Steel For Fuel N° 025 of 56 4 Oct 2024
From SaaS to Robots
by Andy Lubershane, Partner and Head of Research, Energy Impact Partners
In this note
The question
If the software playbook that defined the last fifteen years no longer works, where should climate tech look for its software opportunities?
The answer
Not in software by itself. Lubershane argues that the interesting ground is now where software meets hardware, above all in robotics, with generative AI applied to scientific discovery as a second front. He is explicit that software is a supporting actor rather than a solution to emissions.
03The argument
This is the third part of a series, following one on the end of the beginning for climate tech and one on financing hardware. The software half, he argues, faces a bigger shift than climate tech’s own: the construction of the internet is finishing, and the software-as-a-service and mobile playbooks built on it have become harder to rely on. Applied to the capital-intensive industries that produce most emissions, such as electric power, heavy industry, building materials and transport, those playbooks have underperformed, because building a breakout software business for a niche application inside a risk-averse industry carrying decades of technical debt turns out to be as hard as it sounds. The successes were real but modest and nearly all in electric transport. That yields the turn the piece is built on: the bar for a better software mousetrap in the sectors that matter most for climate may now be even higher than the bar for a better hardware one.
His alternative follows from the diagnosis rather than replacing it. If the obstacle is that electrons and molecules are stubbornly physical, then the opportunities sit where bits meet atoms. Hardware already ships with software embedded, which the last decade of connected devices took care of, but he believes the conditions are now right for a further level of integration, and robotics is his example because every input needed to build a robot has become much cheaper. The robots he has in mind are not humanoids but specialized machines with enough embedded intelligence to sense and navigate messy environments, which makes them cases of what a tech blogger he names as Packy McCormack calls vertical integrators, quoted at length and defined below. His second front is generative AI aimed at foundational science, where he quotes a Microsoft technical fellow, Chris Bishop, claiming that scientific discovery will be AI’s most important use, and offers two papers as early evidence. He states the wider conclusion as a suspicion rather than a finding: that software put to work transforming the physical world could be the next big thing for the whole tech sector, and that climate tech might be at the vanguard of it.
04What you need to know first
- Vertical integrator
- The quoted definition: a company that integrates several cutting-edge but proven technologies, builds deep in-house capability across its stack, modularizes the commoditized components while controlling overall system integration, competes head-on with incumbents, and offers products that are better, faster or cheaper, often all three.
- FFOAK
- He uses the abbreviation without expanding it, in a list of priorities for the sector. From his own framing earlier in the post it refers to financing the first few big manufacturing facilities and major projects for a select handful of solutions.
05Details worth keeping
- His firm’s early investments in electric mobility software, Greenlots and Viriciti, are offered as the example of the modest good outcomes that the software playbook did produce, both sold to strategic acquirers.
- The robots already in his firm’s portfolio are described as mowing lawns, tending greenhouses, fulfilling orders, stringing electric conductors and building solar farms, with Scythe Robotics pictured.
- Quilt is his example of a vertical integrator that does not look like a robot: a two-zone mini-split heat pump whose embedded sensing lets it respond to the pattern of life in a home, the weather, and possibly utility signals.
- The claim that every robot input got much cheaper rests entirely on one chart, sourced to Bloomberg on lidar, Less Wrong on graphics processors, BNEF on battery packs, Stanford’s AI Index on robotic arm prices and Microsoft on sensors, all dated 2018 to 2022. The declines appear only in the image.
- A colleague, Eve Hanson, surfaced the carbon-capture sorbent paper, which the post says came from Professor Susana Garcia, a former collaborator of hers.
- He closes with three priorities for the sector, plus an aside wishing for a carbon tax: financing for the best of the last generation of startups, better mousetraps only where they are much better, and new combinations of software and hardware.
06Claims worth citing
All figures as stated on 2024-10-04. This is an argument rather than a data piece, and its only quantitative evidence sits in a chart.
- Google DeepMind researchers published a paper about a year earlier on using AI to discover hundreds of new crystalline structures with encouraging theoretical properties, a subset of them relevant to climate tech such as battery electrolytes and superconductors. the DeepMind paper, cited by Lubershane
- An AI-driven process identified new candidate materials for carbon capture sorbents, one of which has already been singled out for commercialization. Garcia and colleagues, Nature, July 2024, cited by Lubershane
- Quilt recently launched the most efficient two-zone mini-split heat pump on the market. Lubershane, about a company his firm has invested in
07Where it’s contested
There is no second voice, and he does not argue against himself. What the post does carry is a set of clearly marked confidence levels and one gap.
- He separates what he doubts from what he suspects. He has little doubt that software can accelerate emissions reduction, and none that it will not solve the problem alone. The claim that hardware and software integration is the next big thing for tech generally is offered as a suspicion, and the claim that climate tech could lead it as a belief.
- The diagnosis is not carried through to the prescription. He blames the software disappointment on risk-averse cultures and decades of technical debt in legacy industries, then proposes selling those industries robots, without returning to whether the same buyers behave differently toward hardware.
- The evidence for the timing claim is thin by design. That inputs got cheaper is shown in one chart; that the stars are now aligned for deeper integration is asserted.
- His own position is throughout. The portfolio examples are his firm’s and he identifies them as such, which makes the Quilt performance claim an interested party’s, and makes the observation that his firm keeps falling for robotics companies evidence of his own conviction rather than of the trend.