.”

That’s OpenAI’s way of saying the models don’t just look at images — they can integrate them directly into their reasoning.

You can give them a photo of a messy whiteboard, a physics worksheet, or a napkin sketch, and they’ll figure it out. Tilted, upside down, cluttered — doesn’t matter. They can zoom in, isolate parts, and get to the answer.

This makes the model surprisingly good at Geoguessr:

This is more than a party trick. It means AI can now reason with the kind of inputs humans use every day — not just text boxes and clean screenshots.

Short term, this is useful for education, design, product dev, and anything visual. Long term, it’s what you need if you want AI agents that can operate in the real world, like robots navigating messy environments.

Help Me Improve Failory

That's all of this edition.

Cheers,

Nico