Accelerating Hardware Development

Before co-founding a start-up, I previously tried to spin out a hardware company and failed. This is the story of those learnings.

Between 2015 and 2017, my team developed a solar-powered tracker for shipping containers, offering real-time or gate-triggered information to logistics companies.

We quantified the financial benefits with maritime & overland logistics companies, and tested business models. Ultimately, the margins weren’t there for a pureplay hardware vendor without an integrated software solution (tragically early), so we pivoted.

Here’s where it got hard.

You see, pivoting a hardware play means re-designing hardware for a new customer’s demands. And because several customer interactions are needed to find a product-market fit, you need to re-design hardware many times, for many new customers. If you had believed our whiteboard Gantt charts, each pivot would’ve taken 4 weeks tops. In real life, each pivot could easily take several months, to print a new circuit board, troubleshoot it, integrate it into a device, test it… repeated many times until it actually works… and then present the data to the customer.

In a period of months, customer goalposts change by a lot.

We were too slow.

By summer 2018, our non-dilutive research funding ended, and there was no compelling product-market fit to motivate us spinning out of the lab. Well, that’s a half truth. In reality, there were several compelling product-market fits — but we weren’t fast enough to hit our customers’ windows of opportunity, when business needs and demonstrated technology tentatively align. Very sad.

Accelerating hardware prototyping.

As this effort wound down, MIT graduate student Erin Looney was curious to see how universal our experience was. She agreed to re-direct large part of her Ph.D. thesis toward studying hardware startups. Her goal, was to map the product-development timelines of hardware startups, and to identify where management and technical teams were putting their attention.

Erin called 500+ hardware startups, secured 55 interviews with leadership and technical staff. After assembling an enviable spread of collaborators and co-advisors spanning engineering, design, and sociology to extract quantitative and qualitative data, she mapped key points of the company life cycle, like so:

She also extracted key learnings, shared in her study posted this week to ChemRxiv:

Learning #1. Prototyping is the longest product-development activity, with an average time per prototype of 19 weeks.

Learning #2. Surprisingly, product-development timelines did not scale with product complexity.

Learning #3. Product-development styles correlate with time to prototype.

This insight appears to be crucial to executing successful hardware startups. In the beginning, an “explorative” or “natural” prototyping process often works best. Few rules to slow things down. But the iteration process cannot be aimless — it needs to hunt for a convergence of product, market, and business model. (In our working paper, we show examples of companies that had pretty fast prototyping cycles, but never emerged from the “natural” prototyping process — and thus failed to find a product-market fit.) This convergence process cannot be too fast, nor too slow — lest one lock into the wrong design, or never converge. Gradually, as iterations lead to clarity of product-market fit, a more structured product-development style can be beneficial to drive deliverables. If a pivot is required, then natural prototyping can again return. Awareness of these processes, a strategic intent to transition between them, can be helpful to deploy them effectively.

Changing gears between natural & structured product-dev processes requires buy-in from the entire team. Both technical staff and managers usually excel at one style or the other. Some are great at following their instincts, navigating complexity. Others are great at driving linear, scheduled, KPI-driven processes. Rarely are individuals great at both. But the best managers and technical staff of hardware startups learn when and how to strategically deploy these two processes, as well as the shades of grey in between, to shorten product-development timelines.

An interesting parallel is happening in machine learning. Active-learning algorithms that are too “explorative” or “exploitative” often fail to reach their target if the problem is complex. Adaptive strategies, strategically transitioning between these two modes, often fare better.

We explore many more ideas in our working paper, including examples of companies that did and didn’t do it well, literature examples of other innovators who made similar conclusions, and examples of emergent technologies (machine learning) that can make “transitions between natural and structured styles of product development easier for startups in the future.” (I’m personally bullish about integrating generative design and sentiment analysis in user-centred hardware design processes.) We invite you to read our working paper.

Onward.

Members of the former solar-powered tracking team went off and formed their own companies, including Nasim, Erin, and Ian. Others are applying machine learning to accelerate hardware development, including Marius. The key learning — that acceleration is needed — helped shape the Accelerated Materials Development for Manufacturing Programme in Singapore. Erin put the learnings to work to develop an accelerated solar-cell testing method, to predict energy yield in different geographic locations. Startup Xinterra spun out of this second generation of researchers. We are putting these lessons to work, hopefully to shorten timelines for hardware development, and bring change to market within timeframes that matter.