Hey - It’s Nico.

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

If you only have one, here are the 3 most important things:

• Goterra a maggot-powered food waste startup, has shut down — learn more below.

• The new MVP (Meme Viable Problem), and a 3-step playbook to using it

• OpenAI had a busy week — learn about all the new products below

This Week In Startups

🔗 Resources

The new MVP (Meme Viable Problem), and a 3-step playbook to using it

How to measure the impact of AI search the right way

LLMs are picking winners. Here’s how to become one

The word that kills startups is "Interesting"

📰 News

Apple sues OpenAI over alleged trade secret theft

OpenAI is shutting down Atlas

OpenAI releases a $230 keyboard for Codex

Meta’s new AI chips will begin production in September

💸 Fundraising

Vendelux, an AI platform for measuring B2B event performance, raises ⁠$50M Series B funding

Quaise Energy, a developer of deep geothermal drilling technology, raises ⁠$134M Series B funding

Venus Aerospace, a developer of rotating-detonation rocket engines, raises ⁠$91M Series B funding

Ollama, an open-source platform for running AI models, raises ⁠$65M Series B funding⁠

Fail(St)ory

Waste to Protein

Goterra made big, shipping-container machines that used maggots to turn food waste into useful stuff like protein and fertilizer.

It shut down in 2026 after running out of money. The core problem was pretty simple: there was interest in what they were doing, but every bit of growth meant building more expensive hardware first.

What Was Goterra:

Goterra sold organic-waste processing infrastructure. Its core system, called MIBS, was often described as a “maggot robot”: a modular unit that received food waste, processed it, fed it to larvae, and harvested useful outputs.

The product made most sense for customers with steady organic waste: supermarkets, airports, hotels, city programs, commercial kitchens, and food processors. These buyers were paying to avoid landfill, reduce waste transport, and improve their emissions story.

The unit was designed to sit near the waste source. This was important because wet food waste is heavy, smelly, and expensive to move. A decentralized plant could, in theory, make waste treatment closer to a service network than a large remote facility.

Inside the container, automation handled feeding, material movement, environmental conditions, and parts of harvesting. Larvae consumed the waste over roughly 12 days, turning it into insect biomass and frass.

Goterra had two revenue pools:

• Customers or waste partners paid service fees to divert waste.

• The company could also sell larvae as protein ingredients and frass as fertilizer or soil conditioner.

That second pool made the model attractive on paper. Goterra was paid to receive feedstock, while many insect-protein companies had to buy inputs. If the waste fees covered a big part of processing cost, protein sales could improve the economics.

The operating reality was heavier. Each contract required enough units, site work, maintenance, permits, collection logistics, uptime, and working capital. A customer signing a multi-year deal did not instantly become revenue unless Goterra could build and deploy capacity.

The company said it had contracted demand on both sides: waste-management services and protein offtake. The bottleneck became manufacturing infrastructure quickly enough to activate that demand.

By the end, the business looked like an infrastructure company with startup funding. Revenue was growing, but the gross margin was deeply negative because the system was still being developed, operated, and scaled at the same time.

The Numbers:

• 🗑️ Processed: 35,000+ tonnes

• 📈 Revenue: A$365k → A$1.43M

• 💸 Gross margin: -198%

• 🧾 Net loss: A$11.74M

• 🧱 Spend: ~A$25M

• 🔚 Liquidation: July 8, 2026

Reasons for Failure: 

• Capacity had to be built before revenue could catch up: This wasn’t SaaS, you couldn’t just flip a switch and scale. Every new customer needed real machines, real installs, and real upkeep. Goterra had contracts lined up, but without enough units in the field, that demand stayed stuck on paper. When funding slowed, so did everything else.

• The unit economics were still moving in the wrong direction: Revenue was growing, but costs were sprinting ahead. In 11 months, Goterra made about A$1.43M while spending A$4.26M to do it. The hope was scale would fix it, but the runway ran out first.

• Liquidity depended on tax offsets and bridge funding: A big chunk of the balance sheet came from expected R&D tax credits. That’s fine until you need cash now. Goterra borrowed against those future refunds, which only works if new funding arrives on time. It didn’t.

• The final financing had too many conditions at once: Goterra needed about A$25M to keep scaling, but the deal never quite came together. A key investor didn’t sign in time, which also killed the bridge funding. Plenty of buyers looked, but no one wanted to take on both the tech and the heavy capital bill.

Why It Matters: 

• Demand only matters when you can serve it. Contracts stayed on paper without enough deployed units.

• Growth can make the cash problem worse. Every new customer required more hardware and working capital upfront.

Trend

OpenAI’s New Stack

OpenAI had a busy week filling in the stack around agents. There’s a new model (actually three models), a whole new app, and a pretty impressive new voice mode.

Here’s everything about it.

Why it Matters

• Model choice gets simpler and cheaper. Three tiers map cleanly to task difficulty, so you can match cost to need without overpaying.

• Long tasks and voice both level up. Work handles end-to-end assignments, while Live makes conversations fluid and interruptible.

GPT-5.6 now comes in three sizes

GPT 5.6 is out and the biggest change is the model lineup.

• Sol is the expensive flagship for hard reasoning, coding, science, cybersecurity, and long tasks. Benchmarks put it at or near the top, alongside Fable.

• Terra is the middle option, with performance OpenAI compares to GPT-5.5 at roughly half the token price.

• Luna is the fast, cheap model for high-volume work.

OpenAI also launched Programmatic Tool Calling. GPT-5.6 can generate a small JavaScript program that runs inside OpenAI’s sandbox. That program can orchestrate multiple tool calls, loop over results, filter and deduplicate data, and transform it before returning anything to the model.

In practice, the model offloads structured, repetitive work to code. The script pulls exactly what it needs, cleans it up, and compresses it into a smaller, more relevant payload. The model then reasons over that reduced dataset.

One more thing: OpenAI also announced GPT-Red, an internal model trained to attack other models and agents. It tries prompt injections hidden in emails, webpages, files, and tool outputs, then feeds the successful attacks back into training. GPT-5.6 was one of the models trained against it.

ChatGPT Work

GPT Work is essentially OpenAI’s answer to Claude Cowork: Codex for non-technical work.

You give it a larger assignment, it breaks the work into steps, pulls information from connected apps and files, uses the browser or desktop, and keeps going until it has something finished.

The actual new product is the app around it. OpenAI merged Chat, Work, and Codex into one desktop app for macOS and Windows. Existing Codex users get it through an update, while the previous ChatGPT desktop app is now called ChatGPT Classic.

You can chat, hand off a longer Work task, or jump into Codex from the same app. It is a much cleaner way to understand how OpenAI now sees the product: three modes for three different kinds of work, sitting next to each other.

GPT-Live finally feels live

GPT-Live is pretty cool.

The old voice experience still had the rhythm of a walkie-talkie. You spoke, waited, then the model answered. GPT-Live listens and speaks at the same time, so you can interrupt it, pause mid-sentence, change direction, or keep talking while it figures out when to respond.

It can also send harder work to another model in the background. You might talk through customer notes, ask it to check previous emails, keep discussing the account, and get a structured brief once the deeper work is done.

That combination makes voice much more useful. Before, it seemed more like a toy and I would always avoid it for actual work. This week, I am not ashamed to confess that I have spent more time speaking with it than typing to it.

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico