• Failory
  • Posts
  • Outspent and Outdelivered

Outspent and Outdelivered

How a local grocery survivor faced billion-dollar rivals

Author

Nicolás Cerdeira
September 03, 2026

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:

A big thanks to today’s sponsor…

Can you afford your next hire? AD

Founders learn the answer to this the hard way. Hiring changes your expenses, burn, and the month your cash runs out.

Parallel connects to your accounting and gives you a live forecast built from your real numbers. Add a hire, see the fully loaded cost, and the impact on your runway & cash. So you can hire (and raise) knowing exactly where you stand.

Founders backed by Y Combinator, Khosla Ventures, and Techstars use it.

This Week In Startups

🔗 Resources

Read this before deleting your AGENTS.md

The Creator Economy boom in B2B GTM

Will ChatGPT recommend your brand? Run a free 2-minute AI readiness scan on your site. * 

📰 News

💸 Fundraising

Wonderful, an enterprise platform for deploying and coordinating AI agents and workflows, raises $550M Series C

HiddenLayer, a security platform protecting AI models, agents, and workflows, raises $100M Series B

Superluminal Medicines, a biotechnology startup using machine learning and structural biology to discover small-molecule drugs, raises $60M Series B

Light, a platform that lets businesses offer branded retail-electricity plans, raises $46M Series A

* sponsored

Fail(St)ory

The Second Grocery War

Satvacart began delivering groceries around Gurugram in 2014, during India’s first wave of online grocery startups. It survived India’s first online-grocery shakeout and learned how to make one local cluster economical.

On August 28, 2026, the company stopped operating and disbanded its team after twelve years.

By then, Blinkit, Zepto and Instamart had turned grocery delivery into a race for dense citywide networks. Satvacart’s discipline had produced a carefully tuned local operation. The new contest required far greater coverage and access to much larger pools of capital.

What Was Satvacart:

India’s first online-grocery startups were trying to move the weekly supermarket run onto a phone. Behind each app sat a physical operation that sourced thousands of products, packed every order and sent riders across a city.

Expansion was expensive. A new area needed a warehouse, staff, stock and marketing before enough nearby orders existed. One playbook used investor money to open quickly and absorb early losses. Satvacart, on the other hand, chose to make one small area economical first.

Satvacart launched in early 2015 with 4,000 products in Gurgaon. It planned to reach six cities by year-end. The rollout did not happen, and the company remained local.

Inspired by the online grocer FreshDirect, Satvacart built one warehouse serving homes within roughly five kilometers. It called this a micro-cluster: one fulfillment point, nearby customers and short delivery routes. Shoppers could choose half-hour delivery slots.

The economics improved when nearby households ordered regularly. The warehouse and riders stayed busy, while several deliveries could share a route. Prepaid subscriptions helped create that density. By 2016, Satvacart claimed they had cut delivery costs from ₹300 to ₹25 and brought orders to unit-level break-even.

This discipline appears to have helped Satvacart survive an e-grocery meltdown that removed many better-funded startups. By 2019, it reported a small positive EBITDA with support from a separate B2B marketing activity. The grocery business still occupied one micro-cluster.

Quick commerce then raised the cost of competing. Blinkit, Zepto and Swiggy’s Instamart promised groceries in minutes using many small warehouses near customers. Each “dark store” had to be rented, stocked and staffed before demand filled it.

Satvacart tried to adapt. In 2021, it claimed workers could pick 23 items in two minutes and deliver them within ten minutes inside a two-kilometer radius. It announced thirteen-city expansion and talks for a $50 million round.

The round never appeared, and Satvacart remained Gurgaon-focused. By Q3 FY2026, Instamart operated 1,136 dark stores across 131 cities and still reported a ₹9.08 billion adjusted EBITDA loss.

Satvacart had survived the first shakeout with one efficient cluster. The next phase required a network it could no longer finance.

The Numbers:

Reasons for Failure: 

  • One cluster never became a rollout system: Satvacart learned how to concentrate orders around a single warehouse. Each new area still required rent, staff, inventory and customer acquisition before reaching useful density. Repeated city launches remained plans, leaving the company too localized for the scale investors and buyers wanted.

  • Profitability labels obscured the cash available for growth: Unit break-even, contribution margin, two cash-positive months and small EBITDA described different periods and cost layers. Filed accounts later showed a company-level loss. None of those earlier milestones proved that one mature cluster could finance the next.

  • The large round never arrived: Satvacart appears to have raised roughly $2.2 million to $2.3 million across its lifetime, mostly through angel and pre-Series A rounds. In his shutdown account, the founder said smaller tranches were insufficient, two significant investment discussions failed, and acquisition talks stalled because the company lacked scale.

Why It Matters: 

  • The strategy that helps a startup survive one wave of competition may leave it exposed to the next. Revisit old assumptions when the market changes.

  • A profitability claim needs a clear scope. A profitable order or location can still sit inside a company that is losing money.

  • In network businesses, scale changes the customer experience. A strong local operation may become less attractive once rivals offer broader coverage and faster service.

Trend

AI Video Specializes

Runway and World Labs released two video-capable models within two days. Solaris turns clicks and drags into the next frame of an interface. Atlas takes camera geometry and spatial context, then generates controlled views or explicit 3D output.

They point to a wider split in AI video. The question is shifting from which model produces the best-looking clip to which model gives a product the right controls. Interactive software, navigable worlds and spatial reconstruction each ask the model to hold a different kind of state.

Why it Matters

  • Model choice now starts with the workload: live interaction, camera direction, reconstruction, 3D output or a fixed clip.

  • Value moves toward the context a model can maintain and the controls it can follow. That opens infrastructure budgets in software, games, robotics and simulation, along with recurring inference costs.

  • Aesthetics becomes one metric among several. Teams also need to test action fidelity, state drift, geometry, latency, physics and safety.

Solaris

Runway's Solaris generates an interface frame by frame. An LLM decides how the experience should respond, while a world model based on Gen-4.5 renders the next 720p frame after a click, drag or other action.

A conventional interface has coded screens, components and transitions. Solaris turns that fixed set of states into a running exchange: after every action, the LLM decides what should happen and the visual model draws it.

Runway's demos apply the system to virtual stores, rooms and courseware, where the experience can keep changing as the user moves through it.

That freedom is only useful if the generated interface still feels responsive. Runway tested this with 250 participants, who made nearly 7,500 comparisons across 30 examples. Each comparison put Solaris beside an interface coded by Claude Opus 5.

When judging instruction following, participants preferred Solaris 62% of the time, the coded interface 25%, and rated them equal 13%. For natural behavior, Solaris received 72%, compared with 21% for the coded version and 7% equal. These are Runway's own results, and the test measured how short interactions looked and felt. It did not test whether Solaris could preserve state, finish a task securely or remain coherent through a long session.

Runway says stable text, grounding, accessibility and integration also remain unresolved. Solaris could handle the exploratory, visual layer of an interface, with tested components managing payments, permissions and records behind it. A convincing button is still a picture until something reliable handles what happens when you click it.

Atlas

World Labs' Atlas accepts images or depth maps at explicit camera poses. It can generate up to one minute of 1440p video along a specified camera path, reconstruct a scene from a few images and return explicit 3D representations for editing or simulation.

A user tells Atlas where the camera is and where it should move, while the model tries to keep the scene coherent across the path. World Labs says two or three reference images can be enough for reconstruction, although Atlas may imagine areas the inputs never showed.

Those capabilities matter when a scene needs to live beyond one rendered clip:

  • Game and simulation teams could turn a few reference images into an editable 3D starting point, direct new camera moves through it and generate variations without rebuilding the environment by hand.

  • Robotics teams could use those reconstructed scenes in real-to-sim workflows for training and evaluation.

Together, those applications could reduce the hand-built work between capturing a place and using it inside another system.

The Trend

The AI video category keeps splitting into smaller ones. Some models make clips. Solaris generates interfaces you can interact with. Atlas generates views you can direct with a camera path and turn into 3D. Project Genie and Odyssey generate worlds that keep responding to actions.

Choosing a model now starts with what you want to control: the shot, the interface, the camera or the world. Benchmarks will have to measure interaction, consistency and control alongside visual quality.

Help Me Improve Failory

How useful did you find today’s newsletter?

Your feedback helps me make future issues more relevant and valuable.

Login or Subscribe to participate in polls.

That's all for today’s edition.

Cheers,

Nico